2023-04-26 10:07:06,651 INFO [finetune.py:1046] (4/7) Training started 2023-04-26 10:07:06,651 INFO [finetune.py:1056] (4/7) Device: cuda:4 2023-04-26 10:07:06,653 INFO [finetune.py:1065] (4/7) {'frame_shift_ms': 10.0, 'allowed_excess_duration_ratio': 0.1, 'best_train_loss': inf, 'best_valid_loss': inf, 'best_train_epoch': -1, 'best_valid_epoch': -1, 'batch_idx_train': 0, 'log_interval': 50, 'reset_interval': 200, 'valid_interval': 3000, 'feature_dim': 80, 'subsampling_factor': 4, 'warm_step': 2000, 'env_info': {'k2-version': '1.23.4', 'k2-build-type': 'Release', 'k2-with-cuda': True, 'k2-git-sha1': '62e404dd3f3a811d73e424199b3408e309c06e1a', 'k2-git-date': 'Mon Jan 30 02:26:16 2023', 'lhotse-version': '1.12.0.dev+git.3ccfeb7.clean', 'torch-version': '1.13.0', 'torch-cuda-available': True, 'torch-cuda-version': '11.7', 'python-version': '3.8', 'icefall-git-branch': 'master', 'icefall-git-sha1': 'd74822d-dirty', 'icefall-git-date': 'Tue Mar 21 21:35:32 2023', 'icefall-path': '/home/lishaojie/icefall', 'k2-path': '/home/lishaojie/.conda/envs/env_lishaojie/lib/python3.8/site-packages/k2/__init__.py', 'lhotse-path': '/home/lishaojie/.conda/envs/env_lishaojie/lib/python3.8/site-packages/lhotse/__init__.py', 'hostname': 'cnc533', 'IP address': '127.0.1.1'}, 'world_size': 7, 'master_port': 18181, 'tensorboard': True, 'num_epochs': 30, 'start_epoch': 1, 'start_batch': 0, 'exp_dir': PosixPath('pruned_transducer_stateless7_streaming/exp2'), 'bpe_model': 'data/lang_bpe_500/bpe.model', 'base_lr': 0.004, 'lr_batches': 100000.0, 'lr_epochs': 100.0, 'context_size': 2, 'prune_range': 5, 'lm_scale': 0.25, 'am_scale': 0.0, 'simple_loss_scale': 0.5, 'seed': 42, 'print_diagnostics': False, 'inf_check': False, 'save_every_n': 2000, 'keep_last_k': 30, 'average_period': 200, 'use_fp16': True, 'num_encoder_layers': '2,4,3,2,4', 'feedforward_dims': '1024,1024,2048,2048,1024', 'nhead': '8,8,8,8,8', 'encoder_dims': '384,384,384,384,384', 'attention_dims': '192,192,192,192,192', 'encoder_unmasked_dims': '256,256,256,256,256', 'zipformer_downsampling_factors': '1,2,4,8,2', 'cnn_module_kernels': '31,31,31,31,31', 'decoder_dim': 512, 'joiner_dim': 512, 'do_finetune': True, 'init_modules': 'encoder', 'finetune_ckpt': '/home/lishaojie/icefall/egs/commonvoice_fr/ASR/pruned_transducer_stateless7_streaming/exp/english_pretrain/pretrained.pt', 'manifest_dir': PosixPath('data/fbank'), 'max_duration': 200, 'bucketing_sampler': True, 'num_buckets': 30, 'concatenate_cuts': False, 'duration_factor': 1.0, 'gap': 1.0, 'on_the_fly_feats': False, 'shuffle': True, 'drop_last': True, 'return_cuts': True, 'num_workers': 2, 'enable_spec_aug': True, 'spec_aug_time_warp_factor': 80, 'enable_musan': True, 'input_strategy': 'PrecomputedFeatures', 'blank_id': 0, 'vocab_size': 500} 2023-04-26 10:07:06,653 INFO [finetune.py:1067] (4/7) About to create model 2023-04-26 10:07:07,041 INFO [zipformer.py:405] (4/7) At encoder stack 4, which has downsampling_factor=2, we will combine the outputs of layers 1 and 3, with downsampling_factors=2 and 8. 2023-04-26 10:07:07,051 INFO [finetune.py:1071] (4/7) Number of model parameters: 70369391 2023-04-26 10:07:07,051 INFO [finetune.py:626] (4/7) Loading checkpoint from /home/lishaojie/icefall/egs/commonvoice_fr/ASR/pruned_transducer_stateless7_streaming/exp/english_pretrain/pretrained.pt 2023-04-26 10:07:07,232 INFO [finetune.py:647] (4/7) Loading parameters starting with prefix encoder 2023-04-26 10:07:08,599 INFO [finetune.py:1093] (4/7) Using DDP 2023-04-26 10:07:09,244 INFO [commonvoice_fr.py:392] (4/7) About to get train cuts 2023-04-26 10:07:09,247 INFO [commonvoice_fr.py:218] (4/7) Enable MUSAN 2023-04-26 10:07:09,247 INFO [commonvoice_fr.py:219] (4/7) About to get Musan cuts 2023-04-26 10:07:10,799 INFO [commonvoice_fr.py:243] (4/7) Enable SpecAugment 2023-04-26 10:07:10,799 INFO [commonvoice_fr.py:244] (4/7) Time warp factor: 80 2023-04-26 10:07:10,799 INFO [commonvoice_fr.py:254] (4/7) Num frame mask: 10 2023-04-26 10:07:10,800 INFO [commonvoice_fr.py:267] (4/7) About to create train dataset 2023-04-26 10:07:10,800 INFO [commonvoice_fr.py:294] (4/7) Using DynamicBucketingSampler. 2023-04-26 10:07:13,508 INFO [commonvoice_fr.py:309] (4/7) About to create train dataloader 2023-04-26 10:07:13,508 INFO [commonvoice_fr.py:399] (4/7) About to get dev cuts 2023-04-26 10:07:13,509 INFO [commonvoice_fr.py:340] (4/7) About to create dev dataset 2023-04-26 10:07:13,909 INFO [commonvoice_fr.py:357] (4/7) About to create dev dataloader 2023-04-26 10:07:13,909 INFO [finetune.py:1289] (4/7) Sanity check -- see if any of the batches in epoch 1 would cause OOM. 2023-04-26 10:11:06,931 INFO [finetune.py:1317] (4/7) Maximum memory allocated so far is 5247MB 2023-04-26 10:11:07,629 INFO [finetune.py:1317] (4/7) Maximum memory allocated so far is 5735MB 2023-04-26 10:11:07,877 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=130.85 vs. limit=5.0 2023-04-26 10:11:08,317 INFO [finetune.py:1317] (4/7) Maximum memory allocated so far is 5735MB 2023-04-26 10:11:08,989 INFO [finetune.py:1317] (4/7) Maximum memory allocated so far is 5735MB 2023-04-26 10:11:09,672 INFO [finetune.py:1317] (4/7) Maximum memory allocated so far is 5735MB 2023-04-26 10:11:10,374 INFO [finetune.py:1317] (4/7) Maximum memory allocated so far is 5735MB 2023-04-26 10:11:19,556 INFO [finetune.py:976] (4/7) Epoch 1, batch 0, loss[loss=7.465, simple_loss=6.77, pruned_loss=6.939, over 4904.00 frames. ], tot_loss[loss=7.465, simple_loss=6.77, pruned_loss=6.939, over 4904.00 frames. ], batch size: 38, lr: 2.00e-03, grad_scale: 2.0 2023-04-26 10:11:19,556 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-26 10:11:40,203 INFO [finetune.py:1010] (4/7) Epoch 1, validation: loss=7.31, simple_loss=6.623, pruned_loss=6.857, over 2265189.00 frames. 2023-04-26 10:11:40,203 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 5735MB 2023-04-26 10:11:48,689 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=5.0, num_to_drop=2, layers_to_drop={0, 1} 2023-04-26 10:12:11,185 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=23.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 10:12:13,855 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.8675, 0.8756, 0.8514, 1.4018, 0.9339, 0.6805, 0.6160, 0.5552], device='cuda:4'), covar=tensor([13.1618, 10.1461, 13.1262, 7.3940, 18.1441, 13.2438, 11.6241, 6.3841], device='cuda:4'), in_proj_covar=tensor([0.0521, 0.0593, 0.0670, 0.0646, 0.0574, 0.0631, 0.0664, 0.0653], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 10:12:34,677 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-04-26 10:12:43,286 INFO [finetune.py:976] (4/7) Epoch 1, batch 50, loss[loss=2.606, simple_loss=2.485, pruned_loss=1.241, over 4909.00 frames. ], tot_loss[loss=4.38, simple_loss=3.965, pruned_loss=4.013, over 214893.73 frames. ], batch size: 46, lr: 2.20e-03, grad_scale: 0.00390625 2023-04-26 10:13:19,830 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=83.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 10:13:20,910 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=7.12 vs. limit=5.0 2023-04-26 10:13:39,673 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=7.96 vs. limit=5.0 2023-04-26 10:13:41,114 WARNING [finetune.py:966] (4/7) Grad scale is small: 6.103515625e-05 2023-04-26 10:13:41,114 INFO [finetune.py:976] (4/7) Epoch 1, batch 100, loss[loss=2.252, simple_loss=2.132, pruned_loss=1.198, over 4796.00 frames. ], tot_loss[loss=3.401, simple_loss=3.145, pruned_loss=2.508, over 376935.06 frames. ], batch size: 51, lr: 2.40e-03, grad_scale: 0.0001220703125 2023-04-26 10:14:03,496 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 4.683e+02 1.507e+03 7.833e+03 2.572e+04 3.214e+07, threshold=1.567e+04, percent-clipped=0.0 2023-04-26 10:14:06,183 WARNING [optim.py:389] (4/7) Scaling gradients by 0.014711554162204266, model_norm_threshold=15666.9306640625 2023-04-26 10:14:06,257 INFO [optim.py:451] (4/7) Parameter Dominanting tot_sumsq module.encoder.encoder_embed.conv.0.weight with proportion 0.88, where dominant_sumsq=(grad_sumsq*orig_rms_sq)=1.001e+12, grad_sumsq = 2.310e+12, orig_rms_sq=4.331e-01 2023-04-26 10:14:13,164 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5312, 1.9870, 1.8356, 2.5340, 2.4087, 1.7233, 2.0067, 0.8165], device='cuda:4'), covar=tensor([0.0034, 0.0040, 0.0051, 0.0038, 0.0036, 0.0063, 0.0051, 0.0094], device='cuda:4'), in_proj_covar=tensor([0.0083, 0.0093, 0.0086, 0.0092, 0.0107, 0.0106, 0.0105, 0.0092], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-26 10:14:23,410 WARNING [optim.py:389] (4/7) Scaling gradients by 0.00018281130178365856, model_norm_threshold=15666.9306640625 2023-04-26 10:14:23,484 INFO [optim.py:451] (4/7) Parameter Dominanting tot_sumsq module.encoder.encoder_embed.conv.0.weight with proportion 0.29, where dominant_sumsq=(grad_sumsq*orig_rms_sq)=2.155e+15, grad_sumsq = 4.978e+15, orig_rms_sq=4.329e-01 2023-04-26 10:14:24,120 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=144.0, num_to_drop=2, layers_to_drop={0, 1} 2023-04-26 10:14:27,645 INFO [finetune.py:976] (4/7) Epoch 1, batch 150, loss[loss=1.837, simple_loss=1.677, pruned_loss=1.33, over 4843.00 frames. ], tot_loss[loss=2.849, simple_loss=2.641, pruned_loss=1.995, over 506320.52 frames. ], batch size: 47, lr: 2.60e-03, grad_scale: 3.0517578125e-05 2023-04-26 10:14:28,153 WARNING [optim.py:389] (4/7) Scaling gradients by 0.00022292081848718226, model_norm_threshold=15666.9306640625 2023-04-26 10:14:28,228 INFO [optim.py:451] (4/7) Parameter Dominanting tot_sumsq module.encoder.encoder_embed.conv.0.weight with proportion 0.45, where dominant_sumsq=(grad_sumsq*orig_rms_sq)=2.213e+15, grad_sumsq = 5.111e+15, orig_rms_sq=4.330e-01 2023-04-26 10:14:40,641 WARNING [optim.py:389] (4/7) Scaling gradients by 0.05655747279524803, model_norm_threshold=15666.9306640625 2023-04-26 10:14:40,714 INFO [optim.py:451] (4/7) Parameter Dominanting tot_sumsq module.encoder.encoder_embed.conv.0.weight with proportion 0.84, where dominant_sumsq=(grad_sumsq*orig_rms_sq)=6.482e+10, grad_sumsq = 1.497e+11, orig_rms_sq=4.330e-01 2023-04-26 10:14:45,469 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3695, 0.6554, 0.6031, 1.6684, 1.6518, 1.7223, 0.9764, 1.4814], device='cuda:4'), covar=tensor([0.0261, 0.0363, 0.0312, 0.0201, 0.0201, 0.0180, 0.0220, 0.0186], device='cuda:4'), in_proj_covar=tensor([0.0205, 0.0212, 0.0191, 0.0172, 0.0174, 0.0191, 0.0167, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 10:14:47,695 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=2.32 vs. limit=2.0 2023-04-26 10:14:56,529 WARNING [finetune.py:966] (4/7) Grad scale is small: 3.0517578125e-05 2023-04-26 10:14:56,529 INFO [finetune.py:976] (4/7) Epoch 1, batch 200, loss[loss=1.503, simple_loss=1.304, pruned_loss=1.39, over 4806.00 frames. ], tot_loss[loss=2.371, simple_loss=2.177, pruned_loss=1.739, over 608461.24 frames. ], batch size: 45, lr: 2.80e-03, grad_scale: 6.103515625e-05 2023-04-26 10:15:07,790 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 6.387e+01 5.300e+02 1.840e+03 7.547e+03 8.570e+07, threshold=3.680e+03, percent-clipped=20.0 2023-04-26 10:15:12,965 WARNING [optim.py:389] (4/7) Scaling gradients by 0.011872046627104282, model_norm_threshold=3679.54541015625 2023-04-26 10:15:13,040 INFO [optim.py:451] (4/7) Parameter Dominanting tot_sumsq module.encoder.encoder_embed.conv.0.weight with proportion 0.57, where dominant_sumsq=(grad_sumsq*orig_rms_sq)=5.451e+10, grad_sumsq = 1.259e+11, orig_rms_sq=4.329e-01 2023-04-26 10:15:16,158 WARNING [optim.py:389] (4/7) Scaling gradients by 0.08515117317438126, model_norm_threshold=3679.54541015625 2023-04-26 10:15:16,233 INFO [optim.py:451] (4/7) Parameter Dominanting tot_sumsq module.encoder.encoder_embed.conv.0.weight with proportion 0.79, where dominant_sumsq=(grad_sumsq*orig_rms_sq)=1.483e+09, grad_sumsq = 3.425e+09, orig_rms_sq=4.329e-01 2023-04-26 10:15:16,773 WARNING [optim.py:389] (4/7) Scaling gradients by 0.04552413150668144, model_norm_threshold=3679.54541015625 2023-04-26 10:15:16,847 INFO [optim.py:451] (4/7) Parameter Dominanting tot_sumsq module.encoder.encoder_embed.conv.0.weight with proportion 0.84, where dominant_sumsq=(grad_sumsq*orig_rms_sq)=5.493e+09, grad_sumsq = 1.269e+10, orig_rms_sq=4.329e-01 2023-04-26 10:15:17,223 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=13.39 vs. limit=5.0 2023-04-26 10:15:25,653 INFO [finetune.py:976] (4/7) Epoch 1, batch 250, loss[loss=1.501, simple_loss=1.273, pruned_loss=1.446, over 4923.00 frames. ], tot_loss[loss=2.052, simple_loss=1.858, pruned_loss=1.595, over 686093.48 frames. ], batch size: 38, lr: 3.00e-03, grad_scale: 6.103515625e-05 2023-04-26 10:15:31,494 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=2.51 vs. limit=2.0 2023-04-26 10:15:50,817 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=296.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 10:15:52,836 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=300.0, num_to_drop=2, layers_to_drop={0, 1} 2023-04-26 10:15:53,285 WARNING [finetune.py:966] (4/7) Grad scale is small: 6.103515625e-05 2023-04-26 10:15:53,285 INFO [finetune.py:976] (4/7) Epoch 1, batch 300, loss[loss=1.245, simple_loss=1.034, pruned_loss=1.227, over 4773.00 frames. ], tot_loss[loss=1.849, simple_loss=1.65, pruned_loss=1.51, over 747264.60 frames. ], batch size: 28, lr: 3.20e-03, grad_scale: 0.0001220703125 2023-04-26 10:16:02,530 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 2.002e+01 7.343e+01 2.302e+02 1.070e+03 3.099e+05, threshold=4.604e+02, percent-clipped=16.0 2023-04-26 10:16:21,502 INFO [finetune.py:976] (4/7) Epoch 1, batch 350, loss[loss=1.365, simple_loss=1.125, pruned_loss=1.314, over 4814.00 frames. ], tot_loss[loss=1.706, simple_loss=1.5, pruned_loss=1.443, over 792538.57 frames. ], batch size: 39, lr: 3.40e-03, grad_scale: 0.0001220703125 2023-04-26 10:16:23,647 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1712, 1.8914, 1.3140, 2.3742, 2.1357, 1.7916, 1.6238, 1.4466], device='cuda:4'), covar=tensor([0.0074, 0.0051, 0.0109, 0.0051, 0.0056, 0.0139, 0.0076, 0.0141], device='cuda:4'), in_proj_covar=tensor([0.0082, 0.0092, 0.0086, 0.0091, 0.0107, 0.0106, 0.0104, 0.0091], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-26 10:16:24,660 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=357.0, num_to_drop=2, layers_to_drop={0, 1} 2023-04-26 10:16:42,462 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=387.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 10:16:56,315 WARNING [finetune.py:966] (4/7) Grad scale is small: 0.0001220703125 2023-04-26 10:16:56,316 INFO [finetune.py:976] (4/7) Epoch 1, batch 400, loss[loss=1.264, simple_loss=1.013, pruned_loss=1.256, over 4834.00 frames. ], tot_loss[loss=1.596, simple_loss=1.382, pruned_loss=1.393, over 829272.44 frames. ], batch size: 44, lr: 3.60e-03, grad_scale: 0.000244140625 2023-04-26 10:17:16,622 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.830e+01 2.642e+01 7.119e+01 5.229e+02 3.680e+03, threshold=1.424e+02, percent-clipped=26.0 2023-04-26 10:17:27,879 WARNING [optim.py:389] (4/7) Scaling gradients by 0.02133115753531456, model_norm_threshold=142.37583923339844 2023-04-26 10:17:27,958 INFO [optim.py:451] (4/7) Parameter Dominanting tot_sumsq module.encoder.encoder_embed.conv.0.weight with proportion 0.81, where dominant_sumsq=(grad_sumsq*orig_rms_sq)=3.619e+07, grad_sumsq = 8.362e+07, orig_rms_sq=4.328e-01 2023-04-26 10:17:30,803 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.79 vs. limit=2.0 2023-04-26 10:17:41,634 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=439.0, num_to_drop=2, layers_to_drop={0, 3} 2023-04-26 10:17:52,443 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=448.0, num_to_drop=2, layers_to_drop={0, 2} 2023-04-26 10:17:53,957 INFO [finetune.py:976] (4/7) Epoch 1, batch 450, loss[loss=1.154, simple_loss=0.9017, pruned_loss=1.167, over 4801.00 frames. ], tot_loss[loss=1.498, simple_loss=1.275, pruned_loss=1.345, over 857592.73 frames. ], batch size: 29, lr: 3.80e-03, grad_scale: 0.000244140625 2023-04-26 10:18:14,586 WARNING [optim.py:389] (4/7) Scaling gradients by 0.06225070729851723, model_norm_threshold=142.37583923339844 2023-04-26 10:18:14,660 INFO [optim.py:451] (4/7) Parameter Dominanting tot_sumsq module.encoder.encoder_embed.conv.0.weight with proportion 0.68, where dominant_sumsq=(grad_sumsq*orig_rms_sq)=3.535e+06, grad_sumsq = 8.167e+06, orig_rms_sq=4.328e-01 2023-04-26 10:18:22,849 WARNING [finetune.py:966] (4/7) Grad scale is small: 0.000244140625 2023-04-26 10:18:22,849 INFO [finetune.py:976] (4/7) Epoch 1, batch 500, loss[loss=1.145, simple_loss=0.8764, pruned_loss=1.162, over 4924.00 frames. ], tot_loss[loss=1.408, simple_loss=1.177, pruned_loss=1.296, over 879677.80 frames. ], batch size: 37, lr: 4.00e-03, grad_scale: 0.00048828125 2023-04-26 10:18:28,224 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9496, 1.6372, 1.4458, 2.9675, 1.9686, 1.2127, 1.4063, 4.7829], device='cuda:4'), covar=tensor([0.0233, 0.0298, 0.0124, 0.0187, 0.0124, 0.0166, 0.0343, 0.0023], device='cuda:4'), in_proj_covar=tensor([0.0297, 0.0309, 0.0242, 0.0405, 0.0264, 0.0255, 0.0302, 0.0252], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 10:18:31,738 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.709e+01 2.273e+01 3.115e+01 1.071e+02 6.675e+03, threshold=6.230e+01, percent-clipped=18.0 2023-04-26 10:18:47,426 WARNING [optim.py:389] (4/7) Scaling gradients by 0.017591100186109543, model_norm_threshold=62.30100631713867 2023-04-26 10:18:47,498 INFO [optim.py:451] (4/7) Parameter Dominanting tot_sumsq module.encoder.encoder_embed.conv.0.weight with proportion 0.35, where dominant_sumsq=(grad_sumsq*orig_rms_sq)=4.348e+06, grad_sumsq = 1.005e+07, orig_rms_sq=4.327e-01 2023-04-26 10:18:56,301 WARNING [optim.py:389] (4/7) Scaling gradients by 0.005508477333933115, model_norm_threshold=62.30100631713867 2023-04-26 10:18:56,375 INFO [optim.py:451] (4/7) Parameter Dominanting tot_sumsq module.encoder.encoder_embed.conv.0.weight with proportion 0.80, where dominant_sumsq=(grad_sumsq*orig_rms_sq)=1.024e+08, grad_sumsq = 2.367e+08, orig_rms_sq=4.327e-01 2023-04-26 10:18:58,474 INFO [finetune.py:976] (4/7) Epoch 1, batch 550, loss[loss=1.046, simple_loss=0.78, pruned_loss=1.074, over 4897.00 frames. ], tot_loss[loss=1.33, simple_loss=1.091, pruned_loss=1.248, over 896980.74 frames. ], batch size: 35, lr: 4.00e-03, grad_scale: 0.00048828125 2023-04-26 10:19:04,290 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=562.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 10:19:13,088 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=568.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 10:19:32,579 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=3.50 vs. limit=2.0 2023-04-26 10:19:35,171 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=590.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 10:19:45,898 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=600.0, num_to_drop=2, layers_to_drop={0, 2} 2023-04-26 10:19:46,333 WARNING [finetune.py:966] (4/7) Grad scale is small: 0.00048828125 2023-04-26 10:19:46,333 INFO [finetune.py:976] (4/7) Epoch 1, batch 600, loss[loss=1.222, simple_loss=0.8938, pruned_loss=1.249, over 4835.00 frames. ], tot_loss[loss=1.275, simple_loss=1.026, pruned_loss=1.218, over 910374.19 frames. ], batch size: 47, lr: 4.00e-03, grad_scale: 0.0009765625 2023-04-26 10:20:06,707 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.706e+01 2.227e+01 2.629e+01 6.305e+01 1.131e+04, threshold=5.258e+01, percent-clipped=26.0 2023-04-26 10:20:09,418 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=623.0, num_to_drop=2, layers_to_drop={0, 1} 2023-04-26 10:20:18,149 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=629.0, num_to_drop=2, layers_to_drop={1, 2} 2023-04-26 10:20:29,918 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=648.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 10:20:30,854 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=2.12 vs. limit=2.0 2023-04-26 10:20:31,450 INFO [finetune.py:976] (4/7) Epoch 1, batch 650, loss[loss=1.134, simple_loss=0.83, pruned_loss=1.116, over 4843.00 frames. ], tot_loss[loss=1.246, simple_loss=0.9833, pruned_loss=1.202, over 920221.84 frames. ], batch size: 44, lr: 4.00e-03, grad_scale: 0.0009765625 2023-04-26 10:20:31,537 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=651.0, num_to_drop=2, layers_to_drop={0, 3} 2023-04-26 10:20:31,858 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=2.22 vs. limit=2.0 2023-04-26 10:20:32,031 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=652.0, num_to_drop=2, layers_to_drop={0, 3} 2023-04-26 10:20:41,814 WARNING [optim.py:389] (4/7) Scaling gradients by 0.07653743773698807, model_norm_threshold=52.5806770324707 2023-04-26 10:20:41,891 INFO [optim.py:451] (4/7) Parameter Dominanting tot_sumsq module.encoder.encoder_embed.conv.0.weight with proportion 0.93, where dominant_sumsq=(grad_sumsq*orig_rms_sq)=4.392e+05, grad_sumsq = 1.015e+06, orig_rms_sq=4.327e-01 2023-04-26 10:20:50,153 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=3.61 vs. limit=2.0 2023-04-26 10:20:59,955 WARNING [finetune.py:966] (4/7) Grad scale is small: 0.0009765625 2023-04-26 10:20:59,955 INFO [finetune.py:976] (4/7) Epoch 1, batch 700, loss[loss=1.151, simple_loss=0.8476, pruned_loss=1.088, over 4767.00 frames. ], tot_loss[loss=1.216, simple_loss=0.9428, pruned_loss=1.179, over 928418.11 frames. ], batch size: 59, lr: 4.00e-03, grad_scale: 0.001953125 2023-04-26 10:21:09,242 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.790e+01 2.165e+01 2.490e+01 3.193e+01 6.870e+02, threshold=4.979e+01, percent-clipped=6.0 2023-04-26 10:21:19,346 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=2.39 vs. limit=2.0 2023-04-26 10:21:20,032 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=739.0, num_to_drop=2, layers_to_drop={0, 1} 2023-04-26 10:21:22,086 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=743.0, num_to_drop=2, layers_to_drop={1, 3} 2023-04-26 10:21:28,227 INFO [finetune.py:976] (4/7) Epoch 1, batch 750, loss[loss=0.8732, simple_loss=0.6263, pruned_loss=0.8322, over 4225.00 frames. ], tot_loss[loss=1.185, simple_loss=0.9046, pruned_loss=1.147, over 933086.58 frames. ], batch size: 18, lr: 4.00e-03, grad_scale: 0.001953125 2023-04-26 10:21:34,057 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=21.45 vs. limit=5.0 2023-04-26 10:21:46,873 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=3.33 vs. limit=2.0 2023-04-26 10:21:48,069 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=787.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 10:21:53,224 WARNING [optim.py:389] (4/7) Scaling gradients by 0.039711207151412964, model_norm_threshold=49.79251480102539 2023-04-26 10:21:53,296 INFO [optim.py:451] (4/7) Parameter Dominanting tot_sumsq module.encoder.encoder_embed.conv.0.weight with proportion 0.81, where dominant_sumsq=(grad_sumsq*orig_rms_sq)=1.271e+06, grad_sumsq = 2.939e+06, orig_rms_sq=4.325e-01 2023-04-26 10:21:55,910 WARNING [finetune.py:966] (4/7) Grad scale is small: 0.001953125 2023-04-26 10:21:55,910 INFO [finetune.py:976] (4/7) Epoch 1, batch 800, loss[loss=1.099, simple_loss=0.7821, pruned_loss=1.029, over 4815.00 frames. ], tot_loss[loss=1.16, simple_loss=0.8733, pruned_loss=1.117, over 937971.24 frames. ], batch size: 38, lr: 4.00e-03, grad_scale: 0.00390625 2023-04-26 10:22:02,916 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=2.71 vs. limit=2.0 2023-04-26 10:22:11,115 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 2.049e+01 2.360e+01 2.808e+01 3.468e+01 1.254e+03, threshold=5.615e+01, percent-clipped=6.0 2023-04-26 10:22:26,847 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=847.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 10:22:28,870 INFO [finetune.py:976] (4/7) Epoch 1, batch 850, loss[loss=0.9939, simple_loss=0.7101, pruned_loss=0.9006, over 4837.00 frames. ], tot_loss[loss=1.127, simple_loss=0.8395, pruned_loss=1.076, over 942549.77 frames. ], batch size: 47, lr: 4.00e-03, grad_scale: 0.00390625 2023-04-26 10:23:15,923 WARNING [finetune.py:966] (4/7) Grad scale is small: 0.00390625 2023-04-26 10:23:15,923 INFO [finetune.py:976] (4/7) Epoch 1, batch 900, loss[loss=0.9236, simple_loss=0.6459, pruned_loss=0.8374, over 4741.00 frames. ], tot_loss[loss=1.094, simple_loss=0.8079, pruned_loss=1.033, over 945913.09 frames. ], batch size: 27, lr: 4.00e-03, grad_scale: 0.0078125 2023-04-26 10:23:20,207 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=908.0, num_to_drop=2, layers_to_drop={2, 3} 2023-04-26 10:23:26,894 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 2.132e+01 2.534e+01 2.883e+01 3.456e+01 6.931e+01, threshold=5.766e+01, percent-clipped=4.0 2023-04-26 10:23:26,978 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=918.0, num_to_drop=2, layers_to_drop={1, 3} 2023-04-26 10:23:30,105 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=924.0, num_to_drop=2, layers_to_drop={0, 1} 2023-04-26 10:23:42,109 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=946.0, num_to_drop=2, layers_to_drop={0, 3} 2023-04-26 10:23:44,693 INFO [finetune.py:976] (4/7) Epoch 1, batch 950, loss[loss=0.9883, simple_loss=0.7012, pruned_loss=0.8596, over 4815.00 frames. ], tot_loss[loss=1.07, simple_loss=0.7847, pruned_loss=0.9941, over 948328.97 frames. ], batch size: 38, lr: 4.00e-03, grad_scale: 0.0078125 2023-04-26 10:23:45,274 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=952.0, num_to_drop=2, layers_to_drop={0, 1} 2023-04-26 10:24:11,020 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=13.86 vs. limit=5.0 2023-04-26 10:24:21,278 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=15.92 vs. limit=5.0 2023-04-26 10:24:41,985 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=1000.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 10:24:42,451 WARNING [finetune.py:966] (4/7) Grad scale is small: 0.0078125 2023-04-26 10:24:42,451 INFO [finetune.py:976] (4/7) Epoch 1, batch 1000, loss[loss=1.153, simple_loss=0.8206, pruned_loss=0.9772, over 4849.00 frames. ], tot_loss[loss=1.061, simple_loss=0.7728, pruned_loss=0.9696, over 948801.12 frames. ], batch size: 47, lr: 4.00e-03, grad_scale: 0.015625 2023-04-26 10:25:00,169 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 2.486e+01 3.090e+01 3.594e+01 4.276e+01 7.170e+01, threshold=7.188e+01, percent-clipped=7.0 2023-04-26 10:25:14,557 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=1043.0, num_to_drop=2, layers_to_drop={0, 2} 2023-04-26 10:25:18,745 INFO [finetune.py:976] (4/7) Epoch 1, batch 1050, loss[loss=1.046, simple_loss=0.7466, pruned_loss=0.8652, over 4816.00 frames. ], tot_loss[loss=1.058, simple_loss=0.7668, pruned_loss=0.9499, over 947931.98 frames. ], batch size: 40, lr: 4.00e-03, grad_scale: 0.015625 2023-04-26 10:25:42,488 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=1091.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 10:25:48,290 INFO [finetune.py:976] (4/7) Epoch 1, batch 1100, loss[loss=0.8678, simple_loss=0.6121, pruned_loss=0.7124, over 4731.00 frames. ], tot_loss[loss=1.05, simple_loss=0.7574, pruned_loss=0.9261, over 948883.97 frames. ], batch size: 23, lr: 4.00e-03, grad_scale: 0.03125 2023-04-26 10:25:52,173 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=87.25 vs. limit=5.0 2023-04-26 10:25:57,277 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 2.855e+01 3.563e+01 4.845e+01 6.131e+01 1.271e+02, threshold=9.690e+01, percent-clipped=11.0 2023-04-26 10:26:14,659 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6859, 1.1840, 0.7185, 1.3712, 1.2000, 1.7351, 1.0443, 3.8633], device='cuda:4'), covar=tensor([0.0132, 0.0211, 0.0283, 0.0290, 0.0217, 0.0148, 0.0212, 0.0066], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0043, 0.0043, 0.0048, 0.0044, 0.0041, 0.0042, 0.0068], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0014, 0.0015, 0.0014, 0.0013, 0.0014, 0.0018], device='cuda:4') 2023-04-26 10:26:17,717 INFO [finetune.py:976] (4/7) Epoch 1, batch 1150, loss[loss=1.044, simple_loss=0.7365, pruned_loss=0.84, over 4893.00 frames. ], tot_loss[loss=1.044, simple_loss=0.7491, pruned_loss=0.9046, over 949100.47 frames. ], batch size: 43, lr: 4.00e-03, grad_scale: 0.03125 2023-04-26 10:26:18,901 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=1153.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 10:26:46,977 INFO [finetune.py:976] (4/7) Epoch 1, batch 1200, loss[loss=0.9754, simple_loss=0.6935, pruned_loss=0.7652, over 4810.00 frames. ], tot_loss[loss=1.032, simple_loss=0.7371, pruned_loss=0.8803, over 950857.58 frames. ], batch size: 40, lr: 4.00e-03, grad_scale: 0.0625 2023-04-26 10:26:48,593 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=1203.0, num_to_drop=2, layers_to_drop={0, 1} 2023-04-26 10:26:54,332 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=1214.0, num_to_drop=2, layers_to_drop={2, 3} 2023-04-26 10:26:56,333 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 2.920e+01 3.405e+01 3.864e+01 4.670e+01 1.171e+02, threshold=7.728e+01, percent-clipped=2.0 2023-04-26 10:26:56,441 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=1218.0, num_to_drop=2, layers_to_drop={1, 2} 2023-04-26 10:26:59,577 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=1224.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 10:27:19,101 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=1246.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 10:27:21,700 INFO [finetune.py:976] (4/7) Epoch 1, batch 1250, loss[loss=1.051, simple_loss=0.7336, pruned_loss=0.8255, over 4828.00 frames. ], tot_loss[loss=1.018, simple_loss=0.7239, pruned_loss=0.8549, over 951837.27 frames. ], batch size: 38, lr: 4.00e-03, grad_scale: 0.0625 2023-04-26 10:27:40,182 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=1266.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 10:27:49,531 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=1272.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 10:27:50,803 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-26 10:28:15,540 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=1294.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 10:28:17,431 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.76 vs. limit=5.0 2023-04-26 10:28:24,918 INFO [finetune.py:976] (4/7) Epoch 1, batch 1300, loss[loss=1.033, simple_loss=0.714, pruned_loss=0.8045, over 4756.00 frames. ], tot_loss[loss=1.007, simple_loss=0.7124, pruned_loss=0.8326, over 952079.50 frames. ], batch size: 26, lr: 4.00e-03, grad_scale: 0.125 2023-04-26 10:28:40,244 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 3.024e+01 3.776e+01 4.391e+01 5.836e+01 1.199e+02, threshold=8.782e+01, percent-clipped=8.0 2023-04-26 10:29:00,690 INFO [finetune.py:976] (4/7) Epoch 1, batch 1350, loss[loss=0.922, simple_loss=0.6327, pruned_loss=0.7105, over 4809.00 frames. ], tot_loss[loss=1.007, simple_loss=0.7091, pruned_loss=0.8203, over 953948.48 frames. ], batch size: 25, lr: 4.00e-03, grad_scale: 0.125 2023-04-26 10:29:04,088 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.8006, 4.7936, 3.1452, 5.3479, 4.5662, 4.7748, 1.7374, 4.6299], device='cuda:4'), covar=tensor([0.1325, 0.0598, 0.1989, 0.0863, 0.1665, 0.1334, 0.5439, 0.1741], device='cuda:4'), in_proj_covar=tensor([0.0264, 0.0236, 0.0289, 0.0332, 0.0328, 0.0272, 0.0292, 0.0285], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 10:29:18,093 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.68 vs. limit=2.0 2023-04-26 10:29:49,682 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.1338, 1.1406, 2.0826, 2.1175, 2.1684, 2.0136, 2.3948, 1.6426], device='cuda:4'), covar=tensor([0.0363, 0.1078, 0.0460, 0.0693, 0.0557, 0.0768, 0.0553, 0.0744], device='cuda:4'), in_proj_covar=tensor([0.0398, 0.0405, 0.0390, 0.0362, 0.0423, 0.0450, 0.0387, 0.0428], device='cuda:4'), out_proj_covar=tensor([8.7512e-05, 8.7444e-05, 8.4152e-05, 7.5837e-05, 9.1176e-05, 9.8895e-05, 8.4961e-05, 9.2649e-05], device='cuda:4') 2023-04-26 10:29:49,691 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2154, 2.6001, 1.8506, 1.9359, 1.5496, 1.4877, 1.9440, 1.5518], device='cuda:4'), covar=tensor([0.1911, 0.2637, 0.1794, 0.3007, 0.3156, 0.3398, 0.1269, 0.2135], device='cuda:4'), in_proj_covar=tensor([0.0224, 0.0256, 0.0225, 0.0248, 0.0263, 0.0220, 0.0214, 0.0237], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-26 10:29:52,996 INFO [finetune.py:976] (4/7) Epoch 1, batch 1400, loss[loss=1.024, simple_loss=0.6987, pruned_loss=0.7804, over 4765.00 frames. ], tot_loss[loss=1.019, simple_loss=0.7133, pruned_loss=0.8176, over 955706.42 frames. ], batch size: 27, lr: 4.00e-03, grad_scale: 0.25 2023-04-26 10:30:14,313 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 3.571e+01 4.693e+01 6.414e+01 7.905e+01 1.778e+02, threshold=1.283e+02, percent-clipped=17.0 2023-04-26 10:30:34,165 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=1434.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 10:30:55,953 INFO [finetune.py:976] (4/7) Epoch 1, batch 1450, loss[loss=1.007, simple_loss=0.689, pruned_loss=0.7539, over 4873.00 frames. ], tot_loss[loss=1.023, simple_loss=0.7126, pruned_loss=0.8097, over 956216.53 frames. ], batch size: 31, lr: 4.00e-03, grad_scale: 0.25 2023-04-26 10:31:42,610 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=1495.0, num_to_drop=2, layers_to_drop={0, 3} 2023-04-26 10:31:53,100 INFO [finetune.py:976] (4/7) Epoch 1, batch 1500, loss[loss=0.9875, simple_loss=0.6844, pruned_loss=0.722, over 4819.00 frames. ], tot_loss[loss=1.018, simple_loss=0.7074, pruned_loss=0.7947, over 956895.87 frames. ], batch size: 47, lr: 4.00e-03, grad_scale: 0.5 2023-04-26 10:31:59,746 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=1503.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 10:32:03,094 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=1509.0, num_to_drop=2, layers_to_drop={1, 3} 2023-04-26 10:32:13,079 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 3.937e+01 5.249e+01 6.518e+01 8.001e+01 1.317e+02, threshold=1.304e+02, percent-clipped=3.0 2023-04-26 10:32:48,627 INFO [finetune.py:976] (4/7) Epoch 1, batch 1550, loss[loss=0.9512, simple_loss=0.6602, pruned_loss=0.6855, over 4774.00 frames. ], tot_loss[loss=1.009, simple_loss=0.7006, pruned_loss=0.7755, over 956835.98 frames. ], batch size: 28, lr: 4.00e-03, grad_scale: 0.5 2023-04-26 10:32:48,698 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=1551.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 10:33:08,370 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=1566.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 10:33:50,408 INFO [finetune.py:976] (4/7) Epoch 1, batch 1600, loss[loss=0.8798, simple_loss=0.625, pruned_loss=0.6155, over 4799.00 frames. ], tot_loss[loss=0.9882, simple_loss=0.6886, pruned_loss=0.7471, over 957515.11 frames. ], batch size: 29, lr: 4.00e-03, grad_scale: 1.0 2023-04-26 10:34:04,244 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=1611.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 10:34:13,528 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 5.628e+01 9.315e+01 1.324e+02 1.713e+02 3.757e+02, threshold=2.648e+02, percent-clipped=49.0 2023-04-26 10:34:14,213 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3623, 2.2619, 1.2141, 1.2655, 1.0928, 1.0874, 1.1169, 0.9905], device='cuda:4'), covar=tensor([0.3514, 0.2403, 0.3407, 0.4359, 0.4802, 0.5542, 0.2507, 0.4108], device='cuda:4'), in_proj_covar=tensor([0.0228, 0.0258, 0.0230, 0.0253, 0.0269, 0.0227, 0.0219, 0.0241], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0003, 0.0004], device='cuda:4') 2023-04-26 10:34:23,564 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=19.94 vs. limit=5.0 2023-04-26 10:34:24,076 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=1627.0, num_to_drop=2, layers_to_drop={1, 3} 2023-04-26 10:34:48,566 INFO [finetune.py:976] (4/7) Epoch 1, batch 1650, loss[loss=0.7939, simple_loss=0.5864, pruned_loss=0.5333, over 4906.00 frames. ], tot_loss[loss=0.9582, simple_loss=0.6718, pruned_loss=0.7118, over 958713.35 frames. ], batch size: 36, lr: 4.00e-03, grad_scale: 1.0 2023-04-26 10:35:08,308 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=1672.0, num_to_drop=2, layers_to_drop={0, 3} 2023-04-26 10:35:36,516 INFO [finetune.py:976] (4/7) Epoch 1, batch 1700, loss[loss=0.8468, simple_loss=0.6277, pruned_loss=0.5616, over 4822.00 frames. ], tot_loss[loss=0.9243, simple_loss=0.6537, pruned_loss=0.674, over 957155.54 frames. ], batch size: 33, lr: 4.00e-03, grad_scale: 1.0 2023-04-26 10:35:59,302 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.556e+01 1.723e+02 2.061e+02 2.467e+02 4.417e+02, threshold=4.123e+02, percent-clipped=15.0 2023-04-26 10:36:24,060 INFO [finetune.py:976] (4/7) Epoch 1, batch 1750, loss[loss=0.8166, simple_loss=0.6132, pruned_loss=0.5313, over 4183.00 frames. ], tot_loss[loss=0.9012, simple_loss=0.6456, pruned_loss=0.6432, over 956947.96 frames. ], batch size: 66, lr: 4.00e-03, grad_scale: 1.0 2023-04-26 10:36:49,401 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.90 vs. limit=2.0 2023-04-26 10:36:58,579 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=1790.0, num_to_drop=1, layers_to_drop={3} 2023-04-26 10:37:11,175 INFO [finetune.py:976] (4/7) Epoch 1, batch 1800, loss[loss=0.7501, simple_loss=0.5936, pruned_loss=0.4654, over 4819.00 frames. ], tot_loss[loss=0.8794, simple_loss=0.6397, pruned_loss=0.6138, over 955977.20 frames. ], batch size: 39, lr: 4.00e-03, grad_scale: 1.0 2023-04-26 10:37:21,935 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=1809.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 10:37:28,282 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.499e+02 2.323e+02 2.803e+02 3.399e+02 5.478e+02, threshold=5.607e+02, percent-clipped=9.0 2023-04-26 10:37:40,501 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=1838.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 10:37:47,701 INFO [finetune.py:976] (4/7) Epoch 1, batch 1850, loss[loss=0.745, simple_loss=0.596, pruned_loss=0.4554, over 4869.00 frames. ], tot_loss[loss=0.8446, simple_loss=0.6247, pruned_loss=0.5766, over 956358.83 frames. ], batch size: 34, lr: 4.00e-03, grad_scale: 1.0 2023-04-26 10:37:51,631 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=1857.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 10:37:51,661 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1166, 3.2329, 1.1006, 1.6920, 1.4478, 2.1875, 1.9239, 0.9881], device='cuda:4'), covar=tensor([0.1819, 0.0933, 0.2021, 0.1490, 0.1498, 0.1170, 0.1609, 0.2107], device='cuda:4'), in_proj_covar=tensor([0.0128, 0.0275, 0.0152, 0.0136, 0.0148, 0.0168, 0.0136, 0.0135], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 10:37:51,680 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=1857.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 10:38:00,456 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=1870.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 10:38:17,755 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=1899.0, num_to_drop=2, layers_to_drop={1, 3} 2023-04-26 10:38:18,742 INFO [finetune.py:976] (4/7) Epoch 1, batch 1900, loss[loss=0.6747, simple_loss=0.5637, pruned_loss=0.3966, over 4820.00 frames. ], tot_loss[loss=0.81, simple_loss=0.6097, pruned_loss=0.541, over 955252.85 frames. ], batch size: 39, lr: 4.00e-03, grad_scale: 1.0 2023-04-26 10:38:28,760 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.629e+02 2.504e+02 3.040e+02 3.671e+02 6.110e+02, threshold=6.080e+02, percent-clipped=2.0 2023-04-26 10:38:28,865 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=1918.0, num_to_drop=2, layers_to_drop={2, 3} 2023-04-26 10:38:32,527 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=1922.0, num_to_drop=1, layers_to_drop={3} 2023-04-26 10:38:39,000 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=1931.0, num_to_drop=2, layers_to_drop={0, 1} 2023-04-26 10:38:50,105 INFO [finetune.py:976] (4/7) Epoch 1, batch 1950, loss[loss=0.6377, simple_loss=0.532, pruned_loss=0.3734, over 4916.00 frames. ], tot_loss[loss=0.7707, simple_loss=0.5904, pruned_loss=0.504, over 954783.84 frames. ], batch size: 37, lr: 4.00e-03, grad_scale: 1.0 2023-04-26 10:38:59,747 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=1967.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 10:39:22,528 INFO [finetune.py:976] (4/7) Epoch 1, batch 2000, loss[loss=0.5679, simple_loss=0.4799, pruned_loss=0.3279, over 4792.00 frames. ], tot_loss[loss=0.7308, simple_loss=0.5686, pruned_loss=0.4689, over 954643.46 frames. ], batch size: 29, lr: 4.00e-03, grad_scale: 2.0 2023-04-26 10:39:39,765 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.592e+02 2.794e+02 3.334e+02 3.918e+02 6.535e+02, threshold=6.668e+02, percent-clipped=3.0 2023-04-26 10:39:40,448 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.5794, 4.4636, 3.1080, 5.1253, 4.4244, 4.5571, 1.5729, 4.3571], device='cuda:4'), covar=tensor([0.1333, 0.0758, 0.3218, 0.0875, 0.2741, 0.1494, 0.5912, 0.1786], device='cuda:4'), in_proj_covar=tensor([0.0261, 0.0237, 0.0288, 0.0328, 0.0327, 0.0271, 0.0285, 0.0285], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 10:40:01,843 INFO [finetune.py:976] (4/7) Epoch 1, batch 2050, loss[loss=0.5492, simple_loss=0.4663, pruned_loss=0.316, over 4484.00 frames. ], tot_loss[loss=0.6883, simple_loss=0.5456, pruned_loss=0.433, over 954025.11 frames. ], batch size: 20, lr: 4.00e-03, grad_scale: 2.0 2023-04-26 10:40:37,623 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=2090.0, num_to_drop=2, layers_to_drop={0, 1} 2023-04-26 10:40:50,242 INFO [finetune.py:976] (4/7) Epoch 1, batch 2100, loss[loss=0.5227, simple_loss=0.4501, pruned_loss=0.2977, over 4755.00 frames. ], tot_loss[loss=0.6546, simple_loss=0.5287, pruned_loss=0.4038, over 954523.33 frames. ], batch size: 26, lr: 4.00e-03, grad_scale: 2.0 2023-04-26 10:41:11,501 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.614e+02 2.534e+02 2.914e+02 3.246e+02 6.149e+02, threshold=5.827e+02, percent-clipped=0.0 2023-04-26 10:41:32,455 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=2138.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 10:41:41,039 INFO [finetune.py:976] (4/7) Epoch 1, batch 2150, loss[loss=0.5424, simple_loss=0.4985, pruned_loss=0.2931, over 4793.00 frames. ], tot_loss[loss=0.6323, simple_loss=0.5203, pruned_loss=0.3827, over 953370.13 frames. ], batch size: 29, lr: 4.00e-03, grad_scale: 4.0 2023-04-26 10:41:51,008 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=5.70 vs. limit=5.0 2023-04-26 10:41:53,216 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4803, 2.0326, 1.5384, 1.9235, 1.4370, 1.4580, 1.9558, 1.3925], device='cuda:4'), covar=tensor([0.2052, 0.1332, 0.1543, 0.1511, 0.2801, 0.1627, 0.1547, 0.2217], device='cuda:4'), in_proj_covar=tensor([0.0274, 0.0287, 0.0211, 0.0267, 0.0283, 0.0244, 0.0246, 0.0256], device='cuda:4'), out_proj_covar=tensor([1.1187e-04, 1.1748e-04, 8.6325e-05, 1.0836e-04, 1.1771e-04, 9.8652e-05, 1.0227e-04, 1.0423e-04], device='cuda:4') 2023-04-26 10:42:25,881 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=2194.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 10:42:30,518 INFO [finetune.py:976] (4/7) Epoch 1, batch 2200, loss[loss=0.457, simple_loss=0.4202, pruned_loss=0.2469, over 4666.00 frames. ], tot_loss[loss=0.612, simple_loss=0.5128, pruned_loss=0.3638, over 953754.78 frames. ], batch size: 23, lr: 4.00e-03, grad_scale: 4.0 2023-04-26 10:42:37,619 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=2213.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 10:42:40,431 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.685e+02 2.565e+02 3.033e+02 3.519e+02 5.393e+02, threshold=6.065e+02, percent-clipped=0.0 2023-04-26 10:42:42,846 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=2222.0, num_to_drop=2, layers_to_drop={0, 1} 2023-04-26 10:42:42,865 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1866, 0.9363, 1.5951, 1.5470, 1.1307, 0.9293, 1.0466, 0.6245], device='cuda:4'), covar=tensor([0.1917, 0.1491, 0.0682, 0.1056, 0.1740, 0.2265, 0.1434, 0.1913], device='cuda:4'), in_proj_covar=tensor([0.0080, 0.0089, 0.0082, 0.0086, 0.0103, 0.0104, 0.0102, 0.0089], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-26 10:42:45,137 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=2226.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 10:43:02,985 INFO [finetune.py:976] (4/7) Epoch 1, batch 2250, loss[loss=0.6205, simple_loss=0.5364, pruned_loss=0.3523, over 4817.00 frames. ], tot_loss[loss=0.5924, simple_loss=0.5045, pruned_loss=0.3465, over 954505.57 frames. ], batch size: 40, lr: 4.00e-03, grad_scale: 4.0 2023-04-26 10:43:12,379 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=2267.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 10:43:14,136 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=2270.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 10:43:57,382 INFO [finetune.py:976] (4/7) Epoch 1, batch 2300, loss[loss=0.4091, simple_loss=0.41, pruned_loss=0.2041, over 4751.00 frames. ], tot_loss[loss=0.5684, simple_loss=0.493, pruned_loss=0.3269, over 955814.47 frames. ], batch size: 27, lr: 4.00e-03, grad_scale: 4.0 2023-04-26 10:44:17,313 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=2315.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 10:44:19,012 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.752e+02 2.423e+02 2.781e+02 3.392e+02 7.688e+02, threshold=5.562e+02, percent-clipped=1.0 2023-04-26 10:44:28,648 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=12.09 vs. limit=5.0 2023-04-26 10:44:43,418 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=2340.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 10:44:51,818 INFO [finetune.py:976] (4/7) Epoch 1, batch 2350, loss[loss=0.4659, simple_loss=0.4439, pruned_loss=0.244, over 4739.00 frames. ], tot_loss[loss=0.5438, simple_loss=0.4784, pruned_loss=0.3085, over 955410.94 frames. ], batch size: 26, lr: 4.00e-03, grad_scale: 4.0 2023-04-26 10:45:08,896 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=6.91 vs. limit=5.0 2023-04-26 10:45:32,236 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5427, 1.4404, 1.6652, 1.7999, 1.9259, 1.3702, 1.0665, 1.5312], device='cuda:4'), covar=tensor([0.1299, 0.1451, 0.1012, 0.0908, 0.0754, 0.1213, 0.1434, 0.0954], device='cuda:4'), in_proj_covar=tensor([0.0201, 0.0210, 0.0189, 0.0169, 0.0169, 0.0186, 0.0164, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 10:45:40,785 INFO [finetune.py:976] (4/7) Epoch 1, batch 2400, loss[loss=0.4321, simple_loss=0.4127, pruned_loss=0.2258, over 4767.00 frames. ], tot_loss[loss=0.5217, simple_loss=0.4651, pruned_loss=0.2922, over 954866.15 frames. ], batch size: 27, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 10:45:40,918 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=2401.0, num_to_drop=2, layers_to_drop={2, 3} 2023-04-26 10:45:51,690 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.659e+02 2.375e+02 2.859e+02 3.332e+02 6.118e+02, threshold=5.719e+02, percent-clipped=1.0 2023-04-26 10:46:11,803 INFO [finetune.py:976] (4/7) Epoch 1, batch 2450, loss[loss=0.4326, simple_loss=0.4159, pruned_loss=0.2247, over 4927.00 frames. ], tot_loss[loss=0.5028, simple_loss=0.4533, pruned_loss=0.2785, over 954466.95 frames. ], batch size: 33, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 10:46:11,964 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=5.21 vs. limit=5.0 2023-04-26 10:46:32,698 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4693, 1.3832, 3.9129, 3.7015, 3.5450, 3.7060, 3.8032, 3.4802], device='cuda:4'), covar=tensor([0.6134, 0.5247, 0.1127, 0.1467, 0.1084, 0.1390, 0.1201, 0.1413], device='cuda:4'), in_proj_covar=tensor([0.0342, 0.0316, 0.0453, 0.0463, 0.0379, 0.0432, 0.0347, 0.0401], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0002], device='cuda:4') 2023-04-26 10:46:39,148 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=2494.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 10:46:44,084 INFO [finetune.py:976] (4/7) Epoch 1, batch 2500, loss[loss=0.5449, simple_loss=0.5086, pruned_loss=0.2906, over 4730.00 frames. ], tot_loss[loss=0.494, simple_loss=0.4496, pruned_loss=0.2711, over 953295.98 frames. ], batch size: 59, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 10:46:48,771 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6802, 0.9660, 1.0175, 1.4004, 1.5681, 1.4818, 1.3015, 0.9792], device='cuda:4'), covar=tensor([0.1041, 0.2724, 0.3090, 0.1122, 0.1234, 0.1475, 0.1754, 0.2869], device='cuda:4'), in_proj_covar=tensor([0.0354, 0.0370, 0.0363, 0.0325, 0.0381, 0.0406, 0.0350, 0.0391], device='cuda:4'), out_proj_covar=tensor([7.7739e-05, 7.9745e-05, 7.8367e-05, 6.7819e-05, 8.1468e-05, 8.8865e-05, 7.6505e-05, 8.4720e-05], device='cuda:4') 2023-04-26 10:46:52,659 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=2513.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 10:46:56,031 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.542e+02 2.485e+02 2.798e+02 3.305e+02 6.030e+02, threshold=5.597e+02, percent-clipped=1.0 2023-04-26 10:47:00,940 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=2526.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 10:47:03,969 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5098, 1.2047, 1.9902, 2.0535, 1.3612, 1.1073, 1.4503, 1.0534], device='cuda:4'), covar=tensor([0.1842, 0.1372, 0.0567, 0.0619, 0.1478, 0.2426, 0.1464, 0.1952], device='cuda:4'), in_proj_covar=tensor([0.0080, 0.0088, 0.0081, 0.0085, 0.0102, 0.0104, 0.0102, 0.0089], device='cuda:4'), out_proj_covar=tensor([0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-26 10:47:10,443 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=2542.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 10:47:16,162 INFO [finetune.py:976] (4/7) Epoch 1, batch 2550, loss[loss=0.5266, simple_loss=0.4866, pruned_loss=0.2833, over 4286.00 frames. ], tot_loss[loss=0.4919, simple_loss=0.4529, pruned_loss=0.2668, over 955466.67 frames. ], batch size: 65, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 10:47:35,366 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=2561.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 10:47:50,047 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=2574.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 10:48:06,448 INFO [finetune.py:976] (4/7) Epoch 1, batch 2600, loss[loss=0.4113, simple_loss=0.413, pruned_loss=0.2048, over 4765.00 frames. ], tot_loss[loss=0.4823, simple_loss=0.4489, pruned_loss=0.2589, over 955089.16 frames. ], batch size: 27, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 10:48:06,801 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-26 10:48:18,462 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.395e+02 2.465e+02 2.854e+02 3.439e+02 6.010e+02, threshold=5.707e+02, percent-clipped=1.0 2023-04-26 10:48:21,243 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-04-26 10:48:38,252 INFO [finetune.py:976] (4/7) Epoch 1, batch 2650, loss[loss=0.4242, simple_loss=0.4232, pruned_loss=0.2126, over 4831.00 frames. ], tot_loss[loss=0.4728, simple_loss=0.4444, pruned_loss=0.2514, over 954635.08 frames. ], batch size: 47, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 10:49:01,810 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=2685.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 10:49:03,715 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.73 vs. limit=2.0 2023-04-26 10:49:14,012 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=2696.0, num_to_drop=1, layers_to_drop={3} 2023-04-26 10:49:16,817 INFO [finetune.py:976] (4/7) Epoch 1, batch 2700, loss[loss=0.4687, simple_loss=0.4214, pruned_loss=0.258, over 4105.00 frames. ], tot_loss[loss=0.4613, simple_loss=0.438, pruned_loss=0.243, over 954176.77 frames. ], batch size: 65, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 10:49:39,497 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.764e+02 2.585e+02 2.984e+02 3.485e+02 4.746e+02, threshold=5.968e+02, percent-clipped=0.0 2023-04-26 10:50:04,145 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=2746.0, num_to_drop=2, layers_to_drop={0, 2} 2023-04-26 10:50:13,027 INFO [finetune.py:976] (4/7) Epoch 1, batch 2750, loss[loss=0.4121, simple_loss=0.4151, pruned_loss=0.2046, over 4824.00 frames. ], tot_loss[loss=0.4488, simple_loss=0.4292, pruned_loss=0.2347, over 954320.34 frames. ], batch size: 41, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 10:50:24,480 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4133, 1.4248, 1.6331, 1.6847, 1.7883, 1.2665, 0.8475, 1.5061], device='cuda:4'), covar=tensor([0.1321, 0.1462, 0.0927, 0.1080, 0.0746, 0.1281, 0.1532, 0.0970], device='cuda:4'), in_proj_covar=tensor([0.0204, 0.0211, 0.0191, 0.0173, 0.0173, 0.0189, 0.0167, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 10:51:16,384 INFO [finetune.py:976] (4/7) Epoch 1, batch 2800, loss[loss=0.3344, simple_loss=0.3596, pruned_loss=0.1546, over 4912.00 frames. ], tot_loss[loss=0.4366, simple_loss=0.4205, pruned_loss=0.2267, over 955989.91 frames. ], batch size: 32, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 10:51:38,598 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.584e+02 2.321e+02 2.765e+02 3.400e+02 8.366e+02, threshold=5.531e+02, percent-clipped=2.0 2023-04-26 10:51:49,407 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=2824.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 10:52:12,951 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6701, 3.1755, 1.2813, 2.0505, 2.1218, 2.4316, 2.2765, 1.3477], device='cuda:4'), covar=tensor([0.0835, 0.0723, 0.1395, 0.0925, 0.0639, 0.0739, 0.0960, 0.1364], device='cuda:4'), in_proj_covar=tensor([0.0121, 0.0266, 0.0147, 0.0130, 0.0142, 0.0163, 0.0128, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 10:52:24,239 INFO [finetune.py:976] (4/7) Epoch 1, batch 2850, loss[loss=0.3832, simple_loss=0.3702, pruned_loss=0.198, over 4720.00 frames. ], tot_loss[loss=0.4267, simple_loss=0.4134, pruned_loss=0.2203, over 953665.80 frames. ], batch size: 23, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 10:52:26,735 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3593, 1.6292, 1.1238, 1.6828, 1.4632, 1.2024, 1.5665, 1.0409], device='cuda:4'), covar=tensor([0.2176, 0.1840, 0.1834, 0.1806, 0.3335, 0.1716, 0.1988, 0.3069], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0306, 0.0226, 0.0286, 0.0299, 0.0261, 0.0261, 0.0276], device='cuda:4'), out_proj_covar=tensor([1.1973e-04, 1.2529e-04, 9.2849e-05, 1.1567e-04, 1.2419e-04, 1.0573e-04, 1.0831e-04, 1.1240e-04], device='cuda:4') 2023-04-26 10:53:07,948 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=2885.0, num_to_drop=2, layers_to_drop={1, 3} 2023-04-26 10:53:16,962 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0554, 2.5407, 1.8842, 2.5583, 2.0684, 1.9351, 2.3260, 1.6530], device='cuda:4'), covar=tensor([0.1992, 0.1417, 0.1627, 0.1393, 0.2544, 0.1663, 0.1780, 0.2702], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0307, 0.0227, 0.0286, 0.0300, 0.0262, 0.0262, 0.0277], device='cuda:4'), out_proj_covar=tensor([1.2002e-04, 1.2565e-04, 9.3114e-05, 1.1591e-04, 1.2466e-04, 1.0603e-04, 1.0869e-04, 1.1281e-04], device='cuda:4') 2023-04-26 10:53:29,241 INFO [finetune.py:976] (4/7) Epoch 1, batch 2900, loss[loss=0.3452, simple_loss=0.3702, pruned_loss=0.1601, over 4823.00 frames. ], tot_loss[loss=0.4279, simple_loss=0.4158, pruned_loss=0.2203, over 952608.70 frames. ], batch size: 51, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 10:53:46,010 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.613e+02 2.467e+02 2.930e+02 3.474e+02 6.951e+02, threshold=5.860e+02, percent-clipped=1.0 2023-04-26 10:53:59,784 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=2937.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 10:54:01,477 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8637, 2.2577, 1.3394, 1.5450, 2.2714, 1.7821, 1.8460, 1.8464], device='cuda:4'), covar=tensor([0.0560, 0.0486, 0.0444, 0.0658, 0.0300, 0.0621, 0.0564, 0.0783], device='cuda:4'), in_proj_covar=tensor([0.0032, 0.0026, 0.0024, 0.0031, 0.0021, 0.0031, 0.0030, 0.0033], device='cuda:4'), out_proj_covar=tensor([0.0048, 0.0043, 0.0037, 0.0049, 0.0036, 0.0048, 0.0047, 0.0051], device='cuda:4') 2023-04-26 10:54:08,440 INFO [finetune.py:976] (4/7) Epoch 1, batch 2950, loss[loss=0.3899, simple_loss=0.3978, pruned_loss=0.191, over 4776.00 frames. ], tot_loss[loss=0.4259, simple_loss=0.4168, pruned_loss=0.2177, over 951928.19 frames. ], batch size: 27, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 10:54:16,937 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=5.69 vs. limit=5.0 2023-04-26 10:54:32,085 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=2.00 vs. limit=2.0 2023-04-26 10:54:37,952 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=2996.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 10:54:39,132 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=2998.0, num_to_drop=2, layers_to_drop={0, 1} 2023-04-26 10:54:40,896 INFO [finetune.py:976] (4/7) Epoch 1, batch 3000, loss[loss=0.4392, simple_loss=0.4374, pruned_loss=0.2205, over 4911.00 frames. ], tot_loss[loss=0.4217, simple_loss=0.4153, pruned_loss=0.2142, over 951458.83 frames. ], batch size: 38, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 10:54:40,896 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-26 10:54:51,383 INFO [finetune.py:1010] (4/7) Epoch 1, validation: loss=0.4217, simple_loss=0.4614, pruned_loss=0.191, over 2265189.00 frames. 2023-04-26 10:54:51,383 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 5735MB 2023-04-26 10:55:01,571 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.562e+02 2.517e+02 2.904e+02 3.818e+02 1.122e+03, threshold=5.808e+02, percent-clipped=2.0 2023-04-26 10:55:02,290 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9402, 2.0239, 1.7124, 1.7736, 2.1095, 1.6224, 2.7275, 1.3886], device='cuda:4'), covar=tensor([0.3185, 0.1034, 0.2761, 0.1839, 0.1314, 0.2094, 0.0636, 0.3169], device='cuda:4'), in_proj_covar=tensor([0.0303, 0.0313, 0.0380, 0.0325, 0.0360, 0.0336, 0.0350, 0.0364], device='cuda:4'), out_proj_covar=tensor([9.3763e-05, 9.6476e-05, 1.1777e-04, 1.0144e-04, 1.1046e-04, 1.0258e-04, 1.0566e-04, 1.1280e-04], device='cuda:4') 2023-04-26 10:55:02,293 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=3019.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 10:55:06,742 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.95 vs. limit=5.0 2023-04-26 10:55:09,333 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.7557, 3.6979, 2.9431, 4.2467, 3.7834, 3.6522, 1.7735, 3.6320], device='cuda:4'), covar=tensor([0.1390, 0.0962, 0.2669, 0.1334, 0.2071, 0.1627, 0.4915, 0.1853], device='cuda:4'), in_proj_covar=tensor([0.0261, 0.0233, 0.0283, 0.0326, 0.0321, 0.0271, 0.0284, 0.0284], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 10:55:16,272 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=3041.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 10:55:18,054 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=3044.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 10:55:19,340 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=7.28 vs. limit=5.0 2023-04-26 10:55:23,631 INFO [finetune.py:976] (4/7) Epoch 1, batch 3050, loss[loss=0.4359, simple_loss=0.4254, pruned_loss=0.2232, over 4806.00 frames. ], tot_loss[loss=0.4193, simple_loss=0.4154, pruned_loss=0.2118, over 952908.68 frames. ], batch size: 55, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 10:55:30,770 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1075, 2.4030, 0.9196, 1.4261, 1.5932, 1.2176, 3.1121, 1.6565], device='cuda:4'), covar=tensor([0.0698, 0.0504, 0.0783, 0.1161, 0.0579, 0.0998, 0.0178, 0.0641], device='cuda:4'), in_proj_covar=tensor([0.0057, 0.0072, 0.0053, 0.0049, 0.0054, 0.0055, 0.0087, 0.0053], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0011, 0.0007], device='cuda:4') 2023-04-26 10:55:42,255 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=3080.0, num_to_drop=2, layers_to_drop={1, 3} 2023-04-26 10:55:54,206 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.71 vs. limit=2.0 2023-04-26 10:55:56,457 INFO [finetune.py:976] (4/7) Epoch 1, batch 3100, loss[loss=0.4346, simple_loss=0.4083, pruned_loss=0.2304, over 4836.00 frames. ], tot_loss[loss=0.4094, simple_loss=0.4085, pruned_loss=0.2053, over 954371.55 frames. ], batch size: 47, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 10:56:03,445 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2883, 1.5094, 1.3945, 1.9840, 2.3224, 2.0576, 1.9237, 1.4672], device='cuda:4'), covar=tensor([0.1543, 0.2312, 0.3100, 0.1901, 0.0908, 0.1418, 0.1984, 0.2120], device='cuda:4'), in_proj_covar=tensor([0.0343, 0.0356, 0.0354, 0.0316, 0.0368, 0.0390, 0.0337, 0.0376], device='cuda:4'), out_proj_covar=tensor([7.5182e-05, 7.6612e-05, 7.6375e-05, 6.5942e-05, 7.8437e-05, 8.5393e-05, 7.3651e-05, 8.1238e-05], device='cuda:4') 2023-04-26 10:56:12,964 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.494e+02 2.339e+02 2.765e+02 3.275e+02 7.045e+02, threshold=5.531e+02, percent-clipped=2.0 2023-04-26 10:56:14,915 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=7.00 vs. limit=5.0 2023-04-26 10:56:55,557 INFO [finetune.py:976] (4/7) Epoch 1, batch 3150, loss[loss=0.3783, simple_loss=0.3736, pruned_loss=0.1915, over 4805.00 frames. ], tot_loss[loss=0.4026, simple_loss=0.4029, pruned_loss=0.2012, over 955162.09 frames. ], batch size: 29, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 10:56:57,847 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3584, 2.6503, 1.0148, 1.6814, 1.9916, 1.2876, 3.7418, 1.9599], device='cuda:4'), covar=tensor([0.0666, 0.0689, 0.0948, 0.1123, 0.0609, 0.1025, 0.0186, 0.0620], device='cuda:4'), in_proj_covar=tensor([0.0057, 0.0072, 0.0053, 0.0049, 0.0055, 0.0055, 0.0088, 0.0053], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 10:57:32,979 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=3180.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 10:57:46,343 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.64 vs. limit=5.0 2023-04-26 10:57:46,511 INFO [finetune.py:976] (4/7) Epoch 1, batch 3200, loss[loss=0.392, simple_loss=0.3973, pruned_loss=0.1933, over 4714.00 frames. ], tot_loss[loss=0.3927, simple_loss=0.395, pruned_loss=0.1952, over 953487.53 frames. ], batch size: 23, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 10:57:55,131 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.47 vs. limit=2.0 2023-04-26 10:58:16,454 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.736e+02 2.453e+02 2.807e+02 3.222e+02 7.994e+02, threshold=5.615e+02, percent-clipped=2.0 2023-04-26 10:58:51,399 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5375, 1.0451, 1.1079, 1.6601, 1.0884, 1.0937, 0.9617, 1.0820], device='cuda:4'), covar=tensor([ 6.3138, 7.8342, 3.9581, 13.4620, 10.8959, 6.5145, 9.7445, 10.9065], device='cuda:4'), in_proj_covar=tensor([0.0265, 0.0285, 0.0224, 0.0350, 0.0245, 0.0235, 0.0280, 0.0233], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 10:58:52,918 INFO [finetune.py:976] (4/7) Epoch 1, batch 3250, loss[loss=0.5194, simple_loss=0.5029, pruned_loss=0.2679, over 4817.00 frames. ], tot_loss[loss=0.39, simple_loss=0.3933, pruned_loss=0.1933, over 953004.36 frames. ], batch size: 38, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 10:59:39,011 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8567, 1.3420, 2.3856, 2.7453, 1.5273, 1.1815, 1.6214, 0.9412], device='cuda:4'), covar=tensor([0.1881, 0.1606, 0.0554, 0.0414, 0.1592, 0.2098, 0.1484, 0.2072], device='cuda:4'), in_proj_covar=tensor([0.0079, 0.0085, 0.0079, 0.0082, 0.0099, 0.0101, 0.0100, 0.0086], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0004, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-26 10:59:47,324 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=3288.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 10:59:50,311 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=3293.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:00:00,688 INFO [finetune.py:976] (4/7) Epoch 1, batch 3300, loss[loss=0.3613, simple_loss=0.3899, pruned_loss=0.1664, over 4858.00 frames. ], tot_loss[loss=0.3907, simple_loss=0.3956, pruned_loss=0.1929, over 953786.60 frames. ], batch size: 44, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 11:00:19,087 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.582e+02 2.400e+02 2.649e+02 3.193e+02 6.501e+02, threshold=5.297e+02, percent-clipped=1.0 2023-04-26 11:00:33,901 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=3341.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:00:38,738 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=3349.0, num_to_drop=2, layers_to_drop={0, 3} 2023-04-26 11:00:39,794 INFO [finetune.py:976] (4/7) Epoch 1, batch 3350, loss[loss=0.3294, simple_loss=0.3413, pruned_loss=0.1587, over 4828.00 frames. ], tot_loss[loss=0.3905, simple_loss=0.3971, pruned_loss=0.1919, over 954087.69 frames. ], batch size: 30, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 11:00:54,105 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5583, 1.9350, 1.5020, 1.8910, 1.4523, 1.4491, 1.7638, 1.2463], device='cuda:4'), covar=tensor([0.2314, 0.1542, 0.1693, 0.1699, 0.3240, 0.2020, 0.2232, 0.3051], device='cuda:4'), in_proj_covar=tensor([0.0295, 0.0310, 0.0230, 0.0289, 0.0300, 0.0265, 0.0265, 0.0281], device='cuda:4'), out_proj_covar=tensor([1.2057e-04, 1.2691e-04, 9.4212e-05, 1.1705e-04, 1.2448e-04, 1.0708e-04, 1.0989e-04, 1.1455e-04], device='cuda:4') 2023-04-26 11:00:57,398 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=3375.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:01:05,771 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=3389.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:01:13,019 INFO [finetune.py:976] (4/7) Epoch 1, batch 3400, loss[loss=0.3446, simple_loss=0.3765, pruned_loss=0.1564, over 4818.00 frames. ], tot_loss[loss=0.3897, simple_loss=0.3978, pruned_loss=0.1909, over 954197.04 frames. ], batch size: 41, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 11:01:23,592 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.622e+02 2.298e+02 2.697e+02 3.198e+02 5.981e+02, threshold=5.394e+02, percent-clipped=4.0 2023-04-26 11:01:57,861 INFO [finetune.py:976] (4/7) Epoch 1, batch 3450, loss[loss=0.3723, simple_loss=0.3952, pruned_loss=0.1747, over 4811.00 frames. ], tot_loss[loss=0.3867, simple_loss=0.396, pruned_loss=0.1887, over 953125.13 frames. ], batch size: 40, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 11:02:35,984 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=3480.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:02:49,059 INFO [finetune.py:976] (4/7) Epoch 1, batch 3500, loss[loss=0.3719, simple_loss=0.3859, pruned_loss=0.179, over 4929.00 frames. ], tot_loss[loss=0.38, simple_loss=0.3904, pruned_loss=0.1848, over 953511.89 frames. ], batch size: 38, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 11:03:00,464 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.505e+02 2.494e+02 2.871e+02 4.319e+02 1.287e+03, threshold=5.742e+02, percent-clipped=13.0 2023-04-26 11:03:07,524 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=3528.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:03:15,089 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.2550, 3.1559, 2.7056, 3.6732, 3.0725, 3.2180, 1.7456, 3.1295], device='cuda:4'), covar=tensor([0.1511, 0.1302, 0.3560, 0.2149, 0.2941, 0.1741, 0.4365, 0.2098], device='cuda:4'), in_proj_covar=tensor([0.0263, 0.0234, 0.0285, 0.0329, 0.0323, 0.0274, 0.0287, 0.0286], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 11:03:24,411 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=3546.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 11:03:27,265 INFO [finetune.py:976] (4/7) Epoch 1, batch 3550, loss[loss=0.4048, simple_loss=0.4033, pruned_loss=0.2032, over 4847.00 frames. ], tot_loss[loss=0.3768, simple_loss=0.3868, pruned_loss=0.1834, over 954520.16 frames. ], batch size: 44, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 11:03:46,806 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=6.32 vs. limit=5.0 2023-04-26 11:04:11,237 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=3593.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:04:16,040 INFO [finetune.py:976] (4/7) Epoch 1, batch 3600, loss[loss=0.3417, simple_loss=0.3614, pruned_loss=0.161, over 4768.00 frames. ], tot_loss[loss=0.3707, simple_loss=0.3819, pruned_loss=0.1798, over 957162.07 frames. ], batch size: 28, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 11:04:19,839 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=3607.0, num_to_drop=2, layers_to_drop={1, 3} 2023-04-26 11:04:26,349 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.597e+02 2.647e+02 3.095e+02 3.748e+02 7.550e+02, threshold=6.190e+02, percent-clipped=4.0 2023-04-26 11:04:42,920 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=3641.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:04:43,959 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-26 11:04:44,785 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=3644.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 11:04:48,963 INFO [finetune.py:976] (4/7) Epoch 1, batch 3650, loss[loss=0.3581, simple_loss=0.383, pruned_loss=0.1666, over 4867.00 frames. ], tot_loss[loss=0.3697, simple_loss=0.3814, pruned_loss=0.179, over 955984.27 frames. ], batch size: 31, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 11:05:04,833 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=3675.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:05:05,432 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=3676.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 11:05:23,024 INFO [finetune.py:976] (4/7) Epoch 1, batch 3700, loss[loss=0.3269, simple_loss=0.34, pruned_loss=0.1569, over 4079.00 frames. ], tot_loss[loss=0.3727, simple_loss=0.3858, pruned_loss=0.1798, over 953871.94 frames. ], batch size: 17, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 11:05:44,462 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.676e+02 2.590e+02 3.040e+02 3.768e+02 5.314e+02, threshold=6.080e+02, percent-clipped=0.0 2023-04-26 11:05:53,053 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=3723.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:06:12,554 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=3737.0, num_to_drop=2, layers_to_drop={1, 3} 2023-04-26 11:06:16,671 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6733, 1.7862, 1.2648, 1.4170, 2.1310, 1.5970, 1.5549, 1.5532], device='cuda:4'), covar=tensor([0.0567, 0.0451, 0.0463, 0.0579, 0.0316, 0.0557, 0.0527, 0.0694], device='cuda:4'), in_proj_covar=tensor([0.0032, 0.0027, 0.0025, 0.0031, 0.0021, 0.0031, 0.0030, 0.0033], device='cuda:4'), out_proj_covar=tensor([0.0048, 0.0043, 0.0038, 0.0049, 0.0037, 0.0048, 0.0047, 0.0051], device='cuda:4') 2023-04-26 11:06:21,955 INFO [finetune.py:976] (4/7) Epoch 1, batch 3750, loss[loss=0.3735, simple_loss=0.398, pruned_loss=0.1745, over 4819.00 frames. ], tot_loss[loss=0.3703, simple_loss=0.3858, pruned_loss=0.1774, over 954049.91 frames. ], batch size: 47, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 11:07:07,664 INFO [finetune.py:976] (4/7) Epoch 1, batch 3800, loss[loss=0.3076, simple_loss=0.3562, pruned_loss=0.1295, over 4889.00 frames. ], tot_loss[loss=0.3699, simple_loss=0.3863, pruned_loss=0.1768, over 954933.95 frames. ], batch size: 32, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 11:07:29,602 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.740e+02 2.601e+02 3.098e+02 3.867e+02 7.221e+02, threshold=6.196e+02, percent-clipped=5.0 2023-04-26 11:08:14,239 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.98 vs. limit=2.0 2023-04-26 11:08:14,574 INFO [finetune.py:976] (4/7) Epoch 1, batch 3850, loss[loss=0.3954, simple_loss=0.3954, pruned_loss=0.1976, over 4873.00 frames. ], tot_loss[loss=0.3642, simple_loss=0.3821, pruned_loss=0.1732, over 955954.85 frames. ], batch size: 34, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 11:09:22,875 INFO [finetune.py:976] (4/7) Epoch 1, batch 3900, loss[loss=0.3271, simple_loss=0.3546, pruned_loss=0.1498, over 4933.00 frames. ], tot_loss[loss=0.3592, simple_loss=0.3771, pruned_loss=0.1706, over 954222.00 frames. ], batch size: 33, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 11:09:29,508 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=3902.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 11:09:31,526 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-04-26 11:09:44,796 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.635e+02 2.527e+02 3.007e+02 3.749e+02 8.787e+02, threshold=6.015e+02, percent-clipped=2.0 2023-04-26 11:09:44,909 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4288, 3.4826, 0.8318, 2.0832, 1.9599, 2.5042, 2.2395, 0.9403], device='cuda:4'), covar=tensor([0.1265, 0.0671, 0.1960, 0.1148, 0.0990, 0.1018, 0.1194, 0.2083], device='cuda:4'), in_proj_covar=tensor([0.0123, 0.0272, 0.0150, 0.0132, 0.0144, 0.0167, 0.0130, 0.0134], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 11:09:49,819 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=3926.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:10:02,239 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=3944.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 11:10:07,406 INFO [finetune.py:976] (4/7) Epoch 1, batch 3950, loss[loss=0.3112, simple_loss=0.3475, pruned_loss=0.1374, over 4824.00 frames. ], tot_loss[loss=0.3528, simple_loss=0.3715, pruned_loss=0.1671, over 955468.08 frames. ], batch size: 38, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 11:10:16,434 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1659, 2.5798, 0.9535, 1.3766, 1.7021, 1.2745, 3.4452, 1.7106], device='cuda:4'), covar=tensor([0.0734, 0.0752, 0.1004, 0.1229, 0.0648, 0.1027, 0.0213, 0.0672], device='cuda:4'), in_proj_covar=tensor([0.0057, 0.0073, 0.0053, 0.0050, 0.0055, 0.0056, 0.0088, 0.0054], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0008, 0.0008, 0.0008, 0.0008, 0.0012, 0.0008], device='cuda:4') 2023-04-26 11:10:30,538 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=3987.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:10:34,041 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=3992.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 11:10:42,148 INFO [finetune.py:976] (4/7) Epoch 1, batch 4000, loss[loss=0.3941, simple_loss=0.4024, pruned_loss=0.193, over 4771.00 frames. ], tot_loss[loss=0.3502, simple_loss=0.3689, pruned_loss=0.1657, over 952242.68 frames. ], batch size: 54, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 11:10:45,726 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8088, 2.5272, 1.6422, 1.5563, 1.2698, 1.3840, 1.5845, 1.1877], device='cuda:4'), covar=tensor([0.2311, 0.2184, 0.3143, 0.3676, 0.3844, 0.2938, 0.2467, 0.3360], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0221, 0.0201, 0.0217, 0.0237, 0.0198, 0.0194, 0.0211], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 11:10:51,173 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6837, 0.9583, 1.2293, 1.8446, 1.3108, 1.0300, 0.9195, 1.3775], device='cuda:4'), covar=tensor([1.2703, 1.3401, 0.6558, 2.2797, 1.3050, 1.0420, 2.3025, 1.2980], device='cuda:4'), in_proj_covar=tensor([0.0255, 0.0277, 0.0217, 0.0335, 0.0236, 0.0228, 0.0271, 0.0223], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 11:10:54,137 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.677e+02 2.386e+02 2.816e+02 3.337e+02 7.046e+02, threshold=5.633e+02, percent-clipped=3.0 2023-04-26 11:11:02,812 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=4032.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 11:11:15,866 INFO [finetune.py:976] (4/7) Epoch 1, batch 4050, loss[loss=0.4572, simple_loss=0.4555, pruned_loss=0.2295, over 4159.00 frames. ], tot_loss[loss=0.3508, simple_loss=0.3708, pruned_loss=0.1654, over 950798.99 frames. ], batch size: 65, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 11:11:49,877 INFO [finetune.py:976] (4/7) Epoch 1, batch 4100, loss[loss=0.3581, simple_loss=0.3926, pruned_loss=0.1618, over 4895.00 frames. ], tot_loss[loss=0.3515, simple_loss=0.3727, pruned_loss=0.1652, over 950767.56 frames. ], batch size: 37, lr: 4.00e-03, grad_scale: 8.0 2023-04-26 11:11:51,782 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5677, 0.6264, 0.8993, 0.9080, 1.0076, 1.0973, 0.8018, 0.8795], device='cuda:4'), covar=tensor([ 4.6348, 16.6744, 9.7764, 8.1443, 8.5494, 12.4788, 13.1218, 10.3089], device='cuda:4'), in_proj_covar=tensor([0.0254, 0.0362, 0.0294, 0.0291, 0.0321, 0.0332, 0.0352, 0.0322], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 11:12:08,252 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.683e+02 2.322e+02 2.767e+02 3.338e+02 6.077e+02, threshold=5.534e+02, percent-clipped=1.0 2023-04-26 11:12:34,757 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=4141.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:12:41,351 INFO [finetune.py:976] (4/7) Epoch 1, batch 4150, loss[loss=0.3643, simple_loss=0.384, pruned_loss=0.1723, over 4728.00 frames. ], tot_loss[loss=0.3497, simple_loss=0.3725, pruned_loss=0.1635, over 950363.67 frames. ], batch size: 59, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:12:43,028 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0486, 1.5648, 1.5360, 1.4626, 2.0685, 1.8564, 1.6413, 1.5198], device='cuda:4'), covar=tensor([0.1378, 0.1716, 0.2521, 0.2559, 0.0894, 0.1734, 0.1874, 0.2058], device='cuda:4'), in_proj_covar=tensor([0.0333, 0.0345, 0.0346, 0.0312, 0.0356, 0.0379, 0.0327, 0.0365], device='cuda:4'), out_proj_covar=tensor([7.2872e-05, 7.4244e-05, 7.4742e-05, 6.5168e-05, 7.5975e-05, 8.3110e-05, 7.1257e-05, 7.8951e-05], device='cuda:4') 2023-04-26 11:13:15,084 INFO [finetune.py:976] (4/7) Epoch 1, batch 4200, loss[loss=0.3277, simple_loss=0.3694, pruned_loss=0.143, over 4811.00 frames. ], tot_loss[loss=0.3471, simple_loss=0.3711, pruned_loss=0.1615, over 949402.62 frames. ], batch size: 38, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:13:16,199 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=4202.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 11:13:16,245 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=4202.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:13:28,645 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.716e+02 2.286e+02 2.811e+02 3.257e+02 1.063e+03, threshold=5.622e+02, percent-clipped=1.0 2023-04-26 11:14:05,466 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=4250.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 11:14:05,986 INFO [finetune.py:976] (4/7) Epoch 1, batch 4250, loss[loss=0.3005, simple_loss=0.3223, pruned_loss=0.1393, over 4802.00 frames. ], tot_loss[loss=0.3403, simple_loss=0.3654, pruned_loss=0.1576, over 951940.76 frames. ], batch size: 51, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:14:14,619 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2115, 1.7871, 1.5893, 1.8340, 1.7578, 2.1337, 1.6162, 3.8478], device='cuda:4'), covar=tensor([0.0744, 0.0684, 0.0737, 0.1296, 0.0640, 0.0570, 0.0726, 0.0154], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0040, 0.0041, 0.0046, 0.0041, 0.0040, 0.0040, 0.0064], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 11:14:50,641 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7871, 1.0602, 1.3618, 1.8631, 1.4416, 1.1165, 0.9797, 1.2880], device='cuda:4'), covar=tensor([0.9163, 1.1298, 0.5134, 1.9875, 1.2452, 0.9199, 2.2674, 1.3054], device='cuda:4'), in_proj_covar=tensor([0.0251, 0.0271, 0.0212, 0.0329, 0.0231, 0.0224, 0.0267, 0.0219], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 11:14:51,195 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=4282.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:14:59,325 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=4288.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:15:00,553 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6470, 2.0227, 1.5984, 1.9979, 1.6332, 1.5658, 1.8106, 1.2875], device='cuda:4'), covar=tensor([0.1957, 0.1500, 0.1419, 0.1426, 0.2884, 0.1482, 0.1688, 0.2874], device='cuda:4'), in_proj_covar=tensor([0.0304, 0.0321, 0.0237, 0.0300, 0.0306, 0.0275, 0.0273, 0.0292], device='cuda:4'), out_proj_covar=tensor([1.2444e-04, 1.3130e-04, 9.7307e-05, 1.2160e-04, 1.2697e-04, 1.1117e-04, 1.1332e-04, 1.1925e-04], device='cuda:4') 2023-04-26 11:15:07,280 INFO [finetune.py:976] (4/7) Epoch 1, batch 4300, loss[loss=0.2851, simple_loss=0.3157, pruned_loss=0.1273, over 4829.00 frames. ], tot_loss[loss=0.334, simple_loss=0.3605, pruned_loss=0.1537, over 953201.80 frames. ], batch size: 39, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:15:16,905 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5581, 1.2077, 1.1679, 1.1918, 1.6823, 1.5694, 1.2579, 1.1775], device='cuda:4'), covar=tensor([0.1347, 0.1710, 0.2469, 0.1901, 0.0988, 0.1461, 0.1806, 0.1979], device='cuda:4'), in_proj_covar=tensor([0.0334, 0.0346, 0.0349, 0.0314, 0.0358, 0.0381, 0.0328, 0.0366], device='cuda:4'), out_proj_covar=tensor([7.3202e-05, 7.4654e-05, 7.5475e-05, 6.5657e-05, 7.6425e-05, 8.3442e-05, 7.1500e-05, 7.9233e-05], device='cuda:4') 2023-04-26 11:15:19,787 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.630e+02 2.200e+02 2.676e+02 3.122e+02 6.239e+02, threshold=5.353e+02, percent-clipped=1.0 2023-04-26 11:15:29,486 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=4332.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 11:15:39,939 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=4349.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:15:41,064 INFO [finetune.py:976] (4/7) Epoch 1, batch 4350, loss[loss=0.2912, simple_loss=0.321, pruned_loss=0.1307, over 4773.00 frames. ], tot_loss[loss=0.328, simple_loss=0.3553, pruned_loss=0.1504, over 954206.63 frames. ], batch size: 26, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:15:41,242 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.79 vs. limit=2.0 2023-04-26 11:16:02,411 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=4380.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 11:16:10,449 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6404, 1.5130, 1.3882, 1.4332, 1.4605, 1.5763, 1.6691, 1.4289], device='cuda:4'), covar=tensor([2.4289, 7.0754, 4.3671, 3.7388, 4.3554, 5.3902, 5.5833, 5.0560], device='cuda:4'), in_proj_covar=tensor([0.0253, 0.0358, 0.0290, 0.0288, 0.0317, 0.0328, 0.0347, 0.0318], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 11:16:15,320 INFO [finetune.py:976] (4/7) Epoch 1, batch 4400, loss[loss=0.3309, simple_loss=0.3707, pruned_loss=0.1455, over 4843.00 frames. ], tot_loss[loss=0.3296, simple_loss=0.3569, pruned_loss=0.1511, over 954133.26 frames. ], batch size: 44, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:16:26,706 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.735e+02 2.527e+02 2.839e+02 3.555e+02 1.567e+03, threshold=5.678e+02, percent-clipped=5.0 2023-04-26 11:16:48,905 INFO [finetune.py:976] (4/7) Epoch 1, batch 4450, loss[loss=0.3498, simple_loss=0.3929, pruned_loss=0.1534, over 4837.00 frames. ], tot_loss[loss=0.3347, simple_loss=0.3617, pruned_loss=0.1538, over 953995.26 frames. ], batch size: 49, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:17:19,619 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=4497.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:17:22,032 INFO [finetune.py:976] (4/7) Epoch 1, batch 4500, loss[loss=0.3598, simple_loss=0.394, pruned_loss=0.1628, over 4792.00 frames. ], tot_loss[loss=0.3329, simple_loss=0.361, pruned_loss=0.1525, over 953631.07 frames. ], batch size: 51, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:17:32,938 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.661e+02 2.259e+02 2.800e+02 3.318e+02 7.116e+02, threshold=5.601e+02, percent-clipped=2.0 2023-04-26 11:18:03,008 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5222, 1.1093, 1.1044, 1.0756, 1.6405, 1.5083, 1.1918, 1.1242], device='cuda:4'), covar=tensor([0.1337, 0.1828, 0.2961, 0.2184, 0.0981, 0.1657, 0.1929, 0.2069], device='cuda:4'), in_proj_covar=tensor([0.0331, 0.0344, 0.0348, 0.0313, 0.0356, 0.0380, 0.0327, 0.0365], device='cuda:4'), out_proj_covar=tensor([7.2620e-05, 7.4202e-05, 7.5157e-05, 6.5559e-05, 7.5884e-05, 8.3258e-05, 7.1272e-05, 7.8928e-05], device='cuda:4') 2023-04-26 11:18:15,789 INFO [finetune.py:976] (4/7) Epoch 1, batch 4550, loss[loss=0.3184, simple_loss=0.369, pruned_loss=0.1339, over 4846.00 frames. ], tot_loss[loss=0.3335, simple_loss=0.3625, pruned_loss=0.1522, over 953672.29 frames. ], batch size: 44, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:19:01,297 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=4582.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:19:13,035 INFO [finetune.py:976] (4/7) Epoch 1, batch 4600, loss[loss=0.3094, simple_loss=0.3465, pruned_loss=0.1361, over 4854.00 frames. ], tot_loss[loss=0.3299, simple_loss=0.3601, pruned_loss=0.1498, over 955852.98 frames. ], batch size: 31, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:19:23,473 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.373e+02 2.361e+02 2.771e+02 3.480e+02 8.913e+02, threshold=5.542e+02, percent-clipped=3.0 2023-04-26 11:19:32,362 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-26 11:19:32,636 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=5.02 vs. limit=5.0 2023-04-26 11:19:33,260 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=4630.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:19:42,838 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=4644.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:19:47,128 INFO [finetune.py:976] (4/7) Epoch 1, batch 4650, loss[loss=0.2821, simple_loss=0.3112, pruned_loss=0.1265, over 4887.00 frames. ], tot_loss[loss=0.3244, simple_loss=0.3547, pruned_loss=0.147, over 953864.36 frames. ], batch size: 35, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:20:11,408 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4566, 1.4397, 1.5172, 1.7988, 2.1161, 1.9542, 1.8278, 1.6921], device='cuda:4'), covar=tensor([0.1616, 0.2376, 0.2957, 0.3078, 0.1770, 0.2364, 0.2098, 0.1972], device='cuda:4'), in_proj_covar=tensor([0.0332, 0.0345, 0.0349, 0.0314, 0.0356, 0.0381, 0.0327, 0.0365], device='cuda:4'), out_proj_covar=tensor([7.2859e-05, 7.4328e-05, 7.5330e-05, 6.5752e-05, 7.5855e-05, 8.3363e-05, 7.1330e-05, 7.8903e-05], device='cuda:4') 2023-04-26 11:20:42,427 INFO [finetune.py:976] (4/7) Epoch 1, batch 4700, loss[loss=0.2834, simple_loss=0.3231, pruned_loss=0.1219, over 4826.00 frames. ], tot_loss[loss=0.3183, simple_loss=0.349, pruned_loss=0.1438, over 953542.16 frames. ], batch size: 39, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:21:03,816 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.132e+02 1.981e+02 2.474e+02 3.057e+02 6.452e+02, threshold=4.948e+02, percent-clipped=2.0 2023-04-26 11:21:38,194 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1552, 1.1948, 1.1893, 1.1110, 1.2327, 1.3899, 1.2514, 1.2555], device='cuda:4'), covar=tensor([ 5.2783, 16.2728, 10.0756, 8.6548, 9.0105, 12.3178, 14.5922, 10.8814], device='cuda:4'), in_proj_covar=tensor([0.0256, 0.0360, 0.0292, 0.0289, 0.0318, 0.0332, 0.0350, 0.0320], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 11:21:39,265 INFO [finetune.py:976] (4/7) Epoch 1, batch 4750, loss[loss=0.3006, simple_loss=0.3405, pruned_loss=0.1304, over 4864.00 frames. ], tot_loss[loss=0.3144, simple_loss=0.3456, pruned_loss=0.1416, over 955573.60 frames. ], batch size: 34, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:22:11,031 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=4797.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:22:13,544 INFO [finetune.py:976] (4/7) Epoch 1, batch 4800, loss[loss=0.341, simple_loss=0.3692, pruned_loss=0.1564, over 4768.00 frames. ], tot_loss[loss=0.3191, simple_loss=0.3498, pruned_loss=0.1441, over 956592.12 frames. ], batch size: 28, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:22:24,010 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.522e+02 2.397e+02 2.799e+02 3.310e+02 5.685e+02, threshold=5.598e+02, percent-clipped=2.0 2023-04-26 11:22:40,835 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-26 11:22:43,635 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=4845.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:22:47,740 INFO [finetune.py:976] (4/7) Epoch 1, batch 4850, loss[loss=0.3685, simple_loss=0.391, pruned_loss=0.1731, over 4303.00 frames. ], tot_loss[loss=0.321, simple_loss=0.3531, pruned_loss=0.1445, over 953968.67 frames. ], batch size: 66, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:22:50,319 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4563, 0.5586, 0.8890, 0.9354, 1.0698, 1.1545, 0.8540, 0.9976], device='cuda:4'), covar=tensor([3.4176, 9.9500, 5.8057, 4.8857, 5.3408, 8.0573, 8.1344, 5.9477], device='cuda:4'), in_proj_covar=tensor([0.0256, 0.0360, 0.0292, 0.0289, 0.0317, 0.0332, 0.0349, 0.0319], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 11:23:02,952 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.71 vs. limit=2.0 2023-04-26 11:23:33,605 INFO [finetune.py:976] (4/7) Epoch 1, batch 4900, loss[loss=0.279, simple_loss=0.3238, pruned_loss=0.1171, over 4780.00 frames. ], tot_loss[loss=0.3204, simple_loss=0.3528, pruned_loss=0.144, over 953146.22 frames. ], batch size: 29, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:23:49,557 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.488e+02 2.175e+02 2.572e+02 3.151e+02 5.494e+02, threshold=5.143e+02, percent-clipped=0.0 2023-04-26 11:24:10,520 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5440, 1.0919, 0.3885, 1.1635, 1.0962, 1.4061, 1.2502, 1.2370], device='cuda:4'), covar=tensor([0.0676, 0.0548, 0.0603, 0.0691, 0.0410, 0.0657, 0.0647, 0.0823], device='cuda:4'), in_proj_covar=tensor([0.0032, 0.0027, 0.0025, 0.0031, 0.0022, 0.0031, 0.0030, 0.0033], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0049, 0.0037, 0.0048, 0.0047, 0.0051], device='cuda:4') 2023-04-26 11:24:29,557 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=4944.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:24:41,249 INFO [finetune.py:976] (4/7) Epoch 1, batch 4950, loss[loss=0.2887, simple_loss=0.333, pruned_loss=0.1222, over 4865.00 frames. ], tot_loss[loss=0.3189, simple_loss=0.3521, pruned_loss=0.1428, over 954241.59 frames. ], batch size: 31, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:25:03,886 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6418, 2.3401, 1.4874, 1.3872, 1.1661, 1.2100, 1.5056, 1.0968], device='cuda:4'), covar=tensor([0.2758, 0.2501, 0.3267, 0.4090, 0.4387, 0.3319, 0.2380, 0.3490], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0221, 0.0200, 0.0216, 0.0236, 0.0197, 0.0193, 0.0209], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 11:25:19,476 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=4992.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:25:26,589 INFO [finetune.py:976] (4/7) Epoch 1, batch 5000, loss[loss=0.3604, simple_loss=0.3827, pruned_loss=0.1691, over 4913.00 frames. ], tot_loss[loss=0.3155, simple_loss=0.3497, pruned_loss=0.1407, over 956446.20 frames. ], batch size: 37, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:25:28,079 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.65 vs. limit=5.0 2023-04-26 11:25:37,464 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.472e+02 2.232e+02 2.673e+02 3.232e+02 7.029e+02, threshold=5.346e+02, percent-clipped=5.0 2023-04-26 11:25:44,387 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6933, 1.0097, 1.2918, 1.7173, 1.2639, 1.0069, 0.9232, 1.1988], device='cuda:4'), covar=tensor([0.8031, 0.9736, 0.4574, 1.2408, 1.0639, 0.7733, 1.5086, 1.0280], device='cuda:4'), in_proj_covar=tensor([0.0254, 0.0274, 0.0214, 0.0333, 0.0233, 0.0226, 0.0268, 0.0220], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 11:26:10,645 INFO [finetune.py:976] (4/7) Epoch 1, batch 5050, loss[loss=0.2599, simple_loss=0.3103, pruned_loss=0.1048, over 4763.00 frames. ], tot_loss[loss=0.3124, simple_loss=0.346, pruned_loss=0.1394, over 955175.88 frames. ], batch size: 27, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:26:22,985 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5716, 1.4495, 0.7070, 1.2437, 1.5223, 1.4539, 1.3113, 1.3942], device='cuda:4'), covar=tensor([0.0712, 0.0574, 0.0589, 0.0759, 0.0431, 0.0708, 0.0713, 0.0870], device='cuda:4'), in_proj_covar=tensor([0.0032, 0.0027, 0.0025, 0.0031, 0.0022, 0.0031, 0.0030, 0.0033], device='cuda:4'), out_proj_covar=tensor([0.0048, 0.0043, 0.0038, 0.0049, 0.0037, 0.0048, 0.0047, 0.0051], device='cuda:4') 2023-04-26 11:26:25,430 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4419, 1.2878, 1.9051, 1.7915, 1.3592, 1.0867, 1.4463, 0.9116], device='cuda:4'), covar=tensor([0.0994, 0.1182, 0.0641, 0.0857, 0.1410, 0.1639, 0.1111, 0.1604], device='cuda:4'), in_proj_covar=tensor([0.0073, 0.0080, 0.0076, 0.0076, 0.0091, 0.0096, 0.0092, 0.0081], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0004], device='cuda:4') 2023-04-26 11:26:55,686 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.48 vs. limit=2.0 2023-04-26 11:27:11,304 INFO [finetune.py:976] (4/7) Epoch 1, batch 5100, loss[loss=0.3023, simple_loss=0.3378, pruned_loss=0.1334, over 4817.00 frames. ], tot_loss[loss=0.3074, simple_loss=0.3412, pruned_loss=0.1368, over 954597.78 frames. ], batch size: 38, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:27:40,804 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.449e+02 2.109e+02 2.481e+02 2.814e+02 6.642e+02, threshold=4.963e+02, percent-clipped=1.0 2023-04-26 11:28:17,911 INFO [finetune.py:976] (4/7) Epoch 1, batch 5150, loss[loss=0.273, simple_loss=0.3244, pruned_loss=0.1108, over 4911.00 frames. ], tot_loss[loss=0.305, simple_loss=0.3387, pruned_loss=0.1357, over 953173.44 frames. ], batch size: 36, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:29:05,505 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0382, 1.5550, 1.3543, 1.7354, 1.5871, 1.9307, 1.4500, 3.4761], device='cuda:4'), covar=tensor([0.0789, 0.0772, 0.0809, 0.1336, 0.0714, 0.0628, 0.0770, 0.0186], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0040, 0.0041, 0.0046, 0.0041, 0.0040, 0.0040, 0.0064], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 11:29:07,358 INFO [finetune.py:976] (4/7) Epoch 1, batch 5200, loss[loss=0.3831, simple_loss=0.4166, pruned_loss=0.1748, over 4902.00 frames. ], tot_loss[loss=0.3088, simple_loss=0.3427, pruned_loss=0.1374, over 952563.57 frames. ], batch size: 43, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:29:21,206 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.499e+02 2.511e+02 2.846e+02 3.705e+02 8.558e+02, threshold=5.692e+02, percent-clipped=9.0 2023-04-26 11:29:35,237 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-04-26 11:29:41,594 INFO [finetune.py:976] (4/7) Epoch 1, batch 5250, loss[loss=0.3813, simple_loss=0.3993, pruned_loss=0.1817, over 4746.00 frames. ], tot_loss[loss=0.3093, simple_loss=0.3439, pruned_loss=0.1374, over 952648.05 frames. ], batch size: 59, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:29:59,799 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4892, 2.2182, 1.3094, 1.4362, 1.0867, 1.1362, 1.3202, 0.9982], device='cuda:4'), covar=tensor([0.2691, 0.2440, 0.3362, 0.3847, 0.4496, 0.3197, 0.2610, 0.3686], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0222, 0.0200, 0.0217, 0.0237, 0.0198, 0.0194, 0.0210], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 11:30:06,553 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6822, 1.3574, 1.1853, 1.3271, 2.0314, 1.6029, 1.2797, 1.2419], device='cuda:4'), covar=tensor([0.1993, 0.1934, 0.3218, 0.2257, 0.0981, 0.1969, 0.2448, 0.2357], device='cuda:4'), in_proj_covar=tensor([0.0334, 0.0349, 0.0351, 0.0318, 0.0360, 0.0384, 0.0332, 0.0367], device='cuda:4'), out_proj_covar=tensor([7.3257e-05, 7.5191e-05, 7.5864e-05, 6.6581e-05, 7.6844e-05, 8.4090e-05, 7.2396e-05, 7.9255e-05], device='cuda:4') 2023-04-26 11:30:12,767 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-04-26 11:30:15,015 INFO [finetune.py:976] (4/7) Epoch 1, batch 5300, loss[loss=0.3218, simple_loss=0.3584, pruned_loss=0.1426, over 4811.00 frames. ], tot_loss[loss=0.31, simple_loss=0.3449, pruned_loss=0.1375, over 951963.38 frames. ], batch size: 40, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:30:27,273 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.471e+02 2.212e+02 2.624e+02 3.137e+02 5.522e+02, threshold=5.248e+02, percent-clipped=0.0 2023-04-26 11:30:41,780 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-26 11:30:49,038 INFO [finetune.py:976] (4/7) Epoch 1, batch 5350, loss[loss=0.2649, simple_loss=0.3179, pruned_loss=0.106, over 4849.00 frames. ], tot_loss[loss=0.3104, simple_loss=0.3457, pruned_loss=0.1376, over 952887.45 frames. ], batch size: 49, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:31:35,564 INFO [finetune.py:976] (4/7) Epoch 1, batch 5400, loss[loss=0.2767, simple_loss=0.319, pruned_loss=0.1172, over 4815.00 frames. ], tot_loss[loss=0.3072, simple_loss=0.3423, pruned_loss=0.136, over 953996.42 frames. ], batch size: 40, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:31:52,845 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.437e+02 2.058e+02 2.416e+02 3.015e+02 5.928e+02, threshold=4.831e+02, percent-clipped=1.0 2023-04-26 11:31:53,039 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.59 vs. limit=2.0 2023-04-26 11:32:15,535 INFO [finetune.py:976] (4/7) Epoch 1, batch 5450, loss[loss=0.2408, simple_loss=0.2925, pruned_loss=0.09458, over 4704.00 frames. ], tot_loss[loss=0.3025, simple_loss=0.338, pruned_loss=0.1335, over 954483.21 frames. ], batch size: 23, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:32:49,240 INFO [finetune.py:976] (4/7) Epoch 1, batch 5500, loss[loss=0.3148, simple_loss=0.3345, pruned_loss=0.1476, over 4822.00 frames. ], tot_loss[loss=0.2984, simple_loss=0.3342, pruned_loss=0.1313, over 954565.94 frames. ], batch size: 39, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:32:54,280 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.5697, 1.7108, 1.5974, 1.3288, 1.6968, 1.2888, 2.2528, 1.2366], device='cuda:4'), covar=tensor([0.3542, 0.1322, 0.3930, 0.2260, 0.1618, 0.2219, 0.1169, 0.4069], device='cuda:4'), in_proj_covar=tensor([0.0331, 0.0341, 0.0423, 0.0357, 0.0391, 0.0364, 0.0388, 0.0398], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 11:32:59,728 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.389e+02 2.158e+02 2.515e+02 3.075e+02 5.534e+02, threshold=5.030e+02, percent-clipped=1.0 2023-04-26 11:33:46,363 INFO [finetune.py:976] (4/7) Epoch 1, batch 5550, loss[loss=0.2989, simple_loss=0.3466, pruned_loss=0.1256, over 4877.00 frames. ], tot_loss[loss=0.3035, simple_loss=0.3386, pruned_loss=0.1342, over 954776.09 frames. ], batch size: 34, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:34:10,462 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.88 vs. limit=5.0 2023-04-26 11:34:53,183 INFO [finetune.py:976] (4/7) Epoch 1, batch 5600, loss[loss=0.375, simple_loss=0.3814, pruned_loss=0.1843, over 4191.00 frames. ], tot_loss[loss=0.3052, simple_loss=0.3413, pruned_loss=0.1345, over 953702.64 frames. ], batch size: 65, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:35:13,911 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.486e+02 2.261e+02 2.635e+02 3.434e+02 7.043e+02, threshold=5.269e+02, percent-clipped=4.0 2023-04-26 11:35:43,166 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8009, 1.6988, 2.0383, 2.2135, 2.2773, 1.6565, 1.5024, 1.9767], device='cuda:4'), covar=tensor([0.1202, 0.1160, 0.0773, 0.0896, 0.0718, 0.1283, 0.1291, 0.0771], device='cuda:4'), in_proj_covar=tensor([0.0208, 0.0209, 0.0189, 0.0181, 0.0178, 0.0196, 0.0174, 0.0191], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 11:35:44,916 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=5650.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 11:35:45,418 INFO [finetune.py:976] (4/7) Epoch 1, batch 5650, loss[loss=0.3091, simple_loss=0.3516, pruned_loss=0.1333, over 4869.00 frames. ], tot_loss[loss=0.3072, simple_loss=0.3441, pruned_loss=0.1352, over 953727.87 frames. ], batch size: 31, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:36:13,549 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2737, 2.7512, 0.9967, 1.6344, 1.6669, 2.0471, 1.8828, 1.0413], device='cuda:4'), covar=tensor([0.1255, 0.1058, 0.1852, 0.1202, 0.0984, 0.0972, 0.1483, 0.1531], device='cuda:4'), in_proj_covar=tensor([0.0124, 0.0272, 0.0151, 0.0133, 0.0144, 0.0166, 0.0130, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 11:36:15,235 INFO [finetune.py:976] (4/7) Epoch 1, batch 5700, loss[loss=0.2128, simple_loss=0.252, pruned_loss=0.08677, over 4598.00 frames. ], tot_loss[loss=0.3031, simple_loss=0.3386, pruned_loss=0.1338, over 933538.94 frames. ], batch size: 20, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:36:21,251 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=5711.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 11:36:31,944 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.325e+02 1.970e+02 2.507e+02 3.043e+02 6.051e+02, threshold=5.014e+02, percent-clipped=1.0 2023-04-26 11:36:33,047 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.78 vs. limit=2.0 2023-04-26 11:37:01,401 INFO [finetune.py:976] (4/7) Epoch 2, batch 0, loss[loss=0.2906, simple_loss=0.3386, pruned_loss=0.1214, over 4760.00 frames. ], tot_loss[loss=0.2906, simple_loss=0.3386, pruned_loss=0.1214, over 4760.00 frames. ], batch size: 28, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:37:01,401 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-26 11:37:23,940 INFO [finetune.py:1010] (4/7) Epoch 2, validation: loss=0.2101, simple_loss=0.2777, pruned_loss=0.0712, over 2265189.00 frames. 2023-04-26 11:37:23,940 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6098MB 2023-04-26 11:37:42,878 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5838, 4.3585, 0.8501, 2.0779, 1.9703, 2.6945, 2.8649, 1.0053], device='cuda:4'), covar=tensor([0.1560, 0.1272, 0.2494, 0.1639, 0.1248, 0.1315, 0.1282, 0.1994], device='cuda:4'), in_proj_covar=tensor([0.0123, 0.0269, 0.0149, 0.0131, 0.0142, 0.0164, 0.0128, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 11:37:46,569 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=5755.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:38:02,075 INFO [finetune.py:976] (4/7) Epoch 2, batch 50, loss[loss=0.3144, simple_loss=0.3427, pruned_loss=0.143, over 4902.00 frames. ], tot_loss[loss=0.3037, simple_loss=0.3419, pruned_loss=0.1328, over 213945.89 frames. ], batch size: 46, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:38:10,479 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.61 vs. limit=2.0 2023-04-26 11:38:27,682 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=5816.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:38:29,285 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.431e+02 2.066e+02 2.410e+02 2.890e+02 4.929e+02, threshold=4.819e+02, percent-clipped=0.0 2023-04-26 11:38:29,362 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.3436, 4.1765, 3.2480, 4.8879, 4.2496, 4.2742, 2.0746, 4.1357], device='cuda:4'), covar=tensor([0.1323, 0.0958, 0.3126, 0.0798, 0.2176, 0.1461, 0.4897, 0.1867], device='cuda:4'), in_proj_covar=tensor([0.0262, 0.0230, 0.0280, 0.0325, 0.0321, 0.0270, 0.0286, 0.0287], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 11:38:35,781 INFO [finetune.py:976] (4/7) Epoch 2, batch 100, loss[loss=0.2758, simple_loss=0.3202, pruned_loss=0.1157, over 4755.00 frames. ], tot_loss[loss=0.2924, simple_loss=0.331, pruned_loss=0.1269, over 378173.61 frames. ], batch size: 27, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:38:39,428 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.9223, 2.8800, 2.2671, 3.3034, 2.8469, 2.8491, 1.2098, 2.8668], device='cuda:4'), covar=tensor([0.1934, 0.1502, 0.3284, 0.2622, 0.2668, 0.2195, 0.5365, 0.2340], device='cuda:4'), in_proj_covar=tensor([0.0263, 0.0232, 0.0281, 0.0327, 0.0323, 0.0272, 0.0287, 0.0288], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 11:38:55,163 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=5858.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:39:09,189 INFO [finetune.py:976] (4/7) Epoch 2, batch 150, loss[loss=0.2608, simple_loss=0.3171, pruned_loss=0.1023, over 4831.00 frames. ], tot_loss[loss=0.2867, simple_loss=0.3247, pruned_loss=0.1243, over 506002.54 frames. ], batch size: 30, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:39:34,932 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.317e+02 2.049e+02 2.408e+02 2.939e+02 5.453e+02, threshold=4.816e+02, percent-clipped=1.0 2023-04-26 11:39:35,688 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=5919.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:39:48,162 INFO [finetune.py:976] (4/7) Epoch 2, batch 200, loss[loss=0.3152, simple_loss=0.353, pruned_loss=0.1387, over 4739.00 frames. ], tot_loss[loss=0.2847, simple_loss=0.323, pruned_loss=0.1232, over 606844.66 frames. ], batch size: 59, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:40:02,350 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3017, 1.3256, 1.2622, 1.3049, 1.3738, 1.6081, 1.3925, 1.3269], device='cuda:4'), covar=tensor([ 3.7288, 11.1133, 7.4049, 5.9816, 6.4325, 9.2928, 9.7046, 7.8596], device='cuda:4'), in_proj_covar=tensor([0.0263, 0.0367, 0.0295, 0.0295, 0.0323, 0.0344, 0.0355, 0.0323], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 11:40:10,402 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6051, 1.3689, 4.2725, 3.9730, 3.7802, 3.9793, 3.9576, 3.7255], device='cuda:4'), covar=tensor([0.6722, 0.6015, 0.1083, 0.1792, 0.1152, 0.1837, 0.1470, 0.1580], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0315, 0.0451, 0.0459, 0.0378, 0.0436, 0.0344, 0.0403], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0002], device='cuda:4') 2023-04-26 11:40:52,015 INFO [finetune.py:976] (4/7) Epoch 2, batch 250, loss[loss=0.2948, simple_loss=0.3257, pruned_loss=0.132, over 4906.00 frames. ], tot_loss[loss=0.2884, simple_loss=0.3273, pruned_loss=0.1248, over 682995.34 frames. ], batch size: 32, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:41:06,891 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=5990.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:41:09,793 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5281, 1.0068, 1.2252, 1.2885, 1.1244, 1.0032, 0.5145, 0.8808], device='cuda:4'), covar=tensor([0.7438, 0.8858, 0.3966, 1.0179, 0.8915, 0.6243, 1.3510, 0.9251], device='cuda:4'), in_proj_covar=tensor([0.0260, 0.0279, 0.0219, 0.0341, 0.0236, 0.0232, 0.0272, 0.0223], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 11:41:18,235 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=6006.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 11:41:25,491 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.356e+02 2.171e+02 2.695e+02 3.137e+02 1.432e+03, threshold=5.390e+02, percent-clipped=4.0 2023-04-26 11:41:33,099 INFO [finetune.py:976] (4/7) Epoch 2, batch 300, loss[loss=0.3562, simple_loss=0.3711, pruned_loss=0.1706, over 4911.00 frames. ], tot_loss[loss=0.2945, simple_loss=0.3334, pruned_loss=0.1278, over 744559.19 frames. ], batch size: 35, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:41:49,301 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=6043.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:41:54,647 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=6051.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:42:22,356 INFO [finetune.py:976] (4/7) Epoch 2, batch 350, loss[loss=0.2417, simple_loss=0.2774, pruned_loss=0.103, over 4324.00 frames. ], tot_loss[loss=0.2966, simple_loss=0.3357, pruned_loss=0.1287, over 789650.31 frames. ], batch size: 19, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:42:36,691 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6395, 1.8415, 1.0462, 1.2717, 2.0978, 1.5168, 1.3365, 1.4777], device='cuda:4'), covar=tensor([0.0640, 0.0497, 0.0469, 0.0704, 0.0339, 0.0649, 0.0600, 0.0803], device='cuda:4'), in_proj_covar=tensor([0.0032, 0.0027, 0.0025, 0.0031, 0.0022, 0.0031, 0.0030, 0.0033], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0049, 0.0038, 0.0048, 0.0047, 0.0051], device='cuda:4') 2023-04-26 11:42:47,483 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=6104.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:42:57,936 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=6111.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:43:02,077 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.403e+02 2.085e+02 2.458e+02 2.985e+02 5.837e+02, threshold=4.917e+02, percent-clipped=2.0 2023-04-26 11:43:14,254 INFO [finetune.py:976] (4/7) Epoch 2, batch 400, loss[loss=0.3156, simple_loss=0.3584, pruned_loss=0.1364, over 4812.00 frames. ], tot_loss[loss=0.2972, simple_loss=0.3368, pruned_loss=0.1288, over 824718.71 frames. ], batch size: 39, lr: 4.00e-03, grad_scale: 16.0 2023-04-26 11:43:20,184 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=6135.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:43:40,657 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9248, 2.3580, 1.8191, 2.1699, 1.7901, 1.7119, 2.1152, 1.5781], device='cuda:4'), covar=tensor([0.2458, 0.1842, 0.1510, 0.1954, 0.3091, 0.1878, 0.2371, 0.3227], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0334, 0.0245, 0.0310, 0.0313, 0.0284, 0.0280, 0.0301], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 11:43:43,640 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=6170.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:43:48,411 INFO [finetune.py:976] (4/7) Epoch 2, batch 450, loss[loss=0.2611, simple_loss=0.315, pruned_loss=0.1036, over 4886.00 frames. ], tot_loss[loss=0.2937, simple_loss=0.3336, pruned_loss=0.1269, over 854131.94 frames. ], batch size: 35, lr: 4.00e-03, grad_scale: 32.0 2023-04-26 11:43:54,412 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7948, 1.1433, 1.4767, 1.7567, 1.4087, 1.1353, 0.8212, 1.3062], device='cuda:4'), covar=tensor([0.6728, 0.8668, 0.3942, 1.0169, 0.9665, 0.6473, 1.3314, 0.9582], device='cuda:4'), in_proj_covar=tensor([0.0261, 0.0278, 0.0220, 0.0342, 0.0236, 0.0232, 0.0272, 0.0222], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 11:44:00,947 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=6196.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:44:06,760 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=6203.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:44:11,625 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=6211.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:44:13,447 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=6214.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:44:15,818 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.476e+02 2.104e+02 2.551e+02 3.094e+02 4.755e+02, threshold=5.103e+02, percent-clipped=0.0 2023-04-26 11:44:21,881 INFO [finetune.py:976] (4/7) Epoch 2, batch 500, loss[loss=0.2491, simple_loss=0.3031, pruned_loss=0.09753, over 4845.00 frames. ], tot_loss[loss=0.2897, simple_loss=0.3301, pruned_loss=0.1247, over 877847.95 frames. ], batch size: 47, lr: 4.00e-03, grad_scale: 32.0 2023-04-26 11:44:23,808 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=6231.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:44:38,025 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9488, 1.3102, 1.6561, 2.0259, 1.5774, 1.3093, 1.0101, 1.5239], device='cuda:4'), covar=tensor([0.6328, 0.7460, 0.3704, 0.9962, 0.9009, 0.6013, 1.2850, 0.8450], device='cuda:4'), in_proj_covar=tensor([0.0262, 0.0280, 0.0221, 0.0343, 0.0237, 0.0233, 0.0274, 0.0223], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 11:44:46,835 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=6264.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:44:48,049 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=6266.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:44:51,711 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=6272.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:44:55,250 INFO [finetune.py:976] (4/7) Epoch 2, batch 550, loss[loss=0.2626, simple_loss=0.3089, pruned_loss=0.1081, over 4757.00 frames. ], tot_loss[loss=0.2859, simple_loss=0.3264, pruned_loss=0.1227, over 895141.31 frames. ], batch size: 28, lr: 4.00e-03, grad_scale: 32.0 2023-04-26 11:44:59,710 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1356, 1.2424, 1.2387, 1.3126, 1.3259, 1.5647, 1.4166, 1.3823], device='cuda:4'), covar=tensor([3.1209, 6.6774, 5.0733, 4.2009, 4.6768, 7.1210, 6.2028, 5.4100], device='cuda:4'), in_proj_covar=tensor([0.0266, 0.0369, 0.0296, 0.0297, 0.0325, 0.0347, 0.0357, 0.0325], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 11:45:15,027 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=6306.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 11:45:15,656 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7094, 1.1959, 1.2524, 1.2250, 1.8414, 1.5590, 1.2485, 1.2117], device='cuda:4'), covar=tensor([0.1397, 0.1929, 0.2290, 0.1995, 0.0970, 0.1787, 0.2196, 0.2145], device='cuda:4'), in_proj_covar=tensor([0.0327, 0.0343, 0.0346, 0.0315, 0.0353, 0.0378, 0.0326, 0.0360], device='cuda:4'), out_proj_covar=tensor([7.1726e-05, 7.3988e-05, 7.4760e-05, 6.5942e-05, 7.5332e-05, 8.2868e-05, 7.1087e-05, 7.7814e-05], device='cuda:4') 2023-04-26 11:45:27,157 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0322, 1.0489, 1.2182, 1.2366, 1.1711, 1.0584, 1.0931, 1.0930], device='cuda:4'), covar=tensor([ 6.9546, 10.5255, 12.1518, 10.9504, 8.5735, 13.1044, 13.2745, 9.1078], device='cuda:4'), in_proj_covar=tensor([0.0435, 0.0502, 0.0580, 0.0559, 0.0463, 0.0522, 0.0527, 0.0538], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 11:45:28,237 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.445e+02 2.269e+02 2.692e+02 3.237e+02 6.108e+02, threshold=5.385e+02, percent-clipped=2.0 2023-04-26 11:45:37,842 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=6327.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:45:38,337 INFO [finetune.py:976] (4/7) Epoch 2, batch 600, loss[loss=0.2726, simple_loss=0.3478, pruned_loss=0.09872, over 4921.00 frames. ], tot_loss[loss=0.2883, simple_loss=0.3282, pruned_loss=0.1242, over 904569.62 frames. ], batch size: 42, lr: 4.00e-03, grad_scale: 32.0 2023-04-26 11:46:00,722 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=6346.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:46:12,544 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=6354.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 11:46:39,284 INFO [finetune.py:976] (4/7) Epoch 2, batch 650, loss[loss=0.3525, simple_loss=0.3991, pruned_loss=0.1529, over 4801.00 frames. ], tot_loss[loss=0.2904, simple_loss=0.3306, pruned_loss=0.1251, over 913710.23 frames. ], batch size: 45, lr: 4.00e-03, grad_scale: 32.0 2023-04-26 11:46:53,616 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=6399.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:47:02,618 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=6411.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:47:07,257 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.650e+02 2.377e+02 2.867e+02 3.488e+02 9.630e+02, threshold=5.734e+02, percent-clipped=2.0 2023-04-26 11:47:13,365 INFO [finetune.py:976] (4/7) Epoch 2, batch 700, loss[loss=0.3013, simple_loss=0.3475, pruned_loss=0.1276, over 4932.00 frames. ], tot_loss[loss=0.2912, simple_loss=0.3318, pruned_loss=0.1252, over 923334.38 frames. ], batch size: 33, lr: 4.00e-03, grad_scale: 32.0 2023-04-26 11:47:46,401 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=6459.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:47:48,743 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=6462.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:47:51,709 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=6467.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:47:58,778 INFO [finetune.py:976] (4/7) Epoch 2, batch 750, loss[loss=0.3117, simple_loss=0.3303, pruned_loss=0.1465, over 4690.00 frames. ], tot_loss[loss=0.2904, simple_loss=0.3315, pruned_loss=0.1247, over 928710.17 frames. ], batch size: 23, lr: 4.00e-03, grad_scale: 32.0 2023-04-26 11:48:18,131 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=6491.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:48:50,918 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=6514.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:48:53,698 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.458e+02 2.088e+02 2.383e+02 2.881e+02 6.100e+02, threshold=4.765e+02, percent-clipped=1.0 2023-04-26 11:49:03,668 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=6523.0, num_to_drop=1, layers_to_drop={3} 2023-04-26 11:49:03,752 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.51 vs. limit=2.0 2023-04-26 11:49:05,431 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=6526.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:49:06,607 INFO [finetune.py:976] (4/7) Epoch 2, batch 800, loss[loss=0.2365, simple_loss=0.306, pruned_loss=0.08348, over 4765.00 frames. ], tot_loss[loss=0.2883, simple_loss=0.3297, pruned_loss=0.1234, over 932718.82 frames. ], batch size: 27, lr: 4.00e-03, grad_scale: 32.0 2023-04-26 11:49:06,737 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=6528.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:49:11,615 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0087, 2.0529, 1.8723, 1.7620, 2.1880, 1.7896, 2.8021, 1.5311], device='cuda:4'), covar=tensor([0.4668, 0.1590, 0.4476, 0.2673, 0.1915, 0.2646, 0.1190, 0.4845], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0346, 0.0429, 0.0361, 0.0398, 0.0368, 0.0393, 0.0403], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 11:49:27,196 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=6559.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:49:29,547 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=6562.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:49:33,148 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=6567.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:49:40,229 INFO [finetune.py:976] (4/7) Epoch 2, batch 850, loss[loss=0.2698, simple_loss=0.3138, pruned_loss=0.1129, over 4789.00 frames. ], tot_loss[loss=0.2849, simple_loss=0.3266, pruned_loss=0.1215, over 938268.42 frames. ], batch size: 29, lr: 4.00e-03, grad_scale: 32.0 2023-04-26 11:50:18,615 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.341e+02 2.044e+02 2.405e+02 3.038e+02 6.612e+02, threshold=4.810e+02, percent-clipped=4.0 2023-04-26 11:50:21,172 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=6622.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:50:25,264 INFO [finetune.py:976] (4/7) Epoch 2, batch 900, loss[loss=0.2689, simple_loss=0.3064, pruned_loss=0.1157, over 4849.00 frames. ], tot_loss[loss=0.279, simple_loss=0.3214, pruned_loss=0.1183, over 944197.42 frames. ], batch size: 47, lr: 4.00e-03, grad_scale: 32.0 2023-04-26 11:50:25,988 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=6629.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:50:36,408 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=6646.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:50:37,662 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5911, 1.3728, 1.9586, 1.8008, 1.3893, 1.0985, 1.5387, 1.0745], device='cuda:4'), covar=tensor([0.0841, 0.1070, 0.0564, 0.0952, 0.1387, 0.1600, 0.0975, 0.1384], device='cuda:4'), in_proj_covar=tensor([0.0072, 0.0079, 0.0075, 0.0075, 0.0088, 0.0095, 0.0090, 0.0080], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-26 11:50:44,222 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1917, 0.8620, 1.1350, 1.5163, 1.3948, 1.1959, 1.1622, 1.1502], device='cuda:4'), covar=tensor([ 8.6372, 13.1796, 16.1856, 14.9214, 10.0896, 14.1502, 14.9927, 9.5757], device='cuda:4'), in_proj_covar=tensor([0.0440, 0.0505, 0.0588, 0.0569, 0.0470, 0.0527, 0.0534, 0.0543], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 11:50:58,408 INFO [finetune.py:976] (4/7) Epoch 2, batch 950, loss[loss=0.2287, simple_loss=0.2682, pruned_loss=0.09457, over 4195.00 frames. ], tot_loss[loss=0.2764, simple_loss=0.3183, pruned_loss=0.1172, over 945208.22 frames. ], batch size: 18, lr: 4.00e-03, grad_scale: 32.0 2023-04-26 11:51:06,475 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=6690.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:51:08,886 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=6694.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:51:17,814 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=6699.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:51:41,854 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.328e+02 2.213e+02 2.518e+02 3.022e+02 8.936e+02, threshold=5.037e+02, percent-clipped=4.0 2023-04-26 11:51:48,518 INFO [finetune.py:976] (4/7) Epoch 2, batch 1000, loss[loss=0.368, simple_loss=0.389, pruned_loss=0.1735, over 4846.00 frames. ], tot_loss[loss=0.2779, simple_loss=0.32, pruned_loss=0.1179, over 946492.52 frames. ], batch size: 44, lr: 4.00e-03, grad_scale: 32.0 2023-04-26 11:51:48,629 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4485, 1.2468, 1.7939, 1.6060, 1.3023, 1.0300, 1.4171, 0.9241], device='cuda:4'), covar=tensor([0.0738, 0.0988, 0.0500, 0.0857, 0.1154, 0.1435, 0.0725, 0.1193], device='cuda:4'), in_proj_covar=tensor([0.0071, 0.0079, 0.0075, 0.0075, 0.0088, 0.0094, 0.0089, 0.0080], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-26 11:52:11,813 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=6747.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:52:12,482 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.8729, 1.9279, 1.6048, 1.5078, 1.8836, 1.6185, 2.4940, 1.3463], device='cuda:4'), covar=tensor([0.3959, 0.1372, 0.4970, 0.3044, 0.2062, 0.2423, 0.1078, 0.4514], device='cuda:4'), in_proj_covar=tensor([0.0339, 0.0347, 0.0431, 0.0364, 0.0400, 0.0370, 0.0394, 0.0406], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 11:52:53,043 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-26 11:52:55,094 INFO [finetune.py:976] (4/7) Epoch 2, batch 1050, loss[loss=0.2859, simple_loss=0.3312, pruned_loss=0.1203, over 4739.00 frames. ], tot_loss[loss=0.2818, simple_loss=0.324, pruned_loss=0.1198, over 949148.06 frames. ], batch size: 59, lr: 4.00e-03, grad_scale: 32.0 2023-04-26 11:53:03,727 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=6791.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:53:10,965 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.80 vs. limit=5.0 2023-04-26 11:53:33,222 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.642e+02 2.078e+02 2.516e+02 2.901e+02 4.864e+02, threshold=5.033e+02, percent-clipped=0.0 2023-04-26 11:53:33,308 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=6818.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 11:53:42,755 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=6823.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:53:44,627 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=6826.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:53:51,479 INFO [finetune.py:976] (4/7) Epoch 2, batch 1100, loss[loss=0.325, simple_loss=0.3603, pruned_loss=0.1448, over 4800.00 frames. ], tot_loss[loss=0.2823, simple_loss=0.3254, pruned_loss=0.1196, over 952077.11 frames. ], batch size: 51, lr: 4.00e-03, grad_scale: 32.0 2023-04-26 11:54:04,441 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=6839.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:54:16,595 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=6859.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:54:20,277 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.34 vs. limit=5.0 2023-04-26 11:54:22,472 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=6867.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:54:27,635 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=6874.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:54:30,494 INFO [finetune.py:976] (4/7) Epoch 2, batch 1150, loss[loss=0.2807, simple_loss=0.3457, pruned_loss=0.1079, over 4808.00 frames. ], tot_loss[loss=0.2836, simple_loss=0.3271, pruned_loss=0.1201, over 954451.61 frames. ], batch size: 39, lr: 4.00e-03, grad_scale: 32.0 2023-04-26 11:54:30,602 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=6878.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 11:54:49,255 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=6907.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:54:50,535 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=6909.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:54:54,161 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=6915.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:54:56,910 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.447e+02 2.231e+02 2.530e+02 3.126e+02 7.008e+02, threshold=5.060e+02, percent-clipped=3.0 2023-04-26 11:54:57,935 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-04-26 11:54:59,941 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=6922.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:55:04,516 INFO [finetune.py:976] (4/7) Epoch 2, batch 1200, loss[loss=0.2429, simple_loss=0.3037, pruned_loss=0.091, over 4768.00 frames. ], tot_loss[loss=0.2816, simple_loss=0.325, pruned_loss=0.1191, over 954493.90 frames. ], batch size: 28, lr: 4.00e-03, grad_scale: 32.0 2023-04-26 11:55:11,766 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=6939.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 11:55:18,612 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-04-26 11:55:29,637 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.66 vs. limit=2.0 2023-04-26 11:55:32,155 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=6970.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:55:32,230 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=6970.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:55:38,017 INFO [finetune.py:976] (4/7) Epoch 2, batch 1250, loss[loss=0.2488, simple_loss=0.3006, pruned_loss=0.0985, over 4859.00 frames. ], tot_loss[loss=0.2757, simple_loss=0.3198, pruned_loss=0.1158, over 953887.55 frames. ], batch size: 47, lr: 4.00e-03, grad_scale: 32.0 2023-04-26 11:55:43,402 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=6985.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:55:44,056 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3384, 2.6221, 0.9800, 1.4166, 1.9851, 1.3406, 3.6287, 1.9297], device='cuda:4'), covar=tensor([0.0667, 0.0735, 0.1018, 0.1339, 0.0594, 0.1035, 0.0208, 0.0639], device='cuda:4'), in_proj_covar=tensor([0.0058, 0.0075, 0.0055, 0.0051, 0.0056, 0.0057, 0.0088, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0008, 0.0008, 0.0008, 0.0008, 0.0012, 0.0008], device='cuda:4') 2023-04-26 11:56:00,051 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0755, 0.6076, 0.8504, 0.6444, 1.2144, 0.8978, 0.7397, 0.9986], device='cuda:4'), covar=tensor([0.1701, 0.2129, 0.2366, 0.2126, 0.1165, 0.1693, 0.2016, 0.2295], device='cuda:4'), in_proj_covar=tensor([0.0328, 0.0341, 0.0346, 0.0315, 0.0352, 0.0376, 0.0327, 0.0359], device='cuda:4'), out_proj_covar=tensor([7.1782e-05, 7.3331e-05, 7.4738e-05, 6.5971e-05, 7.5099e-05, 8.2280e-05, 7.1061e-05, 7.7456e-05], device='cuda:4') 2023-04-26 11:56:03,801 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.65 vs. limit=2.0 2023-04-26 11:56:04,204 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.519e+02 2.106e+02 2.592e+02 3.021e+02 5.805e+02, threshold=5.184e+02, percent-clipped=1.0 2023-04-26 11:56:07,254 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5505, 2.2122, 1.4668, 1.2825, 1.1412, 1.1948, 1.4886, 1.1053], device='cuda:4'), covar=tensor([0.2608, 0.2168, 0.3052, 0.3630, 0.4150, 0.3152, 0.2298, 0.3373], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0223, 0.0199, 0.0219, 0.0236, 0.0198, 0.0193, 0.0211], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 11:56:11,785 INFO [finetune.py:976] (4/7) Epoch 2, batch 1300, loss[loss=0.2905, simple_loss=0.3036, pruned_loss=0.1387, over 4160.00 frames. ], tot_loss[loss=0.2746, simple_loss=0.3178, pruned_loss=0.1157, over 953388.80 frames. ], batch size: 18, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 11:56:25,542 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=7048.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 11:56:49,850 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-04-26 11:56:55,870 INFO [finetune.py:976] (4/7) Epoch 2, batch 1350, loss[loss=0.2681, simple_loss=0.3121, pruned_loss=0.1121, over 4901.00 frames. ], tot_loss[loss=0.2766, simple_loss=0.3194, pruned_loss=0.117, over 954954.80 frames. ], batch size: 36, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 11:57:34,466 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=7109.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 11:57:39,844 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.388e+02 2.079e+02 2.494e+02 3.022e+02 7.754e+02, threshold=4.988e+02, percent-clipped=2.0 2023-04-26 11:57:45,825 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=7118.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 11:57:48,842 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=7123.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:57:56,140 INFO [finetune.py:976] (4/7) Epoch 2, batch 1400, loss[loss=0.2497, simple_loss=0.3058, pruned_loss=0.09687, over 4762.00 frames. ], tot_loss[loss=0.2798, simple_loss=0.3234, pruned_loss=0.1181, over 956890.66 frames. ], batch size: 27, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 11:58:09,933 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-26 11:58:36,308 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=7163.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:58:38,104 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=7166.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:58:47,201 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=7171.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:58:51,418 INFO [finetune.py:976] (4/7) Epoch 2, batch 1450, loss[loss=0.267, simple_loss=0.3159, pruned_loss=0.109, over 4818.00 frames. ], tot_loss[loss=0.2818, simple_loss=0.3261, pruned_loss=0.1188, over 957790.49 frames. ], batch size: 39, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 11:59:15,252 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1044, 2.1175, 1.7785, 1.8372, 2.2046, 1.6460, 2.7469, 1.5280], device='cuda:4'), covar=tensor([0.4381, 0.1549, 0.4604, 0.3086, 0.2022, 0.2776, 0.1275, 0.4053], device='cuda:4'), in_proj_covar=tensor([0.0339, 0.0348, 0.0432, 0.0363, 0.0398, 0.0371, 0.0394, 0.0406], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 11:59:19,364 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.461e+02 2.205e+02 2.539e+02 2.992e+02 8.732e+02, threshold=5.079e+02, percent-clipped=2.0 2023-04-26 11:59:23,184 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=7224.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 11:59:24,343 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5152, 2.4799, 2.0711, 2.8665, 2.4334, 2.5423, 1.0561, 2.3759], device='cuda:4'), covar=tensor([0.1586, 0.1296, 0.2436, 0.1932, 0.3016, 0.1791, 0.4795, 0.2108], device='cuda:4'), in_proj_covar=tensor([0.0259, 0.0231, 0.0278, 0.0328, 0.0322, 0.0273, 0.0286, 0.0287], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 11:59:25,501 INFO [finetune.py:976] (4/7) Epoch 2, batch 1500, loss[loss=0.3132, simple_loss=0.3552, pruned_loss=0.1356, over 4806.00 frames. ], tot_loss[loss=0.2841, simple_loss=0.3284, pruned_loss=0.1199, over 957727.95 frames. ], batch size: 40, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 11:59:30,167 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=7234.0, num_to_drop=1, layers_to_drop={3} 2023-04-26 11:59:51,522 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=7265.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:59:57,760 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=7275.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 11:59:59,532 INFO [finetune.py:976] (4/7) Epoch 2, batch 1550, loss[loss=0.2327, simple_loss=0.2929, pruned_loss=0.08628, over 4810.00 frames. ], tot_loss[loss=0.2844, simple_loss=0.3283, pruned_loss=0.1202, over 954329.15 frames. ], batch size: 40, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:00:04,507 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=7285.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:00:04,556 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5296, 0.7463, 1.0739, 1.1697, 1.2436, 1.4040, 1.0557, 1.1411], device='cuda:4'), covar=tensor([2.3285, 5.5234, 3.8869, 3.5891, 3.9786, 6.4741, 5.1912, 4.1698], device='cuda:4'), in_proj_covar=tensor([0.0272, 0.0374, 0.0299, 0.0300, 0.0329, 0.0357, 0.0361, 0.0328], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 12:00:10,752 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.59 vs. limit=5.0 2023-04-26 12:00:27,652 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.588e+02 2.228e+02 2.487e+02 3.016e+02 5.047e+02, threshold=4.974e+02, percent-clipped=0.0 2023-04-26 12:00:33,829 INFO [finetune.py:976] (4/7) Epoch 2, batch 1600, loss[loss=0.2149, simple_loss=0.2728, pruned_loss=0.0785, over 4769.00 frames. ], tot_loss[loss=0.2807, simple_loss=0.3246, pruned_loss=0.1185, over 954466.63 frames. ], batch size: 28, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:00:36,973 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=7333.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:00:38,913 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=7336.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:01:07,698 INFO [finetune.py:976] (4/7) Epoch 2, batch 1650, loss[loss=0.2854, simple_loss=0.3366, pruned_loss=0.1171, over 4806.00 frames. ], tot_loss[loss=0.2773, simple_loss=0.3208, pruned_loss=0.1169, over 953993.91 frames. ], batch size: 29, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:01:09,603 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2205, 1.3396, 1.3896, 1.4805, 1.4526, 1.1570, 0.7701, 1.3342], device='cuda:4'), covar=tensor([0.1125, 0.1412, 0.1029, 0.0845, 0.0752, 0.1095, 0.1329, 0.0797], device='cuda:4'), in_proj_covar=tensor([0.0207, 0.0206, 0.0188, 0.0180, 0.0177, 0.0195, 0.0173, 0.0191], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 12:01:10,620 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-04-26 12:01:25,937 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=7404.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 12:01:28,247 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=7407.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:01:34,837 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.570e+02 2.142e+02 2.529e+02 2.973e+02 5.158e+02, threshold=5.058e+02, percent-clipped=1.0 2023-04-26 12:01:40,967 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2055, 1.2800, 1.3467, 1.4573, 1.5201, 1.1717, 0.7960, 1.3233], device='cuda:4'), covar=tensor([0.1210, 0.1436, 0.1040, 0.0869, 0.0775, 0.1152, 0.1295, 0.0762], device='cuda:4'), in_proj_covar=tensor([0.0209, 0.0207, 0.0189, 0.0181, 0.0179, 0.0196, 0.0174, 0.0192], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 12:01:41,459 INFO [finetune.py:976] (4/7) Epoch 2, batch 1700, loss[loss=0.2985, simple_loss=0.3399, pruned_loss=0.1286, over 4835.00 frames. ], tot_loss[loss=0.2746, simple_loss=0.3179, pruned_loss=0.1157, over 954352.25 frames. ], batch size: 39, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:02:09,805 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=7468.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:02:11,049 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5463, 2.1307, 1.4148, 1.3026, 1.1444, 1.2070, 1.4315, 1.0360], device='cuda:4'), covar=tensor([0.2512, 0.2306, 0.2735, 0.3566, 0.3921, 0.2996, 0.2212, 0.3256], device='cuda:4'), in_proj_covar=tensor([0.0200, 0.0225, 0.0199, 0.0220, 0.0238, 0.0198, 0.0194, 0.0211], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 12:02:22,156 INFO [finetune.py:976] (4/7) Epoch 2, batch 1750, loss[loss=0.2551, simple_loss=0.2991, pruned_loss=0.1056, over 4820.00 frames. ], tot_loss[loss=0.2748, simple_loss=0.3185, pruned_loss=0.1156, over 954546.59 frames. ], batch size: 25, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:03:11,386 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.401e+02 2.126e+02 2.582e+02 2.994e+02 4.895e+02, threshold=5.165e+02, percent-clipped=0.0 2023-04-26 12:03:12,094 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=7519.0, num_to_drop=1, layers_to_drop={3} 2023-04-26 12:03:23,644 INFO [finetune.py:976] (4/7) Epoch 2, batch 1800, loss[loss=0.2915, simple_loss=0.345, pruned_loss=0.1189, over 4902.00 frames. ], tot_loss[loss=0.2781, simple_loss=0.3227, pruned_loss=0.1167, over 955975.22 frames. ], batch size: 43, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:03:33,046 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=7534.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 12:03:38,038 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-04-26 12:04:17,657 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=7565.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:04:31,078 INFO [finetune.py:976] (4/7) Epoch 2, batch 1850, loss[loss=0.3491, simple_loss=0.3589, pruned_loss=0.1696, over 4881.00 frames. ], tot_loss[loss=0.2804, simple_loss=0.3249, pruned_loss=0.1179, over 956452.55 frames. ], batch size: 32, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:04:31,691 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4289, 2.9702, 1.0201, 1.6400, 1.7233, 2.2582, 1.9078, 0.9700], device='cuda:4'), covar=tensor([0.1297, 0.1054, 0.1937, 0.1324, 0.1057, 0.0993, 0.1430, 0.2007], device='cuda:4'), in_proj_covar=tensor([0.0126, 0.0277, 0.0154, 0.0135, 0.0146, 0.0168, 0.0132, 0.0136], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 12:04:34,062 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=7582.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 12:05:07,098 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=7613.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:05:10,536 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.515e+02 2.096e+02 2.728e+02 3.385e+02 6.441e+02, threshold=5.455e+02, percent-clipped=3.0 2023-04-26 12:05:14,375 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=7624.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:05:16,722 INFO [finetune.py:976] (4/7) Epoch 2, batch 1900, loss[loss=0.3071, simple_loss=0.3481, pruned_loss=0.133, over 4915.00 frames. ], tot_loss[loss=0.28, simple_loss=0.3253, pruned_loss=0.1173, over 957283.28 frames. ], batch size: 46, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:05:18,631 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=7631.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:05:22,248 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6721, 2.1594, 1.7257, 2.0671, 1.6167, 1.7136, 1.8051, 1.4748], device='cuda:4'), covar=tensor([0.2381, 0.1633, 0.1301, 0.1595, 0.3219, 0.1676, 0.2215, 0.3082], device='cuda:4'), in_proj_covar=tensor([0.0314, 0.0333, 0.0244, 0.0309, 0.0313, 0.0284, 0.0278, 0.0301], device='cuda:4'), out_proj_covar=tensor([1.2842e-04, 1.3678e-04, 9.9986e-05, 1.2529e-04, 1.2941e-04, 1.1484e-04, 1.1525e-04, 1.2237e-04], device='cuda:4') 2023-04-26 12:05:26,601 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.80 vs. limit=5.0 2023-04-26 12:05:50,037 INFO [finetune.py:976] (4/7) Epoch 2, batch 1950, loss[loss=0.3282, simple_loss=0.3556, pruned_loss=0.1505, over 4830.00 frames. ], tot_loss[loss=0.2765, simple_loss=0.3222, pruned_loss=0.1154, over 958419.32 frames. ], batch size: 33, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:05:54,928 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=7685.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:06:00,450 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6457, 1.9904, 1.5673, 1.1871, 1.2850, 1.3006, 1.5338, 1.2236], device='cuda:4'), covar=tensor([0.2426, 0.2157, 0.2622, 0.3348, 0.3819, 0.2871, 0.2185, 0.3142], device='cuda:4'), in_proj_covar=tensor([0.0200, 0.0225, 0.0199, 0.0220, 0.0237, 0.0198, 0.0194, 0.0211], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 12:06:07,027 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=7704.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 12:06:17,408 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.610e+02 2.113e+02 2.438e+02 2.790e+02 7.164e+02, threshold=4.875e+02, percent-clipped=1.0 2023-04-26 12:06:18,162 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7588, 2.4805, 1.5711, 1.6882, 1.2662, 1.3584, 1.6200, 1.1223], device='cuda:4'), covar=tensor([0.2495, 0.2365, 0.2743, 0.3401, 0.3919, 0.2856, 0.2261, 0.3320], device='cuda:4'), in_proj_covar=tensor([0.0200, 0.0225, 0.0199, 0.0220, 0.0237, 0.0198, 0.0194, 0.0211], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 12:06:20,645 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.48 vs. limit=2.0 2023-04-26 12:06:23,979 INFO [finetune.py:976] (4/7) Epoch 2, batch 2000, loss[loss=0.2324, simple_loss=0.2816, pruned_loss=0.09154, over 4797.00 frames. ], tot_loss[loss=0.2721, simple_loss=0.3178, pruned_loss=0.1132, over 956131.45 frames. ], batch size: 25, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:06:31,871 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6852, 1.7921, 1.0421, 1.3355, 2.0349, 1.6196, 1.4447, 1.5979], device='cuda:4'), covar=tensor([0.0614, 0.0492, 0.0478, 0.0665, 0.0330, 0.0658, 0.0607, 0.0758], device='cuda:4'), in_proj_covar=tensor([0.0032, 0.0027, 0.0025, 0.0031, 0.0022, 0.0031, 0.0030, 0.0033], device='cuda:4'), out_proj_covar=tensor([0.0048, 0.0044, 0.0039, 0.0049, 0.0037, 0.0048, 0.0047, 0.0051], device='cuda:4') 2023-04-26 12:06:33,725 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=7743.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:06:39,112 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=7752.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 12:06:42,691 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7783, 1.4468, 1.3859, 1.4443, 2.1102, 1.7570, 1.3388, 1.3288], device='cuda:4'), covar=tensor([0.1599, 0.1783, 0.2353, 0.1766, 0.0797, 0.1621, 0.2487, 0.1928], device='cuda:4'), in_proj_covar=tensor([0.0328, 0.0343, 0.0347, 0.0318, 0.0354, 0.0375, 0.0327, 0.0358], device='cuda:4'), out_proj_covar=tensor([7.1747e-05, 7.3751e-05, 7.5015e-05, 6.6630e-05, 7.5571e-05, 8.2143e-05, 7.1225e-05, 7.7330e-05], device='cuda:4') 2023-04-26 12:06:47,290 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=7763.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:06:57,780 INFO [finetune.py:976] (4/7) Epoch 2, batch 2050, loss[loss=0.2611, simple_loss=0.3029, pruned_loss=0.1096, over 4919.00 frames. ], tot_loss[loss=0.2666, simple_loss=0.312, pruned_loss=0.1106, over 954401.40 frames. ], batch size: 37, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:07:13,087 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=7802.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:07:14,341 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=7804.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:07:24,633 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.397e+02 2.038e+02 2.548e+02 3.001e+02 7.131e+02, threshold=5.096e+02, percent-clipped=2.0 2023-04-26 12:07:25,337 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=7819.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 12:07:31,147 INFO [finetune.py:976] (4/7) Epoch 2, batch 2100, loss[loss=0.3129, simple_loss=0.349, pruned_loss=0.1384, over 4924.00 frames. ], tot_loss[loss=0.2667, simple_loss=0.3118, pruned_loss=0.1108, over 955702.59 frames. ], batch size: 38, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:07:43,321 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0318, 1.8007, 2.3557, 2.5356, 1.7462, 1.4032, 1.9793, 1.0757], device='cuda:4'), covar=tensor([0.1151, 0.1160, 0.0738, 0.1008, 0.1397, 0.1664, 0.1170, 0.1765], device='cuda:4'), in_proj_covar=tensor([0.0072, 0.0080, 0.0076, 0.0075, 0.0088, 0.0096, 0.0089, 0.0081], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-26 12:07:53,418 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3407, 3.0558, 0.9669, 1.4923, 2.1027, 1.4489, 4.1650, 2.0096], device='cuda:4'), covar=tensor([0.0699, 0.0942, 0.1057, 0.1327, 0.0670, 0.1097, 0.0186, 0.0650], device='cuda:4'), in_proj_covar=tensor([0.0058, 0.0074, 0.0055, 0.0051, 0.0056, 0.0057, 0.0088, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0008, 0.0008, 0.0008, 0.0008, 0.0012, 0.0008], device='cuda:4') 2023-04-26 12:07:54,054 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=7863.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:07:56,932 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=7867.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:08:14,903 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-04-26 12:08:15,065 INFO [finetune.py:976] (4/7) Epoch 2, batch 2150, loss[loss=0.2889, simple_loss=0.3339, pruned_loss=0.122, over 4908.00 frames. ], tot_loss[loss=0.2702, simple_loss=0.3161, pruned_loss=0.1122, over 956294.34 frames. ], batch size: 36, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:08:49,696 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.54 vs. limit=2.0 2023-04-26 12:08:58,805 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.98 vs. limit=2.0 2023-04-26 12:09:01,552 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3031, 1.6411, 1.4033, 1.8838, 1.7791, 2.1860, 1.5395, 3.6586], device='cuda:4'), covar=tensor([0.0703, 0.0754, 0.0852, 0.1293, 0.0662, 0.0492, 0.0785, 0.0153], device='cuda:4'), in_proj_covar=tensor([0.0041, 0.0041, 0.0042, 0.0047, 0.0042, 0.0041, 0.0041, 0.0066], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 12:09:02,673 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.204e+02 2.251e+02 2.703e+02 3.381e+02 5.777e+02, threshold=5.406e+02, percent-clipped=2.0 2023-04-26 12:09:20,986 INFO [finetune.py:976] (4/7) Epoch 2, batch 2200, loss[loss=0.275, simple_loss=0.3133, pruned_loss=0.1183, over 4851.00 frames. ], tot_loss[loss=0.2733, simple_loss=0.3201, pruned_loss=0.1133, over 954745.74 frames. ], batch size: 44, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:09:22,962 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=7931.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:09:31,990 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=5.23 vs. limit=5.0 2023-04-26 12:09:54,033 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0349, 2.0404, 2.2881, 2.5698, 2.3819, 1.8206, 1.4047, 1.9669], device='cuda:4'), covar=tensor([0.1201, 0.0995, 0.0569, 0.0735, 0.0743, 0.1195, 0.1316, 0.0826], device='cuda:4'), in_proj_covar=tensor([0.0213, 0.0212, 0.0192, 0.0185, 0.0182, 0.0201, 0.0178, 0.0196], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 12:10:05,326 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3308, 3.1661, 0.9290, 1.8008, 1.7338, 2.3348, 1.9433, 0.9358], device='cuda:4'), covar=tensor([0.1412, 0.1016, 0.1965, 0.1295, 0.1126, 0.1039, 0.1484, 0.2102], device='cuda:4'), in_proj_covar=tensor([0.0126, 0.0276, 0.0154, 0.0134, 0.0146, 0.0169, 0.0131, 0.0135], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 12:10:05,392 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0921, 0.7945, 1.0193, 1.3904, 1.3134, 1.0354, 1.0431, 0.9998], device='cuda:4'), covar=tensor([4.4055, 6.6010, 6.9472, 7.1264, 4.4283, 7.4240, 7.5857, 5.3993], device='cuda:4'), in_proj_covar=tensor([0.0446, 0.0511, 0.0596, 0.0584, 0.0478, 0.0530, 0.0538, 0.0549], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 12:10:21,353 INFO [finetune.py:976] (4/7) Epoch 2, batch 2250, loss[loss=0.328, simple_loss=0.3709, pruned_loss=0.1426, over 4811.00 frames. ], tot_loss[loss=0.2744, simple_loss=0.3216, pruned_loss=0.1136, over 956480.51 frames. ], batch size: 38, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:10:22,019 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=7979.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:10:23,124 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=7980.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:10:47,701 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.449e+02 2.028e+02 2.420e+02 2.875e+02 5.113e+02, threshold=4.840e+02, percent-clipped=0.0 2023-04-26 12:10:55,295 INFO [finetune.py:976] (4/7) Epoch 2, batch 2300, loss[loss=0.2713, simple_loss=0.3105, pruned_loss=0.116, over 4024.00 frames. ], tot_loss[loss=0.275, simple_loss=0.322, pruned_loss=0.114, over 955767.49 frames. ], batch size: 65, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:11:18,730 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=8063.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:11:28,367 INFO [finetune.py:976] (4/7) Epoch 2, batch 2350, loss[loss=0.2902, simple_loss=0.3262, pruned_loss=0.1271, over 4896.00 frames. ], tot_loss[loss=0.2719, simple_loss=0.3184, pruned_loss=0.1127, over 955943.48 frames. ], batch size: 35, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:11:42,671 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=8099.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:11:44,308 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=2.05 vs. limit=2.0 2023-04-26 12:11:50,390 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=8111.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:11:54,659 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.190e+02 2.063e+02 2.510e+02 2.746e+02 4.758e+02, threshold=5.020e+02, percent-clipped=0.0 2023-04-26 12:12:01,656 INFO [finetune.py:976] (4/7) Epoch 2, batch 2400, loss[loss=0.2421, simple_loss=0.2791, pruned_loss=0.1026, over 4740.00 frames. ], tot_loss[loss=0.2692, simple_loss=0.3148, pruned_loss=0.1118, over 956863.28 frames. ], batch size: 23, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:12:21,842 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=8158.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:12:34,355 INFO [finetune.py:976] (4/7) Epoch 2, batch 2450, loss[loss=0.2521, simple_loss=0.3017, pruned_loss=0.1012, over 4830.00 frames. ], tot_loss[loss=0.2674, simple_loss=0.3124, pruned_loss=0.1112, over 955616.53 frames. ], batch size: 33, lr: 3.99e-03, grad_scale: 64.0 2023-04-26 12:12:49,154 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7394, 2.3484, 1.5222, 1.5239, 1.2882, 1.3168, 1.6220, 1.1816], device='cuda:4'), covar=tensor([0.2189, 0.1896, 0.2619, 0.3032, 0.3534, 0.2700, 0.1997, 0.3019], device='cuda:4'), in_proj_covar=tensor([0.0201, 0.0227, 0.0199, 0.0221, 0.0238, 0.0199, 0.0194, 0.0212], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 12:13:01,851 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.405e+02 2.093e+02 2.421e+02 2.814e+02 5.190e+02, threshold=4.843e+02, percent-clipped=1.0 2023-04-26 12:13:07,998 INFO [finetune.py:976] (4/7) Epoch 2, batch 2500, loss[loss=0.3461, simple_loss=0.3777, pruned_loss=0.1572, over 4822.00 frames. ], tot_loss[loss=0.2676, simple_loss=0.3128, pruned_loss=0.1112, over 956953.09 frames. ], batch size: 40, lr: 3.99e-03, grad_scale: 64.0 2023-04-26 12:13:49,859 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=8265.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:13:51,171 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.24 vs. limit=5.0 2023-04-26 12:14:03,521 INFO [finetune.py:976] (4/7) Epoch 2, batch 2550, loss[loss=0.2812, simple_loss=0.3356, pruned_loss=0.1134, over 4798.00 frames. ], tot_loss[loss=0.2706, simple_loss=0.3171, pruned_loss=0.1121, over 955830.09 frames. ], batch size: 45, lr: 3.99e-03, grad_scale: 64.0 2023-04-26 12:14:10,626 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=8280.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:15:02,136 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.269e+02 2.032e+02 2.420e+02 2.910e+02 5.283e+02, threshold=4.841e+02, percent-clipped=1.0 2023-04-26 12:15:13,563 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=8326.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:15:20,189 INFO [finetune.py:976] (4/7) Epoch 2, batch 2600, loss[loss=0.3038, simple_loss=0.3138, pruned_loss=0.1469, over 4022.00 frames. ], tot_loss[loss=0.2726, simple_loss=0.3189, pruned_loss=0.1132, over 953488.63 frames. ], batch size: 17, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:15:20,256 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=8328.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:15:24,055 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.31 vs. limit=5.0 2023-04-26 12:16:23,381 INFO [finetune.py:976] (4/7) Epoch 2, batch 2650, loss[loss=0.2589, simple_loss=0.3164, pruned_loss=0.1008, over 4806.00 frames. ], tot_loss[loss=0.2745, simple_loss=0.321, pruned_loss=0.114, over 954469.35 frames. ], batch size: 45, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:16:49,117 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=8399.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:17:13,447 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.592e+02 2.071e+02 2.411e+02 3.038e+02 9.260e+02, threshold=4.823e+02, percent-clipped=4.0 2023-04-26 12:17:19,413 INFO [finetune.py:976] (4/7) Epoch 2, batch 2700, loss[loss=0.3248, simple_loss=0.3529, pruned_loss=0.1483, over 4843.00 frames. ], tot_loss[loss=0.2719, simple_loss=0.3188, pruned_loss=0.1124, over 952926.07 frames. ], batch size: 44, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:17:23,799 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0384, 2.9160, 2.2095, 2.5453, 2.0298, 2.3276, 2.6287, 1.7935], device='cuda:4'), covar=tensor([0.2697, 0.1428, 0.1288, 0.1642, 0.3592, 0.1586, 0.1971, 0.3308], device='cuda:4'), in_proj_covar=tensor([0.0315, 0.0335, 0.0246, 0.0312, 0.0317, 0.0286, 0.0279, 0.0303], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 12:17:30,559 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-26 12:17:33,006 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=8447.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:17:40,764 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=8458.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:17:53,352 INFO [finetune.py:976] (4/7) Epoch 2, batch 2750, loss[loss=0.2644, simple_loss=0.3084, pruned_loss=0.1103, over 4870.00 frames. ], tot_loss[loss=0.267, simple_loss=0.3139, pruned_loss=0.11, over 954329.50 frames. ], batch size: 31, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:18:12,774 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=8506.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:18:21,018 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.402e+02 1.950e+02 2.341e+02 2.831e+02 5.180e+02, threshold=4.681e+02, percent-clipped=1.0 2023-04-26 12:18:26,542 INFO [finetune.py:976] (4/7) Epoch 2, batch 2800, loss[loss=0.2403, simple_loss=0.296, pruned_loss=0.09232, over 4904.00 frames. ], tot_loss[loss=0.2639, simple_loss=0.3103, pruned_loss=0.1087, over 954665.71 frames. ], batch size: 32, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:18:36,762 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6993, 1.8342, 1.0390, 1.3942, 2.0912, 1.6226, 1.5175, 1.5561], device='cuda:4'), covar=tensor([0.0621, 0.0458, 0.0445, 0.0620, 0.0282, 0.0619, 0.0567, 0.0721], device='cuda:4'), in_proj_covar=tensor([0.0032, 0.0027, 0.0025, 0.0031, 0.0022, 0.0031, 0.0030, 0.0032], device='cuda:4'), out_proj_covar=tensor([0.0048, 0.0044, 0.0039, 0.0049, 0.0037, 0.0048, 0.0047, 0.0051], device='cuda:4') 2023-04-26 12:18:44,369 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9345, 1.8683, 2.1195, 2.2559, 2.2443, 1.8052, 1.3861, 1.9401], device='cuda:4'), covar=tensor([0.1280, 0.1181, 0.0706, 0.0854, 0.0746, 0.1238, 0.1325, 0.0817], device='cuda:4'), in_proj_covar=tensor([0.0211, 0.0211, 0.0190, 0.0183, 0.0181, 0.0199, 0.0176, 0.0194], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 12:19:00,280 INFO [finetune.py:976] (4/7) Epoch 2, batch 2850, loss[loss=0.3299, simple_loss=0.3656, pruned_loss=0.1471, over 4906.00 frames. ], tot_loss[loss=0.2617, simple_loss=0.3084, pruned_loss=0.1075, over 955487.52 frames. ], batch size: 43, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:19:05,688 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1818, 1.3593, 1.4364, 1.4704, 1.4764, 1.1585, 0.7408, 1.3486], device='cuda:4'), covar=tensor([0.1137, 0.1235, 0.0856, 0.0771, 0.0763, 0.1147, 0.1300, 0.0748], device='cuda:4'), in_proj_covar=tensor([0.0211, 0.0210, 0.0190, 0.0183, 0.0180, 0.0199, 0.0176, 0.0194], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 12:19:45,299 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.328e+02 2.166e+02 2.478e+02 2.838e+02 4.479e+02, threshold=4.956e+02, percent-clipped=0.0 2023-04-26 12:19:52,098 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=8621.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:19:56,367 INFO [finetune.py:976] (4/7) Epoch 2, batch 2900, loss[loss=0.3144, simple_loss=0.3523, pruned_loss=0.1383, over 4858.00 frames. ], tot_loss[loss=0.2622, simple_loss=0.3093, pruned_loss=0.1075, over 953003.24 frames. ], batch size: 44, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:20:56,462 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9631, 2.8362, 2.0720, 2.5380, 2.1926, 2.2261, 2.5484, 1.7731], device='cuda:4'), covar=tensor([0.3326, 0.1864, 0.1554, 0.2104, 0.3381, 0.1800, 0.2274, 0.3598], device='cuda:4'), in_proj_covar=tensor([0.0317, 0.0336, 0.0247, 0.0313, 0.0317, 0.0287, 0.0281, 0.0304], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 12:21:07,304 INFO [finetune.py:976] (4/7) Epoch 2, batch 2950, loss[loss=0.238, simple_loss=0.304, pruned_loss=0.086, over 4758.00 frames. ], tot_loss[loss=0.2658, simple_loss=0.3133, pruned_loss=0.1092, over 953273.78 frames. ], batch size: 27, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:22:03,512 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.340e+02 2.015e+02 2.424e+02 2.946e+02 5.788e+02, threshold=4.848e+02, percent-clipped=1.0 2023-04-26 12:22:15,450 INFO [finetune.py:976] (4/7) Epoch 2, batch 3000, loss[loss=0.3159, simple_loss=0.3482, pruned_loss=0.1418, over 4243.00 frames. ], tot_loss[loss=0.2669, simple_loss=0.3145, pruned_loss=0.1096, over 950857.78 frames. ], batch size: 65, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:22:15,450 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-26 12:22:20,847 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5967, 1.7620, 1.8626, 2.0132, 1.9694, 1.5387, 1.2291, 1.7952], device='cuda:4'), covar=tensor([0.1228, 0.1142, 0.0717, 0.0758, 0.0730, 0.1165, 0.1269, 0.0680], device='cuda:4'), in_proj_covar=tensor([0.0210, 0.0209, 0.0189, 0.0182, 0.0180, 0.0198, 0.0176, 0.0193], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 12:22:32,547 INFO [finetune.py:1010] (4/7) Epoch 2, validation: loss=0.1863, simple_loss=0.2571, pruned_loss=0.0578, over 2265189.00 frames. 2023-04-26 12:22:32,547 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6259MB 2023-04-26 12:22:42,138 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1911, 2.3583, 1.8616, 1.9940, 2.2530, 1.7980, 2.9518, 1.3115], device='cuda:4'), covar=tensor([0.3613, 0.1274, 0.3224, 0.2304, 0.1755, 0.2362, 0.0902, 0.4014], device='cuda:4'), in_proj_covar=tensor([0.0344, 0.0350, 0.0434, 0.0368, 0.0404, 0.0374, 0.0398, 0.0410], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 12:23:09,087 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=8760.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:23:21,913 INFO [finetune.py:976] (4/7) Epoch 2, batch 3050, loss[loss=0.2644, simple_loss=0.3096, pruned_loss=0.1096, over 4882.00 frames. ], tot_loss[loss=0.2659, simple_loss=0.3147, pruned_loss=0.1086, over 952782.23 frames. ], batch size: 35, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:23:27,627 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.80 vs. limit=5.0 2023-04-26 12:23:27,888 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1508, 1.5517, 5.4796, 5.1373, 4.8071, 5.0905, 4.7827, 4.8837], device='cuda:4'), covar=tensor([0.6412, 0.6000, 0.0915, 0.1596, 0.0890, 0.1368, 0.1147, 0.1747], device='cuda:4'), in_proj_covar=tensor([0.0336, 0.0315, 0.0449, 0.0457, 0.0380, 0.0434, 0.0343, 0.0401], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0002], device='cuda:4') 2023-04-26 12:23:49,151 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.298e+02 2.073e+02 2.370e+02 3.091e+02 5.908e+02, threshold=4.740e+02, percent-clipped=3.0 2023-04-26 12:23:50,999 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=8821.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:23:56,176 INFO [finetune.py:976] (4/7) Epoch 2, batch 3100, loss[loss=0.2744, simple_loss=0.3126, pruned_loss=0.1181, over 4795.00 frames. ], tot_loss[loss=0.2632, simple_loss=0.3119, pruned_loss=0.1072, over 952397.43 frames. ], batch size: 45, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:24:23,905 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.94 vs. limit=5.0 2023-04-26 12:24:29,317 INFO [finetune.py:976] (4/7) Epoch 2, batch 3150, loss[loss=0.2663, simple_loss=0.3046, pruned_loss=0.114, over 4772.00 frames. ], tot_loss[loss=0.2617, simple_loss=0.3094, pruned_loss=0.107, over 953801.45 frames. ], batch size: 28, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:24:40,619 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1200, 2.4855, 0.9830, 1.3261, 1.7860, 1.2773, 3.0169, 1.6110], device='cuda:4'), covar=tensor([0.0693, 0.0528, 0.0819, 0.1187, 0.0525, 0.0979, 0.0223, 0.0645], device='cuda:4'), in_proj_covar=tensor([0.0057, 0.0074, 0.0055, 0.0051, 0.0056, 0.0057, 0.0087, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0008, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 12:24:56,258 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.408e+02 2.116e+02 2.518e+02 3.010e+02 6.088e+02, threshold=5.037e+02, percent-clipped=1.0 2023-04-26 12:24:57,597 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=8921.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:25:01,753 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=8927.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:25:02,222 INFO [finetune.py:976] (4/7) Epoch 2, batch 3200, loss[loss=0.2958, simple_loss=0.3343, pruned_loss=0.1287, over 4935.00 frames. ], tot_loss[loss=0.2569, simple_loss=0.3045, pruned_loss=0.1047, over 955290.76 frames. ], batch size: 38, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:25:18,829 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8115, 1.8435, 1.9546, 2.1446, 2.1931, 1.6910, 1.3649, 1.8528], device='cuda:4'), covar=tensor([0.1110, 0.1109, 0.0690, 0.0750, 0.0602, 0.1140, 0.1259, 0.0712], device='cuda:4'), in_proj_covar=tensor([0.0209, 0.0208, 0.0188, 0.0182, 0.0179, 0.0197, 0.0175, 0.0192], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 12:25:29,670 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4198, 2.6564, 1.1516, 1.4903, 2.2003, 1.5303, 3.6412, 1.8939], device='cuda:4'), covar=tensor([0.0631, 0.0657, 0.0839, 0.1301, 0.0523, 0.0920, 0.0190, 0.0649], device='cuda:4'), in_proj_covar=tensor([0.0057, 0.0074, 0.0055, 0.0051, 0.0056, 0.0057, 0.0087, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0008, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 12:25:30,261 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=8969.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:25:41,812 INFO [finetune.py:976] (4/7) Epoch 2, batch 3250, loss[loss=0.2792, simple_loss=0.3227, pruned_loss=0.1178, over 4765.00 frames. ], tot_loss[loss=0.2575, simple_loss=0.305, pruned_loss=0.105, over 954674.33 frames. ], batch size: 27, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:25:53,825 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=8986.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:25:55,567 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=8988.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:25:56,535 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-26 12:26:27,412 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=9013.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:26:37,526 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.210e+02 2.159e+02 2.478e+02 3.003e+02 4.851e+02, threshold=4.955e+02, percent-clipped=0.0 2023-04-26 12:26:39,308 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-26 12:26:48,555 INFO [finetune.py:976] (4/7) Epoch 2, batch 3300, loss[loss=0.2622, simple_loss=0.3174, pruned_loss=0.1035, over 4906.00 frames. ], tot_loss[loss=0.2618, simple_loss=0.3095, pruned_loss=0.1071, over 953576.28 frames. ], batch size: 37, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:27:05,887 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7466, 1.5564, 1.6952, 2.1120, 2.1033, 1.6252, 1.3928, 1.7931], device='cuda:4'), covar=tensor([0.0981, 0.1211, 0.0743, 0.0625, 0.0550, 0.1000, 0.1104, 0.0745], device='cuda:4'), in_proj_covar=tensor([0.0210, 0.0210, 0.0189, 0.0184, 0.0180, 0.0199, 0.0176, 0.0194], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 12:27:08,780 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=9047.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:27:26,233 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=9074.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:27:28,529 INFO [finetune.py:976] (4/7) Epoch 2, batch 3350, loss[loss=0.2722, simple_loss=0.3198, pruned_loss=0.1122, over 4857.00 frames. ], tot_loss[loss=0.2641, simple_loss=0.3117, pruned_loss=0.1083, over 953432.14 frames. ], batch size: 34, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:27:58,927 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0243, 2.1489, 1.7850, 1.8200, 2.0748, 1.7586, 2.7921, 1.3796], device='cuda:4'), covar=tensor([0.4575, 0.1533, 0.4572, 0.3069, 0.2101, 0.2765, 0.1186, 0.4948], device='cuda:4'), in_proj_covar=tensor([0.0346, 0.0351, 0.0436, 0.0368, 0.0404, 0.0374, 0.0398, 0.0412], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 12:28:00,088 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=9116.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:28:07,187 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.418e+02 2.168e+02 2.587e+02 3.094e+02 5.996e+02, threshold=5.175e+02, percent-clipped=1.0 2023-04-26 12:28:15,832 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5406, 1.5823, 1.7320, 1.8708, 1.8498, 1.4651, 1.0388, 1.5813], device='cuda:4'), covar=tensor([0.1078, 0.1158, 0.0656, 0.0721, 0.0642, 0.1051, 0.1325, 0.0746], device='cuda:4'), in_proj_covar=tensor([0.0212, 0.0211, 0.0190, 0.0185, 0.0181, 0.0200, 0.0178, 0.0195], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 12:28:17,532 INFO [finetune.py:976] (4/7) Epoch 2, batch 3400, loss[loss=0.277, simple_loss=0.3212, pruned_loss=0.1164, over 4886.00 frames. ], tot_loss[loss=0.2665, simple_loss=0.3137, pruned_loss=0.1096, over 952833.48 frames. ], batch size: 32, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:29:01,795 INFO [finetune.py:976] (4/7) Epoch 2, batch 3450, loss[loss=0.2905, simple_loss=0.3312, pruned_loss=0.1249, over 4891.00 frames. ], tot_loss[loss=0.2667, simple_loss=0.3141, pruned_loss=0.1097, over 953732.77 frames. ], batch size: 37, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:29:16,217 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.75 vs. limit=5.0 2023-04-26 12:29:29,510 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.432e+02 2.133e+02 2.511e+02 2.934e+02 6.196e+02, threshold=5.021e+02, percent-clipped=2.0 2023-04-26 12:29:35,076 INFO [finetune.py:976] (4/7) Epoch 2, batch 3500, loss[loss=0.1728, simple_loss=0.2339, pruned_loss=0.05585, over 4895.00 frames. ], tot_loss[loss=0.2624, simple_loss=0.3098, pruned_loss=0.1075, over 955356.43 frames. ], batch size: 43, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:29:41,698 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.75 vs. limit=2.0 2023-04-26 12:30:08,951 INFO [finetune.py:976] (4/7) Epoch 2, batch 3550, loss[loss=0.2055, simple_loss=0.2617, pruned_loss=0.07466, over 4763.00 frames. ], tot_loss[loss=0.258, simple_loss=0.3056, pruned_loss=0.1052, over 955289.26 frames. ], batch size: 28, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:30:12,102 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=9283.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:30:37,460 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.026e+02 2.002e+02 2.435e+02 2.997e+02 6.904e+02, threshold=4.871e+02, percent-clipped=3.0 2023-04-26 12:30:48,982 INFO [finetune.py:976] (4/7) Epoch 2, batch 3600, loss[loss=0.2478, simple_loss=0.2931, pruned_loss=0.1013, over 4863.00 frames. ], tot_loss[loss=0.2551, simple_loss=0.3026, pruned_loss=0.1038, over 956798.23 frames. ], batch size: 31, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:31:03,419 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=9342.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:31:22,338 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.53 vs. limit=2.0 2023-04-26 12:31:23,248 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4379, 2.9371, 1.3687, 1.6166, 2.2371, 1.5712, 3.5220, 1.8642], device='cuda:4'), covar=tensor([0.0585, 0.0984, 0.0935, 0.0996, 0.0481, 0.0874, 0.0201, 0.0579], device='cuda:4'), in_proj_covar=tensor([0.0057, 0.0074, 0.0054, 0.0051, 0.0056, 0.0056, 0.0087, 0.0054], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0008, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 12:31:34,836 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-04-26 12:31:44,877 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=9369.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:31:55,597 INFO [finetune.py:976] (4/7) Epoch 2, batch 3650, loss[loss=0.246, simple_loss=0.2975, pruned_loss=0.09727, over 4768.00 frames. ], tot_loss[loss=0.2552, simple_loss=0.3032, pruned_loss=0.1037, over 955425.16 frames. ], batch size: 28, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:31:59,552 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-26 12:32:25,410 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=9414.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:32:27,087 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=9416.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:32:28,804 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.507e+02 2.166e+02 2.483e+02 2.970e+02 6.745e+02, threshold=4.965e+02, percent-clipped=1.0 2023-04-26 12:32:34,828 INFO [finetune.py:976] (4/7) Epoch 2, batch 3700, loss[loss=0.27, simple_loss=0.3159, pruned_loss=0.112, over 4817.00 frames. ], tot_loss[loss=0.2581, simple_loss=0.3066, pruned_loss=0.1048, over 954585.15 frames. ], batch size: 33, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:32:58,819 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=9464.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:33:00,747 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.69 vs. limit=5.0 2023-04-26 12:33:06,999 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=9475.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:33:08,729 INFO [finetune.py:976] (4/7) Epoch 2, batch 3750, loss[loss=0.2589, simple_loss=0.3179, pruned_loss=0.09997, over 4739.00 frames. ], tot_loss[loss=0.2592, simple_loss=0.3079, pruned_loss=0.1053, over 954103.96 frames. ], batch size: 54, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:33:41,751 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0460, 2.3172, 1.1199, 1.2861, 1.7388, 1.2404, 3.0193, 1.5416], device='cuda:4'), covar=tensor([0.0676, 0.0600, 0.0771, 0.1333, 0.0553, 0.1075, 0.0274, 0.0727], device='cuda:4'), in_proj_covar=tensor([0.0056, 0.0073, 0.0054, 0.0050, 0.0055, 0.0056, 0.0086, 0.0054], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0008, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 12:33:47,550 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.577e+02 2.140e+02 2.529e+02 3.098e+02 5.505e+02, threshold=5.058e+02, percent-clipped=1.0 2023-04-26 12:33:59,189 INFO [finetune.py:976] (4/7) Epoch 2, batch 3800, loss[loss=0.2617, simple_loss=0.3228, pruned_loss=0.1003, over 4922.00 frames. ], tot_loss[loss=0.2631, simple_loss=0.3117, pruned_loss=0.1072, over 953561.41 frames. ], batch size: 42, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:34:55,618 INFO [finetune.py:976] (4/7) Epoch 2, batch 3850, loss[loss=0.1938, simple_loss=0.2475, pruned_loss=0.07001, over 4701.00 frames. ], tot_loss[loss=0.2615, simple_loss=0.3104, pruned_loss=0.1063, over 952823.70 frames. ], batch size: 23, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:34:58,833 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=9583.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:34:59,480 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=9584.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:35:03,831 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1903, 1.4095, 1.8771, 2.5635, 1.7341, 1.4948, 1.2491, 1.6815], device='cuda:4'), covar=tensor([0.5893, 0.7616, 0.3567, 0.6546, 0.8008, 0.5390, 0.9999, 0.7287], device='cuda:4'), in_proj_covar=tensor([0.0268, 0.0277, 0.0222, 0.0347, 0.0236, 0.0235, 0.0270, 0.0218], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 12:35:22,633 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.190e+02 2.119e+02 2.508e+02 2.913e+02 5.368e+02, threshold=5.017e+02, percent-clipped=1.0 2023-04-26 12:35:29,153 INFO [finetune.py:976] (4/7) Epoch 2, batch 3900, loss[loss=0.3137, simple_loss=0.3349, pruned_loss=0.1462, over 4247.00 frames. ], tot_loss[loss=0.2597, simple_loss=0.3076, pruned_loss=0.1058, over 952612.36 frames. ], batch size: 65, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:35:38,164 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=9631.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:35:49,614 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7766, 1.2284, 1.5635, 1.8246, 1.4200, 1.1814, 0.8812, 1.2431], device='cuda:4'), covar=tensor([0.5738, 0.6919, 0.3154, 0.5404, 0.7043, 0.5492, 0.9589, 0.6753], device='cuda:4'), in_proj_covar=tensor([0.0267, 0.0277, 0.0222, 0.0346, 0.0235, 0.0234, 0.0270, 0.0217], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 12:35:50,772 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=9642.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:35:52,642 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9945, 1.9288, 1.7401, 1.5941, 2.0823, 1.6414, 2.6885, 1.5861], device='cuda:4'), covar=tensor([0.5075, 0.2124, 0.5465, 0.3606, 0.2054, 0.3171, 0.1299, 0.4590], device='cuda:4'), in_proj_covar=tensor([0.0345, 0.0348, 0.0435, 0.0368, 0.0402, 0.0371, 0.0397, 0.0411], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 12:35:52,654 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=9645.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:36:12,998 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=9658.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:36:26,765 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=9669.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:36:38,399 INFO [finetune.py:976] (4/7) Epoch 2, batch 3950, loss[loss=0.2718, simple_loss=0.3043, pruned_loss=0.1196, over 4747.00 frames. ], tot_loss[loss=0.2544, simple_loss=0.3026, pruned_loss=0.103, over 955275.91 frames. ], batch size: 54, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:36:48,484 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-04-26 12:36:55,195 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=9690.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:37:29,749 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=9717.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:37:36,035 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.382e+02 1.910e+02 2.412e+02 2.849e+02 4.927e+02, threshold=4.824e+02, percent-clipped=0.0 2023-04-26 12:37:36,690 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=9719.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:37:47,958 INFO [finetune.py:976] (4/7) Epoch 2, batch 4000, loss[loss=0.2338, simple_loss=0.2729, pruned_loss=0.09736, over 4223.00 frames. ], tot_loss[loss=0.2535, simple_loss=0.3017, pruned_loss=0.1027, over 955190.95 frames. ], batch size: 18, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:37:58,324 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=9743.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:38:15,574 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=9770.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:38:20,827 INFO [finetune.py:976] (4/7) Epoch 2, batch 4050, loss[loss=0.2681, simple_loss=0.328, pruned_loss=0.1041, over 4777.00 frames. ], tot_loss[loss=0.2559, simple_loss=0.3043, pruned_loss=0.1038, over 953992.98 frames. ], batch size: 54, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:38:38,421 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=9804.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:38:47,819 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.265e+02 2.123e+02 2.473e+02 2.970e+02 7.758e+02, threshold=4.946e+02, percent-clipped=3.0 2023-04-26 12:38:54,762 INFO [finetune.py:976] (4/7) Epoch 2, batch 4100, loss[loss=0.2673, simple_loss=0.3196, pruned_loss=0.1075, over 4905.00 frames. ], tot_loss[loss=0.2596, simple_loss=0.3083, pruned_loss=0.1054, over 952061.18 frames. ], batch size: 37, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:39:34,298 INFO [finetune.py:976] (4/7) Epoch 2, batch 4150, loss[loss=0.3242, simple_loss=0.3542, pruned_loss=0.1471, over 4185.00 frames. ], tot_loss[loss=0.2633, simple_loss=0.3114, pruned_loss=0.1076, over 947695.43 frames. ], batch size: 66, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:39:38,566 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4598, 1.2485, 1.0977, 1.3972, 1.7274, 1.4572, 1.2767, 1.0440], device='cuda:4'), covar=tensor([0.2119, 0.1765, 0.2449, 0.1645, 0.0991, 0.2086, 0.2313, 0.2076], device='cuda:4'), in_proj_covar=tensor([0.0327, 0.0342, 0.0351, 0.0319, 0.0356, 0.0373, 0.0325, 0.0358], device='cuda:4'), out_proj_covar=tensor([7.1409e-05, 7.3518e-05, 7.5938e-05, 6.6900e-05, 7.5941e-05, 8.1569e-05, 7.0826e-05, 7.7355e-05], device='cuda:4') 2023-04-26 12:40:06,637 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-26 12:40:10,748 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3981, 2.3800, 2.7223, 2.8614, 2.7140, 2.2819, 1.6505, 2.4097], device='cuda:4'), covar=tensor([0.1179, 0.1009, 0.0592, 0.0749, 0.0683, 0.1130, 0.1399, 0.0769], device='cuda:4'), in_proj_covar=tensor([0.0214, 0.0214, 0.0192, 0.0186, 0.0183, 0.0203, 0.0180, 0.0198], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 12:40:18,085 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.356e+02 2.059e+02 2.438e+02 3.061e+02 5.791e+02, threshold=4.877e+02, percent-clipped=3.0 2023-04-26 12:40:29,030 INFO [finetune.py:976] (4/7) Epoch 2, batch 4200, loss[loss=0.2346, simple_loss=0.2894, pruned_loss=0.08989, over 4895.00 frames. ], tot_loss[loss=0.2617, simple_loss=0.3107, pruned_loss=0.1064, over 947961.52 frames. ], batch size: 35, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:40:41,577 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5879, 1.9071, 1.6226, 1.8631, 1.5767, 1.5380, 1.7463, 1.2795], device='cuda:4'), covar=tensor([0.2170, 0.1532, 0.1286, 0.1710, 0.3658, 0.1689, 0.2084, 0.3239], device='cuda:4'), in_proj_covar=tensor([0.0319, 0.0340, 0.0249, 0.0314, 0.0317, 0.0290, 0.0282, 0.0305], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 12:40:49,801 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=9940.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:41:36,131 INFO [finetune.py:976] (4/7) Epoch 2, batch 4250, loss[loss=0.239, simple_loss=0.2978, pruned_loss=0.09012, over 4792.00 frames. ], tot_loss[loss=0.2585, simple_loss=0.3077, pruned_loss=0.1046, over 948779.53 frames. ], batch size: 51, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:42:19,442 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8312, 1.2788, 1.6182, 1.8569, 1.4605, 1.2482, 0.8250, 1.2818], device='cuda:4'), covar=tensor([0.5291, 0.6422, 0.2878, 0.5220, 0.6454, 0.4869, 0.8838, 0.6298], device='cuda:4'), in_proj_covar=tensor([0.0268, 0.0278, 0.0223, 0.0348, 0.0235, 0.0234, 0.0269, 0.0217], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 12:42:30,341 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=10014.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:42:33,325 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.150e+02 1.966e+02 2.336e+02 2.838e+02 4.911e+02, threshold=4.671e+02, percent-clipped=1.0 2023-04-26 12:42:43,920 INFO [finetune.py:976] (4/7) Epoch 2, batch 4300, loss[loss=0.2556, simple_loss=0.3001, pruned_loss=0.1056, over 4811.00 frames. ], tot_loss[loss=0.2562, simple_loss=0.3049, pruned_loss=0.1037, over 950086.74 frames. ], batch size: 51, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:42:50,006 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8522, 1.2988, 1.5829, 1.8002, 1.4540, 1.2298, 0.8550, 1.2022], device='cuda:4'), covar=tensor([0.5806, 0.6939, 0.3258, 0.5422, 0.6862, 0.5121, 0.9302, 0.6992], device='cuda:4'), in_proj_covar=tensor([0.0269, 0.0278, 0.0223, 0.0349, 0.0236, 0.0235, 0.0270, 0.0218], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 12:43:02,760 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=10039.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:43:18,156 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=10062.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 12:43:23,023 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=10070.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:43:27,737 INFO [finetune.py:976] (4/7) Epoch 2, batch 4350, loss[loss=0.2769, simple_loss=0.3118, pruned_loss=0.121, over 4746.00 frames. ], tot_loss[loss=0.2536, simple_loss=0.3018, pruned_loss=0.1027, over 951650.01 frames. ], batch size: 23, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:43:32,579 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6025, 1.9681, 1.5305, 1.8982, 1.5841, 1.5475, 1.7387, 1.2228], device='cuda:4'), covar=tensor([0.2224, 0.1542, 0.1484, 0.1585, 0.3526, 0.1798, 0.2080, 0.3328], device='cuda:4'), in_proj_covar=tensor([0.0318, 0.0338, 0.0248, 0.0312, 0.0316, 0.0289, 0.0280, 0.0304], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 12:43:42,004 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=10099.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:43:43,140 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=10100.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 12:43:54,963 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=10118.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:43:55,492 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.384e+02 2.038e+02 2.484e+02 2.915e+02 5.390e+02, threshold=4.968e+02, percent-clipped=2.0 2023-04-26 12:43:58,081 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=10123.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 12:44:00,980 INFO [finetune.py:976] (4/7) Epoch 2, batch 4400, loss[loss=0.2724, simple_loss=0.3314, pruned_loss=0.1067, over 4817.00 frames. ], tot_loss[loss=0.2544, simple_loss=0.3028, pruned_loss=0.103, over 952605.47 frames. ], batch size: 39, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:44:34,792 INFO [finetune.py:976] (4/7) Epoch 2, batch 4450, loss[loss=0.2165, simple_loss=0.2668, pruned_loss=0.08312, over 4417.00 frames. ], tot_loss[loss=0.2554, simple_loss=0.3048, pruned_loss=0.103, over 951971.95 frames. ], batch size: 19, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:44:49,278 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-04-26 12:44:50,198 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6403, 1.9856, 1.0177, 1.3445, 2.2308, 1.5545, 1.4380, 1.5656], device='cuda:4'), covar=tensor([0.0616, 0.0423, 0.0427, 0.0618, 0.0283, 0.0594, 0.0570, 0.0699], device='cuda:4'), in_proj_covar=tensor([0.0032, 0.0027, 0.0025, 0.0031, 0.0022, 0.0031, 0.0030, 0.0032], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0039, 0.0049, 0.0038, 0.0048, 0.0047, 0.0051], device='cuda:4') 2023-04-26 12:45:03,113 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.522e+02 2.117e+02 2.495e+02 3.123e+02 6.793e+02, threshold=4.991e+02, percent-clipped=1.0 2023-04-26 12:45:08,638 INFO [finetune.py:976] (4/7) Epoch 2, batch 4500, loss[loss=0.2304, simple_loss=0.268, pruned_loss=0.09641, over 4312.00 frames. ], tot_loss[loss=0.2561, simple_loss=0.3059, pruned_loss=0.1031, over 950588.55 frames. ], batch size: 18, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:45:16,000 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=10240.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:45:42,230 INFO [finetune.py:976] (4/7) Epoch 2, batch 4550, loss[loss=0.2591, simple_loss=0.3117, pruned_loss=0.1033, over 4766.00 frames. ], tot_loss[loss=0.257, simple_loss=0.3071, pruned_loss=0.1035, over 951656.50 frames. ], batch size: 51, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:45:48,411 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=10288.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:45:48,458 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0272, 1.3201, 1.2330, 1.6070, 1.4122, 1.6088, 1.2685, 2.4820], device='cuda:4'), covar=tensor([0.0728, 0.0915, 0.0909, 0.1378, 0.0774, 0.0570, 0.0832, 0.0260], device='cuda:4'), in_proj_covar=tensor([0.0040, 0.0041, 0.0042, 0.0046, 0.0042, 0.0041, 0.0041, 0.0065], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 12:46:21,468 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=10314.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:46:28,150 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5206, 1.2380, 0.3153, 1.2136, 1.2146, 1.3970, 1.2874, 1.2960], device='cuda:4'), covar=tensor([0.0731, 0.0445, 0.0541, 0.0721, 0.0350, 0.0737, 0.0727, 0.0770], device='cuda:4'), in_proj_covar=tensor([0.0032, 0.0027, 0.0025, 0.0031, 0.0022, 0.0031, 0.0030, 0.0032], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0049, 0.0038, 0.0048, 0.0047, 0.0051], device='cuda:4') 2023-04-26 12:46:29,829 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.123e+02 1.925e+02 2.197e+02 2.697e+02 5.130e+02, threshold=4.395e+02, percent-clipped=1.0 2023-04-26 12:46:30,601 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3485, 1.8160, 2.2083, 2.5700, 2.0816, 1.7686, 1.5935, 1.9132], device='cuda:4'), covar=tensor([0.5459, 0.6488, 0.3069, 0.5639, 0.6775, 0.4994, 0.8110, 0.6828], device='cuda:4'), in_proj_covar=tensor([0.0269, 0.0277, 0.0223, 0.0347, 0.0235, 0.0235, 0.0269, 0.0216], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 12:46:41,911 INFO [finetune.py:976] (4/7) Epoch 2, batch 4600, loss[loss=0.2617, simple_loss=0.3072, pruned_loss=0.108, over 4929.00 frames. ], tot_loss[loss=0.255, simple_loss=0.3055, pruned_loss=0.1023, over 949361.49 frames. ], batch size: 33, lr: 3.99e-03, grad_scale: 64.0 2023-04-26 12:47:09,969 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=10362.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:47:19,243 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0650, 1.5004, 1.3254, 1.7471, 1.5681, 1.9934, 1.3858, 3.3961], device='cuda:4'), covar=tensor([0.0721, 0.0798, 0.0817, 0.1292, 0.0691, 0.0581, 0.0771, 0.0169], device='cuda:4'), in_proj_covar=tensor([0.0040, 0.0041, 0.0042, 0.0047, 0.0042, 0.0041, 0.0041, 0.0065], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 12:47:29,149 INFO [finetune.py:976] (4/7) Epoch 2, batch 4650, loss[loss=0.2104, simple_loss=0.2754, pruned_loss=0.07268, over 4786.00 frames. ], tot_loss[loss=0.2532, simple_loss=0.3039, pruned_loss=0.1012, over 953082.95 frames. ], batch size: 29, lr: 3.99e-03, grad_scale: 64.0 2023-04-26 12:47:50,978 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=10395.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 12:47:58,962 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=10399.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:48:12,475 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=10409.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:48:24,001 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=10418.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 12:48:25,137 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.154e+02 2.119e+02 2.411e+02 2.762e+02 5.438e+02, threshold=4.823e+02, percent-clipped=2.0 2023-04-26 12:48:35,916 INFO [finetune.py:976] (4/7) Epoch 2, batch 4700, loss[loss=0.2256, simple_loss=0.2735, pruned_loss=0.08888, over 3115.00 frames. ], tot_loss[loss=0.2518, simple_loss=0.3013, pruned_loss=0.1011, over 950686.50 frames. ], batch size: 13, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:48:53,256 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=10447.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:48:57,648 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=10454.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:49:07,975 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.97 vs. limit=2.0 2023-04-26 12:49:10,017 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=10470.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:49:15,192 INFO [finetune.py:976] (4/7) Epoch 2, batch 4750, loss[loss=0.2385, simple_loss=0.283, pruned_loss=0.09701, over 4813.00 frames. ], tot_loss[loss=0.2492, simple_loss=0.2986, pruned_loss=0.09987, over 952690.98 frames. ], batch size: 51, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:49:40,436 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=10515.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 12:49:43,424 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.363e+02 1.957e+02 2.292e+02 2.912e+02 6.222e+02, threshold=4.583e+02, percent-clipped=3.0 2023-04-26 12:49:49,341 INFO [finetune.py:976] (4/7) Epoch 2, batch 4800, loss[loss=0.2159, simple_loss=0.2749, pruned_loss=0.07844, over 4832.00 frames. ], tot_loss[loss=0.2518, simple_loss=0.3014, pruned_loss=0.1011, over 953339.82 frames. ], batch size: 30, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:49:55,672 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.61 vs. limit=2.0 2023-04-26 12:50:23,196 INFO [finetune.py:976] (4/7) Epoch 2, batch 4850, loss[loss=0.2416, simple_loss=0.3038, pruned_loss=0.08964, over 4927.00 frames. ], tot_loss[loss=0.2543, simple_loss=0.3049, pruned_loss=0.1019, over 953154.00 frames. ], batch size: 38, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:50:50,776 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.143e+02 2.100e+02 2.315e+02 2.802e+02 4.090e+02, threshold=4.629e+02, percent-clipped=1.0 2023-04-26 12:50:56,077 INFO [finetune.py:976] (4/7) Epoch 2, batch 4900, loss[loss=0.261, simple_loss=0.314, pruned_loss=0.104, over 4815.00 frames. ], tot_loss[loss=0.255, simple_loss=0.3056, pruned_loss=0.1022, over 953250.02 frames. ], batch size: 40, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:51:46,566 INFO [finetune.py:976] (4/7) Epoch 2, batch 4950, loss[loss=0.266, simple_loss=0.3196, pruned_loss=0.1062, over 4889.00 frames. ], tot_loss[loss=0.2552, simple_loss=0.3062, pruned_loss=0.102, over 953859.52 frames. ], batch size: 35, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:52:10,652 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=10695.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:52:42,694 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=10718.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 12:52:43,792 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.438e+02 1.913e+02 2.320e+02 2.882e+02 5.075e+02, threshold=4.640e+02, percent-clipped=3.0 2023-04-26 12:52:54,986 INFO [finetune.py:976] (4/7) Epoch 2, batch 5000, loss[loss=0.213, simple_loss=0.2658, pruned_loss=0.08009, over 4877.00 frames. ], tot_loss[loss=0.2526, simple_loss=0.3038, pruned_loss=0.1007, over 953857.75 frames. ], batch size: 31, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:53:16,594 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=10743.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:53:17,261 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=10744.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:53:29,936 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=10765.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:53:30,548 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=10766.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 12:53:42,295 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=10774.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:53:45,586 INFO [finetune.py:976] (4/7) Epoch 2, batch 5050, loss[loss=0.246, simple_loss=0.2989, pruned_loss=0.09657, over 4855.00 frames. ], tot_loss[loss=0.2517, simple_loss=0.3024, pruned_loss=0.1005, over 955389.78 frames. ], batch size: 44, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:53:53,131 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0159, 1.0888, 1.2889, 1.4732, 1.4699, 1.6273, 1.3446, 1.4202], device='cuda:4'), covar=tensor([1.6579, 2.7692, 2.2828, 2.0632, 2.2093, 3.7939, 2.7551, 2.3723], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0392, 0.0311, 0.0315, 0.0344, 0.0388, 0.0378, 0.0341], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 12:54:08,864 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-04-26 12:54:25,688 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=10805.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:54:28,748 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=10810.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 12:54:40,509 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.434e+02 1.942e+02 2.267e+02 2.719e+02 4.388e+02, threshold=4.533e+02, percent-clipped=0.0 2023-04-26 12:54:51,193 INFO [finetune.py:976] (4/7) Epoch 2, batch 5100, loss[loss=0.2247, simple_loss=0.2745, pruned_loss=0.08749, over 4776.00 frames. ], tot_loss[loss=0.2477, simple_loss=0.2988, pruned_loss=0.09836, over 955305.33 frames. ], batch size: 26, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:55:01,978 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=10835.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:55:30,149 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4981, 1.0614, 1.3329, 1.0630, 1.6465, 1.4236, 1.0963, 1.3710], device='cuda:4'), covar=tensor([0.1617, 0.1667, 0.2045, 0.2036, 0.0991, 0.1707, 0.2199, 0.2015], device='cuda:4'), in_proj_covar=tensor([0.0321, 0.0338, 0.0345, 0.0314, 0.0349, 0.0364, 0.0319, 0.0352], device='cuda:4'), out_proj_covar=tensor([7.0154e-05, 7.2646e-05, 7.4679e-05, 6.6005e-05, 7.4351e-05, 7.9693e-05, 6.9447e-05, 7.5991e-05], device='cuda:4') 2023-04-26 12:55:45,712 INFO [finetune.py:976] (4/7) Epoch 2, batch 5150, loss[loss=0.25, simple_loss=0.2904, pruned_loss=0.1048, over 4693.00 frames. ], tot_loss[loss=0.2474, simple_loss=0.2983, pruned_loss=0.09831, over 955947.58 frames. ], batch size: 23, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:55:46,413 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3558, 2.4958, 1.0347, 1.3856, 2.0115, 1.3915, 3.3939, 1.8001], device='cuda:4'), covar=tensor([0.0616, 0.0692, 0.0874, 0.1351, 0.0519, 0.1031, 0.0329, 0.0652], device='cuda:4'), in_proj_covar=tensor([0.0056, 0.0073, 0.0054, 0.0050, 0.0055, 0.0056, 0.0086, 0.0054], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0008, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 12:55:50,628 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=10885.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 12:55:53,608 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4900, 1.3932, 0.7564, 1.1778, 1.6175, 1.3805, 1.2738, 1.3232], device='cuda:4'), covar=tensor([0.0604, 0.0476, 0.0493, 0.0634, 0.0359, 0.0594, 0.0590, 0.0714], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0027, 0.0025, 0.0030, 0.0022, 0.0031, 0.0030, 0.0032], device='cuda:4'), out_proj_covar=tensor([0.0048, 0.0044, 0.0038, 0.0049, 0.0037, 0.0047, 0.0047, 0.0050], device='cuda:4') 2023-04-26 12:56:13,138 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.326e+02 2.074e+02 2.408e+02 2.882e+02 5.966e+02, threshold=4.815e+02, percent-clipped=3.0 2023-04-26 12:56:16,895 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3041, 2.3383, 2.6731, 2.8817, 2.8612, 2.2074, 1.7537, 2.4324], device='cuda:4'), covar=tensor([0.1329, 0.1117, 0.0642, 0.0874, 0.0701, 0.1312, 0.1494, 0.0781], device='cuda:4'), in_proj_covar=tensor([0.0215, 0.0215, 0.0192, 0.0188, 0.0184, 0.0204, 0.0181, 0.0199], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 12:56:18,492 INFO [finetune.py:976] (4/7) Epoch 2, batch 5200, loss[loss=0.2392, simple_loss=0.3078, pruned_loss=0.08531, over 4814.00 frames. ], tot_loss[loss=0.2512, simple_loss=0.3026, pruned_loss=0.09991, over 955040.16 frames. ], batch size: 38, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:56:31,986 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=10946.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 12:56:57,991 INFO [finetune.py:976] (4/7) Epoch 2, batch 5250, loss[loss=0.1989, simple_loss=0.255, pruned_loss=0.0714, over 4793.00 frames. ], tot_loss[loss=0.2502, simple_loss=0.3019, pruned_loss=0.09926, over 954724.16 frames. ], batch size: 23, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:57:08,231 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6293, 1.3678, 1.9676, 1.9357, 1.3701, 1.1123, 1.6844, 1.1209], device='cuda:4'), covar=tensor([0.1018, 0.1359, 0.0564, 0.1032, 0.1447, 0.1790, 0.0960, 0.1304], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0079, 0.0076, 0.0073, 0.0085, 0.0096, 0.0089, 0.0080], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-26 12:57:21,691 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3962, 1.2776, 1.5678, 1.6290, 1.6057, 1.3496, 1.4566, 1.4796], device='cuda:4'), covar=tensor([3.4497, 4.6337, 5.7868, 5.8178, 3.5002, 6.0187, 6.0739, 4.4958], device='cuda:4'), in_proj_covar=tensor([0.0449, 0.0507, 0.0601, 0.0599, 0.0482, 0.0524, 0.0534, 0.0546], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 12:57:27,044 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5974, 2.3081, 1.7685, 2.1421, 2.4649, 2.0245, 2.9938, 1.5430], device='cuda:4'), covar=tensor([0.3996, 0.2074, 0.4927, 0.3248, 0.1888, 0.2617, 0.2306, 0.4860], device='cuda:4'), in_proj_covar=tensor([0.0346, 0.0351, 0.0434, 0.0368, 0.0402, 0.0374, 0.0399, 0.0413], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 12:57:40,210 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.533e+02 2.106e+02 2.454e+02 2.928e+02 5.017e+02, threshold=4.909e+02, percent-clipped=1.0 2023-04-26 12:57:51,419 INFO [finetune.py:976] (4/7) Epoch 2, batch 5300, loss[loss=0.2707, simple_loss=0.3262, pruned_loss=0.1076, over 4820.00 frames. ], tot_loss[loss=0.2516, simple_loss=0.3033, pruned_loss=0.09995, over 953008.25 frames. ], batch size: 39, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:58:12,796 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-04-26 12:58:45,112 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=11065.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:58:57,625 INFO [finetune.py:976] (4/7) Epoch 2, batch 5350, loss[loss=0.2184, simple_loss=0.2722, pruned_loss=0.08228, over 4789.00 frames. ], tot_loss[loss=0.2518, simple_loss=0.3042, pruned_loss=0.09973, over 953282.75 frames. ], batch size: 51, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 12:59:29,510 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=11100.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 12:59:47,703 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=11110.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 12:59:50,021 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=11113.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:00:00,706 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.273e+02 1.923e+02 2.240e+02 2.858e+02 6.419e+02, threshold=4.480e+02, percent-clipped=4.0 2023-04-26 13:00:11,122 INFO [finetune.py:976] (4/7) Epoch 2, batch 5400, loss[loss=0.2409, simple_loss=0.289, pruned_loss=0.09644, over 4900.00 frames. ], tot_loss[loss=0.2498, simple_loss=0.3015, pruned_loss=0.09909, over 954352.99 frames. ], batch size: 35, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 13:00:12,410 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=11130.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:00:32,456 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=11158.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:00:42,745 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=11174.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:00:45,047 INFO [finetune.py:976] (4/7) Epoch 2, batch 5450, loss[loss=0.2646, simple_loss=0.3178, pruned_loss=0.1057, over 4940.00 frames. ], tot_loss[loss=0.2463, simple_loss=0.2976, pruned_loss=0.09747, over 955238.94 frames. ], batch size: 33, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 13:01:18,381 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.436e+02 1.901e+02 2.286e+02 2.895e+02 5.867e+02, threshold=4.573e+02, percent-clipped=4.0 2023-04-26 13:01:23,720 INFO [finetune.py:976] (4/7) Epoch 2, batch 5500, loss[loss=0.2172, simple_loss=0.2775, pruned_loss=0.07839, over 4768.00 frames. ], tot_loss[loss=0.2434, simple_loss=0.2946, pruned_loss=0.09615, over 955549.16 frames. ], batch size: 28, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 13:01:28,098 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=11235.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:01:31,696 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=11241.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 13:01:40,586 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7501, 2.3715, 1.8966, 2.0990, 1.6499, 1.8380, 2.0199, 1.4936], device='cuda:4'), covar=tensor([0.2529, 0.1516, 0.1234, 0.1482, 0.3168, 0.1839, 0.2199, 0.3115], device='cuda:4'), in_proj_covar=tensor([0.0315, 0.0335, 0.0245, 0.0308, 0.0314, 0.0287, 0.0276, 0.0300], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 13:01:54,038 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6902, 3.6589, 0.7983, 1.8911, 2.0669, 2.5105, 2.2235, 1.0534], device='cuda:4'), covar=tensor([0.1294, 0.0941, 0.2299, 0.1465, 0.1019, 0.1084, 0.1353, 0.1968], device='cuda:4'), in_proj_covar=tensor([0.0125, 0.0272, 0.0153, 0.0132, 0.0144, 0.0166, 0.0129, 0.0134], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 13:01:57,530 INFO [finetune.py:976] (4/7) Epoch 2, batch 5550, loss[loss=0.1944, simple_loss=0.2432, pruned_loss=0.07281, over 4735.00 frames. ], tot_loss[loss=0.2463, simple_loss=0.2973, pruned_loss=0.09764, over 954796.48 frames. ], batch size: 23, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 13:02:35,254 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.528e+02 2.026e+02 2.364e+02 2.674e+02 3.741e+02, threshold=4.728e+02, percent-clipped=0.0 2023-04-26 13:02:37,671 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0604, 1.2396, 1.3581, 1.4610, 1.4929, 1.1043, 0.8171, 1.3448], device='cuda:4'), covar=tensor([0.1403, 0.1600, 0.0973, 0.0862, 0.0856, 0.1317, 0.1521, 0.0823], device='cuda:4'), in_proj_covar=tensor([0.0212, 0.0212, 0.0191, 0.0186, 0.0182, 0.0201, 0.0178, 0.0197], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 13:02:39,953 INFO [finetune.py:976] (4/7) Epoch 2, batch 5600, loss[loss=0.2861, simple_loss=0.3336, pruned_loss=0.1193, over 4931.00 frames. ], tot_loss[loss=0.2509, simple_loss=0.3027, pruned_loss=0.09955, over 956131.17 frames. ], batch size: 42, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 13:02:46,672 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.51 vs. limit=5.0 2023-04-26 13:02:47,229 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6306, 2.2524, 1.5780, 1.3227, 1.2149, 1.2501, 1.5182, 1.2214], device='cuda:4'), covar=tensor([0.2560, 0.2112, 0.2627, 0.3291, 0.3928, 0.3033, 0.2142, 0.3221], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0224, 0.0192, 0.0215, 0.0231, 0.0195, 0.0188, 0.0207], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 13:02:57,757 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3332, 1.0908, 3.8185, 3.5319, 3.3656, 3.6557, 3.6210, 3.3511], device='cuda:4'), covar=tensor([0.6999, 0.6351, 0.1219, 0.1853, 0.1259, 0.2125, 0.1677, 0.1529], device='cuda:4'), in_proj_covar=tensor([0.0330, 0.0311, 0.0443, 0.0448, 0.0377, 0.0430, 0.0339, 0.0398], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0002], device='cuda:4') 2023-04-26 13:03:42,989 INFO [finetune.py:976] (4/7) Epoch 2, batch 5650, loss[loss=0.2491, simple_loss=0.3058, pruned_loss=0.09625, over 4381.00 frames. ], tot_loss[loss=0.2527, simple_loss=0.3053, pruned_loss=0.1001, over 955361.22 frames. ], batch size: 65, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 13:04:13,886 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=11400.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:04:31,189 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.082e+02 1.867e+02 2.284e+02 2.863e+02 5.268e+02, threshold=4.568e+02, percent-clipped=1.0 2023-04-26 13:04:31,682 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.59 vs. limit=2.0 2023-04-26 13:04:35,906 INFO [finetune.py:976] (4/7) Epoch 2, batch 5700, loss[loss=0.2399, simple_loss=0.2725, pruned_loss=0.1036, over 4212.00 frames. ], tot_loss[loss=0.2492, simple_loss=0.2999, pruned_loss=0.09924, over 936391.00 frames. ], batch size: 18, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 13:04:37,186 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=11430.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:04:46,233 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5413, 1.8164, 2.1661, 3.0603, 2.2149, 1.7867, 1.6737, 2.2921], device='cuda:4'), covar=tensor([0.5893, 0.7041, 0.3309, 0.5798, 0.6897, 0.5056, 0.8554, 0.6214], device='cuda:4'), in_proj_covar=tensor([0.0272, 0.0278, 0.0225, 0.0348, 0.0235, 0.0237, 0.0269, 0.0216], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 13:04:47,961 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=11448.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:04:50,138 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=5.01 vs. limit=5.0 2023-04-26 13:05:07,854 INFO [finetune.py:976] (4/7) Epoch 3, batch 0, loss[loss=0.35, simple_loss=0.3762, pruned_loss=0.1619, over 4750.00 frames. ], tot_loss[loss=0.35, simple_loss=0.3762, pruned_loss=0.1619, over 4750.00 frames. ], batch size: 54, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 13:05:07,855 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-26 13:05:24,858 INFO [finetune.py:1010] (4/7) Epoch 3, validation: loss=0.1779, simple_loss=0.251, pruned_loss=0.05243, over 2265189.00 frames. 2023-04-26 13:05:24,859 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6259MB 2023-04-26 13:05:42,225 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=11478.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:05:44,183 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=5.09 vs. limit=5.0 2023-04-26 13:06:01,722 INFO [finetune.py:976] (4/7) Epoch 3, batch 50, loss[loss=0.2799, simple_loss=0.3158, pruned_loss=0.122, over 4912.00 frames. ], tot_loss[loss=0.2524, simple_loss=0.305, pruned_loss=0.09993, over 215513.72 frames. ], batch size: 33, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 13:06:11,129 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.469e+02 1.955e+02 2.272e+02 2.719e+02 4.720e+02, threshold=4.545e+02, percent-clipped=1.0 2023-04-26 13:06:17,274 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=11530.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:06:18,526 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=11532.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:06:24,400 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=11541.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 13:06:28,879 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-04-26 13:06:34,900 INFO [finetune.py:976] (4/7) Epoch 3, batch 100, loss[loss=0.2519, simple_loss=0.303, pruned_loss=0.1004, over 4900.00 frames. ], tot_loss[loss=0.2482, simple_loss=0.2986, pruned_loss=0.09884, over 378130.61 frames. ], batch size: 36, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 13:06:55,969 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=11589.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 13:06:58,892 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=11593.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:07:08,543 INFO [finetune.py:976] (4/7) Epoch 3, batch 150, loss[loss=0.1969, simple_loss=0.2587, pruned_loss=0.06761, over 4871.00 frames. ], tot_loss[loss=0.2409, simple_loss=0.292, pruned_loss=0.09493, over 506572.44 frames. ], batch size: 32, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 13:07:18,013 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.434e+02 1.865e+02 2.222e+02 2.672e+02 4.139e+02, threshold=4.445e+02, percent-clipped=0.0 2023-04-26 13:07:26,715 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5830, 1.4687, 0.5469, 1.2521, 1.6863, 1.4439, 1.3577, 1.3964], device='cuda:4'), covar=tensor([0.0643, 0.0440, 0.0532, 0.0626, 0.0323, 0.0625, 0.0579, 0.0670], device='cuda:4'), in_proj_covar=tensor([0.0032, 0.0027, 0.0025, 0.0031, 0.0022, 0.0031, 0.0030, 0.0032], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0049, 0.0038, 0.0048, 0.0047, 0.0050], device='cuda:4') 2023-04-26 13:07:27,313 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=11635.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:07:29,847 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9071, 1.2299, 1.7113, 2.2983, 1.6067, 1.2698, 1.0675, 1.4288], device='cuda:4'), covar=tensor([0.5881, 0.6925, 0.3225, 0.5394, 0.6607, 0.5114, 0.9359, 0.6492], device='cuda:4'), in_proj_covar=tensor([0.0271, 0.0278, 0.0225, 0.0347, 0.0234, 0.0237, 0.0268, 0.0215], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 13:07:42,010 INFO [finetune.py:976] (4/7) Epoch 3, batch 200, loss[loss=0.2967, simple_loss=0.3448, pruned_loss=0.1243, over 4703.00 frames. ], tot_loss[loss=0.2395, simple_loss=0.2907, pruned_loss=0.09414, over 608797.96 frames. ], batch size: 59, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 13:08:29,895 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=11696.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:08:31,694 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=11699.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:08:42,575 INFO [finetune.py:976] (4/7) Epoch 3, batch 250, loss[loss=0.2626, simple_loss=0.3167, pruned_loss=0.1042, over 4840.00 frames. ], tot_loss[loss=0.2419, simple_loss=0.2936, pruned_loss=0.09509, over 687651.09 frames. ], batch size: 49, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 13:09:02,117 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.384e+02 2.012e+02 2.336e+02 2.910e+02 4.662e+02, threshold=4.672e+02, percent-clipped=2.0 2023-04-26 13:09:23,967 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0656, 2.5312, 1.0206, 1.3198, 2.0770, 1.2660, 3.5426, 1.7724], device='cuda:4'), covar=tensor([0.0726, 0.0759, 0.0946, 0.1431, 0.0591, 0.1140, 0.0234, 0.0706], device='cuda:4'), in_proj_covar=tensor([0.0056, 0.0073, 0.0054, 0.0051, 0.0056, 0.0057, 0.0086, 0.0054], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0008, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 13:09:28,231 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2953, 3.0374, 0.9962, 1.4177, 2.5053, 1.3376, 4.1934, 1.8910], device='cuda:4'), covar=tensor([0.0729, 0.0793, 0.1032, 0.1482, 0.0570, 0.1183, 0.0266, 0.0731], device='cuda:4'), in_proj_covar=tensor([0.0056, 0.0074, 0.0054, 0.0051, 0.0056, 0.0057, 0.0086, 0.0054], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0008, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 13:09:36,497 INFO [finetune.py:976] (4/7) Epoch 3, batch 300, loss[loss=0.2373, simple_loss=0.3029, pruned_loss=0.08581, over 4930.00 frames. ], tot_loss[loss=0.2485, simple_loss=0.3008, pruned_loss=0.09813, over 749353.06 frames. ], batch size: 42, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 13:09:38,485 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.71 vs. limit=5.0 2023-04-26 13:09:40,561 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=11760.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:09:52,918 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5404, 1.6197, 0.7456, 1.2497, 1.7382, 1.4156, 1.3326, 1.4774], device='cuda:4'), covar=tensor([0.0562, 0.0421, 0.0487, 0.0596, 0.0337, 0.0567, 0.0528, 0.0646], device='cuda:4'), in_proj_covar=tensor([0.0032, 0.0027, 0.0025, 0.0031, 0.0022, 0.0031, 0.0030, 0.0032], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0049, 0.0038, 0.0048, 0.0047, 0.0050], device='cuda:4') 2023-04-26 13:10:10,605 INFO [finetune.py:976] (4/7) Epoch 3, batch 350, loss[loss=0.3431, simple_loss=0.385, pruned_loss=0.1506, over 4808.00 frames. ], tot_loss[loss=0.2509, simple_loss=0.3033, pruned_loss=0.09921, over 795763.12 frames. ], batch size: 39, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 13:10:18,387 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=11816.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:10:21,167 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.314e+02 2.028e+02 2.307e+02 2.756e+02 7.160e+02, threshold=4.615e+02, percent-clipped=2.0 2023-04-26 13:10:28,338 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=11830.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:10:38,186 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=11846.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:10:50,159 INFO [finetune.py:976] (4/7) Epoch 3, batch 400, loss[loss=0.2776, simple_loss=0.3228, pruned_loss=0.1162, over 4889.00 frames. ], tot_loss[loss=0.2513, simple_loss=0.3038, pruned_loss=0.09939, over 831910.24 frames. ], batch size: 32, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 13:11:20,864 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=11877.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:11:21,407 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=11878.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:11:33,330 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=11888.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:11:44,734 INFO [finetune.py:976] (4/7) Epoch 3, batch 450, loss[loss=0.2066, simple_loss=0.2563, pruned_loss=0.07842, over 4800.00 frames. ], tot_loss[loss=0.2494, simple_loss=0.3022, pruned_loss=0.09826, over 860989.58 frames. ], batch size: 25, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 13:11:45,504 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=11907.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:11:54,737 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.367e+02 1.957e+02 2.306e+02 2.695e+02 4.680e+02, threshold=4.613e+02, percent-clipped=1.0 2023-04-26 13:11:54,878 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=11920.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:12:13,342 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=11948.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:12:18,148 INFO [finetune.py:976] (4/7) Epoch 3, batch 500, loss[loss=0.2354, simple_loss=0.2804, pruned_loss=0.09524, over 4933.00 frames. ], tot_loss[loss=0.2473, simple_loss=0.2993, pruned_loss=0.09766, over 882167.03 frames. ], batch size: 38, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 13:12:25,301 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2139, 1.3950, 1.4057, 1.5617, 1.4724, 1.6281, 1.4846, 1.5323], device='cuda:4'), covar=tensor([1.7894, 2.9305, 2.4475, 1.9969, 2.4185, 3.8210, 3.0814, 2.4813], device='cuda:4'), in_proj_covar=tensor([0.0295, 0.0395, 0.0312, 0.0318, 0.0346, 0.0392, 0.0379, 0.0342], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 13:12:36,373 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=11981.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:12:38,151 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=11984.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:12:42,344 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=11991.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:12:52,575 INFO [finetune.py:976] (4/7) Epoch 3, batch 550, loss[loss=0.2329, simple_loss=0.2892, pruned_loss=0.08829, over 4827.00 frames. ], tot_loss[loss=0.2435, simple_loss=0.295, pruned_loss=0.09605, over 898828.56 frames. ], batch size: 33, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 13:12:55,032 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=12009.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:13:02,614 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.347e+02 1.896e+02 2.113e+02 2.653e+02 4.840e+02, threshold=4.226e+02, percent-clipped=1.0 2023-04-26 13:13:19,329 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=12045.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:13:30,908 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=12055.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:13:31,437 INFO [finetune.py:976] (4/7) Epoch 3, batch 600, loss[loss=0.2274, simple_loss=0.2862, pruned_loss=0.08426, over 4799.00 frames. ], tot_loss[loss=0.2431, simple_loss=0.2946, pruned_loss=0.09584, over 912059.61 frames. ], batch size: 51, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 13:14:01,481 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=12076.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:14:24,997 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.43 vs. limit=2.0 2023-04-26 13:14:35,142 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6887, 1.4901, 2.0151, 1.9889, 1.4939, 1.2686, 1.6890, 1.0814], device='cuda:4'), covar=tensor([0.0950, 0.1271, 0.0586, 0.0999, 0.1478, 0.1677, 0.1045, 0.1287], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0079, 0.0075, 0.0071, 0.0085, 0.0096, 0.0088, 0.0079], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-26 13:14:37,620 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6239, 1.1589, 1.2268, 1.0764, 1.7849, 1.4678, 1.0744, 1.1834], device='cuda:4'), covar=tensor([0.1753, 0.1786, 0.2263, 0.1919, 0.1031, 0.1655, 0.2600, 0.2118], device='cuda:4'), in_proj_covar=tensor([0.0322, 0.0339, 0.0348, 0.0315, 0.0349, 0.0365, 0.0321, 0.0354], device='cuda:4'), out_proj_covar=tensor([7.0136e-05, 7.2849e-05, 7.5477e-05, 6.6034e-05, 7.4414e-05, 7.9720e-05, 6.9814e-05, 7.6582e-05], device='cuda:4') 2023-04-26 13:14:38,092 INFO [finetune.py:976] (4/7) Epoch 3, batch 650, loss[loss=0.2737, simple_loss=0.3236, pruned_loss=0.1119, over 4931.00 frames. ], tot_loss[loss=0.2461, simple_loss=0.2986, pruned_loss=0.09685, over 921571.78 frames. ], batch size: 33, lr: 3.99e-03, grad_scale: 32.0 2023-04-26 13:14:47,759 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8325, 1.2317, 1.4444, 1.3308, 1.9353, 1.7203, 1.3067, 1.3549], device='cuda:4'), covar=tensor([0.1751, 0.1830, 0.2131, 0.1763, 0.1003, 0.1668, 0.2381, 0.2136], device='cuda:4'), in_proj_covar=tensor([0.0322, 0.0339, 0.0348, 0.0315, 0.0349, 0.0365, 0.0320, 0.0354], device='cuda:4'), out_proj_covar=tensor([7.0107e-05, 7.2830e-05, 7.5509e-05, 6.6018e-05, 7.4423e-05, 7.9756e-05, 6.9803e-05, 7.6571e-05], device='cuda:4') 2023-04-26 13:14:55,294 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.228e+02 1.956e+02 2.261e+02 2.773e+02 5.813e+02, threshold=4.521e+02, percent-clipped=3.0 2023-04-26 13:15:18,810 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=12137.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:15:30,237 INFO [finetune.py:976] (4/7) Epoch 3, batch 700, loss[loss=0.2591, simple_loss=0.2988, pruned_loss=0.1097, over 4160.00 frames. ], tot_loss[loss=0.247, simple_loss=0.2994, pruned_loss=0.09732, over 926349.08 frames. ], batch size: 65, lr: 3.98e-03, grad_scale: 32.0 2023-04-26 13:15:40,478 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=12172.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:15:52,207 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=12188.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:16:00,819 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=12202.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:16:03,160 INFO [finetune.py:976] (4/7) Epoch 3, batch 750, loss[loss=0.2651, simple_loss=0.3269, pruned_loss=0.1017, over 4796.00 frames. ], tot_loss[loss=0.248, simple_loss=0.301, pruned_loss=0.09751, over 933822.31 frames. ], batch size: 45, lr: 3.98e-03, grad_scale: 32.0 2023-04-26 13:16:11,581 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 2.083e+02 2.562e+02 2.927e+02 7.910e+02, threshold=5.125e+02, percent-clipped=5.0 2023-04-26 13:16:22,132 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8723, 1.4729, 1.3034, 1.4276, 2.0234, 1.7591, 1.3969, 1.3036], device='cuda:4'), covar=tensor([0.1744, 0.2052, 0.2470, 0.1628, 0.0934, 0.1706, 0.2555, 0.2357], device='cuda:4'), in_proj_covar=tensor([0.0321, 0.0339, 0.0350, 0.0314, 0.0350, 0.0365, 0.0320, 0.0354], device='cuda:4'), out_proj_covar=tensor([6.9941e-05, 7.2789e-05, 7.5786e-05, 6.5895e-05, 7.4507e-05, 7.9800e-05, 6.9726e-05, 7.6432e-05], device='cuda:4') 2023-04-26 13:16:23,275 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=12236.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:16:42,546 INFO [finetune.py:976] (4/7) Epoch 3, batch 800, loss[loss=0.2018, simple_loss=0.2586, pruned_loss=0.07247, over 4749.00 frames. ], tot_loss[loss=0.2467, simple_loss=0.3, pruned_loss=0.0967, over 936496.74 frames. ], batch size: 59, lr: 3.98e-03, grad_scale: 32.0 2023-04-26 13:17:00,569 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=12276.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:17:11,621 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=12291.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:17:19,898 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=12304.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:17:21,019 INFO [finetune.py:976] (4/7) Epoch 3, batch 850, loss[loss=0.2674, simple_loss=0.3119, pruned_loss=0.1114, over 4752.00 frames. ], tot_loss[loss=0.2447, simple_loss=0.2977, pruned_loss=0.09589, over 940468.51 frames. ], batch size: 59, lr: 3.98e-03, grad_scale: 32.0 2023-04-26 13:17:29,559 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.293e+02 1.971e+02 2.371e+02 2.612e+02 4.896e+02, threshold=4.741e+02, percent-clipped=0.0 2023-04-26 13:17:43,091 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=12339.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:17:43,713 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=12340.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:17:54,286 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=12355.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:17:54,808 INFO [finetune.py:976] (4/7) Epoch 3, batch 900, loss[loss=0.2315, simple_loss=0.286, pruned_loss=0.0885, over 4663.00 frames. ], tot_loss[loss=0.2418, simple_loss=0.2946, pruned_loss=0.09445, over 944656.70 frames. ], batch size: 23, lr: 3.98e-03, grad_scale: 32.0 2023-04-26 13:18:26,968 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=12403.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:18:28,775 INFO [finetune.py:976] (4/7) Epoch 3, batch 950, loss[loss=0.2196, simple_loss=0.281, pruned_loss=0.07909, over 4779.00 frames. ], tot_loss[loss=0.241, simple_loss=0.2933, pruned_loss=0.09436, over 947056.84 frames. ], batch size: 29, lr: 3.98e-03, grad_scale: 64.0 2023-04-26 13:18:37,307 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.210e+02 1.930e+02 2.195e+02 2.798e+02 5.251e+02, threshold=4.389e+02, percent-clipped=2.0 2023-04-26 13:18:50,722 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=12432.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:19:13,292 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=12446.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:19:20,768 INFO [finetune.py:976] (4/7) Epoch 3, batch 1000, loss[loss=0.3008, simple_loss=0.3511, pruned_loss=0.1253, over 4818.00 frames. ], tot_loss[loss=0.2439, simple_loss=0.2965, pruned_loss=0.09563, over 950592.63 frames. ], batch size: 40, lr: 3.98e-03, grad_scale: 64.0 2023-04-26 13:19:30,628 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=12472.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:19:52,935 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-26 13:20:03,632 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=12502.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:20:06,939 INFO [finetune.py:976] (4/7) Epoch 3, batch 1050, loss[loss=0.2715, simple_loss=0.3267, pruned_loss=0.1081, over 4917.00 frames. ], tot_loss[loss=0.2463, simple_loss=0.2997, pruned_loss=0.09646, over 951010.71 frames. ], batch size: 38, lr: 3.98e-03, grad_scale: 64.0 2023-04-26 13:20:07,659 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=12507.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:20:25,976 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.380e+02 2.009e+02 2.317e+02 2.704e+02 5.500e+02, threshold=4.634e+02, percent-clipped=2.0 2023-04-26 13:20:26,053 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=12520.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:20:39,077 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-26 13:21:02,677 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=12550.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:21:12,471 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-04-26 13:21:12,629 INFO [finetune.py:976] (4/7) Epoch 3, batch 1100, loss[loss=0.2543, simple_loss=0.3108, pruned_loss=0.09894, over 4898.00 frames. ], tot_loss[loss=0.2467, simple_loss=0.3003, pruned_loss=0.09657, over 951817.63 frames. ], batch size: 37, lr: 3.98e-03, grad_scale: 64.0 2023-04-26 13:21:15,545 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5490, 1.3625, 4.4585, 4.1091, 3.8992, 4.2069, 4.1734, 3.9380], device='cuda:4'), covar=tensor([0.6833, 0.5766, 0.0856, 0.1687, 0.1150, 0.1139, 0.1065, 0.1385], device='cuda:4'), in_proj_covar=tensor([0.0326, 0.0310, 0.0438, 0.0444, 0.0375, 0.0425, 0.0335, 0.0394], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0002], device='cuda:4') 2023-04-26 13:21:18,368 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.51 vs. limit=2.0 2023-04-26 13:21:24,130 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-04-26 13:21:25,862 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=12576.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:21:35,469 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.93 vs. limit=2.0 2023-04-26 13:21:44,436 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=12604.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:21:46,054 INFO [finetune.py:976] (4/7) Epoch 3, batch 1150, loss[loss=0.2639, simple_loss=0.316, pruned_loss=0.1059, over 4844.00 frames. ], tot_loss[loss=0.246, simple_loss=0.3002, pruned_loss=0.09592, over 953558.52 frames. ], batch size: 44, lr: 3.98e-03, grad_scale: 64.0 2023-04-26 13:21:54,484 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6784, 1.2094, 1.4847, 1.4152, 1.3386, 1.1262, 0.6315, 1.0818], device='cuda:4'), covar=tensor([0.5714, 0.6570, 0.3091, 0.5057, 0.5779, 0.4714, 0.8108, 0.5294], device='cuda:4'), in_proj_covar=tensor([0.0270, 0.0274, 0.0223, 0.0343, 0.0232, 0.0235, 0.0264, 0.0211], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 13:21:55,573 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.367e+02 1.845e+02 2.160e+02 2.613e+02 5.486e+02, threshold=4.320e+02, percent-clipped=1.0 2023-04-26 13:21:58,101 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=12624.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:22:19,253 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=12640.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:22:29,672 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-04-26 13:22:34,212 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=12652.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:22:36,635 INFO [finetune.py:976] (4/7) Epoch 3, batch 1200, loss[loss=0.2356, simple_loss=0.2902, pruned_loss=0.09044, over 4716.00 frames. ], tot_loss[loss=0.244, simple_loss=0.2986, pruned_loss=0.09475, over 954976.94 frames. ], batch size: 23, lr: 3.98e-03, grad_scale: 64.0 2023-04-26 13:22:38,483 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5895, 2.4351, 2.9419, 3.0327, 2.6733, 2.3249, 1.8433, 2.5841], device='cuda:4'), covar=tensor([0.1159, 0.1103, 0.0548, 0.0795, 0.0818, 0.1224, 0.1280, 0.0683], device='cuda:4'), in_proj_covar=tensor([0.0211, 0.0211, 0.0190, 0.0184, 0.0184, 0.0201, 0.0177, 0.0197], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 13:22:47,278 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4332, 1.6595, 1.2023, 0.8319, 1.1581, 1.1307, 1.1926, 1.0664], device='cuda:4'), covar=tensor([0.2143, 0.1663, 0.2289, 0.2652, 0.3342, 0.2486, 0.1752, 0.2617], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0223, 0.0190, 0.0215, 0.0229, 0.0194, 0.0186, 0.0205], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 13:22:57,555 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=12688.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:23:09,102 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.72 vs. limit=5.0 2023-04-26 13:23:09,931 INFO [finetune.py:976] (4/7) Epoch 3, batch 1250, loss[loss=0.234, simple_loss=0.2881, pruned_loss=0.09, over 4929.00 frames. ], tot_loss[loss=0.242, simple_loss=0.2959, pruned_loss=0.0941, over 954725.25 frames. ], batch size: 38, lr: 3.98e-03, grad_scale: 64.0 2023-04-26 13:23:19,407 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.393e+02 1.929e+02 2.277e+02 2.741e+02 5.638e+02, threshold=4.554e+02, percent-clipped=3.0 2023-04-26 13:23:24,999 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1094, 0.7811, 0.8724, 0.6767, 1.2745, 1.0301, 0.8692, 0.9862], device='cuda:4'), covar=tensor([0.1597, 0.1564, 0.2034, 0.1603, 0.0970, 0.1475, 0.1881, 0.2023], device='cuda:4'), in_proj_covar=tensor([0.0319, 0.0334, 0.0345, 0.0310, 0.0345, 0.0359, 0.0317, 0.0350], device='cuda:4'), out_proj_covar=tensor([6.9436e-05, 7.1711e-05, 7.4640e-05, 6.4968e-05, 7.3481e-05, 7.8575e-05, 6.9036e-05, 7.5566e-05], device='cuda:4') 2023-04-26 13:23:26,755 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=12732.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:23:32,263 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4323, 1.0566, 1.1798, 0.9823, 1.6357, 1.3061, 1.0111, 1.1495], device='cuda:4'), covar=tensor([0.1638, 0.1470, 0.2011, 0.1527, 0.0799, 0.1507, 0.2072, 0.1927], device='cuda:4'), in_proj_covar=tensor([0.0319, 0.0334, 0.0345, 0.0310, 0.0346, 0.0360, 0.0318, 0.0350], device='cuda:4'), out_proj_covar=tensor([6.9525e-05, 7.1745e-05, 7.4670e-05, 6.5033e-05, 7.3559e-05, 7.8651e-05, 6.9151e-05, 7.5671e-05], device='cuda:4') 2023-04-26 13:23:39,950 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.9150, 3.6966, 2.8097, 4.4502, 3.9015, 3.8617, 1.6862, 3.8026], device='cuda:4'), covar=tensor([0.1699, 0.1153, 0.3400, 0.1261, 0.3240, 0.1828, 0.5621, 0.2170], device='cuda:4'), in_proj_covar=tensor([0.0253, 0.0227, 0.0268, 0.0319, 0.0314, 0.0264, 0.0278, 0.0280], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 13:23:42,759 INFO [finetune.py:976] (4/7) Epoch 3, batch 1300, loss[loss=0.22, simple_loss=0.2744, pruned_loss=0.08283, over 4926.00 frames. ], tot_loss[loss=0.239, simple_loss=0.2925, pruned_loss=0.09271, over 956662.53 frames. ], batch size: 37, lr: 3.98e-03, grad_scale: 64.0 2023-04-26 13:23:47,594 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=12763.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:24:09,820 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=12780.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:24:35,998 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=12802.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:24:39,384 INFO [finetune.py:976] (4/7) Epoch 3, batch 1350, loss[loss=0.2227, simple_loss=0.2862, pruned_loss=0.07958, over 4795.00 frames. ], tot_loss[loss=0.2394, simple_loss=0.2928, pruned_loss=0.09298, over 958804.99 frames. ], batch size: 29, lr: 3.98e-03, grad_scale: 64.0 2023-04-26 13:24:47,681 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-26 13:24:48,912 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.232e+02 1.880e+02 2.141e+02 2.633e+02 4.297e+02, threshold=4.283e+02, percent-clipped=1.0 2023-04-26 13:24:51,954 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=12824.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 13:24:54,939 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.5526, 4.3258, 3.0844, 5.0981, 4.4096, 4.4322, 1.9634, 4.3058], device='cuda:4'), covar=tensor([0.1345, 0.0881, 0.3161, 0.0755, 0.3037, 0.1296, 0.5273, 0.2020], device='cuda:4'), in_proj_covar=tensor([0.0253, 0.0227, 0.0269, 0.0320, 0.0314, 0.0264, 0.0279, 0.0281], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 13:25:12,232 INFO [finetune.py:976] (4/7) Epoch 3, batch 1400, loss[loss=0.2756, simple_loss=0.3231, pruned_loss=0.114, over 4829.00 frames. ], tot_loss[loss=0.2416, simple_loss=0.2959, pruned_loss=0.0936, over 958420.67 frames. ], batch size: 33, lr: 3.98e-03, grad_scale: 64.0 2023-04-26 13:25:27,653 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8009, 2.2810, 1.8691, 2.1220, 1.6012, 1.7340, 1.8808, 1.5133], device='cuda:4'), covar=tensor([0.2313, 0.1492, 0.1088, 0.1497, 0.3197, 0.1594, 0.1991, 0.2956], device='cuda:4'), in_proj_covar=tensor([0.0315, 0.0337, 0.0247, 0.0309, 0.0315, 0.0288, 0.0277, 0.0300], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 13:25:56,156 INFO [finetune.py:976] (4/7) Epoch 3, batch 1450, loss[loss=0.2264, simple_loss=0.2858, pruned_loss=0.08347, over 4751.00 frames. ], tot_loss[loss=0.243, simple_loss=0.298, pruned_loss=0.09401, over 957708.46 frames. ], batch size: 26, lr: 3.98e-03, grad_scale: 64.0 2023-04-26 13:26:17,005 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.305e+02 1.925e+02 2.434e+02 2.951e+02 7.906e+02, threshold=4.868e+02, percent-clipped=4.0 2023-04-26 13:27:02,917 INFO [finetune.py:976] (4/7) Epoch 3, batch 1500, loss[loss=0.2106, simple_loss=0.2793, pruned_loss=0.07097, over 4922.00 frames. ], tot_loss[loss=0.2441, simple_loss=0.2989, pruned_loss=0.09459, over 957855.65 frames. ], batch size: 42, lr: 3.98e-03, grad_scale: 64.0 2023-04-26 13:27:23,378 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.74 vs. limit=2.0 2023-04-26 13:27:41,571 INFO [finetune.py:976] (4/7) Epoch 3, batch 1550, loss[loss=0.2202, simple_loss=0.2767, pruned_loss=0.08187, over 4755.00 frames. ], tot_loss[loss=0.2425, simple_loss=0.2978, pruned_loss=0.09358, over 958873.34 frames. ], batch size: 28, lr: 3.98e-03, grad_scale: 64.0 2023-04-26 13:27:49,878 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2570, 3.8735, 0.6662, 1.9785, 1.9258, 2.3222, 2.2394, 1.0971], device='cuda:4'), covar=tensor([0.1863, 0.1652, 0.2716, 0.1800, 0.1296, 0.1620, 0.1526, 0.2442], device='cuda:4'), in_proj_covar=tensor([0.0125, 0.0270, 0.0152, 0.0132, 0.0143, 0.0166, 0.0129, 0.0134], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 13:27:53,402 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-26 13:28:02,730 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.270e+02 2.021e+02 2.296e+02 2.726e+02 4.661e+02, threshold=4.592e+02, percent-clipped=0.0 2023-04-26 13:28:47,302 INFO [finetune.py:976] (4/7) Epoch 3, batch 1600, loss[loss=0.2281, simple_loss=0.2844, pruned_loss=0.08594, over 4757.00 frames. ], tot_loss[loss=0.2394, simple_loss=0.2945, pruned_loss=0.09212, over 956356.19 frames. ], batch size: 54, lr: 3.98e-03, grad_scale: 64.0 2023-04-26 13:29:22,134 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-26 13:29:23,821 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=13102.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:29:26,216 INFO [finetune.py:976] (4/7) Epoch 3, batch 1650, loss[loss=0.182, simple_loss=0.2511, pruned_loss=0.05642, over 4802.00 frames. ], tot_loss[loss=0.2371, simple_loss=0.2912, pruned_loss=0.09143, over 956272.90 frames. ], batch size: 51, lr: 3.98e-03, grad_scale: 64.0 2023-04-26 13:29:31,052 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8041, 2.4598, 1.9992, 2.2489, 1.6670, 1.9356, 2.0552, 1.6608], device='cuda:4'), covar=tensor([0.2374, 0.1289, 0.0961, 0.1581, 0.3105, 0.1356, 0.1989, 0.2850], device='cuda:4'), in_proj_covar=tensor([0.0320, 0.0341, 0.0249, 0.0312, 0.0319, 0.0292, 0.0281, 0.0304], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 13:29:34,728 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=13119.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 13:29:35,233 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.220e+02 1.899e+02 2.180e+02 2.609e+02 5.027e+02, threshold=4.359e+02, percent-clipped=1.0 2023-04-26 13:29:44,656 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1426, 2.9727, 0.8741, 1.5571, 1.6628, 2.0631, 1.8601, 0.9642], device='cuda:4'), covar=tensor([0.1627, 0.1241, 0.2109, 0.1486, 0.1220, 0.1183, 0.1546, 0.1960], device='cuda:4'), in_proj_covar=tensor([0.0125, 0.0271, 0.0152, 0.0132, 0.0143, 0.0166, 0.0129, 0.0134], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 13:29:53,534 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1596, 1.3015, 5.2357, 4.6180, 4.6065, 4.8472, 4.4641, 4.4181], device='cuda:4'), covar=tensor([0.8478, 0.8436, 0.1110, 0.2565, 0.1809, 0.3132, 0.2452, 0.2510], device='cuda:4'), in_proj_covar=tensor([0.0329, 0.0313, 0.0441, 0.0446, 0.0376, 0.0429, 0.0339, 0.0395], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0002], device='cuda:4') 2023-04-26 13:29:55,914 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=13150.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:29:55,965 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4996, 1.2666, 0.7007, 1.2173, 1.4932, 1.3964, 1.3108, 1.3486], device='cuda:4'), covar=tensor([0.0605, 0.0475, 0.0483, 0.0632, 0.0342, 0.0592, 0.0618, 0.0677], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0027, 0.0025, 0.0031, 0.0022, 0.0031, 0.0030, 0.0032], device='cuda:4'), out_proj_covar=tensor([0.0048, 0.0044, 0.0038, 0.0049, 0.0037, 0.0048, 0.0047, 0.0050], device='cuda:4') 2023-04-26 13:29:59,486 INFO [finetune.py:976] (4/7) Epoch 3, batch 1700, loss[loss=0.2466, simple_loss=0.3006, pruned_loss=0.09623, over 4914.00 frames. ], tot_loss[loss=0.2358, simple_loss=0.2896, pruned_loss=0.09103, over 957126.79 frames. ], batch size: 36, lr: 3.98e-03, grad_scale: 64.0 2023-04-26 13:30:00,206 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6008, 2.0823, 1.7561, 1.9388, 1.5044, 1.6432, 1.6900, 1.3948], device='cuda:4'), covar=tensor([0.2198, 0.1312, 0.0981, 0.1236, 0.3092, 0.1277, 0.1773, 0.2468], device='cuda:4'), in_proj_covar=tensor([0.0320, 0.0341, 0.0249, 0.0312, 0.0319, 0.0292, 0.0281, 0.0304], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 13:30:33,506 INFO [finetune.py:976] (4/7) Epoch 3, batch 1750, loss[loss=0.2869, simple_loss=0.3544, pruned_loss=0.1097, over 4852.00 frames. ], tot_loss[loss=0.2389, simple_loss=0.2927, pruned_loss=0.0926, over 956987.31 frames. ], batch size: 49, lr: 3.98e-03, grad_scale: 64.0 2023-04-26 13:30:43,127 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.228e+02 1.834e+02 2.276e+02 2.863e+02 4.766e+02, threshold=4.553e+02, percent-clipped=3.0 2023-04-26 13:31:07,405 INFO [finetune.py:976] (4/7) Epoch 3, batch 1800, loss[loss=0.2568, simple_loss=0.3107, pruned_loss=0.1014, over 4925.00 frames. ], tot_loss[loss=0.2419, simple_loss=0.2966, pruned_loss=0.09358, over 957321.23 frames. ], batch size: 33, lr: 3.98e-03, grad_scale: 32.0 2023-04-26 13:31:20,366 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.95 vs. limit=2.0 2023-04-26 13:31:47,120 INFO [finetune.py:976] (4/7) Epoch 3, batch 1850, loss[loss=0.3099, simple_loss=0.3442, pruned_loss=0.1377, over 4867.00 frames. ], tot_loss[loss=0.2429, simple_loss=0.2975, pruned_loss=0.09411, over 958087.42 frames. ], batch size: 34, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:31:56,813 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.349e+02 1.904e+02 2.289e+02 2.801e+02 9.204e+02, threshold=4.579e+02, percent-clipped=2.0 2023-04-26 13:32:30,328 INFO [finetune.py:976] (4/7) Epoch 3, batch 1900, loss[loss=0.2518, simple_loss=0.3144, pruned_loss=0.0946, over 4776.00 frames. ], tot_loss[loss=0.2433, simple_loss=0.2984, pruned_loss=0.09417, over 957932.60 frames. ], batch size: 28, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:32:38,716 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6906, 2.3520, 1.6715, 1.6047, 1.2415, 1.3642, 1.7559, 1.2740], device='cuda:4'), covar=tensor([0.2034, 0.1881, 0.2230, 0.2583, 0.3215, 0.2325, 0.1689, 0.2685], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0225, 0.0191, 0.0216, 0.0229, 0.0195, 0.0186, 0.0206], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 13:32:56,187 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.40 vs. limit=5.0 2023-04-26 13:33:07,829 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.57 vs. limit=2.0 2023-04-26 13:33:26,629 INFO [finetune.py:976] (4/7) Epoch 3, batch 1950, loss[loss=0.2222, simple_loss=0.2673, pruned_loss=0.08857, over 4900.00 frames. ], tot_loss[loss=0.2401, simple_loss=0.2953, pruned_loss=0.09241, over 956011.52 frames. ], batch size: 32, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:33:39,969 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=13419.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:33:41,756 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.072e+02 1.766e+02 2.138e+02 2.602e+02 4.545e+02, threshold=4.277e+02, percent-clipped=0.0 2023-04-26 13:34:17,446 INFO [finetune.py:976] (4/7) Epoch 3, batch 2000, loss[loss=0.1785, simple_loss=0.2473, pruned_loss=0.05485, over 4846.00 frames. ], tot_loss[loss=0.2389, simple_loss=0.2935, pruned_loss=0.09219, over 954663.45 frames. ], batch size: 47, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:34:24,602 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=13467.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:34:25,250 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1787, 1.5609, 1.3442, 1.8408, 1.6534, 2.3275, 1.4029, 3.6346], device='cuda:4'), covar=tensor([0.0740, 0.0781, 0.0858, 0.1235, 0.0675, 0.0490, 0.0749, 0.0165], device='cuda:4'), in_proj_covar=tensor([0.0040, 0.0041, 0.0042, 0.0046, 0.0042, 0.0041, 0.0041, 0.0065], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 13:34:30,356 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-04-26 13:34:51,090 INFO [finetune.py:976] (4/7) Epoch 3, batch 2050, loss[loss=0.2276, simple_loss=0.2689, pruned_loss=0.09317, over 4739.00 frames. ], tot_loss[loss=0.2366, simple_loss=0.2906, pruned_loss=0.09134, over 954551.43 frames. ], batch size: 54, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:34:51,856 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5594, 1.8141, 1.6089, 1.7350, 1.6752, 1.8704, 1.7466, 1.6839], device='cuda:4'), covar=tensor([1.2623, 2.4208, 1.9815, 1.6105, 1.8310, 2.9692, 2.4590, 2.2232], device='cuda:4'), in_proj_covar=tensor([0.0297, 0.0397, 0.0313, 0.0318, 0.0346, 0.0398, 0.0381, 0.0341], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 13:35:01,245 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.294e+02 1.882e+02 2.225e+02 2.584e+02 5.744e+02, threshold=4.451e+02, percent-clipped=3.0 2023-04-26 13:35:01,950 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8642, 1.5857, 2.0447, 2.0928, 1.6548, 1.3543, 1.8318, 1.2982], device='cuda:4'), covar=tensor([0.0873, 0.1450, 0.0800, 0.1145, 0.1203, 0.1766, 0.1107, 0.1396], device='cuda:4'), in_proj_covar=tensor([0.0071, 0.0080, 0.0077, 0.0072, 0.0085, 0.0097, 0.0088, 0.0081], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-26 13:35:24,228 INFO [finetune.py:976] (4/7) Epoch 3, batch 2100, loss[loss=0.2816, simple_loss=0.3309, pruned_loss=0.1161, over 4802.00 frames. ], tot_loss[loss=0.2366, simple_loss=0.2905, pruned_loss=0.09139, over 954875.62 frames. ], batch size: 51, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:35:33,925 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1990, 2.8977, 0.9830, 1.6277, 1.6445, 2.1368, 1.8260, 1.0348], device='cuda:4'), covar=tensor([0.1440, 0.1031, 0.1840, 0.1356, 0.1152, 0.1036, 0.1496, 0.1997], device='cuda:4'), in_proj_covar=tensor([0.0125, 0.0271, 0.0152, 0.0132, 0.0143, 0.0166, 0.0129, 0.0133], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 13:35:57,671 INFO [finetune.py:976] (4/7) Epoch 3, batch 2150, loss[loss=0.3234, simple_loss=0.3541, pruned_loss=0.1464, over 4748.00 frames. ], tot_loss[loss=0.2398, simple_loss=0.2942, pruned_loss=0.0927, over 954847.22 frames. ], batch size: 54, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:35:59,571 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7368, 2.2168, 1.8493, 2.0572, 1.5609, 1.7993, 1.7700, 1.5663], device='cuda:4'), covar=tensor([0.2152, 0.1320, 0.0992, 0.1464, 0.3378, 0.1340, 0.2183, 0.2812], device='cuda:4'), in_proj_covar=tensor([0.0320, 0.0342, 0.0250, 0.0313, 0.0321, 0.0294, 0.0282, 0.0305], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 13:36:07,905 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.340e+02 1.924e+02 2.294e+02 2.856e+02 4.572e+02, threshold=4.587e+02, percent-clipped=1.0 2023-04-26 13:36:30,856 INFO [finetune.py:976] (4/7) Epoch 3, batch 2200, loss[loss=0.2708, simple_loss=0.3282, pruned_loss=0.1066, over 4821.00 frames. ], tot_loss[loss=0.2425, simple_loss=0.2971, pruned_loss=0.09393, over 954260.71 frames. ], batch size: 38, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:37:08,211 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-26 13:37:21,694 INFO [finetune.py:976] (4/7) Epoch 3, batch 2250, loss[loss=0.2, simple_loss=0.2681, pruned_loss=0.06601, over 4870.00 frames. ], tot_loss[loss=0.2433, simple_loss=0.2984, pruned_loss=0.09413, over 956091.77 frames. ], batch size: 31, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:37:29,515 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9473, 1.4462, 1.4865, 1.5247, 2.1126, 1.8275, 1.4053, 1.3852], device='cuda:4'), covar=tensor([0.1997, 0.1910, 0.2270, 0.1863, 0.1019, 0.1625, 0.2965, 0.2464], device='cuda:4'), in_proj_covar=tensor([0.0318, 0.0337, 0.0347, 0.0311, 0.0348, 0.0358, 0.0319, 0.0350], device='cuda:4'), out_proj_covar=tensor([6.9286e-05, 7.2281e-05, 7.5044e-05, 6.5076e-05, 7.4020e-05, 7.8232e-05, 6.9437e-05, 7.5542e-05], device='cuda:4') 2023-04-26 13:37:31,818 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.241e+02 1.916e+02 2.386e+02 2.867e+02 7.645e+02, threshold=4.772e+02, percent-clipped=4.0 2023-04-26 13:37:43,748 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=13737.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:38:02,149 INFO [finetune.py:976] (4/7) Epoch 3, batch 2300, loss[loss=0.262, simple_loss=0.309, pruned_loss=0.1075, over 4775.00 frames. ], tot_loss[loss=0.2411, simple_loss=0.297, pruned_loss=0.09257, over 955084.59 frames. ], batch size: 51, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:38:57,222 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6517, 1.2599, 1.4937, 1.8871, 1.7954, 1.5148, 1.5539, 1.5233], device='cuda:4'), covar=tensor([2.1514, 2.8446, 3.4654, 3.6589, 2.3438, 3.2466, 3.4323, 2.5987], device='cuda:4'), in_proj_covar=tensor([0.0451, 0.0503, 0.0595, 0.0605, 0.0482, 0.0518, 0.0529, 0.0541], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 13:38:57,787 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=13798.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 13:39:09,160 INFO [finetune.py:976] (4/7) Epoch 3, batch 2350, loss[loss=0.2482, simple_loss=0.2978, pruned_loss=0.09926, over 4868.00 frames. ], tot_loss[loss=0.2396, simple_loss=0.2943, pruned_loss=0.09242, over 955485.65 frames. ], batch size: 31, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:39:26,429 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1706, 2.5275, 1.0994, 1.2949, 1.9473, 1.1870, 3.1656, 1.6452], device='cuda:4'), covar=tensor([0.0662, 0.0712, 0.0854, 0.1210, 0.0497, 0.1013, 0.0226, 0.0635], device='cuda:4'), in_proj_covar=tensor([0.0056, 0.0073, 0.0054, 0.0050, 0.0055, 0.0056, 0.0085, 0.0054], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0008, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 13:39:28,286 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1412, 2.1743, 1.7863, 1.8287, 2.3347, 1.8967, 2.7810, 1.5696], device='cuda:4'), covar=tensor([0.4756, 0.1857, 0.5357, 0.3809, 0.1992, 0.2628, 0.1488, 0.4932], device='cuda:4'), in_proj_covar=tensor([0.0355, 0.0357, 0.0438, 0.0375, 0.0406, 0.0383, 0.0402, 0.0420], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 13:39:28,306 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0135, 1.3710, 1.9172, 2.3282, 1.7544, 1.3525, 1.1380, 1.6598], device='cuda:4'), covar=tensor([0.5165, 0.6061, 0.2658, 0.5001, 0.5882, 0.4762, 0.8156, 0.5420], device='cuda:4'), in_proj_covar=tensor([0.0270, 0.0273, 0.0224, 0.0343, 0.0232, 0.0235, 0.0263, 0.0209], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 13:39:37,980 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.317e+02 2.025e+02 2.480e+02 2.975e+02 5.089e+02, threshold=4.960e+02, percent-clipped=2.0 2023-04-26 13:40:20,272 INFO [finetune.py:976] (4/7) Epoch 3, batch 2400, loss[loss=0.2418, simple_loss=0.2913, pruned_loss=0.09612, over 4844.00 frames. ], tot_loss[loss=0.2358, simple_loss=0.29, pruned_loss=0.0908, over 954240.63 frames. ], batch size: 47, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:40:57,680 INFO [finetune.py:976] (4/7) Epoch 3, batch 2450, loss[loss=0.2165, simple_loss=0.2684, pruned_loss=0.0823, over 4708.00 frames. ], tot_loss[loss=0.2344, simple_loss=0.2881, pruned_loss=0.0904, over 956518.22 frames. ], batch size: 23, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:41:09,357 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.373e+02 2.035e+02 2.305e+02 2.759e+02 5.668e+02, threshold=4.609e+02, percent-clipped=2.0 2023-04-26 13:41:31,084 INFO [finetune.py:976] (4/7) Epoch 3, batch 2500, loss[loss=0.2087, simple_loss=0.2651, pruned_loss=0.07615, over 4757.00 frames. ], tot_loss[loss=0.2353, simple_loss=0.2892, pruned_loss=0.09071, over 956453.50 frames. ], batch size: 26, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:42:05,884 INFO [finetune.py:976] (4/7) Epoch 3, batch 2550, loss[loss=0.2035, simple_loss=0.2547, pruned_loss=0.07614, over 4717.00 frames. ], tot_loss[loss=0.2379, simple_loss=0.2928, pruned_loss=0.09153, over 955130.87 frames. ], batch size: 23, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:42:26,619 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.53 vs. limit=5.0 2023-04-26 13:42:28,109 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.240e+02 1.992e+02 2.434e+02 2.841e+02 4.921e+02, threshold=4.868e+02, percent-clipped=1.0 2023-04-26 13:43:13,248 INFO [finetune.py:976] (4/7) Epoch 3, batch 2600, loss[loss=0.1856, simple_loss=0.251, pruned_loss=0.06006, over 4103.00 frames. ], tot_loss[loss=0.2368, simple_loss=0.2925, pruned_loss=0.09059, over 956911.11 frames. ], batch size: 65, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:43:37,584 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=14072.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:43:55,881 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5731, 2.5139, 2.8538, 3.0875, 2.8312, 2.3073, 2.0559, 2.5049], device='cuda:4'), covar=tensor([0.1195, 0.0994, 0.0587, 0.0725, 0.0745, 0.1301, 0.1174, 0.0798], device='cuda:4'), in_proj_covar=tensor([0.0210, 0.0211, 0.0190, 0.0183, 0.0183, 0.0201, 0.0175, 0.0196], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 13:44:06,898 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=14093.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 13:44:18,988 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=14103.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:44:25,130 INFO [finetune.py:976] (4/7) Epoch 3, batch 2650, loss[loss=0.2664, simple_loss=0.3123, pruned_loss=0.1103, over 4858.00 frames. ], tot_loss[loss=0.2383, simple_loss=0.2939, pruned_loss=0.09137, over 957048.79 frames. ], batch size: 31, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:44:40,396 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=14120.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 13:44:47,129 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.256e+02 1.867e+02 2.323e+02 2.768e+02 1.192e+03, threshold=4.647e+02, percent-clipped=2.0 2023-04-26 13:45:00,931 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=14133.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:45:30,810 INFO [finetune.py:976] (4/7) Epoch 3, batch 2700, loss[loss=0.2708, simple_loss=0.2981, pruned_loss=0.1218, over 4821.00 frames. ], tot_loss[loss=0.2374, simple_loss=0.2933, pruned_loss=0.09077, over 958699.37 frames. ], batch size: 25, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:45:41,241 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=14164.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:46:04,148 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=14181.0, num_to_drop=1, layers_to_drop={3} 2023-04-26 13:46:30,713 INFO [finetune.py:976] (4/7) Epoch 3, batch 2750, loss[loss=0.2184, simple_loss=0.2661, pruned_loss=0.08536, over 4728.00 frames. ], tot_loss[loss=0.2359, simple_loss=0.2906, pruned_loss=0.09057, over 956221.50 frames. ], batch size: 23, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:46:38,686 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1164, 0.9994, 1.2271, 1.1848, 1.0740, 0.8887, 1.0210, 0.6528], device='cuda:4'), covar=tensor([0.0677, 0.0713, 0.0648, 0.0649, 0.0810, 0.1416, 0.0573, 0.1125], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0079, 0.0076, 0.0071, 0.0084, 0.0096, 0.0088, 0.0080], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-26 13:46:40,382 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.352e+02 1.823e+02 2.218e+02 2.622e+02 5.684e+02, threshold=4.435e+02, percent-clipped=2.0 2023-04-26 13:46:48,680 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4172, 3.0487, 0.7684, 1.6485, 1.7738, 2.1899, 1.9305, 1.0225], device='cuda:4'), covar=tensor([0.1361, 0.0978, 0.2289, 0.1404, 0.1111, 0.1066, 0.1402, 0.1839], device='cuda:4'), in_proj_covar=tensor([0.0123, 0.0266, 0.0149, 0.0130, 0.0141, 0.0164, 0.0127, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 13:46:58,365 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6986, 0.8723, 1.2407, 1.4071, 1.4273, 1.5997, 1.2677, 1.3204], device='cuda:4'), covar=tensor([1.0872, 2.0070, 1.5992, 1.4366, 1.7174, 2.7522, 1.9420, 1.6051], device='cuda:4'), in_proj_covar=tensor([0.0303, 0.0402, 0.0317, 0.0323, 0.0350, 0.0404, 0.0386, 0.0344], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 13:47:02,849 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.95 vs. limit=2.0 2023-04-26 13:47:03,042 INFO [finetune.py:976] (4/7) Epoch 3, batch 2800, loss[loss=0.2371, simple_loss=0.2766, pruned_loss=0.09882, over 4824.00 frames. ], tot_loss[loss=0.2337, simple_loss=0.2875, pruned_loss=0.08993, over 955945.57 frames. ], batch size: 33, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:47:08,492 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8741, 1.3741, 1.6741, 1.9152, 1.5483, 1.2707, 0.9572, 1.3436], device='cuda:4'), covar=tensor([0.4653, 0.5537, 0.2669, 0.4018, 0.4868, 0.4251, 0.7377, 0.4918], device='cuda:4'), in_proj_covar=tensor([0.0270, 0.0271, 0.0223, 0.0341, 0.0230, 0.0234, 0.0261, 0.0208], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 13:47:24,593 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4717, 3.4191, 1.0464, 1.8079, 2.0193, 2.2412, 2.1400, 0.9349], device='cuda:4'), covar=tensor([0.1370, 0.0920, 0.1972, 0.1424, 0.0996, 0.1242, 0.1435, 0.2236], device='cuda:4'), in_proj_covar=tensor([0.0123, 0.0266, 0.0149, 0.0130, 0.0141, 0.0164, 0.0126, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 13:47:35,420 INFO [finetune.py:976] (4/7) Epoch 3, batch 2850, loss[loss=0.2732, simple_loss=0.3232, pruned_loss=0.1117, over 4197.00 frames. ], tot_loss[loss=0.2327, simple_loss=0.2866, pruned_loss=0.08934, over 954327.62 frames. ], batch size: 65, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:47:36,122 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6796, 1.7803, 0.8924, 1.2429, 1.8824, 1.5472, 1.3751, 1.4837], device='cuda:4'), covar=tensor([0.0584, 0.0454, 0.0451, 0.0660, 0.0303, 0.0613, 0.0641, 0.0723], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0027, 0.0025, 0.0031, 0.0022, 0.0031, 0.0030, 0.0032], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0049, 0.0038, 0.0048, 0.0047, 0.0050], device='cuda:4') 2023-04-26 13:47:45,429 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.273e+02 1.919e+02 2.285e+02 2.703e+02 8.098e+02, threshold=4.570e+02, percent-clipped=1.0 2023-04-26 13:48:07,418 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.20 vs. limit=5.0 2023-04-26 13:48:08,417 INFO [finetune.py:976] (4/7) Epoch 3, batch 2900, loss[loss=0.2316, simple_loss=0.3051, pruned_loss=0.07902, over 4809.00 frames. ], tot_loss[loss=0.2361, simple_loss=0.2901, pruned_loss=0.09104, over 953050.86 frames. ], batch size: 41, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:48:12,770 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0506, 1.3586, 1.2420, 1.5803, 1.4465, 1.5535, 1.3329, 2.4155], device='cuda:4'), covar=tensor([0.0606, 0.0731, 0.0762, 0.1196, 0.0564, 0.0510, 0.0716, 0.0243], device='cuda:4'), in_proj_covar=tensor([0.0040, 0.0041, 0.0042, 0.0047, 0.0042, 0.0041, 0.0041, 0.0064], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 13:48:33,700 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=14393.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:48:41,649 INFO [finetune.py:976] (4/7) Epoch 3, batch 2950, loss[loss=0.2451, simple_loss=0.3011, pruned_loss=0.09454, over 4913.00 frames. ], tot_loss[loss=0.2389, simple_loss=0.2935, pruned_loss=0.09215, over 954775.10 frames. ], batch size: 38, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:48:46,074 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=14413.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:48:51,912 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.499e+02 2.004e+02 2.390e+02 2.807e+02 6.198e+02, threshold=4.780e+02, percent-clipped=3.0 2023-04-26 13:48:55,645 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=14428.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:49:04,516 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=14441.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:49:14,497 INFO [finetune.py:976] (4/7) Epoch 3, batch 3000, loss[loss=0.2754, simple_loss=0.3046, pruned_loss=0.1231, over 4421.00 frames. ], tot_loss[loss=0.2391, simple_loss=0.294, pruned_loss=0.09213, over 953595.89 frames. ], batch size: 19, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:49:14,497 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-26 13:49:19,193 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4861, 2.9829, 0.9960, 1.6929, 1.9178, 2.2449, 2.0410, 1.0570], device='cuda:4'), covar=tensor([0.1259, 0.0938, 0.1915, 0.1430, 0.0981, 0.1001, 0.1359, 0.1733], device='cuda:4'), in_proj_covar=tensor([0.0123, 0.0266, 0.0149, 0.0131, 0.0141, 0.0164, 0.0127, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 13:49:19,485 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5143, 0.9184, 1.3195, 1.7949, 1.6870, 1.3614, 1.3880, 1.3992], device='cuda:4'), covar=tensor([2.0633, 2.7609, 2.8398, 3.6037, 2.2856, 3.1819, 3.3623, 2.5500], device='cuda:4'), in_proj_covar=tensor([0.0448, 0.0497, 0.0591, 0.0600, 0.0479, 0.0513, 0.0525, 0.0535], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 13:49:25,022 INFO [finetune.py:1010] (4/7) Epoch 3, validation: loss=0.1699, simple_loss=0.2433, pruned_loss=0.04821, over 2265189.00 frames. 2023-04-26 13:49:25,022 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6259MB 2023-04-26 13:49:26,927 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=14459.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:49:36,017 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=14474.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:49:37,155 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=14476.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 13:49:43,717 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6365, 1.7472, 0.8424, 1.2302, 1.8152, 1.5343, 1.3515, 1.4584], device='cuda:4'), covar=tensor([0.0575, 0.0430, 0.0448, 0.0642, 0.0304, 0.0597, 0.0594, 0.0682], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0027, 0.0025, 0.0031, 0.0022, 0.0031, 0.0030, 0.0032], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0049, 0.0038, 0.0048, 0.0047, 0.0050], device='cuda:4') 2023-04-26 13:49:44,332 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=14487.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:50:13,196 INFO [finetune.py:976] (4/7) Epoch 3, batch 3050, loss[loss=0.2428, simple_loss=0.2966, pruned_loss=0.09448, over 4809.00 frames. ], tot_loss[loss=0.2405, simple_loss=0.2955, pruned_loss=0.09276, over 955111.51 frames. ], batch size: 33, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:50:18,964 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.61 vs. limit=2.0 2023-04-26 13:50:20,660 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1064, 1.5569, 1.4312, 1.7788, 1.6129, 2.0721, 1.4688, 3.6111], device='cuda:4'), covar=tensor([0.0698, 0.0807, 0.0792, 0.1256, 0.0664, 0.0545, 0.0753, 0.0123], device='cuda:4'), in_proj_covar=tensor([0.0041, 0.0041, 0.0042, 0.0047, 0.0042, 0.0041, 0.0041, 0.0064], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 13:50:24,170 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.447e+02 2.059e+02 2.417e+02 2.833e+02 5.136e+02, threshold=4.834e+02, percent-clipped=1.0 2023-04-26 13:50:32,123 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8590, 1.3304, 1.6634, 1.9423, 1.5260, 1.2478, 0.8888, 1.4263], device='cuda:4'), covar=tensor([0.4612, 0.5411, 0.2556, 0.4105, 0.4916, 0.4149, 0.7295, 0.4781], device='cuda:4'), in_proj_covar=tensor([0.0270, 0.0271, 0.0223, 0.0341, 0.0229, 0.0234, 0.0260, 0.0208], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 13:50:43,469 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-26 13:50:52,337 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=14548.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:51:03,204 INFO [finetune.py:976] (4/7) Epoch 3, batch 3100, loss[loss=0.2628, simple_loss=0.3012, pruned_loss=0.1122, over 4734.00 frames. ], tot_loss[loss=0.2367, simple_loss=0.2913, pruned_loss=0.09102, over 954732.06 frames. ], batch size: 54, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:52:09,675 INFO [finetune.py:976] (4/7) Epoch 3, batch 3150, loss[loss=0.2394, simple_loss=0.2952, pruned_loss=0.09183, over 4841.00 frames. ], tot_loss[loss=0.2338, simple_loss=0.288, pruned_loss=0.08984, over 955827.27 frames. ], batch size: 47, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:52:30,363 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.439e+02 2.002e+02 2.381e+02 3.061e+02 6.553e+02, threshold=4.761e+02, percent-clipped=3.0 2023-04-26 13:52:51,774 INFO [finetune.py:976] (4/7) Epoch 3, batch 3200, loss[loss=0.2364, simple_loss=0.2907, pruned_loss=0.09103, over 4909.00 frames. ], tot_loss[loss=0.2304, simple_loss=0.2845, pruned_loss=0.08812, over 957651.39 frames. ], batch size: 35, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:53:06,084 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-26 13:53:24,169 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-04-26 13:53:25,510 INFO [finetune.py:976] (4/7) Epoch 3, batch 3250, loss[loss=0.2318, simple_loss=0.278, pruned_loss=0.09284, over 4767.00 frames. ], tot_loss[loss=0.2313, simple_loss=0.285, pruned_loss=0.08882, over 956390.13 frames. ], batch size: 26, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:53:37,682 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.274e+02 1.942e+02 2.271e+02 2.763e+02 5.504e+02, threshold=4.542e+02, percent-clipped=2.0 2023-04-26 13:53:41,982 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=14728.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:53:49,353 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=14740.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:53:59,443 INFO [finetune.py:976] (4/7) Epoch 3, batch 3300, loss[loss=0.3261, simple_loss=0.3655, pruned_loss=0.1434, over 4899.00 frames. ], tot_loss[loss=0.2353, simple_loss=0.2897, pruned_loss=0.09042, over 954350.24 frames. ], batch size: 36, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:54:01,387 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=14759.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:54:09,428 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=14769.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:54:14,674 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=14776.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:54:14,699 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=14776.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 13:54:30,606 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=14801.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:54:33,538 INFO [finetune.py:976] (4/7) Epoch 3, batch 3350, loss[loss=0.2011, simple_loss=0.2728, pruned_loss=0.0647, over 4754.00 frames. ], tot_loss[loss=0.2364, simple_loss=0.2915, pruned_loss=0.09062, over 955185.47 frames. ], batch size: 27, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:54:34,179 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=14807.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:54:44,575 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.421e+02 1.946e+02 2.255e+02 2.721e+02 6.180e+02, threshold=4.510e+02, percent-clipped=3.0 2023-04-26 13:54:46,351 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=14824.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 13:54:58,945 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=14843.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:55:13,040 INFO [finetune.py:976] (4/7) Epoch 3, batch 3400, loss[loss=0.2491, simple_loss=0.3114, pruned_loss=0.09336, over 4848.00 frames. ], tot_loss[loss=0.237, simple_loss=0.2926, pruned_loss=0.09069, over 954972.03 frames. ], batch size: 31, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:56:08,499 INFO [finetune.py:976] (4/7) Epoch 3, batch 3450, loss[loss=0.2384, simple_loss=0.2981, pruned_loss=0.0894, over 4885.00 frames. ], tot_loss[loss=0.2355, simple_loss=0.2918, pruned_loss=0.08959, over 954101.17 frames. ], batch size: 43, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:56:18,847 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.286e+02 1.992e+02 2.282e+02 2.746e+02 5.093e+02, threshold=4.564e+02, percent-clipped=2.0 2023-04-26 13:56:42,395 INFO [finetune.py:976] (4/7) Epoch 3, batch 3500, loss[loss=0.2361, simple_loss=0.2967, pruned_loss=0.08774, over 4862.00 frames. ], tot_loss[loss=0.2329, simple_loss=0.2892, pruned_loss=0.08829, over 955745.87 frames. ], batch size: 44, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:56:56,148 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.9472, 3.8484, 2.9447, 4.5057, 3.8894, 3.9248, 1.7824, 3.8346], device='cuda:4'), covar=tensor([0.1509, 0.1116, 0.3032, 0.1241, 0.2351, 0.1594, 0.5726, 0.1988], device='cuda:4'), in_proj_covar=tensor([0.0255, 0.0226, 0.0267, 0.0319, 0.0312, 0.0263, 0.0280, 0.0280], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 13:57:38,372 INFO [finetune.py:976] (4/7) Epoch 3, batch 3550, loss[loss=0.1946, simple_loss=0.2617, pruned_loss=0.06382, over 4762.00 frames. ], tot_loss[loss=0.2314, simple_loss=0.2873, pruned_loss=0.08776, over 956990.68 frames. ], batch size: 28, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:57:54,036 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.153e+02 1.782e+02 2.249e+02 2.630e+02 4.772e+02, threshold=4.498e+02, percent-clipped=1.0 2023-04-26 13:58:24,505 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.95 vs. limit=2.0 2023-04-26 13:58:28,222 INFO [finetune.py:976] (4/7) Epoch 3, batch 3600, loss[loss=0.1991, simple_loss=0.2631, pruned_loss=0.06755, over 4932.00 frames. ], tot_loss[loss=0.2287, simple_loss=0.2837, pruned_loss=0.08682, over 955145.90 frames. ], batch size: 33, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:58:36,851 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=15069.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:58:56,110 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=15096.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:59:02,224 INFO [finetune.py:976] (4/7) Epoch 3, batch 3650, loss[loss=0.2799, simple_loss=0.3301, pruned_loss=0.1149, over 4912.00 frames. ], tot_loss[loss=0.2319, simple_loss=0.2866, pruned_loss=0.08862, over 953667.45 frames. ], batch size: 37, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:59:09,045 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=15117.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:59:12,568 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.417e+02 2.035e+02 2.405e+02 3.001e+02 5.155e+02, threshold=4.810e+02, percent-clipped=2.0 2023-04-26 13:59:28,523 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=15143.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 13:59:30,400 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4087, 1.1293, 1.3944, 1.6950, 1.5897, 1.3383, 1.3986, 1.3808], device='cuda:4'), covar=tensor([2.0895, 2.7988, 3.0101, 3.3009, 2.0963, 3.4235, 3.2933, 2.5303], device='cuda:4'), in_proj_covar=tensor([0.0447, 0.0498, 0.0590, 0.0599, 0.0480, 0.0512, 0.0525, 0.0534], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 13:59:36,235 INFO [finetune.py:976] (4/7) Epoch 3, batch 3700, loss[loss=0.2517, simple_loss=0.3155, pruned_loss=0.09395, over 4810.00 frames. ], tot_loss[loss=0.2351, simple_loss=0.2905, pruned_loss=0.08987, over 954735.82 frames. ], batch size: 45, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 13:59:59,374 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=15191.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:00:02,740 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=15195.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:00:10,003 INFO [finetune.py:976] (4/7) Epoch 3, batch 3750, loss[loss=0.2444, simple_loss=0.3084, pruned_loss=0.09022, over 4850.00 frames. ], tot_loss[loss=0.236, simple_loss=0.2918, pruned_loss=0.09005, over 951549.16 frames. ], batch size: 44, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 14:00:11,314 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.8959, 1.9013, 1.6586, 1.5687, 1.9361, 1.6154, 2.4526, 1.4405], device='cuda:4'), covar=tensor([0.3976, 0.1563, 0.4625, 0.2762, 0.1796, 0.2427, 0.1349, 0.4131], device='cuda:4'), in_proj_covar=tensor([0.0355, 0.0357, 0.0439, 0.0374, 0.0406, 0.0383, 0.0402, 0.0421], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 14:00:24,833 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.429e+02 1.976e+02 2.277e+02 2.735e+02 5.558e+02, threshold=4.554e+02, percent-clipped=2.0 2023-04-26 14:00:55,182 INFO [finetune.py:976] (4/7) Epoch 3, batch 3800, loss[loss=0.2437, simple_loss=0.303, pruned_loss=0.09222, over 4806.00 frames. ], tot_loss[loss=0.2355, simple_loss=0.2919, pruned_loss=0.08958, over 952205.90 frames. ], batch size: 39, lr: 3.98e-03, grad_scale: 16.0 2023-04-26 14:00:55,310 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=15256.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 14:00:55,921 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4848, 2.3665, 2.2473, 2.2068, 2.5300, 2.1677, 3.0759, 1.9386], device='cuda:4'), covar=tensor([0.3696, 0.1713, 0.3695, 0.2729, 0.1700, 0.2157, 0.1485, 0.3487], device='cuda:4'), in_proj_covar=tensor([0.0356, 0.0358, 0.0440, 0.0374, 0.0408, 0.0383, 0.0403, 0.0421], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 14:01:23,461 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3742, 0.9274, 1.1903, 1.6735, 1.5522, 1.2658, 1.2885, 1.2745], device='cuda:4'), covar=tensor([1.7130, 2.2683, 2.5268, 2.8338, 1.7957, 2.5893, 2.7091, 2.2638], device='cuda:4'), in_proj_covar=tensor([0.0446, 0.0496, 0.0589, 0.0598, 0.0479, 0.0510, 0.0523, 0.0532], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 14:01:29,010 INFO [finetune.py:976] (4/7) Epoch 3, batch 3850, loss[loss=0.1898, simple_loss=0.2564, pruned_loss=0.06163, over 4844.00 frames. ], tot_loss[loss=0.234, simple_loss=0.2901, pruned_loss=0.08896, over 951579.24 frames. ], batch size: 44, lr: 3.98e-03, grad_scale: 32.0 2023-04-26 14:01:30,050 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.13 vs. limit=2.0 2023-04-26 14:01:38,747 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.117e+02 1.876e+02 2.244e+02 2.688e+02 1.462e+03, threshold=4.488e+02, percent-clipped=2.0 2023-04-26 14:02:01,732 INFO [finetune.py:976] (4/7) Epoch 3, batch 3900, loss[loss=0.2419, simple_loss=0.2826, pruned_loss=0.1006, over 4826.00 frames. ], tot_loss[loss=0.2329, simple_loss=0.2885, pruned_loss=0.08868, over 952904.89 frames. ], batch size: 30, lr: 3.98e-03, grad_scale: 32.0 2023-04-26 14:02:14,278 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7971, 1.8386, 1.1411, 1.3744, 2.2961, 1.6739, 1.5186, 1.6127], device='cuda:4'), covar=tensor([0.0567, 0.0404, 0.0378, 0.0600, 0.0244, 0.0567, 0.0553, 0.0643], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0027, 0.0025, 0.0031, 0.0022, 0.0031, 0.0030, 0.0032], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0049, 0.0038, 0.0048, 0.0047, 0.0050], device='cuda:4') 2023-04-26 14:02:19,181 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7058, 1.8227, 1.0503, 1.3156, 2.2495, 1.5584, 1.4448, 1.5189], device='cuda:4'), covar=tensor([0.0586, 0.0409, 0.0391, 0.0610, 0.0257, 0.0572, 0.0570, 0.0645], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0027, 0.0025, 0.0031, 0.0022, 0.0031, 0.0030, 0.0032], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0049, 0.0038, 0.0048, 0.0047, 0.0050], device='cuda:4') 2023-04-26 14:02:50,598 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=15396.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:03:01,151 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.85 vs. limit=2.0 2023-04-26 14:03:03,041 INFO [finetune.py:976] (4/7) Epoch 3, batch 3950, loss[loss=0.223, simple_loss=0.2725, pruned_loss=0.08679, over 4723.00 frames. ], tot_loss[loss=0.2272, simple_loss=0.2831, pruned_loss=0.08567, over 954808.61 frames. ], batch size: 23, lr: 3.98e-03, grad_scale: 32.0 2023-04-26 14:03:11,752 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1488, 4.6107, 1.1434, 2.5695, 2.9763, 3.0202, 2.9471, 1.3658], device='cuda:4'), covar=tensor([0.1236, 0.0824, 0.2164, 0.1315, 0.0814, 0.1160, 0.1225, 0.1844], device='cuda:4'), in_proj_covar=tensor([0.0124, 0.0269, 0.0149, 0.0131, 0.0141, 0.0165, 0.0128, 0.0132], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 14:03:24,413 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.293e+02 1.749e+02 2.086e+02 2.518e+02 3.780e+02, threshold=4.171e+02, percent-clipped=0.0 2023-04-26 14:03:54,152 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=15444.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:04:08,305 INFO [finetune.py:976] (4/7) Epoch 3, batch 4000, loss[loss=0.3179, simple_loss=0.3511, pruned_loss=0.1424, over 4816.00 frames. ], tot_loss[loss=0.2285, simple_loss=0.2837, pruned_loss=0.08663, over 956555.91 frames. ], batch size: 41, lr: 3.98e-03, grad_scale: 32.0 2023-04-26 14:05:00,689 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3511, 1.7199, 2.2184, 2.7656, 2.0883, 1.6210, 1.4740, 2.1019], device='cuda:4'), covar=tensor([0.4861, 0.5696, 0.2694, 0.5082, 0.5771, 0.4591, 0.7283, 0.4880], device='cuda:4'), in_proj_covar=tensor([0.0271, 0.0270, 0.0222, 0.0340, 0.0228, 0.0233, 0.0258, 0.0207], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 14:05:04,589 INFO [finetune.py:976] (4/7) Epoch 3, batch 4050, loss[loss=0.1741, simple_loss=0.2485, pruned_loss=0.0498, over 4825.00 frames. ], tot_loss[loss=0.2309, simple_loss=0.2868, pruned_loss=0.08746, over 955275.70 frames. ], batch size: 39, lr: 3.98e-03, grad_scale: 32.0 2023-04-26 14:05:14,663 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5380, 2.5117, 2.2460, 2.3527, 2.6981, 2.3494, 3.5726, 1.9648], device='cuda:4'), covar=tensor([0.4592, 0.2074, 0.4666, 0.3675, 0.2093, 0.2732, 0.1577, 0.4458], device='cuda:4'), in_proj_covar=tensor([0.0355, 0.0358, 0.0441, 0.0374, 0.0408, 0.0383, 0.0404, 0.0421], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 14:05:15,730 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.134e+02 1.986e+02 2.201e+02 2.641e+02 4.376e+02, threshold=4.402e+02, percent-clipped=3.0 2023-04-26 14:05:33,872 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=15551.0, num_to_drop=1, layers_to_drop={3} 2023-04-26 14:05:36,856 INFO [finetune.py:976] (4/7) Epoch 3, batch 4100, loss[loss=0.2752, simple_loss=0.3142, pruned_loss=0.1181, over 4734.00 frames. ], tot_loss[loss=0.2333, simple_loss=0.2897, pruned_loss=0.08847, over 954704.74 frames. ], batch size: 59, lr: 3.98e-03, grad_scale: 32.0 2023-04-26 14:06:00,715 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.85 vs. limit=2.0 2023-04-26 14:06:10,674 INFO [finetune.py:976] (4/7) Epoch 3, batch 4150, loss[loss=0.226, simple_loss=0.2877, pruned_loss=0.08221, over 4802.00 frames. ], tot_loss[loss=0.2333, simple_loss=0.2901, pruned_loss=0.08826, over 954206.32 frames. ], batch size: 39, lr: 3.98e-03, grad_scale: 32.0 2023-04-26 14:06:23,083 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.264e+02 1.923e+02 2.257e+02 2.741e+02 5.009e+02, threshold=4.514e+02, percent-clipped=1.0 2023-04-26 14:06:44,251 INFO [finetune.py:976] (4/7) Epoch 3, batch 4200, loss[loss=0.2246, simple_loss=0.2836, pruned_loss=0.08275, over 4819.00 frames. ], tot_loss[loss=0.2335, simple_loss=0.2906, pruned_loss=0.08826, over 954488.91 frames. ], batch size: 33, lr: 3.98e-03, grad_scale: 32.0 2023-04-26 14:06:59,799 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.54 vs. limit=2.0 2023-04-26 14:07:17,985 INFO [finetune.py:976] (4/7) Epoch 3, batch 4250, loss[loss=0.2504, simple_loss=0.2979, pruned_loss=0.1015, over 4759.00 frames. ], tot_loss[loss=0.2321, simple_loss=0.2887, pruned_loss=0.08776, over 952993.39 frames. ], batch size: 26, lr: 3.98e-03, grad_scale: 32.0 2023-04-26 14:07:29,539 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.151e+02 1.926e+02 2.195e+02 2.674e+02 5.290e+02, threshold=4.390e+02, percent-clipped=1.0 2023-04-26 14:07:37,363 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2422, 1.4721, 1.3714, 1.7678, 1.5466, 1.8883, 1.4168, 3.4662], device='cuda:4'), covar=tensor([0.0652, 0.0768, 0.0816, 0.1216, 0.0645, 0.0590, 0.0756, 0.0170], device='cuda:4'), in_proj_covar=tensor([0.0040, 0.0040, 0.0042, 0.0046, 0.0041, 0.0041, 0.0040, 0.0064], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 14:07:43,432 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0375, 2.6857, 2.0066, 1.8991, 1.5158, 1.4969, 2.1370, 1.4720], device='cuda:4'), covar=tensor([0.2058, 0.1878, 0.2008, 0.2633, 0.3354, 0.2492, 0.1601, 0.2649], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0221, 0.0186, 0.0211, 0.0224, 0.0191, 0.0181, 0.0201], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 14:07:50,580 INFO [finetune.py:976] (4/7) Epoch 3, batch 4300, loss[loss=0.2231, simple_loss=0.2848, pruned_loss=0.08075, over 4931.00 frames. ], tot_loss[loss=0.2298, simple_loss=0.2858, pruned_loss=0.08688, over 953910.18 frames. ], batch size: 37, lr: 3.98e-03, grad_scale: 32.0 2023-04-26 14:08:59,719 INFO [finetune.py:976] (4/7) Epoch 3, batch 4350, loss[loss=0.2243, simple_loss=0.2608, pruned_loss=0.0939, over 4156.00 frames. ], tot_loss[loss=0.2258, simple_loss=0.2817, pruned_loss=0.08501, over 952833.53 frames. ], batch size: 18, lr: 3.98e-03, grad_scale: 32.0 2023-04-26 14:09:21,816 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.851e+02 2.117e+02 2.553e+02 4.240e+02, threshold=4.233e+02, percent-clipped=0.0 2023-04-26 14:10:01,712 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=15851.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:10:04,718 INFO [finetune.py:976] (4/7) Epoch 3, batch 4400, loss[loss=0.3408, simple_loss=0.3681, pruned_loss=0.1567, over 4793.00 frames. ], tot_loss[loss=0.2285, simple_loss=0.2837, pruned_loss=0.08662, over 954273.64 frames. ], batch size: 51, lr: 3.97e-03, grad_scale: 32.0 2023-04-26 14:10:05,401 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2851, 1.1996, 3.8444, 3.5722, 3.4061, 3.6249, 3.6257, 3.3482], device='cuda:4'), covar=tensor([0.7264, 0.6064, 0.1170, 0.1835, 0.1104, 0.1812, 0.1834, 0.1636], device='cuda:4'), in_proj_covar=tensor([0.0327, 0.0312, 0.0436, 0.0441, 0.0372, 0.0425, 0.0334, 0.0392], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-26 14:10:39,588 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=15883.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:10:49,611 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=15899.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:10:53,840 INFO [finetune.py:976] (4/7) Epoch 3, batch 4450, loss[loss=0.2736, simple_loss=0.333, pruned_loss=0.1071, over 4840.00 frames. ], tot_loss[loss=0.233, simple_loss=0.2889, pruned_loss=0.08858, over 954162.68 frames. ], batch size: 47, lr: 3.97e-03, grad_scale: 32.0 2023-04-26 14:11:04,060 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.361e+02 2.045e+02 2.492e+02 3.065e+02 6.089e+02, threshold=4.985e+02, percent-clipped=5.0 2023-04-26 14:11:20,079 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=15944.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:11:26,262 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=5.05 vs. limit=5.0 2023-04-26 14:11:27,288 INFO [finetune.py:976] (4/7) Epoch 3, batch 4500, loss[loss=0.2305, simple_loss=0.275, pruned_loss=0.09304, over 4787.00 frames. ], tot_loss[loss=0.2322, simple_loss=0.2887, pruned_loss=0.08779, over 953564.86 frames. ], batch size: 25, lr: 3.97e-03, grad_scale: 32.0 2023-04-26 14:11:59,083 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=16001.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:12:02,032 INFO [finetune.py:976] (4/7) Epoch 3, batch 4550, loss[loss=0.2407, simple_loss=0.3103, pruned_loss=0.0856, over 4889.00 frames. ], tot_loss[loss=0.2342, simple_loss=0.2905, pruned_loss=0.08894, over 953702.09 frames. ], batch size: 37, lr: 3.97e-03, grad_scale: 32.0 2023-04-26 14:12:11,702 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.282e+02 1.831e+02 2.180e+02 2.572e+02 4.192e+02, threshold=4.359e+02, percent-clipped=0.0 2023-04-26 14:12:35,169 INFO [finetune.py:976] (4/7) Epoch 3, batch 4600, loss[loss=0.2329, simple_loss=0.2875, pruned_loss=0.08914, over 4710.00 frames. ], tot_loss[loss=0.2337, simple_loss=0.2902, pruned_loss=0.08861, over 951577.40 frames. ], batch size: 59, lr: 3.97e-03, grad_scale: 32.0 2023-04-26 14:12:38,917 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=16062.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:12:46,831 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0238, 1.5254, 1.9024, 2.2176, 1.7923, 1.4312, 1.1068, 1.5579], device='cuda:4'), covar=tensor([0.5219, 0.5923, 0.2877, 0.4628, 0.5222, 0.4802, 0.7315, 0.5069], device='cuda:4'), in_proj_covar=tensor([0.0271, 0.0269, 0.0223, 0.0340, 0.0228, 0.0233, 0.0258, 0.0206], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 14:12:49,478 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.77 vs. limit=5.0 2023-04-26 14:12:53,906 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=16085.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:13:08,933 INFO [finetune.py:976] (4/7) Epoch 3, batch 4650, loss[loss=0.2605, simple_loss=0.3012, pruned_loss=0.1099, over 4794.00 frames. ], tot_loss[loss=0.2311, simple_loss=0.2873, pruned_loss=0.08744, over 953666.58 frames. ], batch size: 51, lr: 3.97e-03, grad_scale: 32.0 2023-04-26 14:13:18,703 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.295e+02 1.883e+02 2.189e+02 2.586e+02 3.625e+02, threshold=4.377e+02, percent-clipped=0.0 2023-04-26 14:13:47,604 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=16146.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:13:49,911 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=16149.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:13:52,277 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5679, 2.4561, 2.1876, 2.3213, 2.6347, 2.3877, 3.5382, 1.9727], device='cuda:4'), covar=tensor([0.4930, 0.2517, 0.4948, 0.4683, 0.2519, 0.2975, 0.1805, 0.4890], device='cuda:4'), in_proj_covar=tensor([0.0357, 0.0360, 0.0443, 0.0376, 0.0412, 0.0388, 0.0405, 0.0424], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 14:13:53,995 INFO [finetune.py:976] (4/7) Epoch 3, batch 4700, loss[loss=0.2592, simple_loss=0.3082, pruned_loss=0.1052, over 4843.00 frames. ], tot_loss[loss=0.2279, simple_loss=0.2839, pruned_loss=0.08595, over 954298.41 frames. ], batch size: 44, lr: 3.97e-03, grad_scale: 32.0 2023-04-26 14:13:58,624 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=2.05 vs. limit=2.0 2023-04-26 14:14:38,783 INFO [finetune.py:976] (4/7) Epoch 3, batch 4750, loss[loss=0.2593, simple_loss=0.3039, pruned_loss=0.1074, over 4727.00 frames. ], tot_loss[loss=0.2255, simple_loss=0.2809, pruned_loss=0.08499, over 953847.33 frames. ], batch size: 59, lr: 3.97e-03, grad_scale: 32.0 2023-04-26 14:14:47,158 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=16210.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:15:00,405 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.300e+02 1.819e+02 2.117e+02 2.422e+02 6.375e+02, threshold=4.235e+02, percent-clipped=3.0 2023-04-26 14:15:03,004 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4470, 1.2223, 1.5289, 1.7303, 1.6254, 1.4537, 1.5222, 1.4824], device='cuda:4'), covar=tensor([1.6817, 2.2672, 2.5355, 2.9030, 1.7948, 2.6511, 2.5533, 2.1084], device='cuda:4'), in_proj_covar=tensor([0.0444, 0.0493, 0.0583, 0.0596, 0.0475, 0.0505, 0.0517, 0.0527], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 14:15:10,822 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=16239.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:15:35,035 INFO [finetune.py:976] (4/7) Epoch 3, batch 4800, loss[loss=0.1789, simple_loss=0.2272, pruned_loss=0.06532, over 4199.00 frames. ], tot_loss[loss=0.2271, simple_loss=0.2829, pruned_loss=0.08565, over 954457.63 frames. ], batch size: 18, lr: 3.97e-03, grad_scale: 32.0 2023-04-26 14:16:42,026 INFO [finetune.py:976] (4/7) Epoch 3, batch 4850, loss[loss=0.2546, simple_loss=0.3051, pruned_loss=0.102, over 4903.00 frames. ], tot_loss[loss=0.2298, simple_loss=0.2862, pruned_loss=0.08666, over 954413.78 frames. ], batch size: 37, lr: 3.97e-03, grad_scale: 32.0 2023-04-26 14:17:03,614 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.365e+02 1.946e+02 2.319e+02 2.886e+02 4.382e+02, threshold=4.637e+02, percent-clipped=1.0 2023-04-26 14:17:05,620 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0019, 1.3692, 1.8636, 2.3856, 1.7732, 1.3647, 1.1157, 1.6749], device='cuda:4'), covar=tensor([0.4929, 0.6025, 0.2843, 0.4541, 0.5896, 0.4621, 0.7418, 0.4928], device='cuda:4'), in_proj_covar=tensor([0.0272, 0.0271, 0.0224, 0.0341, 0.0229, 0.0234, 0.0259, 0.0206], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 14:17:31,158 INFO [finetune.py:976] (4/7) Epoch 3, batch 4900, loss[loss=0.2602, simple_loss=0.3064, pruned_loss=0.107, over 4917.00 frames. ], tot_loss[loss=0.2323, simple_loss=0.2885, pruned_loss=0.08798, over 952650.61 frames. ], batch size: 38, lr: 3.97e-03, grad_scale: 32.0 2023-04-26 14:17:32,328 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=16357.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:17:40,878 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-26 14:17:42,339 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=16371.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:17:44,163 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4541, 4.7103, 1.2958, 2.8156, 3.1327, 3.2446, 3.1577, 1.5985], device='cuda:4'), covar=tensor([0.1124, 0.0909, 0.2027, 0.1118, 0.0769, 0.1037, 0.1077, 0.1743], device='cuda:4'), in_proj_covar=tensor([0.0124, 0.0267, 0.0149, 0.0131, 0.0141, 0.0164, 0.0128, 0.0131], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 14:18:04,092 INFO [finetune.py:976] (4/7) Epoch 3, batch 4950, loss[loss=0.2576, simple_loss=0.3136, pruned_loss=0.1008, over 4808.00 frames. ], tot_loss[loss=0.2332, simple_loss=0.2895, pruned_loss=0.08841, over 950832.93 frames. ], batch size: 45, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:18:15,329 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.97 vs. limit=2.0 2023-04-26 14:18:16,360 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.286e+02 1.926e+02 2.230e+02 2.725e+02 6.217e+02, threshold=4.460e+02, percent-clipped=3.0 2023-04-26 14:18:21,905 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=16432.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:18:24,945 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8216, 2.6438, 2.3713, 2.5855, 3.0214, 2.4873, 3.6825, 2.1043], device='cuda:4'), covar=tensor([0.4283, 0.2198, 0.3874, 0.3543, 0.1777, 0.2619, 0.1596, 0.4220], device='cuda:4'), in_proj_covar=tensor([0.0360, 0.0362, 0.0446, 0.0378, 0.0413, 0.0390, 0.0407, 0.0427], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 14:18:26,163 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1418, 1.3357, 1.4089, 1.5343, 1.4538, 1.6094, 1.4702, 1.4535], device='cuda:4'), covar=tensor([1.1329, 1.8659, 1.4962, 1.3115, 1.5462, 2.4137, 1.8762, 1.6752], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0403, 0.0321, 0.0326, 0.0353, 0.0411, 0.0389, 0.0346], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 14:18:27,290 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=16441.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:18:37,160 INFO [finetune.py:976] (4/7) Epoch 3, batch 5000, loss[loss=0.2563, simple_loss=0.299, pruned_loss=0.1068, over 4818.00 frames. ], tot_loss[loss=0.2318, simple_loss=0.2882, pruned_loss=0.08769, over 952822.76 frames. ], batch size: 39, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:18:47,767 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.50 vs. limit=5.0 2023-04-26 14:19:01,594 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6775, 2.1699, 1.7895, 1.9521, 1.5042, 1.7241, 1.9201, 1.3962], device='cuda:4'), covar=tensor([0.2431, 0.1835, 0.1538, 0.1940, 0.3415, 0.1790, 0.1996, 0.2908], device='cuda:4'), in_proj_covar=tensor([0.0318, 0.0335, 0.0246, 0.0308, 0.0321, 0.0288, 0.0278, 0.0301], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 14:19:09,883 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=16505.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:19:10,436 INFO [finetune.py:976] (4/7) Epoch 3, batch 5050, loss[loss=0.2168, simple_loss=0.2727, pruned_loss=0.0805, over 4780.00 frames. ], tot_loss[loss=0.2282, simple_loss=0.2842, pruned_loss=0.08608, over 952319.64 frames. ], batch size: 29, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:19:23,712 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.221e+02 1.760e+02 2.084e+02 2.475e+02 5.733e+02, threshold=4.169e+02, percent-clipped=2.0 2023-04-26 14:19:26,049 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-04-26 14:19:33,534 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=16539.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:19:35,348 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1413, 0.9788, 1.2272, 1.1814, 1.0732, 0.9036, 0.9182, 0.5613], device='cuda:4'), covar=tensor([0.0748, 0.1069, 0.0718, 0.0953, 0.1067, 0.1592, 0.0825, 0.1285], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0080, 0.0078, 0.0072, 0.0085, 0.0099, 0.0088, 0.0082], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-26 14:19:43,695 INFO [finetune.py:976] (4/7) Epoch 3, batch 5100, loss[loss=0.2341, simple_loss=0.2817, pruned_loss=0.0933, over 4824.00 frames. ], tot_loss[loss=0.2225, simple_loss=0.2787, pruned_loss=0.08309, over 954127.15 frames. ], batch size: 38, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:20:10,604 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=16587.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:20:11,254 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6747, 1.2919, 4.1063, 3.8057, 3.6197, 3.8034, 3.7610, 3.6811], device='cuda:4'), covar=tensor([0.6848, 0.5920, 0.0981, 0.1632, 0.1105, 0.1479, 0.1858, 0.1513], device='cuda:4'), in_proj_covar=tensor([0.0327, 0.0314, 0.0439, 0.0440, 0.0372, 0.0426, 0.0333, 0.0393], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-26 14:20:13,162 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3322, 0.8583, 1.1696, 1.6627, 1.4892, 1.2546, 1.2634, 1.2708], device='cuda:4'), covar=tensor([1.6622, 2.2075, 2.3160, 2.7713, 1.8271, 2.6201, 2.6087, 2.1377], device='cuda:4'), in_proj_covar=tensor([0.0443, 0.0491, 0.0582, 0.0594, 0.0474, 0.0505, 0.0517, 0.0526], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 14:20:33,135 INFO [finetune.py:976] (4/7) Epoch 3, batch 5150, loss[loss=0.2185, simple_loss=0.2714, pruned_loss=0.08277, over 4790.00 frames. ], tot_loss[loss=0.2248, simple_loss=0.2807, pruned_loss=0.08452, over 951529.69 frames. ], batch size: 29, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:20:56,490 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.263e+02 1.945e+02 2.270e+02 2.709e+02 5.209e+02, threshold=4.540e+02, percent-clipped=2.0 2023-04-26 14:21:30,490 INFO [finetune.py:976] (4/7) Epoch 3, batch 5200, loss[loss=0.2111, simple_loss=0.2751, pruned_loss=0.07358, over 4771.00 frames. ], tot_loss[loss=0.2285, simple_loss=0.2851, pruned_loss=0.086, over 951275.02 frames. ], batch size: 26, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:21:31,188 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=16657.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:22:02,898 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=16705.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:22:03,454 INFO [finetune.py:976] (4/7) Epoch 3, batch 5250, loss[loss=0.2293, simple_loss=0.2966, pruned_loss=0.08099, over 4889.00 frames. ], tot_loss[loss=0.2308, simple_loss=0.2879, pruned_loss=0.08681, over 951215.17 frames. ], batch size: 37, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:22:14,743 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.228e+02 1.971e+02 2.303e+02 2.786e+02 5.327e+02, threshold=4.606e+02, percent-clipped=1.0 2023-04-26 14:22:18,689 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=16727.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:22:21,693 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.30 vs. limit=5.0 2023-04-26 14:22:22,218 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1090, 1.5999, 1.4526, 1.7789, 1.6086, 2.2142, 1.4637, 3.7624], device='cuda:4'), covar=tensor([0.0715, 0.0762, 0.0857, 0.1223, 0.0674, 0.0513, 0.0767, 0.0142], device='cuda:4'), in_proj_covar=tensor([0.0040, 0.0040, 0.0042, 0.0046, 0.0042, 0.0041, 0.0040, 0.0064], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 14:22:33,183 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=16741.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:22:34,908 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-26 14:22:47,372 INFO [finetune.py:976] (4/7) Epoch 3, batch 5300, loss[loss=0.2125, simple_loss=0.2724, pruned_loss=0.07632, over 4886.00 frames. ], tot_loss[loss=0.2343, simple_loss=0.2915, pruned_loss=0.0885, over 954011.45 frames. ], batch size: 32, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:22:55,382 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-26 14:23:10,430 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=16789.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:23:20,244 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=16805.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:23:20,766 INFO [finetune.py:976] (4/7) Epoch 3, batch 5350, loss[loss=0.1692, simple_loss=0.2392, pruned_loss=0.04959, over 4763.00 frames. ], tot_loss[loss=0.2341, simple_loss=0.2915, pruned_loss=0.08839, over 953684.18 frames. ], batch size: 28, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:23:23,283 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.7922, 1.7738, 1.6019, 1.3740, 1.8471, 1.4066, 2.2474, 1.4365], device='cuda:4'), covar=tensor([0.4265, 0.1605, 0.5235, 0.3246, 0.1722, 0.2569, 0.1572, 0.4730], device='cuda:4'), in_proj_covar=tensor([0.0356, 0.0357, 0.0441, 0.0372, 0.0410, 0.0386, 0.0403, 0.0422], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 14:23:31,441 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.249e+02 1.831e+02 2.243e+02 2.677e+02 3.612e+02, threshold=4.486e+02, percent-clipped=0.0 2023-04-26 14:23:35,052 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6005, 1.7462, 0.8840, 1.2581, 1.8547, 1.4732, 1.4071, 1.4257], device='cuda:4'), covar=tensor([0.0580, 0.0378, 0.0443, 0.0592, 0.0312, 0.0537, 0.0499, 0.0642], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0027, 0.0025, 0.0031, 0.0022, 0.0031, 0.0030, 0.0032], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0045, 0.0038, 0.0049, 0.0038, 0.0048, 0.0047, 0.0050], device='cuda:4') 2023-04-26 14:23:52,046 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=16853.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:23:53,793 INFO [finetune.py:976] (4/7) Epoch 3, batch 5400, loss[loss=0.2514, simple_loss=0.3044, pruned_loss=0.09926, over 4811.00 frames. ], tot_loss[loss=0.2329, simple_loss=0.2896, pruned_loss=0.08816, over 954628.81 frames. ], batch size: 39, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:24:01,175 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=16868.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 14:24:06,225 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.93 vs. limit=5.0 2023-04-26 14:24:07,612 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2898, 2.2049, 2.4346, 2.7641, 2.7609, 2.1767, 1.8235, 2.3098], device='cuda:4'), covar=tensor([0.0944, 0.1006, 0.0519, 0.0603, 0.0568, 0.0999, 0.1120, 0.0625], device='cuda:4'), in_proj_covar=tensor([0.0209, 0.0211, 0.0188, 0.0184, 0.0182, 0.0200, 0.0174, 0.0193], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 14:24:23,882 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.47 vs. limit=5.0 2023-04-26 14:24:27,335 INFO [finetune.py:976] (4/7) Epoch 3, batch 5450, loss[loss=0.1869, simple_loss=0.2488, pruned_loss=0.06254, over 4773.00 frames. ], tot_loss[loss=0.2288, simple_loss=0.2853, pruned_loss=0.08616, over 956540.11 frames. ], batch size: 28, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:24:27,442 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=16906.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:24:37,664 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.358e+02 1.908e+02 2.254e+02 2.695e+02 6.070e+02, threshold=4.507e+02, percent-clipped=3.0 2023-04-26 14:24:41,866 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=16929.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 14:25:00,558 INFO [finetune.py:976] (4/7) Epoch 3, batch 5500, loss[loss=0.2139, simple_loss=0.2746, pruned_loss=0.07658, over 4739.00 frames. ], tot_loss[loss=0.2245, simple_loss=0.281, pruned_loss=0.08404, over 955476.13 frames. ], batch size: 54, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:25:07,409 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=16967.0, num_to_drop=1, layers_to_drop={3} 2023-04-26 14:25:32,655 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2419, 1.4988, 1.4338, 1.9161, 1.5674, 1.9479, 1.4691, 3.1377], device='cuda:4'), covar=tensor([0.0683, 0.0764, 0.0813, 0.1152, 0.0647, 0.0525, 0.0726, 0.0214], device='cuda:4'), in_proj_covar=tensor([0.0040, 0.0040, 0.0042, 0.0046, 0.0042, 0.0041, 0.0041, 0.0064], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 14:25:33,761 INFO [finetune.py:976] (4/7) Epoch 3, batch 5550, loss[loss=0.3058, simple_loss=0.3493, pruned_loss=0.1311, over 4221.00 frames. ], tot_loss[loss=0.2274, simple_loss=0.283, pruned_loss=0.08594, over 953341.81 frames. ], batch size: 65, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:25:54,585 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.318e+02 1.795e+02 2.241e+02 2.790e+02 4.152e+02, threshold=4.483e+02, percent-clipped=0.0 2023-04-26 14:26:02,723 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=17027.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:26:35,598 INFO [finetune.py:976] (4/7) Epoch 3, batch 5600, loss[loss=0.2413, simple_loss=0.2907, pruned_loss=0.09597, over 4766.00 frames. ], tot_loss[loss=0.2297, simple_loss=0.286, pruned_loss=0.08676, over 949297.91 frames. ], batch size: 26, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:26:59,009 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=17075.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:27:38,805 INFO [finetune.py:976] (4/7) Epoch 3, batch 5650, loss[loss=0.2292, simple_loss=0.3007, pruned_loss=0.07885, over 4928.00 frames. ], tot_loss[loss=0.2325, simple_loss=0.2894, pruned_loss=0.08775, over 950908.88 frames. ], batch size: 38, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:27:55,218 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.353e+02 1.900e+02 2.198e+02 2.594e+02 5.346e+02, threshold=4.396e+02, percent-clipped=1.0 2023-04-26 14:28:26,725 INFO [finetune.py:976] (4/7) Epoch 3, batch 5700, loss[loss=0.2219, simple_loss=0.2617, pruned_loss=0.0911, over 3962.00 frames. ], tot_loss[loss=0.23, simple_loss=0.2853, pruned_loss=0.08735, over 932671.27 frames. ], batch size: 17, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:28:27,730 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-26 14:29:04,048 INFO [finetune.py:976] (4/7) Epoch 4, batch 0, loss[loss=0.2805, simple_loss=0.3324, pruned_loss=0.1143, over 4812.00 frames. ], tot_loss[loss=0.2805, simple_loss=0.3324, pruned_loss=0.1143, over 4812.00 frames. ], batch size: 39, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:29:04,048 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-26 14:29:11,496 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4666, 1.1306, 1.2502, 1.1220, 1.7351, 1.3908, 1.1018, 1.2489], device='cuda:4'), covar=tensor([0.1934, 0.1599, 0.2506, 0.1646, 0.0898, 0.1771, 0.2355, 0.2393], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0334, 0.0348, 0.0309, 0.0345, 0.0353, 0.0314, 0.0351], device='cuda:4'), out_proj_covar=tensor([6.7823e-05, 7.1689e-05, 7.5527e-05, 6.4718e-05, 7.3169e-05, 7.7032e-05, 6.8408e-05, 7.5909e-05], device='cuda:4') 2023-04-26 14:29:26,719 INFO [finetune.py:1010] (4/7) Epoch 4, validation: loss=0.1686, simple_loss=0.2415, pruned_loss=0.04788, over 2265189.00 frames. 2023-04-26 14:29:26,719 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6259MB 2023-04-26 14:29:31,445 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=17189.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:29:52,774 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.151e+02 1.687e+02 2.032e+02 2.492e+02 4.364e+02, threshold=4.064e+02, percent-clipped=0.0 2023-04-26 14:29:53,450 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=17224.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 14:29:58,842 INFO [finetune.py:976] (4/7) Epoch 4, batch 50, loss[loss=0.2234, simple_loss=0.2648, pruned_loss=0.09102, over 4721.00 frames. ], tot_loss[loss=0.2355, simple_loss=0.2918, pruned_loss=0.08957, over 216728.71 frames. ], batch size: 59, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:29:59,566 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0169, 1.9024, 2.0801, 2.3314, 2.3726, 1.9713, 1.5938, 2.0098], device='cuda:4'), covar=tensor([0.0948, 0.1030, 0.0658, 0.0635, 0.0661, 0.0928, 0.1122, 0.0694], device='cuda:4'), in_proj_covar=tensor([0.0209, 0.0211, 0.0188, 0.0184, 0.0183, 0.0199, 0.0174, 0.0193], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 14:30:11,295 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=17250.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:30:18,608 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=17262.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 14:30:31,860 INFO [finetune.py:976] (4/7) Epoch 4, batch 100, loss[loss=0.189, simple_loss=0.2497, pruned_loss=0.06411, over 4758.00 frames. ], tot_loss[loss=0.222, simple_loss=0.2787, pruned_loss=0.08265, over 380128.39 frames. ], batch size: 54, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:30:48,706 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4555, 1.3324, 1.7240, 1.6167, 1.3488, 1.1165, 1.4346, 0.9920], device='cuda:4'), covar=tensor([0.0977, 0.0906, 0.0607, 0.0922, 0.0995, 0.1567, 0.0882, 0.1043], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0078, 0.0076, 0.0071, 0.0083, 0.0097, 0.0086, 0.0080], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-26 14:30:58,842 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.312e+02 1.837e+02 2.342e+02 2.858e+02 5.078e+02, threshold=4.684e+02, percent-clipped=3.0 2023-04-26 14:31:02,700 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2905, 1.4653, 1.5406, 1.6800, 1.6169, 1.7950, 1.6152, 1.5967], device='cuda:4'), covar=tensor([0.9525, 1.5493, 1.3914, 1.2262, 1.4244, 2.1461, 1.6797, 1.4913], device='cuda:4'), in_proj_covar=tensor([0.0306, 0.0400, 0.0319, 0.0325, 0.0350, 0.0411, 0.0387, 0.0342], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 14:31:03,297 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8717, 2.7843, 1.9457, 1.8697, 1.4213, 1.3953, 2.0108, 1.3490], device='cuda:4'), covar=tensor([0.2102, 0.1931, 0.2146, 0.2620, 0.3436, 0.2521, 0.1636, 0.2632], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0220, 0.0184, 0.0210, 0.0222, 0.0190, 0.0179, 0.0200], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 14:31:04,969 INFO [finetune.py:976] (4/7) Epoch 4, batch 150, loss[loss=0.2134, simple_loss=0.2701, pruned_loss=0.07839, over 4820.00 frames. ], tot_loss[loss=0.2222, simple_loss=0.2776, pruned_loss=0.08342, over 508053.41 frames. ], batch size: 38, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:31:12,518 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8249, 4.2573, 0.7861, 2.3938, 2.3352, 2.7513, 2.7074, 1.1128], device='cuda:4'), covar=tensor([0.1399, 0.0891, 0.2391, 0.1248, 0.0983, 0.1190, 0.1325, 0.2093], device='cuda:4'), in_proj_covar=tensor([0.0123, 0.0264, 0.0149, 0.0130, 0.0140, 0.0162, 0.0127, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 14:31:24,759 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=17362.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:31:33,394 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.53 vs. limit=2.0 2023-04-26 14:31:38,037 INFO [finetune.py:976] (4/7) Epoch 4, batch 200, loss[loss=0.1796, simple_loss=0.2458, pruned_loss=0.05673, over 4893.00 frames. ], tot_loss[loss=0.2224, simple_loss=0.2769, pruned_loss=0.08395, over 604507.44 frames. ], batch size: 35, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:31:38,177 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=17383.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 14:32:05,047 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.188e+02 2.027e+02 2.314e+02 2.814e+02 1.019e+03, threshold=4.629e+02, percent-clipped=4.0 2023-04-26 14:32:05,196 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=17423.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:32:11,142 INFO [finetune.py:976] (4/7) Epoch 4, batch 250, loss[loss=0.2689, simple_loss=0.3282, pruned_loss=0.1048, over 4803.00 frames. ], tot_loss[loss=0.2268, simple_loss=0.2819, pruned_loss=0.08582, over 681238.28 frames. ], batch size: 41, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:32:25,182 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=17444.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 14:32:34,073 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6952, 1.4417, 1.8698, 1.9600, 1.5557, 1.1560, 1.6003, 0.9794], device='cuda:4'), covar=tensor([0.0998, 0.0991, 0.0672, 0.0858, 0.1069, 0.2348, 0.0974, 0.1505], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0078, 0.0076, 0.0071, 0.0083, 0.0097, 0.0086, 0.0080], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-26 14:33:16,610 INFO [finetune.py:976] (4/7) Epoch 4, batch 300, loss[loss=0.2551, simple_loss=0.309, pruned_loss=0.1006, over 4820.00 frames. ], tot_loss[loss=0.2312, simple_loss=0.2872, pruned_loss=0.08761, over 742824.39 frames. ], batch size: 49, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:34:04,408 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.014e+02 1.938e+02 2.299e+02 2.709e+02 4.777e+02, threshold=4.598e+02, percent-clipped=1.0 2023-04-26 14:34:10,653 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=17524.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 14:34:22,587 INFO [finetune.py:976] (4/7) Epoch 4, batch 350, loss[loss=0.2455, simple_loss=0.3131, pruned_loss=0.08896, over 4809.00 frames. ], tot_loss[loss=0.2339, simple_loss=0.2905, pruned_loss=0.08862, over 792011.52 frames. ], batch size: 38, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:34:35,181 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.9137, 2.8243, 2.2228, 3.3008, 2.8731, 2.9035, 1.1841, 2.8080], device='cuda:4'), covar=tensor([0.1997, 0.1509, 0.3055, 0.3041, 0.2599, 0.2184, 0.5502, 0.2543], device='cuda:4'), in_proj_covar=tensor([0.0253, 0.0224, 0.0265, 0.0319, 0.0313, 0.0263, 0.0278, 0.0281], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 14:34:36,258 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=17545.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:34:48,313 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0920, 2.8869, 2.4012, 2.5114, 2.2327, 2.3117, 2.5360, 1.9095], device='cuda:4'), covar=tensor([0.3007, 0.1655, 0.1084, 0.1806, 0.3305, 0.1665, 0.2611, 0.3348], device='cuda:4'), in_proj_covar=tensor([0.0317, 0.0335, 0.0245, 0.0309, 0.0320, 0.0288, 0.0276, 0.0299], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 14:34:59,108 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=17562.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 14:35:05,484 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9145, 1.9282, 1.6206, 1.5447, 2.1086, 1.5615, 2.5665, 1.4367], device='cuda:4'), covar=tensor([0.4043, 0.1615, 0.5188, 0.3091, 0.1654, 0.2540, 0.1258, 0.4849], device='cuda:4'), in_proj_covar=tensor([0.0357, 0.0359, 0.0446, 0.0375, 0.0411, 0.0386, 0.0404, 0.0425], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 14:35:10,347 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=17572.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 14:35:17,455 INFO [finetune.py:976] (4/7) Epoch 4, batch 400, loss[loss=0.1983, simple_loss=0.2653, pruned_loss=0.06561, over 4740.00 frames. ], tot_loss[loss=0.2333, simple_loss=0.2907, pruned_loss=0.08791, over 829504.12 frames. ], batch size: 54, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:35:37,142 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=17610.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:35:45,052 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.309e+02 1.823e+02 2.162e+02 2.599e+02 8.047e+02, threshold=4.324e+02, percent-clipped=1.0 2023-04-26 14:35:51,211 INFO [finetune.py:976] (4/7) Epoch 4, batch 450, loss[loss=0.2205, simple_loss=0.2809, pruned_loss=0.08008, over 4910.00 frames. ], tot_loss[loss=0.2302, simple_loss=0.2872, pruned_loss=0.08659, over 853986.04 frames. ], batch size: 43, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:36:25,013 INFO [finetune.py:976] (4/7) Epoch 4, batch 500, loss[loss=0.2192, simple_loss=0.2733, pruned_loss=0.08251, over 4823.00 frames. ], tot_loss[loss=0.228, simple_loss=0.2845, pruned_loss=0.08573, over 877772.26 frames. ], batch size: 39, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:36:35,748 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=17700.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:36:49,493 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=17718.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:36:52,485 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.156e+02 1.778e+02 2.181e+02 2.733e+02 7.061e+02, threshold=4.362e+02, percent-clipped=3.0 2023-04-26 14:36:56,857 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=17730.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:36:58,589 INFO [finetune.py:976] (4/7) Epoch 4, batch 550, loss[loss=0.2143, simple_loss=0.2723, pruned_loss=0.07817, over 4838.00 frames. ], tot_loss[loss=0.2266, simple_loss=0.2824, pruned_loss=0.08541, over 895471.90 frames. ], batch size: 30, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:37:02,831 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=17739.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 14:37:18,654 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=17761.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:37:43,917 INFO [finetune.py:976] (4/7) Epoch 4, batch 600, loss[loss=0.3005, simple_loss=0.3466, pruned_loss=0.1272, over 4746.00 frames. ], tot_loss[loss=0.2273, simple_loss=0.2828, pruned_loss=0.08589, over 907860.86 frames. ], batch size: 54, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:37:49,323 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=17791.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:38:23,101 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.336e+02 1.937e+02 2.249e+02 2.763e+02 6.989e+02, threshold=4.497e+02, percent-clipped=2.0 2023-04-26 14:38:34,867 INFO [finetune.py:976] (4/7) Epoch 4, batch 650, loss[loss=0.2322, simple_loss=0.2868, pruned_loss=0.08881, over 4902.00 frames. ], tot_loss[loss=0.2318, simple_loss=0.2878, pruned_loss=0.0879, over 920240.53 frames. ], batch size: 36, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:38:52,791 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=17845.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:39:11,530 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6348, 1.6589, 1.7458, 1.9843, 1.9671, 1.5341, 1.1920, 1.7170], device='cuda:4'), covar=tensor([0.0956, 0.1125, 0.0707, 0.0631, 0.0654, 0.1034, 0.1154, 0.0707], device='cuda:4'), in_proj_covar=tensor([0.0210, 0.0212, 0.0189, 0.0183, 0.0184, 0.0200, 0.0174, 0.0194], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 14:39:22,791 INFO [finetune.py:976] (4/7) Epoch 4, batch 700, loss[loss=0.2222, simple_loss=0.2906, pruned_loss=0.07692, over 4883.00 frames. ], tot_loss[loss=0.2323, simple_loss=0.289, pruned_loss=0.08782, over 928389.31 frames. ], batch size: 32, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:39:28,941 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=17893.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:40:07,834 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.332e+02 1.898e+02 2.217e+02 2.664e+02 7.094e+02, threshold=4.434e+02, percent-clipped=3.0 2023-04-26 14:40:17,313 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.8109, 2.8537, 2.2222, 3.2516, 2.9060, 2.8158, 1.1965, 2.7547], device='cuda:4'), covar=tensor([0.2220, 0.1705, 0.3370, 0.2961, 0.2622, 0.2239, 0.5778, 0.2754], device='cuda:4'), in_proj_covar=tensor([0.0250, 0.0224, 0.0262, 0.0316, 0.0309, 0.0260, 0.0276, 0.0279], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 14:40:19,659 INFO [finetune.py:976] (4/7) Epoch 4, batch 750, loss[loss=0.26, simple_loss=0.3126, pruned_loss=0.1037, over 4899.00 frames. ], tot_loss[loss=0.2317, simple_loss=0.2892, pruned_loss=0.08705, over 935757.92 frames. ], batch size: 46, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:40:50,352 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.84 vs. limit=2.0 2023-04-26 14:40:53,834 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=17960.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:41:24,416 INFO [finetune.py:976] (4/7) Epoch 4, batch 800, loss[loss=0.2394, simple_loss=0.2735, pruned_loss=0.1026, over 3957.00 frames. ], tot_loss[loss=0.2304, simple_loss=0.2883, pruned_loss=0.08627, over 940112.20 frames. ], batch size: 17, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:41:48,941 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=18018.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:41:49,030 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-26 14:41:51,312 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=18021.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:41:52,027 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.78 vs. limit=5.0 2023-04-26 14:41:52,409 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.136e+02 1.866e+02 2.231e+02 2.895e+02 6.246e+02, threshold=4.462e+02, percent-clipped=2.0 2023-04-26 14:41:57,864 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1092, 1.3190, 4.8022, 4.4642, 4.2055, 4.4044, 4.2170, 4.2077], device='cuda:4'), covar=tensor([0.6128, 0.6046, 0.0888, 0.1650, 0.1103, 0.1332, 0.1729, 0.1429], device='cuda:4'), in_proj_covar=tensor([0.0321, 0.0312, 0.0430, 0.0432, 0.0369, 0.0421, 0.0329, 0.0388], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-26 14:41:58,983 INFO [finetune.py:976] (4/7) Epoch 4, batch 850, loss[loss=0.2183, simple_loss=0.2715, pruned_loss=0.08258, over 4864.00 frames. ], tot_loss[loss=0.2275, simple_loss=0.2853, pruned_loss=0.08485, over 944751.79 frames. ], batch size: 34, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:42:02,669 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=18039.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 14:42:13,539 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=18056.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:42:20,540 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=18066.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:42:22,433 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5083, 3.2498, 2.7705, 2.8526, 2.4136, 2.7817, 2.8490, 2.1577], device='cuda:4'), covar=tensor([0.2626, 0.1666, 0.0944, 0.1699, 0.3287, 0.1679, 0.2394, 0.3653], device='cuda:4'), in_proj_covar=tensor([0.0317, 0.0336, 0.0246, 0.0309, 0.0323, 0.0288, 0.0277, 0.0301], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 14:42:26,239 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-04-26 14:42:32,277 INFO [finetune.py:976] (4/7) Epoch 4, batch 900, loss[loss=0.2228, simple_loss=0.2675, pruned_loss=0.08906, over 4840.00 frames. ], tot_loss[loss=0.224, simple_loss=0.2813, pruned_loss=0.08332, over 946952.41 frames. ], batch size: 33, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:42:34,140 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=18086.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:42:34,730 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=18087.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 14:42:58,116 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.220e+02 1.952e+02 2.266e+02 2.638e+02 7.997e+02, threshold=4.531e+02, percent-clipped=4.0 2023-04-26 14:43:05,177 INFO [finetune.py:976] (4/7) Epoch 4, batch 950, loss[loss=0.2649, simple_loss=0.3252, pruned_loss=0.1023, over 4814.00 frames. ], tot_loss[loss=0.2243, simple_loss=0.2812, pruned_loss=0.08372, over 949482.87 frames. ], batch size: 41, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:43:13,673 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.77 vs. limit=2.0 2023-04-26 14:43:44,150 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=18174.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:43:45,261 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=18175.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:43:53,135 INFO [finetune.py:976] (4/7) Epoch 4, batch 1000, loss[loss=0.2507, simple_loss=0.3046, pruned_loss=0.09834, over 4816.00 frames. ], tot_loss[loss=0.2282, simple_loss=0.2855, pruned_loss=0.08542, over 951639.78 frames. ], batch size: 38, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:44:18,523 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9202, 1.4359, 1.4926, 1.6400, 2.1859, 1.7971, 1.5057, 1.4302], device='cuda:4'), covar=tensor([0.1719, 0.1913, 0.2469, 0.1511, 0.0898, 0.1758, 0.2322, 0.2155], device='cuda:4'), in_proj_covar=tensor([0.0316, 0.0336, 0.0350, 0.0309, 0.0346, 0.0354, 0.0315, 0.0350], device='cuda:4'), out_proj_covar=tensor([6.8310e-05, 7.2082e-05, 7.5908e-05, 6.4724e-05, 7.3571e-05, 7.7196e-05, 6.8544e-05, 7.5689e-05], device='cuda:4') 2023-04-26 14:44:30,497 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.196e+02 1.905e+02 2.183e+02 2.540e+02 5.721e+02, threshold=4.367e+02, percent-clipped=1.0 2023-04-26 14:44:38,565 INFO [finetune.py:976] (4/7) Epoch 4, batch 1050, loss[loss=0.1823, simple_loss=0.2271, pruned_loss=0.06874, over 4236.00 frames. ], tot_loss[loss=0.2274, simple_loss=0.2858, pruned_loss=0.08452, over 949921.07 frames. ], batch size: 18, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:44:39,841 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=18235.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:44:39,865 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7024, 2.5434, 1.6486, 1.5636, 1.2263, 1.2630, 1.6751, 1.1752], device='cuda:4'), covar=tensor([0.2127, 0.1580, 0.2058, 0.2548, 0.3115, 0.2474, 0.1563, 0.2562], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0222, 0.0185, 0.0211, 0.0223, 0.0191, 0.0179, 0.0201], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 14:44:40,481 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=18236.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:45:12,453 INFO [finetune.py:976] (4/7) Epoch 4, batch 1100, loss[loss=0.241, simple_loss=0.2964, pruned_loss=0.09285, over 4878.00 frames. ], tot_loss[loss=0.2286, simple_loss=0.2869, pruned_loss=0.08521, over 950802.31 frames. ], batch size: 32, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:45:50,512 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=18316.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:45:59,380 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6709, 1.8468, 1.1924, 1.3496, 2.2148, 1.5707, 1.4700, 1.5211], device='cuda:4'), covar=tensor([0.0542, 0.0396, 0.0359, 0.0596, 0.0232, 0.0587, 0.0567, 0.0641], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0027, 0.0025, 0.0031, 0.0022, 0.0031, 0.0030, 0.0032], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0045, 0.0038, 0.0049, 0.0038, 0.0048, 0.0047, 0.0050], device='cuda:4') 2023-04-26 14:46:00,507 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.371e+02 2.039e+02 2.386e+02 2.843e+02 4.542e+02, threshold=4.773e+02, percent-clipped=1.0 2023-04-26 14:46:04,252 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0912, 2.8019, 2.3756, 2.3794, 2.2548, 2.3560, 2.5144, 1.7809], device='cuda:4'), covar=tensor([0.3022, 0.1785, 0.1151, 0.2010, 0.3362, 0.1847, 0.2531, 0.3838], device='cuda:4'), in_proj_covar=tensor([0.0320, 0.0339, 0.0247, 0.0312, 0.0326, 0.0290, 0.0279, 0.0303], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 14:46:08,166 INFO [finetune.py:976] (4/7) Epoch 4, batch 1150, loss[loss=0.2088, simple_loss=0.2723, pruned_loss=0.07262, over 4899.00 frames. ], tot_loss[loss=0.2285, simple_loss=0.2874, pruned_loss=0.08484, over 952339.05 frames. ], batch size: 36, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:46:23,818 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=18356.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:46:40,894 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4484, 2.3492, 2.6004, 3.0435, 2.6658, 2.2502, 2.0205, 2.4109], device='cuda:4'), covar=tensor([0.1038, 0.1036, 0.0593, 0.0599, 0.0703, 0.1171, 0.1123, 0.0670], device='cuda:4'), in_proj_covar=tensor([0.0209, 0.0210, 0.0187, 0.0183, 0.0183, 0.0199, 0.0173, 0.0193], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 14:46:52,607 INFO [finetune.py:976] (4/7) Epoch 4, batch 1200, loss[loss=0.1904, simple_loss=0.2599, pruned_loss=0.06047, over 4787.00 frames. ], tot_loss[loss=0.2267, simple_loss=0.285, pruned_loss=0.08415, over 952259.83 frames. ], batch size: 29, lr: 3.97e-03, grad_scale: 16.0 2023-04-26 14:46:55,534 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=18386.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:47:18,678 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=18404.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:47:36,009 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.337e+02 1.888e+02 2.199e+02 2.566e+02 6.503e+02, threshold=4.398e+02, percent-clipped=1.0 2023-04-26 14:47:38,590 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7244, 1.8135, 0.8249, 1.3593, 1.9099, 1.6070, 1.5012, 1.5491], device='cuda:4'), covar=tensor([0.0561, 0.0405, 0.0416, 0.0612, 0.0271, 0.0575, 0.0556, 0.0646], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0027, 0.0025, 0.0031, 0.0022, 0.0031, 0.0030, 0.0032], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0045, 0.0038, 0.0049, 0.0038, 0.0048, 0.0047, 0.0050], device='cuda:4') 2023-04-26 14:47:42,651 INFO [finetune.py:976] (4/7) Epoch 4, batch 1250, loss[loss=0.2424, simple_loss=0.297, pruned_loss=0.09388, over 4913.00 frames. ], tot_loss[loss=0.2242, simple_loss=0.2825, pruned_loss=0.08301, over 952699.87 frames. ], batch size: 36, lr: 3.97e-03, grad_scale: 32.0 2023-04-26 14:47:43,322 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=18434.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:47:59,752 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4762, 2.3414, 2.6577, 3.0789, 2.6896, 2.2261, 1.8564, 2.3688], device='cuda:4'), covar=tensor([0.0927, 0.0986, 0.0548, 0.0581, 0.0694, 0.1034, 0.1149, 0.0672], device='cuda:4'), in_proj_covar=tensor([0.0209, 0.0210, 0.0187, 0.0183, 0.0183, 0.0199, 0.0173, 0.0193], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 14:48:02,189 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1721, 1.7252, 1.5105, 1.9012, 1.7204, 2.1719, 1.5200, 3.7928], device='cuda:4'), covar=tensor([0.0681, 0.0721, 0.0798, 0.1208, 0.0636, 0.0527, 0.0721, 0.0150], device='cuda:4'), in_proj_covar=tensor([0.0040, 0.0040, 0.0042, 0.0046, 0.0042, 0.0041, 0.0041, 0.0064], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0014, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 14:48:06,609 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-26 14:48:15,942 INFO [finetune.py:976] (4/7) Epoch 4, batch 1300, loss[loss=0.2509, simple_loss=0.2993, pruned_loss=0.1013, over 4826.00 frames. ], tot_loss[loss=0.2216, simple_loss=0.2792, pruned_loss=0.08205, over 953829.06 frames. ], batch size: 33, lr: 3.97e-03, grad_scale: 32.0 2023-04-26 14:48:21,727 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=18490.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:48:34,028 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=18508.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 14:48:43,005 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.221e+02 1.895e+02 2.151e+02 2.623e+02 4.474e+02, threshold=4.301e+02, percent-clipped=1.0 2023-04-26 14:48:47,802 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=18530.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:48:48,427 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=18531.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:48:49,576 INFO [finetune.py:976] (4/7) Epoch 4, batch 1350, loss[loss=0.1866, simple_loss=0.2442, pruned_loss=0.06445, over 4833.00 frames. ], tot_loss[loss=0.2231, simple_loss=0.28, pruned_loss=0.08314, over 953873.16 frames. ], batch size: 30, lr: 3.97e-03, grad_scale: 32.0 2023-04-26 14:49:08,074 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=18551.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:49:31,998 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=18569.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 14:49:45,979 INFO [finetune.py:976] (4/7) Epoch 4, batch 1400, loss[loss=0.2265, simple_loss=0.2888, pruned_loss=0.08212, over 4908.00 frames. ], tot_loss[loss=0.2267, simple_loss=0.2843, pruned_loss=0.08456, over 955724.35 frames. ], batch size: 36, lr: 3.97e-03, grad_scale: 32.0 2023-04-26 14:50:09,108 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=18616.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:50:13,227 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.222e+02 1.882e+02 2.176e+02 2.822e+02 8.396e+02, threshold=4.353e+02, percent-clipped=2.0 2023-04-26 14:50:16,014 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.75 vs. limit=2.0 2023-04-26 14:50:19,722 INFO [finetune.py:976] (4/7) Epoch 4, batch 1450, loss[loss=0.2154, simple_loss=0.2761, pruned_loss=0.0774, over 4822.00 frames. ], tot_loss[loss=0.2279, simple_loss=0.2859, pruned_loss=0.08497, over 957587.39 frames. ], batch size: 33, lr: 3.97e-03, grad_scale: 32.0 2023-04-26 14:50:40,787 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=18664.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:50:52,628 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=18674.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:51:01,637 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=5.02 vs. limit=5.0 2023-04-26 14:51:03,066 INFO [finetune.py:976] (4/7) Epoch 4, batch 1500, loss[loss=0.2913, simple_loss=0.3325, pruned_loss=0.1251, over 4902.00 frames. ], tot_loss[loss=0.2283, simple_loss=0.286, pruned_loss=0.08528, over 956117.49 frames. ], batch size: 37, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 14:51:23,300 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8388, 1.8431, 2.0581, 1.9434, 1.9147, 1.7125, 1.8476, 1.8188], device='cuda:4'), covar=tensor([2.0904, 2.4153, 3.1261, 3.6378, 2.0547, 3.8363, 3.8158, 2.9951], device='cuda:4'), in_proj_covar=tensor([0.0434, 0.0480, 0.0567, 0.0585, 0.0464, 0.0493, 0.0506, 0.0514], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 14:51:24,318 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0830, 3.8725, 1.3109, 2.3398, 2.5511, 2.8646, 2.4602, 1.5826], device='cuda:4'), covar=tensor([0.1104, 0.1004, 0.1985, 0.1103, 0.0787, 0.0973, 0.1212, 0.1594], device='cuda:4'), in_proj_covar=tensor([0.0123, 0.0265, 0.0149, 0.0130, 0.0140, 0.0162, 0.0127, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 14:51:35,766 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.7294, 3.6409, 2.8491, 4.2893, 3.6569, 3.6830, 1.5731, 3.7029], device='cuda:4'), covar=tensor([0.1557, 0.1341, 0.3844, 0.1255, 0.2794, 0.2032, 0.5499, 0.2146], device='cuda:4'), in_proj_covar=tensor([0.0253, 0.0225, 0.0264, 0.0318, 0.0312, 0.0262, 0.0278, 0.0281], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 14:51:55,742 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6902, 2.2072, 1.5896, 1.4679, 1.2885, 1.2996, 1.5427, 1.2200], device='cuda:4'), covar=tensor([0.1991, 0.1566, 0.1948, 0.2364, 0.3027, 0.2325, 0.1563, 0.2500], device='cuda:4'), in_proj_covar=tensor([0.0193, 0.0221, 0.0183, 0.0209, 0.0220, 0.0189, 0.0176, 0.0199], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 14:51:57,996 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.134e+02 1.902e+02 2.274e+02 2.776e+02 4.497e+02, threshold=4.548e+02, percent-clipped=1.0 2023-04-26 14:52:09,751 INFO [finetune.py:976] (4/7) Epoch 4, batch 1550, loss[loss=0.2517, simple_loss=0.2966, pruned_loss=0.1034, over 4878.00 frames. ], tot_loss[loss=0.2286, simple_loss=0.2865, pruned_loss=0.0853, over 955402.68 frames. ], batch size: 35, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 14:52:11,084 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=18735.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:52:50,425 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2493, 3.6377, 1.0045, 1.9104, 1.8269, 2.5880, 2.0357, 1.0956], device='cuda:4'), covar=tensor([0.1552, 0.0770, 0.2042, 0.1283, 0.1155, 0.1020, 0.1544, 0.2116], device='cuda:4'), in_proj_covar=tensor([0.0123, 0.0265, 0.0149, 0.0129, 0.0140, 0.0162, 0.0127, 0.0129], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 14:53:12,332 INFO [finetune.py:976] (4/7) Epoch 4, batch 1600, loss[loss=0.2409, simple_loss=0.2854, pruned_loss=0.09818, over 4154.00 frames. ], tot_loss[loss=0.2267, simple_loss=0.2843, pruned_loss=0.08454, over 956950.87 frames. ], batch size: 65, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 14:53:44,154 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.258e+02 2.049e+02 2.445e+02 2.756e+02 4.210e+02, threshold=4.889e+02, percent-clipped=0.0 2023-04-26 14:53:48,512 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=18830.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:53:49,095 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=18831.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:53:50,214 INFO [finetune.py:976] (4/7) Epoch 4, batch 1650, loss[loss=0.1913, simple_loss=0.2565, pruned_loss=0.06305, over 4793.00 frames. ], tot_loss[loss=0.2233, simple_loss=0.2807, pruned_loss=0.08296, over 956992.79 frames. ], batch size: 29, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 14:53:58,586 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=18846.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:54:11,008 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=18864.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 14:54:20,475 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=18878.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:54:21,069 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=18879.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:54:23,387 INFO [finetune.py:976] (4/7) Epoch 4, batch 1700, loss[loss=0.2039, simple_loss=0.2608, pruned_loss=0.07354, over 4838.00 frames. ], tot_loss[loss=0.2226, simple_loss=0.2793, pruned_loss=0.08301, over 956753.08 frames. ], batch size: 33, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 14:55:05,869 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-04-26 14:55:17,073 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.323e+02 1.925e+02 2.287e+02 2.733e+02 5.363e+02, threshold=4.574e+02, percent-clipped=1.0 2023-04-26 14:55:21,934 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=18930.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:55:23,643 INFO [finetune.py:976] (4/7) Epoch 4, batch 1750, loss[loss=0.2563, simple_loss=0.3086, pruned_loss=0.102, over 4908.00 frames. ], tot_loss[loss=0.2254, simple_loss=0.2821, pruned_loss=0.08436, over 955984.25 frames. ], batch size: 36, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 14:55:30,512 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0022, 1.4899, 1.9585, 2.2430, 1.8158, 1.4140, 1.0820, 1.5922], device='cuda:4'), covar=tensor([0.4736, 0.5517, 0.2385, 0.4388, 0.4828, 0.4184, 0.7054, 0.4231], device='cuda:4'), in_proj_covar=tensor([0.0270, 0.0264, 0.0220, 0.0333, 0.0223, 0.0230, 0.0251, 0.0200], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 14:55:53,837 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-26 14:55:56,822 INFO [finetune.py:976] (4/7) Epoch 4, batch 1800, loss[loss=0.1694, simple_loss=0.2351, pruned_loss=0.05184, over 4694.00 frames. ], tot_loss[loss=0.2274, simple_loss=0.2849, pruned_loss=0.08498, over 955788.73 frames. ], batch size: 23, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 14:56:08,187 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=18991.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:56:44,254 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.520e+02 1.971e+02 2.444e+02 3.087e+02 4.882e+02, threshold=4.887e+02, percent-clipped=3.0 2023-04-26 14:56:49,159 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=19030.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:56:50,956 INFO [finetune.py:976] (4/7) Epoch 4, batch 1850, loss[loss=0.2073, simple_loss=0.2517, pruned_loss=0.08151, over 4343.00 frames. ], tot_loss[loss=0.2286, simple_loss=0.2862, pruned_loss=0.08545, over 955045.12 frames. ], batch size: 19, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 14:57:16,501 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3141, 1.5891, 1.5010, 1.6678, 1.5245, 1.7433, 1.6702, 1.5691], device='cuda:4'), covar=tensor([0.9624, 1.4651, 1.2814, 1.1039, 1.3039, 1.9516, 1.4895, 1.3894], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0401, 0.0320, 0.0325, 0.0350, 0.0411, 0.0385, 0.0341], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 14:57:24,083 INFO [finetune.py:976] (4/7) Epoch 4, batch 1900, loss[loss=0.2076, simple_loss=0.2893, pruned_loss=0.0629, over 4744.00 frames. ], tot_loss[loss=0.2304, simple_loss=0.2884, pruned_loss=0.0862, over 957664.28 frames. ], batch size: 54, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 14:57:50,674 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.323e+02 1.880e+02 2.172e+02 2.548e+02 4.187e+02, threshold=4.344e+02, percent-clipped=0.0 2023-04-26 14:57:54,927 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=19129.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 14:57:57,174 INFO [finetune.py:976] (4/7) Epoch 4, batch 1950, loss[loss=0.1533, simple_loss=0.2149, pruned_loss=0.04583, over 4723.00 frames. ], tot_loss[loss=0.2265, simple_loss=0.2847, pruned_loss=0.08416, over 959105.98 frames. ], batch size: 23, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 14:58:05,534 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=19146.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:58:17,029 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=19164.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 14:58:30,346 INFO [finetune.py:976] (4/7) Epoch 4, batch 2000, loss[loss=0.1947, simple_loss=0.2532, pruned_loss=0.0681, over 4901.00 frames. ], tot_loss[loss=0.2226, simple_loss=0.2807, pruned_loss=0.08227, over 957395.55 frames. ], batch size: 46, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 14:58:40,467 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=19190.0, num_to_drop=1, layers_to_drop={3} 2023-04-26 14:58:48,412 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=19194.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 14:59:11,327 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=19212.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 14:59:24,206 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.131e+02 1.742e+02 2.062e+02 2.571e+02 4.374e+02, threshold=4.123e+02, percent-clipped=1.0 2023-04-26 14:59:25,580 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6774, 2.1839, 1.6529, 1.4037, 1.2806, 1.2924, 1.5812, 1.2456], device='cuda:4'), covar=tensor([0.2312, 0.1959, 0.2498, 0.2847, 0.3641, 0.2796, 0.1797, 0.2915], device='cuda:4'), in_proj_covar=tensor([0.0193, 0.0220, 0.0182, 0.0209, 0.0219, 0.0188, 0.0176, 0.0199], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 14:59:36,723 INFO [finetune.py:976] (4/7) Epoch 4, batch 2050, loss[loss=0.2477, simple_loss=0.2744, pruned_loss=0.1105, over 4157.00 frames. ], tot_loss[loss=0.2191, simple_loss=0.2768, pruned_loss=0.08071, over 955226.16 frames. ], batch size: 18, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 14:59:51,103 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.4194, 1.2605, 1.3638, 1.0553, 1.3455, 1.1916, 1.6868, 1.1716], device='cuda:4'), covar=tensor([0.3510, 0.1728, 0.5357, 0.2626, 0.1608, 0.2141, 0.1661, 0.5075], device='cuda:4'), in_proj_covar=tensor([0.0353, 0.0356, 0.0442, 0.0373, 0.0405, 0.0382, 0.0399, 0.0420], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 14:59:59,542 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3796, 1.3041, 0.6456, 1.1970, 1.1868, 1.2739, 1.2151, 1.2376], device='cuda:4'), covar=tensor([0.0541, 0.0304, 0.0459, 0.0520, 0.0358, 0.0553, 0.0541, 0.0530], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0027, 0.0025, 0.0031, 0.0022, 0.0031, 0.0030, 0.0032], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0045, 0.0038, 0.0049, 0.0038, 0.0048, 0.0047, 0.0050], device='cuda:4') 2023-04-26 15:00:15,256 INFO [finetune.py:976] (4/7) Epoch 4, batch 2100, loss[loss=0.1891, simple_loss=0.2599, pruned_loss=0.0591, over 4753.00 frames. ], tot_loss[loss=0.2192, simple_loss=0.2764, pruned_loss=0.08099, over 952352.28 frames. ], batch size: 27, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 15:00:17,153 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=19286.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:01:01,582 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.066e+02 1.821e+02 2.202e+02 2.671e+02 7.978e+02, threshold=4.404e+02, percent-clipped=3.0 2023-04-26 15:01:07,001 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=19330.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:01:14,462 INFO [finetune.py:976] (4/7) Epoch 4, batch 2150, loss[loss=0.1985, simple_loss=0.273, pruned_loss=0.06198, over 4783.00 frames. ], tot_loss[loss=0.2215, simple_loss=0.2797, pruned_loss=0.08165, over 951813.01 frames. ], batch size: 29, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 15:02:07,705 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=19378.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:02:16,425 INFO [finetune.py:976] (4/7) Epoch 4, batch 2200, loss[loss=0.2287, simple_loss=0.287, pruned_loss=0.08519, over 4779.00 frames. ], tot_loss[loss=0.2246, simple_loss=0.283, pruned_loss=0.08312, over 952698.13 frames. ], batch size: 26, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 15:02:30,476 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=19394.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:02:55,168 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9932, 1.5003, 1.9814, 2.2186, 1.8069, 1.4751, 1.2208, 1.7318], device='cuda:4'), covar=tensor([0.4858, 0.5191, 0.2448, 0.3976, 0.4849, 0.3981, 0.6180, 0.3941], device='cuda:4'), in_proj_covar=tensor([0.0275, 0.0269, 0.0223, 0.0339, 0.0226, 0.0233, 0.0253, 0.0203], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 15:03:06,439 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.252e+02 1.815e+02 2.166e+02 2.683e+02 4.330e+02, threshold=4.331e+02, percent-clipped=0.0 2023-04-26 15:03:18,525 INFO [finetune.py:976] (4/7) Epoch 4, batch 2250, loss[loss=0.2302, simple_loss=0.2968, pruned_loss=0.08182, over 4808.00 frames. ], tot_loss[loss=0.2268, simple_loss=0.2852, pruned_loss=0.08424, over 951824.85 frames. ], batch size: 45, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 15:03:51,633 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=19455.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:04:21,503 INFO [finetune.py:976] (4/7) Epoch 4, batch 2300, loss[loss=0.1901, simple_loss=0.2677, pruned_loss=0.05621, over 4748.00 frames. ], tot_loss[loss=0.2247, simple_loss=0.2837, pruned_loss=0.08286, over 953316.88 frames. ], batch size: 27, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 15:04:23,292 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=19485.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 15:04:51,782 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6169, 1.2404, 4.2746, 3.9857, 3.7397, 4.0587, 3.9631, 3.7634], device='cuda:4'), covar=tensor([0.6446, 0.5962, 0.1025, 0.1660, 0.1174, 0.1460, 0.1226, 0.1637], device='cuda:4'), in_proj_covar=tensor([0.0319, 0.0308, 0.0426, 0.0431, 0.0366, 0.0418, 0.0328, 0.0386], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-26 15:04:53,662 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4695, 2.3556, 1.9748, 2.2277, 2.5432, 1.9790, 3.3277, 1.7311], device='cuda:4'), covar=tensor([0.4453, 0.2527, 0.5528, 0.4018, 0.2144, 0.3023, 0.1925, 0.5126], device='cuda:4'), in_proj_covar=tensor([0.0356, 0.0359, 0.0447, 0.0375, 0.0406, 0.0386, 0.0402, 0.0425], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 15:04:58,955 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.086e+02 1.751e+02 2.123e+02 2.573e+02 6.035e+02, threshold=4.246e+02, percent-clipped=1.0 2023-04-26 15:05:05,527 INFO [finetune.py:976] (4/7) Epoch 4, batch 2350, loss[loss=0.2038, simple_loss=0.2741, pruned_loss=0.06679, over 4765.00 frames. ], tot_loss[loss=0.2225, simple_loss=0.2812, pruned_loss=0.08194, over 954098.83 frames. ], batch size: 27, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 15:05:31,200 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-04-26 15:05:39,107 INFO [finetune.py:976] (4/7) Epoch 4, batch 2400, loss[loss=0.1859, simple_loss=0.2307, pruned_loss=0.07053, over 3967.00 frames. ], tot_loss[loss=0.2197, simple_loss=0.2778, pruned_loss=0.08083, over 954872.54 frames. ], batch size: 17, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 15:05:41,021 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=19586.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:06:03,660 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=19618.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:06:06,614 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.143e+02 1.798e+02 2.118e+02 2.540e+02 5.281e+02, threshold=4.235e+02, percent-clipped=1.0 2023-04-26 15:06:09,283 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.66 vs. limit=2.0 2023-04-26 15:06:12,782 INFO [finetune.py:976] (4/7) Epoch 4, batch 2450, loss[loss=0.2474, simple_loss=0.2941, pruned_loss=0.1004, over 4794.00 frames. ], tot_loss[loss=0.2166, simple_loss=0.2744, pruned_loss=0.07938, over 955005.28 frames. ], batch size: 29, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 15:06:13,453 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=19634.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:06:18,923 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6901, 2.0549, 1.7705, 1.9504, 1.6181, 1.6609, 1.7141, 1.3906], device='cuda:4'), covar=tensor([0.1929, 0.1392, 0.0963, 0.1261, 0.3194, 0.1418, 0.1945, 0.2727], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0327, 0.0239, 0.0302, 0.0316, 0.0281, 0.0271, 0.0292], device='cuda:4'), out_proj_covar=tensor([1.2597e-04, 1.3338e-04, 9.7335e-05, 1.2168e-04, 1.3056e-04, 1.1339e-04, 1.1211e-04, 1.1824e-04], device='cuda:4') 2023-04-26 15:06:34,608 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.2718, 4.2455, 3.0450, 4.9380, 4.3337, 4.2494, 1.9097, 4.2098], device='cuda:4'), covar=tensor([0.1664, 0.0946, 0.3437, 0.1101, 0.3212, 0.1815, 0.6012, 0.2153], device='cuda:4'), in_proj_covar=tensor([0.0253, 0.0225, 0.0262, 0.0317, 0.0311, 0.0261, 0.0278, 0.0281], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 15:06:44,233 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=19679.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:06:46,565 INFO [finetune.py:976] (4/7) Epoch 4, batch 2500, loss[loss=0.1637, simple_loss=0.2348, pruned_loss=0.04636, over 4711.00 frames. ], tot_loss[loss=0.22, simple_loss=0.2772, pruned_loss=0.08138, over 954293.86 frames. ], batch size: 23, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 15:07:07,671 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5385, 1.1991, 1.5139, 1.7838, 1.6510, 1.4943, 1.5350, 1.6030], device='cuda:4'), covar=tensor([1.5162, 1.9506, 2.2791, 2.5356, 1.6278, 2.4634, 2.4809, 1.9163], device='cuda:4'), in_proj_covar=tensor([0.0432, 0.0477, 0.0563, 0.0582, 0.0463, 0.0491, 0.0502, 0.0511], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 15:07:19,924 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2868, 2.2067, 2.0958, 2.1111, 2.5129, 1.9833, 3.0369, 1.6992], device='cuda:4'), covar=tensor([0.3854, 0.1734, 0.4425, 0.2914, 0.1585, 0.2472, 0.1181, 0.4478], device='cuda:4'), in_proj_covar=tensor([0.0354, 0.0358, 0.0442, 0.0373, 0.0403, 0.0383, 0.0399, 0.0420], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 15:07:21,784 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5625, 1.4300, 1.8654, 1.8724, 1.4410, 1.1860, 1.5949, 1.0397], device='cuda:4'), covar=tensor([0.0954, 0.1193, 0.0617, 0.1082, 0.1216, 0.1570, 0.1005, 0.1092], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0079, 0.0077, 0.0071, 0.0083, 0.0098, 0.0087, 0.0080], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-26 15:07:31,980 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.364e+02 1.985e+02 2.362e+02 2.892e+02 4.639e+02, threshold=4.724e+02, percent-clipped=3.0 2023-04-26 15:07:43,942 INFO [finetune.py:976] (4/7) Epoch 4, batch 2550, loss[loss=0.2231, simple_loss=0.2831, pruned_loss=0.0815, over 4879.00 frames. ], tot_loss[loss=0.2219, simple_loss=0.2801, pruned_loss=0.08181, over 954664.78 frames. ], batch size: 34, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 15:08:07,137 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=19750.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:08:39,161 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3663, 1.3524, 3.7455, 3.4931, 3.3270, 3.4562, 3.4652, 3.3388], device='cuda:4'), covar=tensor([0.6860, 0.5301, 0.1075, 0.1700, 0.1185, 0.1829, 0.2605, 0.1486], device='cuda:4'), in_proj_covar=tensor([0.0319, 0.0307, 0.0427, 0.0433, 0.0366, 0.0419, 0.0329, 0.0386], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-26 15:08:50,442 INFO [finetune.py:976] (4/7) Epoch 4, batch 2600, loss[loss=0.2001, simple_loss=0.2576, pruned_loss=0.07131, over 4793.00 frames. ], tot_loss[loss=0.2239, simple_loss=0.282, pruned_loss=0.08288, over 954180.45 frames. ], batch size: 29, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 15:08:57,321 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=19785.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 15:09:23,391 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.319e+02 1.754e+02 2.126e+02 2.494e+02 3.908e+02, threshold=4.253e+02, percent-clipped=0.0 2023-04-26 15:09:28,489 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.80 vs. limit=2.0 2023-04-26 15:09:29,904 INFO [finetune.py:976] (4/7) Epoch 4, batch 2650, loss[loss=0.175, simple_loss=0.2406, pruned_loss=0.0547, over 4830.00 frames. ], tot_loss[loss=0.2264, simple_loss=0.2841, pruned_loss=0.08433, over 954859.17 frames. ], batch size: 49, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 15:09:29,968 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=19833.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 15:10:35,942 INFO [finetune.py:976] (4/7) Epoch 4, batch 2700, loss[loss=0.1709, simple_loss=0.2332, pruned_loss=0.05431, over 4780.00 frames. ], tot_loss[loss=0.2233, simple_loss=0.2815, pruned_loss=0.08255, over 956322.78 frames. ], batch size: 51, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 15:11:13,668 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9022, 1.3742, 1.3537, 1.5577, 2.1326, 1.7262, 1.4825, 1.3523], device='cuda:4'), covar=tensor([0.2062, 0.1850, 0.2346, 0.1473, 0.0912, 0.1866, 0.2292, 0.2146], device='cuda:4'), in_proj_covar=tensor([0.0316, 0.0338, 0.0352, 0.0311, 0.0351, 0.0353, 0.0316, 0.0355], device='cuda:4'), out_proj_covar=tensor([6.8369e-05, 7.2432e-05, 7.6195e-05, 6.5029e-05, 7.4632e-05, 7.7065e-05, 6.8701e-05, 7.6609e-05], device='cuda:4') 2023-04-26 15:11:15,311 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.121e+02 1.807e+02 2.180e+02 2.734e+02 4.615e+02, threshold=4.361e+02, percent-clipped=4.0 2023-04-26 15:11:21,338 INFO [finetune.py:976] (4/7) Epoch 4, batch 2750, loss[loss=0.2356, simple_loss=0.297, pruned_loss=0.08709, over 4814.00 frames. ], tot_loss[loss=0.2214, simple_loss=0.2791, pruned_loss=0.08184, over 955690.34 frames. ], batch size: 38, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 15:11:28,130 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-04-26 15:11:36,443 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4743, 1.2558, 1.6429, 1.6008, 1.3916, 1.2408, 1.3737, 0.8909], device='cuda:4'), covar=tensor([0.0722, 0.1071, 0.0720, 0.0897, 0.0977, 0.1549, 0.0843, 0.1174], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0078, 0.0077, 0.0070, 0.0082, 0.0097, 0.0086, 0.0080], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0004, 0.0003], device='cuda:4') 2023-04-26 15:11:49,857 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=19974.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:11:55,248 INFO [finetune.py:976] (4/7) Epoch 4, batch 2800, loss[loss=0.1831, simple_loss=0.2408, pruned_loss=0.06277, over 4383.00 frames. ], tot_loss[loss=0.2188, simple_loss=0.276, pruned_loss=0.08082, over 957992.31 frames. ], batch size: 19, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 15:12:15,092 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8329, 1.3350, 1.6622, 1.5332, 1.5281, 1.2794, 0.6532, 1.2003], device='cuda:4'), covar=tensor([0.4881, 0.5651, 0.2698, 0.4021, 0.4819, 0.4000, 0.6727, 0.4274], device='cuda:4'), in_proj_covar=tensor([0.0272, 0.0265, 0.0221, 0.0335, 0.0224, 0.0231, 0.0251, 0.0201], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 15:12:21,324 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=20020.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:12:23,467 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.089e+02 1.778e+02 2.116e+02 2.453e+02 4.080e+02, threshold=4.233e+02, percent-clipped=0.0 2023-04-26 15:12:30,049 INFO [finetune.py:976] (4/7) Epoch 4, batch 2850, loss[loss=0.1708, simple_loss=0.2401, pruned_loss=0.05074, over 4811.00 frames. ], tot_loss[loss=0.2187, simple_loss=0.2758, pruned_loss=0.08076, over 956214.15 frames. ], batch size: 38, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 15:12:40,890 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=20050.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:13:02,650 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=20081.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:13:03,755 INFO [finetune.py:976] (4/7) Epoch 4, batch 2900, loss[loss=0.2468, simple_loss=0.3186, pruned_loss=0.08748, over 4806.00 frames. ], tot_loss[loss=0.2211, simple_loss=0.2787, pruned_loss=0.08168, over 955953.80 frames. ], batch size: 41, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 15:13:07,522 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=20089.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:13:13,381 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=20098.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:13:22,364 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=20112.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:13:29,803 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.335e+02 1.818e+02 2.275e+02 2.799e+02 4.560e+02, threshold=4.551e+02, percent-clipped=2.0 2023-04-26 15:13:47,509 INFO [finetune.py:976] (4/7) Epoch 4, batch 2950, loss[loss=0.2575, simple_loss=0.3034, pruned_loss=0.1059, over 4905.00 frames. ], tot_loss[loss=0.2226, simple_loss=0.2813, pruned_loss=0.08195, over 955655.19 frames. ], batch size: 37, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 15:14:09,944 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=20150.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:14:35,357 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=20173.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:14:54,004 INFO [finetune.py:976] (4/7) Epoch 4, batch 3000, loss[loss=0.2393, simple_loss=0.2826, pruned_loss=0.09799, over 4714.00 frames. ], tot_loss[loss=0.223, simple_loss=0.2819, pruned_loss=0.0821, over 951546.57 frames. ], batch size: 23, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 15:14:54,004 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-26 15:15:10,814 INFO [finetune.py:1010] (4/7) Epoch 4, validation: loss=0.1632, simple_loss=0.2363, pruned_loss=0.04509, over 2265189.00 frames. 2023-04-26 15:15:10,814 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6259MB 2023-04-26 15:15:29,321 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.6576, 3.6508, 2.7804, 4.2579, 3.7306, 3.6455, 1.6443, 3.5970], device='cuda:4'), covar=tensor([0.1895, 0.1065, 0.3487, 0.1684, 0.2790, 0.1943, 0.5694, 0.2413], device='cuda:4'), in_proj_covar=tensor([0.0252, 0.0223, 0.0261, 0.0313, 0.0309, 0.0259, 0.0277, 0.0280], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 15:15:35,626 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=20204.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:15:47,601 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.225e+02 1.833e+02 2.218e+02 2.696e+02 4.809e+02, threshold=4.435e+02, percent-clipped=1.0 2023-04-26 15:16:05,214 INFO [finetune.py:976] (4/7) Epoch 4, batch 3050, loss[loss=0.2404, simple_loss=0.2934, pruned_loss=0.0937, over 4803.00 frames. ], tot_loss[loss=0.2241, simple_loss=0.2833, pruned_loss=0.08245, over 953275.40 frames. ], batch size: 45, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 15:16:15,854 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7495, 1.3788, 1.3596, 1.4322, 1.9521, 1.5772, 1.2711, 1.3049], device='cuda:4'), covar=tensor([0.2046, 0.1694, 0.2452, 0.1776, 0.0939, 0.2069, 0.2910, 0.2490], device='cuda:4'), in_proj_covar=tensor([0.0321, 0.0342, 0.0357, 0.0314, 0.0356, 0.0358, 0.0320, 0.0359], device='cuda:4'), out_proj_covar=tensor([6.9294e-05, 7.3435e-05, 7.7272e-05, 6.5803e-05, 7.5779e-05, 7.8029e-05, 6.9466e-05, 7.7495e-05], device='cuda:4') 2023-04-26 15:16:43,905 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=20265.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:16:49,408 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=20274.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:16:55,395 INFO [finetune.py:976] (4/7) Epoch 4, batch 3100, loss[loss=0.2376, simple_loss=0.2822, pruned_loss=0.09648, over 4151.00 frames. ], tot_loss[loss=0.2218, simple_loss=0.281, pruned_loss=0.0813, over 954406.28 frames. ], batch size: 65, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 15:17:13,337 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9900, 2.7090, 1.9655, 1.8890, 1.5123, 1.4779, 2.0381, 1.4508], device='cuda:4'), covar=tensor([0.1933, 0.1846, 0.1944, 0.2336, 0.2974, 0.2365, 0.1466, 0.2537], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0224, 0.0184, 0.0213, 0.0222, 0.0190, 0.0178, 0.0201], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 15:17:21,575 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=20322.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:17:22,124 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.364e+02 1.832e+02 2.083e+02 2.367e+02 4.071e+02, threshold=4.166e+02, percent-clipped=0.0 2023-04-26 15:17:28,234 INFO [finetune.py:976] (4/7) Epoch 4, batch 3150, loss[loss=0.1999, simple_loss=0.2664, pruned_loss=0.0667, over 4784.00 frames. ], tot_loss[loss=0.22, simple_loss=0.2784, pruned_loss=0.08076, over 954156.95 frames. ], batch size: 29, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 15:17:57,231 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=20376.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:18:01,415 INFO [finetune.py:976] (4/7) Epoch 4, batch 3200, loss[loss=0.1885, simple_loss=0.2568, pruned_loss=0.06008, over 4752.00 frames. ], tot_loss[loss=0.2174, simple_loss=0.2755, pruned_loss=0.07964, over 955212.89 frames. ], batch size: 27, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 15:18:04,550 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=20388.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:18:25,092 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=20418.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:18:28,511 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.209e+02 1.902e+02 2.131e+02 2.743e+02 4.123e+02, threshold=4.262e+02, percent-clipped=0.0 2023-04-26 15:18:34,645 INFO [finetune.py:976] (4/7) Epoch 4, batch 3250, loss[loss=0.2593, simple_loss=0.3142, pruned_loss=0.1021, over 4819.00 frames. ], tot_loss[loss=0.2181, simple_loss=0.2761, pruned_loss=0.08007, over 955435.34 frames. ], batch size: 40, lr: 3.96e-03, grad_scale: 64.0 2023-04-26 15:18:42,469 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=20445.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:18:45,906 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=20449.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:19:04,241 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=20468.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:19:11,507 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=20479.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:19:13,810 INFO [finetune.py:976] (4/7) Epoch 4, batch 3300, loss[loss=0.2275, simple_loss=0.2925, pruned_loss=0.08121, over 4811.00 frames. ], tot_loss[loss=0.2204, simple_loss=0.2786, pruned_loss=0.08111, over 954002.27 frames. ], batch size: 45, lr: 3.96e-03, grad_scale: 32.0 2023-04-26 15:19:47,322 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.231e+02 1.894e+02 2.274e+02 2.897e+02 7.817e+02, threshold=4.548e+02, percent-clipped=5.0 2023-04-26 15:19:58,660 INFO [finetune.py:976] (4/7) Epoch 4, batch 3350, loss[loss=0.1751, simple_loss=0.2528, pruned_loss=0.04874, over 4796.00 frames. ], tot_loss[loss=0.2217, simple_loss=0.2801, pruned_loss=0.08163, over 952593.12 frames. ], batch size: 29, lr: 3.96e-03, grad_scale: 16.0 2023-04-26 15:20:15,420 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0823, 2.7254, 1.0602, 1.4065, 2.0299, 1.3084, 3.4709, 1.8330], device='cuda:4'), covar=tensor([0.0676, 0.0713, 0.0877, 0.1225, 0.0530, 0.0994, 0.0277, 0.0590], device='cuda:4'), in_proj_covar=tensor([0.0054, 0.0071, 0.0053, 0.0050, 0.0055, 0.0055, 0.0083, 0.0053], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0008, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 15:20:18,966 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0680, 2.9406, 2.3350, 2.5600, 1.9632, 2.2799, 2.6187, 1.9387], device='cuda:4'), covar=tensor([0.3006, 0.1589, 0.1055, 0.1692, 0.3802, 0.1670, 0.2110, 0.3248], device='cuda:4'), in_proj_covar=tensor([0.0315, 0.0335, 0.0244, 0.0309, 0.0323, 0.0287, 0.0274, 0.0298], device='cuda:4'), out_proj_covar=tensor([1.2834e-04, 1.3617e-04, 9.9303e-05, 1.2453e-04, 1.3337e-04, 1.1610e-04, 1.1320e-04, 1.2047e-04], device='cuda:4') 2023-04-26 15:20:39,484 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=20560.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:20:55,813 INFO [finetune.py:976] (4/7) Epoch 4, batch 3400, loss[loss=0.2182, simple_loss=0.2966, pruned_loss=0.06989, over 4918.00 frames. ], tot_loss[loss=0.2218, simple_loss=0.2808, pruned_loss=0.08144, over 954458.70 frames. ], batch size: 33, lr: 3.96e-03, grad_scale: 16.0 2023-04-26 15:21:04,539 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=20588.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:21:10,323 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-26 15:21:29,471 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.140e+02 1.806e+02 2.102e+02 2.435e+02 3.817e+02, threshold=4.203e+02, percent-clipped=0.0 2023-04-26 15:21:34,362 INFO [finetune.py:976] (4/7) Epoch 4, batch 3450, loss[loss=0.2062, simple_loss=0.262, pruned_loss=0.07523, over 4211.00 frames. ], tot_loss[loss=0.2207, simple_loss=0.2803, pruned_loss=0.08057, over 953567.48 frames. ], batch size: 18, lr: 3.96e-03, grad_scale: 16.0 2023-04-26 15:21:43,421 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7170, 1.6938, 1.8707, 2.1078, 2.0525, 1.6501, 1.3153, 1.8065], device='cuda:4'), covar=tensor([0.0929, 0.1191, 0.0734, 0.0596, 0.0635, 0.1057, 0.1032, 0.0632], device='cuda:4'), in_proj_covar=tensor([0.0210, 0.0211, 0.0188, 0.0184, 0.0184, 0.0200, 0.0172, 0.0195], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 15:21:44,628 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=20649.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:21:49,343 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.6067, 1.7659, 1.7872, 1.4864, 1.7984, 1.3813, 2.2717, 1.3705], device='cuda:4'), covar=tensor([0.4005, 0.1689, 0.4416, 0.2618, 0.1605, 0.2563, 0.1511, 0.4793], device='cuda:4'), in_proj_covar=tensor([0.0359, 0.0361, 0.0446, 0.0375, 0.0408, 0.0387, 0.0402, 0.0427], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 15:22:19,739 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=20676.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:22:23,908 INFO [finetune.py:976] (4/7) Epoch 4, batch 3500, loss[loss=0.2542, simple_loss=0.307, pruned_loss=0.1007, over 4814.00 frames. ], tot_loss[loss=0.2195, simple_loss=0.2784, pruned_loss=0.0803, over 954712.60 frames. ], batch size: 39, lr: 3.96e-03, grad_scale: 16.0 2023-04-26 15:22:58,020 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=20724.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:22:58,550 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.140e+02 1.751e+02 2.119e+02 2.772e+02 4.921e+02, threshold=4.238e+02, percent-clipped=1.0 2023-04-26 15:23:03,442 INFO [finetune.py:976] (4/7) Epoch 4, batch 3550, loss[loss=0.2061, simple_loss=0.2587, pruned_loss=0.07681, over 4855.00 frames. ], tot_loss[loss=0.2179, simple_loss=0.2761, pruned_loss=0.07989, over 954487.30 frames. ], batch size: 34, lr: 3.96e-03, grad_scale: 16.0 2023-04-26 15:23:15,952 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=20744.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:23:16,586 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=20745.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:23:48,041 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=20768.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:23:58,640 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=20774.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:24:04,024 INFO [finetune.py:976] (4/7) Epoch 4, batch 3600, loss[loss=0.1604, simple_loss=0.2279, pruned_loss=0.04643, over 4774.00 frames. ], tot_loss[loss=0.2158, simple_loss=0.2739, pruned_loss=0.07887, over 954071.88 frames. ], batch size: 26, lr: 3.96e-03, grad_scale: 16.0 2023-04-26 15:24:10,255 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=20793.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:24:15,846 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3158, 1.5374, 1.6039, 1.7303, 1.6493, 1.7722, 1.6629, 1.6838], device='cuda:4'), covar=tensor([0.8224, 1.3455, 1.1278, 1.0687, 1.2322, 1.9246, 1.4204, 1.2341], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0399, 0.0321, 0.0327, 0.0349, 0.0412, 0.0386, 0.0341], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 15:24:27,766 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4688, 2.4765, 2.7101, 3.1202, 2.6689, 2.2569, 1.9466, 2.4853], device='cuda:4'), covar=tensor([0.1100, 0.0889, 0.0535, 0.0655, 0.0729, 0.1124, 0.1082, 0.0698], device='cuda:4'), in_proj_covar=tensor([0.0212, 0.0212, 0.0189, 0.0185, 0.0184, 0.0202, 0.0173, 0.0195], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 15:24:31,211 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=20816.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:24:33,649 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.14 vs. limit=2.0 2023-04-26 15:24:38,615 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.149e+02 1.726e+02 2.072e+02 2.545e+02 4.959e+02, threshold=4.144e+02, percent-clipped=1.0 2023-04-26 15:24:43,528 INFO [finetune.py:976] (4/7) Epoch 4, batch 3650, loss[loss=0.296, simple_loss=0.3556, pruned_loss=0.1182, over 4825.00 frames. ], tot_loss[loss=0.2188, simple_loss=0.2767, pruned_loss=0.08047, over 954846.56 frames. ], batch size: 40, lr: 3.96e-03, grad_scale: 16.0 2023-04-26 15:25:00,498 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=20860.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:25:16,801 INFO [finetune.py:976] (4/7) Epoch 4, batch 3700, loss[loss=0.2493, simple_loss=0.3104, pruned_loss=0.0941, over 4897.00 frames. ], tot_loss[loss=0.2219, simple_loss=0.2803, pruned_loss=0.08176, over 955886.20 frames. ], batch size: 43, lr: 3.96e-03, grad_scale: 16.0 2023-04-26 15:25:31,862 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1571, 1.3540, 1.2591, 1.7126, 1.4882, 1.7154, 1.2933, 3.0883], device='cuda:4'), covar=tensor([0.0808, 0.0945, 0.0997, 0.1453, 0.0812, 0.0628, 0.0959, 0.0250], device='cuda:4'), in_proj_covar=tensor([0.0040, 0.0040, 0.0042, 0.0046, 0.0041, 0.0041, 0.0040, 0.0063], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 15:25:31,891 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5111, 1.1161, 1.2804, 1.0530, 1.6820, 1.3364, 1.0400, 1.2848], device='cuda:4'), covar=tensor([0.2028, 0.1710, 0.2518, 0.1856, 0.1010, 0.1872, 0.2710, 0.2107], device='cuda:4'), in_proj_covar=tensor([0.0316, 0.0337, 0.0352, 0.0309, 0.0350, 0.0352, 0.0314, 0.0352], device='cuda:4'), out_proj_covar=tensor([6.8111e-05, 7.2261e-05, 7.6301e-05, 6.4765e-05, 7.4276e-05, 7.6599e-05, 6.8271e-05, 7.5997e-05], device='cuda:4') 2023-04-26 15:25:32,434 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=20908.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:25:33,112 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1610, 2.1721, 2.0223, 1.9130, 2.4093, 1.8978, 2.9158, 1.7386], device='cuda:4'), covar=tensor([0.4184, 0.1759, 0.4476, 0.3011, 0.1709, 0.2511, 0.1121, 0.4412], device='cuda:4'), in_proj_covar=tensor([0.0354, 0.0356, 0.0440, 0.0370, 0.0402, 0.0381, 0.0397, 0.0423], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 15:25:50,202 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.325e+02 1.979e+02 2.308e+02 2.768e+02 4.690e+02, threshold=4.617e+02, percent-clipped=4.0 2023-04-26 15:26:01,986 INFO [finetune.py:976] (4/7) Epoch 4, batch 3750, loss[loss=0.2521, simple_loss=0.2977, pruned_loss=0.1032, over 4771.00 frames. ], tot_loss[loss=0.2242, simple_loss=0.2824, pruned_loss=0.08301, over 952455.58 frames. ], batch size: 26, lr: 3.96e-03, grad_scale: 16.0 2023-04-26 15:26:14,218 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=20944.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:27:02,470 INFO [finetune.py:976] (4/7) Epoch 4, batch 3800, loss[loss=0.2147, simple_loss=0.2855, pruned_loss=0.07198, over 4821.00 frames. ], tot_loss[loss=0.2248, simple_loss=0.2838, pruned_loss=0.08295, over 953428.76 frames. ], batch size: 30, lr: 3.96e-03, grad_scale: 16.0 2023-04-26 15:27:21,413 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.84 vs. limit=2.0 2023-04-26 15:27:46,873 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.129e+02 1.703e+02 2.110e+02 2.594e+02 5.214e+02, threshold=4.219e+02, percent-clipped=1.0 2023-04-26 15:28:05,806 INFO [finetune.py:976] (4/7) Epoch 4, batch 3850, loss[loss=0.2251, simple_loss=0.2704, pruned_loss=0.08987, over 4821.00 frames. ], tot_loss[loss=0.2244, simple_loss=0.2825, pruned_loss=0.08317, over 951603.47 frames. ], batch size: 39, lr: 3.96e-03, grad_scale: 16.0 2023-04-26 15:28:18,866 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=21044.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:28:29,632 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=21052.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:28:53,238 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-04-26 15:28:54,709 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=21074.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:29:11,949 INFO [finetune.py:976] (4/7) Epoch 4, batch 3900, loss[loss=0.1804, simple_loss=0.246, pruned_loss=0.05738, over 4811.00 frames. ], tot_loss[loss=0.2212, simple_loss=0.279, pruned_loss=0.08171, over 951968.08 frames. ], batch size: 25, lr: 3.96e-03, grad_scale: 16.0 2023-04-26 15:29:18,470 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=21092.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:29:31,300 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=21113.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:29:34,795 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=21118.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:29:37,129 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=21122.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:29:38,881 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.064e+02 1.817e+02 2.053e+02 2.461e+02 7.075e+02, threshold=4.105e+02, percent-clipped=2.0 2023-04-26 15:29:43,720 INFO [finetune.py:976] (4/7) Epoch 4, batch 3950, loss[loss=0.1935, simple_loss=0.2509, pruned_loss=0.06805, over 4909.00 frames. ], tot_loss[loss=0.2186, simple_loss=0.276, pruned_loss=0.08062, over 954412.76 frames. ], batch size: 36, lr: 3.96e-03, grad_scale: 16.0 2023-04-26 15:30:14,107 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=21179.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:30:16,453 INFO [finetune.py:976] (4/7) Epoch 4, batch 4000, loss[loss=0.249, simple_loss=0.2956, pruned_loss=0.1012, over 4133.00 frames. ], tot_loss[loss=0.2174, simple_loss=0.275, pruned_loss=0.07997, over 954882.32 frames. ], batch size: 65, lr: 3.96e-03, grad_scale: 16.0 2023-04-26 15:30:45,240 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.385e+02 1.961e+02 2.328e+02 2.721e+02 5.358e+02, threshold=4.656e+02, percent-clipped=3.0 2023-04-26 15:30:45,346 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3360, 3.0871, 0.8814, 1.6803, 1.6787, 2.2473, 1.7744, 0.9575], device='cuda:4'), covar=tensor([0.1398, 0.1052, 0.2084, 0.1425, 0.1201, 0.1064, 0.1552, 0.2221], device='cuda:4'), in_proj_covar=tensor([0.0123, 0.0266, 0.0149, 0.0130, 0.0141, 0.0162, 0.0127, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 15:30:50,171 INFO [finetune.py:976] (4/7) Epoch 4, batch 4050, loss[loss=0.168, simple_loss=0.2258, pruned_loss=0.05509, over 4239.00 frames. ], tot_loss[loss=0.2201, simple_loss=0.2777, pruned_loss=0.08129, over 951886.73 frames. ], batch size: 18, lr: 3.96e-03, grad_scale: 16.0 2023-04-26 15:30:58,876 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=21244.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:31:23,490 INFO [finetune.py:976] (4/7) Epoch 4, batch 4100, loss[loss=0.228, simple_loss=0.2969, pruned_loss=0.07955, over 4744.00 frames. ], tot_loss[loss=0.2225, simple_loss=0.2812, pruned_loss=0.08194, over 951428.71 frames. ], batch size: 59, lr: 3.96e-03, grad_scale: 16.0 2023-04-26 15:31:26,644 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1081, 1.3980, 1.2895, 1.6386, 1.4836, 1.5909, 1.3638, 3.0182], device='cuda:4'), covar=tensor([0.0734, 0.0812, 0.0846, 0.1324, 0.0711, 0.0609, 0.0792, 0.0194], device='cuda:4'), in_proj_covar=tensor([0.0040, 0.0040, 0.0042, 0.0046, 0.0041, 0.0041, 0.0040, 0.0063], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 15:31:30,024 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=21292.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:31:43,173 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=21312.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:31:50,917 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.243e+02 1.937e+02 2.313e+02 2.737e+02 5.004e+02, threshold=4.625e+02, percent-clipped=1.0 2023-04-26 15:31:52,886 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=21328.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:31:55,795 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1022, 1.9132, 2.1110, 2.4033, 2.3348, 1.7901, 1.5291, 1.9824], device='cuda:4'), covar=tensor([0.0957, 0.1145, 0.0686, 0.0623, 0.0613, 0.1084, 0.1011, 0.0678], device='cuda:4'), in_proj_covar=tensor([0.0208, 0.0208, 0.0185, 0.0181, 0.0181, 0.0197, 0.0170, 0.0191], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 15:31:56,304 INFO [finetune.py:976] (4/7) Epoch 4, batch 4150, loss[loss=0.2197, simple_loss=0.2816, pruned_loss=0.07886, over 4790.00 frames. ], tot_loss[loss=0.2235, simple_loss=0.2821, pruned_loss=0.08245, over 953345.45 frames. ], batch size: 29, lr: 3.95e-03, grad_scale: 16.0 2023-04-26 15:32:40,014 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=21373.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:32:46,015 INFO [finetune.py:976] (4/7) Epoch 4, batch 4200, loss[loss=0.1945, simple_loss=0.2672, pruned_loss=0.06091, over 4757.00 frames. ], tot_loss[loss=0.223, simple_loss=0.2823, pruned_loss=0.08179, over 952896.19 frames. ], batch size: 27, lr: 3.95e-03, grad_scale: 16.0 2023-04-26 15:32:50,241 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=21389.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:33:21,603 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=21408.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:33:31,333 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8527, 2.3638, 2.0681, 2.2913, 1.7829, 1.8736, 2.0839, 1.5366], device='cuda:4'), covar=tensor([0.2432, 0.1607, 0.1068, 0.1415, 0.3384, 0.1764, 0.2109, 0.3086], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0333, 0.0241, 0.0305, 0.0321, 0.0285, 0.0273, 0.0296], device='cuda:4'), out_proj_covar=tensor([1.2712e-04, 1.3550e-04, 9.8140e-05, 1.2289e-04, 1.3251e-04, 1.1521e-04, 1.1273e-04, 1.1971e-04], device='cuda:4') 2023-04-26 15:33:42,774 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.105e+02 1.624e+02 1.995e+02 2.490e+02 4.928e+02, threshold=3.989e+02, percent-clipped=1.0 2023-04-26 15:33:52,193 INFO [finetune.py:976] (4/7) Epoch 4, batch 4250, loss[loss=0.2121, simple_loss=0.2692, pruned_loss=0.07747, over 4907.00 frames. ], tot_loss[loss=0.2225, simple_loss=0.2813, pruned_loss=0.08186, over 955241.78 frames. ], batch size: 36, lr: 3.95e-03, grad_scale: 16.0 2023-04-26 15:33:55,948 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2148, 1.3950, 1.4018, 1.6028, 1.5283, 1.6767, 1.5036, 1.5481], device='cuda:4'), covar=tensor([0.8409, 1.2207, 1.0883, 0.9194, 1.1627, 1.7670, 1.3413, 1.1330], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0399, 0.0320, 0.0327, 0.0350, 0.0413, 0.0386, 0.0340], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 15:34:30,953 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=21474.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:34:41,941 INFO [finetune.py:976] (4/7) Epoch 4, batch 4300, loss[loss=0.2256, simple_loss=0.286, pruned_loss=0.08265, over 4904.00 frames. ], tot_loss[loss=0.2201, simple_loss=0.2783, pruned_loss=0.08097, over 954233.05 frames. ], batch size: 43, lr: 3.95e-03, grad_scale: 16.0 2023-04-26 15:35:01,527 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.44 vs. limit=2.0 2023-04-26 15:35:15,639 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6481, 1.9079, 0.7899, 1.4043, 1.9640, 1.5632, 1.4580, 1.5836], device='cuda:4'), covar=tensor([0.0569, 0.0386, 0.0435, 0.0590, 0.0301, 0.0580, 0.0588, 0.0627], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0027, 0.0025, 0.0031, 0.0022, 0.0030, 0.0030, 0.0031], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0049, 0.0038, 0.0048, 0.0047, 0.0050], device='cuda:4') 2023-04-26 15:35:22,267 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2774, 1.3274, 3.8730, 3.6023, 3.3938, 3.6690, 3.6953, 3.3475], device='cuda:4'), covar=tensor([0.6798, 0.5574, 0.1211, 0.1991, 0.1281, 0.1544, 0.1364, 0.1593], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0304, 0.0418, 0.0424, 0.0360, 0.0413, 0.0323, 0.0381], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-26 15:35:27,095 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 1.707e+02 2.018e+02 2.403e+02 4.975e+02, threshold=4.035e+02, percent-clipped=3.0 2023-04-26 15:35:31,900 INFO [finetune.py:976] (4/7) Epoch 4, batch 4350, loss[loss=0.2326, simple_loss=0.2859, pruned_loss=0.08962, over 4901.00 frames. ], tot_loss[loss=0.2165, simple_loss=0.2747, pruned_loss=0.07916, over 956958.02 frames. ], batch size: 35, lr: 3.95e-03, grad_scale: 16.0 2023-04-26 15:36:37,738 INFO [finetune.py:976] (4/7) Epoch 4, batch 4400, loss[loss=0.2832, simple_loss=0.3378, pruned_loss=0.1143, over 4901.00 frames. ], tot_loss[loss=0.219, simple_loss=0.2769, pruned_loss=0.08055, over 954491.91 frames. ], batch size: 43, lr: 3.95e-03, grad_scale: 16.0 2023-04-26 15:36:56,212 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=21595.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:37:09,362 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=21613.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:37:17,115 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.164e+02 1.848e+02 2.167e+02 2.564e+02 6.946e+02, threshold=4.335e+02, percent-clipped=3.0 2023-04-26 15:37:22,022 INFO [finetune.py:976] (4/7) Epoch 4, batch 4450, loss[loss=0.2661, simple_loss=0.3226, pruned_loss=0.1048, over 4900.00 frames. ], tot_loss[loss=0.2237, simple_loss=0.282, pruned_loss=0.08267, over 955832.71 frames. ], batch size: 43, lr: 3.95e-03, grad_scale: 16.0 2023-04-26 15:37:23,351 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1085, 1.4609, 1.3074, 1.6339, 1.4442, 1.8100, 1.3526, 3.4135], device='cuda:4'), covar=tensor([0.0737, 0.0785, 0.0848, 0.1271, 0.0705, 0.0664, 0.0772, 0.0158], device='cuda:4'), in_proj_covar=tensor([0.0040, 0.0040, 0.0042, 0.0046, 0.0041, 0.0041, 0.0041, 0.0063], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 15:37:36,764 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=21656.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:37:46,472 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=21668.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:37:50,693 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=21674.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:37:56,118 INFO [finetune.py:976] (4/7) Epoch 4, batch 4500, loss[loss=0.1621, simple_loss=0.2161, pruned_loss=0.05401, over 4069.00 frames. ], tot_loss[loss=0.2241, simple_loss=0.2826, pruned_loss=0.08282, over 953246.75 frames. ], batch size: 17, lr: 3.95e-03, grad_scale: 16.0 2023-04-26 15:37:56,778 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=21684.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:38:11,964 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=21708.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:38:15,897 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-04-26 15:38:24,950 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.135e+02 1.777e+02 2.074e+02 2.624e+02 7.543e+02, threshold=4.148e+02, percent-clipped=2.0 2023-04-26 15:38:29,903 INFO [finetune.py:976] (4/7) Epoch 4, batch 4550, loss[loss=0.2462, simple_loss=0.3004, pruned_loss=0.09603, over 4841.00 frames. ], tot_loss[loss=0.2251, simple_loss=0.2839, pruned_loss=0.08316, over 953009.41 frames. ], batch size: 49, lr: 3.95e-03, grad_scale: 16.0 2023-04-26 15:38:44,653 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=21756.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:38:49,056 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=21763.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:38:58,614 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=21774.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:39:04,077 INFO [finetune.py:976] (4/7) Epoch 4, batch 4600, loss[loss=0.2185, simple_loss=0.2721, pruned_loss=0.08244, over 4854.00 frames. ], tot_loss[loss=0.2235, simple_loss=0.2827, pruned_loss=0.08216, over 954628.40 frames. ], batch size: 44, lr: 3.95e-03, grad_scale: 16.0 2023-04-26 15:39:13,943 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6950, 2.6972, 2.2729, 2.4384, 2.7534, 2.1897, 3.7400, 2.0811], device='cuda:4'), covar=tensor([0.4325, 0.2095, 0.4081, 0.3760, 0.2123, 0.3063, 0.1361, 0.4448], device='cuda:4'), in_proj_covar=tensor([0.0354, 0.0356, 0.0437, 0.0369, 0.0399, 0.0382, 0.0395, 0.0420], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 15:39:30,457 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=21822.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:39:36,940 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=21824.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:39:37,233 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-26 15:39:37,418 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.786e+02 2.120e+02 2.552e+02 5.015e+02, threshold=4.240e+02, percent-clipped=1.0 2023-04-26 15:39:41,058 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0148, 1.0440, 3.1804, 2.7666, 2.8494, 2.9442, 3.0186, 2.6515], device='cuda:4'), covar=tensor([0.8662, 0.7454, 0.2455, 0.3701, 0.2752, 0.3593, 0.3798, 0.3984], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0302, 0.0418, 0.0421, 0.0359, 0.0411, 0.0321, 0.0378], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-26 15:39:48,032 INFO [finetune.py:976] (4/7) Epoch 4, batch 4650, loss[loss=0.1743, simple_loss=0.241, pruned_loss=0.05382, over 4852.00 frames. ], tot_loss[loss=0.22, simple_loss=0.2791, pruned_loss=0.08041, over 954316.31 frames. ], batch size: 44, lr: 3.95e-03, grad_scale: 16.0 2023-04-26 15:40:04,071 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=21850.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:40:49,526 INFO [finetune.py:976] (4/7) Epoch 4, batch 4700, loss[loss=0.2313, simple_loss=0.2715, pruned_loss=0.09559, over 4833.00 frames. ], tot_loss[loss=0.2178, simple_loss=0.2764, pruned_loss=0.07959, over 955887.21 frames. ], batch size: 30, lr: 3.95e-03, grad_scale: 16.0 2023-04-26 15:40:50,308 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9918, 1.4410, 1.8165, 1.9853, 1.7285, 1.3732, 0.9400, 1.4837], device='cuda:4'), covar=tensor([0.4457, 0.4929, 0.2223, 0.3524, 0.4132, 0.3873, 0.6172, 0.3854], device='cuda:4'), in_proj_covar=tensor([0.0271, 0.0262, 0.0220, 0.0332, 0.0221, 0.0229, 0.0248, 0.0198], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 15:41:07,136 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-26 15:41:08,895 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.91 vs. limit=5.0 2023-04-26 15:41:12,783 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=21911.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:41:21,737 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.201e+02 1.822e+02 2.098e+02 2.414e+02 4.975e+02, threshold=4.197e+02, percent-clipped=3.0 2023-04-26 15:41:28,679 INFO [finetune.py:976] (4/7) Epoch 4, batch 4750, loss[loss=0.2324, simple_loss=0.2824, pruned_loss=0.09117, over 4830.00 frames. ], tot_loss[loss=0.2168, simple_loss=0.275, pruned_loss=0.0793, over 956205.76 frames. ], batch size: 33, lr: 3.95e-03, grad_scale: 16.0 2023-04-26 15:41:40,251 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=21951.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:41:50,660 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=21968.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:41:51,759 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=21969.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:41:52,448 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=21970.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:42:02,271 INFO [finetune.py:976] (4/7) Epoch 4, batch 4800, loss[loss=0.202, simple_loss=0.2648, pruned_loss=0.06962, over 4699.00 frames. ], tot_loss[loss=0.2199, simple_loss=0.2777, pruned_loss=0.08106, over 953730.82 frames. ], batch size: 23, lr: 3.95e-03, grad_scale: 16.0 2023-04-26 15:42:02,997 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=21984.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:42:39,325 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=22016.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:42:50,415 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.187e+02 1.864e+02 2.140e+02 2.595e+02 5.123e+02, threshold=4.279e+02, percent-clipped=1.0 2023-04-26 15:42:54,749 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=22031.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:42:55,829 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=22032.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:42:56,374 INFO [finetune.py:976] (4/7) Epoch 4, batch 4850, loss[loss=0.1818, simple_loss=0.2401, pruned_loss=0.06178, over 4750.00 frames. ], tot_loss[loss=0.2215, simple_loss=0.2807, pruned_loss=0.0811, over 953643.77 frames. ], batch size: 23, lr: 3.95e-03, grad_scale: 16.0 2023-04-26 15:43:45,552 INFO [finetune.py:976] (4/7) Epoch 4, batch 4900, loss[loss=0.2204, simple_loss=0.2784, pruned_loss=0.08125, over 4091.00 frames. ], tot_loss[loss=0.2228, simple_loss=0.2823, pruned_loss=0.08161, over 952647.10 frames. ], batch size: 65, lr: 3.95e-03, grad_scale: 16.0 2023-04-26 15:44:15,735 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=22119.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:44:19,291 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.258e+02 1.858e+02 2.268e+02 2.601e+02 5.071e+02, threshold=4.535e+02, percent-clipped=2.0 2023-04-26 15:44:22,974 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=22130.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:44:24,747 INFO [finetune.py:976] (4/7) Epoch 4, batch 4950, loss[loss=0.2175, simple_loss=0.2836, pruned_loss=0.07573, over 4811.00 frames. ], tot_loss[loss=0.2228, simple_loss=0.2827, pruned_loss=0.08145, over 951572.04 frames. ], batch size: 39, lr: 3.95e-03, grad_scale: 16.0 2023-04-26 15:44:26,258 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.34 vs. limit=2.0 2023-04-26 15:44:35,940 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3758, 3.1771, 2.6748, 2.8990, 2.4660, 2.6676, 2.8179, 2.1622], device='cuda:4'), covar=tensor([0.2426, 0.1371, 0.0938, 0.1433, 0.2859, 0.1278, 0.1939, 0.2839], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0334, 0.0242, 0.0306, 0.0325, 0.0288, 0.0276, 0.0298], device='cuda:4'), out_proj_covar=tensor([1.2733e-04, 1.3577e-04, 9.8627e-05, 1.2341e-04, 1.3402e-04, 1.1634e-04, 1.1369e-04, 1.2031e-04], device='cuda:4') 2023-04-26 15:44:43,310 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.0529, 3.9824, 3.1420, 4.8002, 4.0081, 4.1092, 1.7732, 3.9838], device='cuda:4'), covar=tensor([0.1605, 0.1031, 0.3459, 0.1050, 0.3054, 0.1676, 0.5518, 0.2151], device='cuda:4'), in_proj_covar=tensor([0.0253, 0.0226, 0.0264, 0.0319, 0.0314, 0.0262, 0.0280, 0.0283], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 15:44:53,671 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=22176.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:44:58,359 INFO [finetune.py:976] (4/7) Epoch 4, batch 5000, loss[loss=0.1852, simple_loss=0.2451, pruned_loss=0.06264, over 4748.00 frames. ], tot_loss[loss=0.2211, simple_loss=0.2806, pruned_loss=0.08082, over 952679.15 frames. ], batch size: 27, lr: 3.95e-03, grad_scale: 16.0 2023-04-26 15:45:01,562 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3710, 3.3756, 0.9275, 1.8591, 1.9470, 2.3081, 2.0419, 0.9142], device='cuda:4'), covar=tensor([0.1556, 0.1588, 0.2229, 0.1436, 0.1164, 0.1314, 0.1418, 0.2225], device='cuda:4'), in_proj_covar=tensor([0.0121, 0.0261, 0.0147, 0.0127, 0.0137, 0.0159, 0.0124, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 15:45:04,358 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=22191.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:45:15,023 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=22206.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:45:20,506 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=22215.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:45:31,711 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.348e+02 1.793e+02 1.980e+02 2.407e+02 3.890e+02, threshold=3.961e+02, percent-clipped=0.0 2023-04-26 15:45:42,551 INFO [finetune.py:976] (4/7) Epoch 4, batch 5050, loss[loss=0.197, simple_loss=0.2609, pruned_loss=0.06658, over 4795.00 frames. ], tot_loss[loss=0.2181, simple_loss=0.2771, pruned_loss=0.07951, over 954470.66 frames. ], batch size: 51, lr: 3.95e-03, grad_scale: 16.0 2023-04-26 15:45:45,104 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=22237.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:45:45,117 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=22237.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:46:06,567 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=22251.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:46:28,159 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=22269.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:46:38,172 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=22276.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:46:42,377 INFO [finetune.py:976] (4/7) Epoch 4, batch 5100, loss[loss=0.2013, simple_loss=0.256, pruned_loss=0.07325, over 4923.00 frames. ], tot_loss[loss=0.2158, simple_loss=0.274, pruned_loss=0.07884, over 954051.92 frames. ], batch size: 38, lr: 3.95e-03, grad_scale: 16.0 2023-04-26 15:46:52,557 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=22298.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:46:53,598 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=22299.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:47:02,358 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3667, 3.5379, 0.8324, 1.7517, 1.7678, 2.5002, 1.9044, 0.9759], device='cuda:4'), covar=tensor([0.1462, 0.0849, 0.2147, 0.1401, 0.1112, 0.0987, 0.1473, 0.1937], device='cuda:4'), in_proj_covar=tensor([0.0122, 0.0263, 0.0148, 0.0129, 0.0138, 0.0161, 0.0125, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 15:47:03,604 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5065, 3.6578, 0.9942, 1.7982, 1.9728, 2.5054, 2.0100, 1.0306], device='cuda:4'), covar=tensor([0.1398, 0.0990, 0.2028, 0.1436, 0.1067, 0.1056, 0.1573, 0.1873], device='cuda:4'), in_proj_covar=tensor([0.0122, 0.0263, 0.0148, 0.0128, 0.0138, 0.0161, 0.0125, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 15:47:06,011 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=22317.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:47:08,495 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1354, 1.5739, 1.9463, 2.3867, 1.8727, 1.4663, 1.2266, 1.6649], device='cuda:4'), covar=tensor([0.4668, 0.5237, 0.2511, 0.3754, 0.4803, 0.4088, 0.6161, 0.3881], device='cuda:4'), in_proj_covar=tensor([0.0273, 0.0263, 0.0221, 0.0333, 0.0223, 0.0230, 0.0249, 0.0199], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 15:47:10,809 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.090e+02 1.713e+02 2.102e+02 2.590e+02 4.893e+02, threshold=4.205e+02, percent-clipped=2.0 2023-04-26 15:47:11,502 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=22326.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:47:15,715 INFO [finetune.py:976] (4/7) Epoch 4, batch 5150, loss[loss=0.214, simple_loss=0.2707, pruned_loss=0.07865, over 4841.00 frames. ], tot_loss[loss=0.2157, simple_loss=0.2734, pruned_loss=0.07896, over 954197.74 frames. ], batch size: 25, lr: 3.95e-03, grad_scale: 16.0 2023-04-26 15:47:15,867 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7617, 0.9849, 1.2653, 1.4393, 1.4166, 1.5895, 1.3129, 1.3146], device='cuda:4'), covar=tensor([0.7871, 1.2037, 0.9845, 0.9146, 1.1620, 1.6928, 1.1806, 1.0855], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0397, 0.0320, 0.0325, 0.0349, 0.0411, 0.0384, 0.0339], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 15:47:53,214 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0387, 1.3539, 1.2906, 1.6703, 1.4143, 1.7255, 1.3290, 3.0472], device='cuda:4'), covar=tensor([0.0709, 0.0803, 0.0814, 0.1295, 0.0718, 0.0529, 0.0798, 0.0180], device='cuda:4'), in_proj_covar=tensor([0.0040, 0.0040, 0.0042, 0.0046, 0.0041, 0.0041, 0.0040, 0.0063], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 15:47:54,331 INFO [finetune.py:976] (4/7) Epoch 4, batch 5200, loss[loss=0.2622, simple_loss=0.3282, pruned_loss=0.09813, over 4843.00 frames. ], tot_loss[loss=0.2181, simple_loss=0.2765, pruned_loss=0.07988, over 954152.65 frames. ], batch size: 47, lr: 3.95e-03, grad_scale: 16.0 2023-04-26 15:48:01,225 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0135, 2.6046, 0.8990, 1.3470, 1.9289, 1.2095, 3.6195, 1.6582], device='cuda:4'), covar=tensor([0.0712, 0.0734, 0.0922, 0.1331, 0.0596, 0.1083, 0.0238, 0.0674], device='cuda:4'), in_proj_covar=tensor([0.0054, 0.0070, 0.0052, 0.0049, 0.0054, 0.0054, 0.0082, 0.0053], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 15:48:01,255 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=22385.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:48:30,910 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=22419.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:48:34,448 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.341e+02 1.984e+02 2.270e+02 2.624e+02 8.595e+02, threshold=4.540e+02, percent-clipped=4.0 2023-04-26 15:48:38,930 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5059, 1.1566, 1.5186, 1.8064, 1.6050, 1.4735, 1.5017, 1.5478], device='cuda:4'), covar=tensor([1.2066, 1.6684, 1.8231, 2.0117, 1.3714, 1.9439, 1.9532, 1.5828], device='cuda:4'), in_proj_covar=tensor([0.0427, 0.0468, 0.0553, 0.0573, 0.0459, 0.0484, 0.0496, 0.0499], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 15:48:39,406 INFO [finetune.py:976] (4/7) Epoch 4, batch 5250, loss[loss=0.2392, simple_loss=0.2979, pruned_loss=0.0903, over 4811.00 frames. ], tot_loss[loss=0.2203, simple_loss=0.2792, pruned_loss=0.08073, over 954283.50 frames. ], batch size: 39, lr: 3.95e-03, grad_scale: 16.0 2023-04-26 15:48:47,970 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=22446.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:49:09,424 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=22467.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:49:25,139 INFO [finetune.py:976] (4/7) Epoch 4, batch 5300, loss[loss=0.2237, simple_loss=0.2803, pruned_loss=0.08356, over 4886.00 frames. ], tot_loss[loss=0.2226, simple_loss=0.2817, pruned_loss=0.08176, over 953562.92 frames. ], batch size: 35, lr: 3.95e-03, grad_scale: 16.0 2023-04-26 15:49:32,901 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=22486.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:49:56,903 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=22506.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:50:15,639 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.302e+02 1.836e+02 2.188e+02 2.701e+02 5.070e+02, threshold=4.376e+02, percent-clipped=2.0 2023-04-26 15:50:19,966 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=22532.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:50:20,501 INFO [finetune.py:976] (4/7) Epoch 4, batch 5350, loss[loss=0.1825, simple_loss=0.2429, pruned_loss=0.06101, over 4789.00 frames. ], tot_loss[loss=0.2219, simple_loss=0.2814, pruned_loss=0.08116, over 953278.02 frames. ], batch size: 25, lr: 3.95e-03, grad_scale: 32.0 2023-04-26 15:50:33,852 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=22554.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:50:46,634 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=22571.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:50:52,124 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0932, 2.5322, 2.1037, 2.4027, 2.0477, 2.0785, 2.2556, 1.8106], device='cuda:4'), covar=tensor([0.2262, 0.1240, 0.1091, 0.1365, 0.3137, 0.1307, 0.2031, 0.3004], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0331, 0.0241, 0.0304, 0.0322, 0.0285, 0.0273, 0.0295], device='cuda:4'), out_proj_covar=tensor([1.2614e-04, 1.3459e-04, 9.8388e-05, 1.2223e-04, 1.3266e-04, 1.1513e-04, 1.1259e-04, 1.1920e-04], device='cuda:4') 2023-04-26 15:50:53,882 INFO [finetune.py:976] (4/7) Epoch 4, batch 5400, loss[loss=0.2016, simple_loss=0.2539, pruned_loss=0.07468, over 4840.00 frames. ], tot_loss[loss=0.2195, simple_loss=0.279, pruned_loss=0.08002, over 954200.39 frames. ], batch size: 30, lr: 3.95e-03, grad_scale: 32.0 2023-04-26 15:50:57,665 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-26 15:50:59,992 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=22593.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:51:30,048 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7556, 1.4209, 1.3076, 1.5664, 1.9923, 1.6653, 1.3859, 1.3151], device='cuda:4'), covar=tensor([0.1631, 0.1570, 0.2341, 0.1586, 0.0849, 0.1722, 0.1996, 0.1875], device='cuda:4'), in_proj_covar=tensor([0.0315, 0.0334, 0.0352, 0.0308, 0.0347, 0.0349, 0.0314, 0.0354], device='cuda:4'), out_proj_covar=tensor([6.7897e-05, 7.1715e-05, 7.6150e-05, 6.4372e-05, 7.3647e-05, 7.6078e-05, 6.8226e-05, 7.6442e-05], device='cuda:4') 2023-04-26 15:51:38,540 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.141e+02 1.846e+02 2.198e+02 2.575e+02 4.196e+02, threshold=4.395e+02, percent-clipped=1.0 2023-04-26 15:51:39,240 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=22626.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:51:47,644 INFO [finetune.py:976] (4/7) Epoch 4, batch 5450, loss[loss=0.1645, simple_loss=0.2297, pruned_loss=0.04967, over 4752.00 frames. ], tot_loss[loss=0.2157, simple_loss=0.2749, pruned_loss=0.07827, over 954667.21 frames. ], batch size: 26, lr: 3.95e-03, grad_scale: 32.0 2023-04-26 15:51:51,438 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=22639.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:51:52,102 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1294, 1.4444, 1.4937, 1.6261, 1.5837, 1.7322, 1.5981, 1.5466], device='cuda:4'), covar=tensor([0.8066, 1.2174, 1.1066, 0.9289, 1.1154, 1.6870, 1.2842, 1.1901], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0399, 0.0320, 0.0326, 0.0351, 0.0414, 0.0385, 0.0340], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 15:52:39,600 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=22674.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:52:50,911 INFO [finetune.py:976] (4/7) Epoch 4, batch 5500, loss[loss=0.2418, simple_loss=0.2904, pruned_loss=0.09665, over 4838.00 frames. ], tot_loss[loss=0.2139, simple_loss=0.2725, pruned_loss=0.07764, over 955445.37 frames. ], batch size: 47, lr: 3.95e-03, grad_scale: 32.0 2023-04-26 15:53:12,617 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=22700.0, num_to_drop=1, layers_to_drop={3} 2023-04-26 15:53:13,891 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6505, 1.3732, 1.6924, 1.9688, 1.7812, 1.6381, 1.7368, 1.7387], device='cuda:4'), covar=tensor([1.2208, 1.5375, 1.7532, 1.9013, 1.3179, 1.8683, 1.8971, 1.4969], device='cuda:4'), in_proj_covar=tensor([0.0426, 0.0467, 0.0554, 0.0575, 0.0459, 0.0484, 0.0496, 0.0498], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 15:53:32,316 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6517, 1.5697, 0.6173, 1.3795, 1.6876, 1.5160, 1.4158, 1.4766], device='cuda:4'), covar=tensor([0.0532, 0.0417, 0.0459, 0.0603, 0.0308, 0.0581, 0.0551, 0.0647], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0027, 0.0024, 0.0031, 0.0022, 0.0030, 0.0030, 0.0031], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0049, 0.0038, 0.0048, 0.0047, 0.0050], device='cuda:4') 2023-04-26 15:53:34,589 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.236e+02 2.045e+02 2.332e+02 2.799e+02 5.603e+02, threshold=4.665e+02, percent-clipped=2.0 2023-04-26 15:53:40,509 INFO [finetune.py:976] (4/7) Epoch 4, batch 5550, loss[loss=0.2177, simple_loss=0.2715, pruned_loss=0.08199, over 4877.00 frames. ], tot_loss[loss=0.216, simple_loss=0.2742, pruned_loss=0.07885, over 954354.32 frames. ], batch size: 31, lr: 3.95e-03, grad_scale: 32.0 2023-04-26 15:53:43,053 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5279, 1.5015, 0.5998, 1.2471, 1.4029, 1.3889, 1.2954, 1.3332], device='cuda:4'), covar=tensor([0.0539, 0.0384, 0.0449, 0.0585, 0.0336, 0.0555, 0.0525, 0.0627], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0027, 0.0025, 0.0031, 0.0022, 0.0031, 0.0030, 0.0031], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0049, 0.0038, 0.0048, 0.0047, 0.0050], device='cuda:4') 2023-04-26 15:53:45,458 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=22741.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:54:11,072 INFO [finetune.py:976] (4/7) Epoch 4, batch 5600, loss[loss=0.2188, simple_loss=0.2648, pruned_loss=0.08642, over 4689.00 frames. ], tot_loss[loss=0.2177, simple_loss=0.2768, pruned_loss=0.07934, over 955634.05 frames. ], batch size: 23, lr: 3.95e-03, grad_scale: 32.0 2023-04-26 15:54:12,922 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=22786.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:54:18,054 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=22794.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:54:37,297 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.226e+02 1.792e+02 2.075e+02 2.605e+02 4.713e+02, threshold=4.150e+02, percent-clipped=1.0 2023-04-26 15:54:37,396 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=22825.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:54:41,494 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=22832.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:54:42,027 INFO [finetune.py:976] (4/7) Epoch 4, batch 5650, loss[loss=0.1777, simple_loss=0.2413, pruned_loss=0.05704, over 4729.00 frames. ], tot_loss[loss=0.2209, simple_loss=0.2806, pruned_loss=0.08058, over 956209.72 frames. ], batch size: 23, lr: 3.95e-03, grad_scale: 32.0 2023-04-26 15:54:42,646 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=22834.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:54:46,939 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3101, 1.7666, 2.0839, 2.6587, 2.0848, 1.6024, 1.4096, 2.0149], device='cuda:4'), covar=tensor([0.4191, 0.4970, 0.2163, 0.3482, 0.4656, 0.3763, 0.5938, 0.3706], device='cuda:4'), in_proj_covar=tensor([0.0273, 0.0263, 0.0221, 0.0332, 0.0222, 0.0231, 0.0248, 0.0198], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 15:54:55,169 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=22855.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:54:57,522 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5769, 2.9615, 1.3881, 1.9621, 2.3150, 1.6796, 3.9592, 2.3326], device='cuda:4'), covar=tensor([0.0587, 0.0697, 0.0804, 0.1153, 0.0514, 0.0899, 0.0304, 0.0513], device='cuda:4'), in_proj_covar=tensor([0.0054, 0.0070, 0.0053, 0.0049, 0.0054, 0.0055, 0.0082, 0.0053], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 15:55:15,930 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=22871.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:55:27,438 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=22880.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:55:29,216 INFO [finetune.py:976] (4/7) Epoch 4, batch 5700, loss[loss=0.1557, simple_loss=0.2084, pruned_loss=0.05154, over 4563.00 frames. ], tot_loss[loss=0.2176, simple_loss=0.2759, pruned_loss=0.07968, over 936887.69 frames. ], batch size: 20, lr: 3.95e-03, grad_scale: 32.0 2023-04-26 15:55:31,132 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=22886.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:55:40,627 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=22893.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:55:51,837 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-26 15:56:19,323 INFO [finetune.py:976] (4/7) Epoch 5, batch 0, loss[loss=0.2659, simple_loss=0.3198, pruned_loss=0.106, over 4817.00 frames. ], tot_loss[loss=0.2659, simple_loss=0.3198, pruned_loss=0.106, over 4817.00 frames. ], batch size: 38, lr: 3.95e-03, grad_scale: 32.0 2023-04-26 15:56:19,323 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-26 15:56:29,253 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4432, 1.2481, 1.6944, 1.5432, 1.3302, 1.1765, 1.3868, 0.8697], device='cuda:4'), covar=tensor([0.0875, 0.1287, 0.0647, 0.0972, 0.1037, 0.1775, 0.0798, 0.1190], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0077, 0.0075, 0.0069, 0.0080, 0.0096, 0.0083, 0.0078], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 15:56:30,075 INFO [finetune.py:1010] (4/7) Epoch 5, validation: loss=0.1632, simple_loss=0.2369, pruned_loss=0.04473, over 2265189.00 frames. 2023-04-26 15:56:30,076 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6259MB 2023-04-26 15:56:33,904 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5902, 1.0127, 1.4910, 1.8955, 1.7036, 1.4937, 1.5593, 1.5809], device='cuda:4'), covar=tensor([1.1411, 1.6011, 1.5909, 1.7602, 1.3135, 1.8615, 1.7981, 1.4557], device='cuda:4'), in_proj_covar=tensor([0.0426, 0.0466, 0.0554, 0.0573, 0.0458, 0.0483, 0.0495, 0.0498], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 15:56:35,033 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=22919.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:56:38,532 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.121e+02 1.609e+02 1.975e+02 2.369e+02 5.506e+02, threshold=3.950e+02, percent-clipped=1.0 2023-04-26 15:56:44,773 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5174, 1.3017, 1.5936, 1.8934, 1.7795, 1.4712, 1.5664, 1.5858], device='cuda:4'), covar=tensor([1.1088, 1.4630, 1.5113, 1.7032, 1.1885, 1.8417, 1.7658, 1.5346], device='cuda:4'), in_proj_covar=tensor([0.0425, 0.0465, 0.0554, 0.0572, 0.0458, 0.0482, 0.0494, 0.0498], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 15:56:48,362 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=22941.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:56:50,211 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.3217, 3.1938, 2.4992, 3.8223, 3.3333, 3.2952, 1.3706, 3.2201], device='cuda:4'), covar=tensor([0.2080, 0.1382, 0.3549, 0.2293, 0.2512, 0.2112, 0.5664, 0.2791], device='cuda:4'), in_proj_covar=tensor([0.0250, 0.0223, 0.0261, 0.0314, 0.0309, 0.0258, 0.0277, 0.0278], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 15:57:02,120 INFO [finetune.py:976] (4/7) Epoch 5, batch 50, loss[loss=0.195, simple_loss=0.2643, pruned_loss=0.06288, over 4777.00 frames. ], tot_loss[loss=0.2267, simple_loss=0.2836, pruned_loss=0.0849, over 214996.28 frames. ], batch size: 29, lr: 3.95e-03, grad_scale: 32.0 2023-04-26 15:57:23,647 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0068, 1.3908, 1.2617, 1.6294, 1.4660, 1.5175, 1.3054, 2.3805], device='cuda:4'), covar=tensor([0.0684, 0.0763, 0.0795, 0.1209, 0.0663, 0.0562, 0.0745, 0.0257], device='cuda:4'), in_proj_covar=tensor([0.0040, 0.0040, 0.0042, 0.0046, 0.0041, 0.0041, 0.0040, 0.0062], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 15:57:30,162 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=22995.0, num_to_drop=1, layers_to_drop={3} 2023-04-26 15:57:52,332 INFO [finetune.py:976] (4/7) Epoch 5, batch 100, loss[loss=0.1951, simple_loss=0.2349, pruned_loss=0.07758, over 4010.00 frames. ], tot_loss[loss=0.215, simple_loss=0.2728, pruned_loss=0.0786, over 378324.89 frames. ], batch size: 17, lr: 3.95e-03, grad_scale: 32.0 2023-04-26 15:58:02,671 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.276e+01 1.854e+02 2.138e+02 2.662e+02 6.636e+02, threshold=4.277e+02, percent-clipped=4.0 2023-04-26 15:58:12,530 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=23041.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:58:17,442 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2943, 1.6286, 1.4598, 1.5709, 1.4739, 1.6958, 1.6380, 1.5942], device='cuda:4'), covar=tensor([0.7954, 1.1685, 1.0881, 0.9746, 1.1552, 1.7703, 1.2426, 1.1484], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0398, 0.0319, 0.0326, 0.0350, 0.0413, 0.0385, 0.0339], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 15:58:25,549 INFO [finetune.py:976] (4/7) Epoch 5, batch 150, loss[loss=0.2215, simple_loss=0.2799, pruned_loss=0.08152, over 4903.00 frames. ], tot_loss[loss=0.2135, simple_loss=0.2702, pruned_loss=0.07835, over 504852.69 frames. ], batch size: 37, lr: 3.95e-03, grad_scale: 32.0 2023-04-26 15:58:44,620 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=23089.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:58:57,366 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=23108.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:58:59,551 INFO [finetune.py:976] (4/7) Epoch 5, batch 200, loss[loss=0.1719, simple_loss=0.2355, pruned_loss=0.05414, over 4900.00 frames. ], tot_loss[loss=0.2091, simple_loss=0.2667, pruned_loss=0.07576, over 603548.82 frames. ], batch size: 32, lr: 3.95e-03, grad_scale: 32.0 2023-04-26 15:59:10,064 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.278e+02 1.706e+02 2.083e+02 2.458e+02 7.322e+02, threshold=4.166e+02, percent-clipped=1.0 2023-04-26 15:59:25,391 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=23150.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:59:33,030 INFO [finetune.py:976] (4/7) Epoch 5, batch 250, loss[loss=0.2188, simple_loss=0.2838, pruned_loss=0.07694, over 4928.00 frames. ], tot_loss[loss=0.2104, simple_loss=0.2695, pruned_loss=0.07565, over 682735.14 frames. ], batch size: 38, lr: 3.95e-03, grad_scale: 32.0 2023-04-26 15:59:38,563 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=23169.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 15:59:47,773 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=23181.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:00:00,762 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5969, 1.2631, 1.6965, 1.7754, 1.6654, 1.4960, 1.6069, 1.6425], device='cuda:4'), covar=tensor([1.5813, 2.1197, 2.5301, 3.0606, 1.8210, 2.6106, 2.6776, 2.1578], device='cuda:4'), in_proj_covar=tensor([0.0427, 0.0466, 0.0555, 0.0575, 0.0459, 0.0484, 0.0496, 0.0498], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 16:00:06,065 INFO [finetune.py:976] (4/7) Epoch 5, batch 300, loss[loss=0.2427, simple_loss=0.3063, pruned_loss=0.0896, over 4896.00 frames. ], tot_loss[loss=0.2145, simple_loss=0.2749, pruned_loss=0.0771, over 744411.38 frames. ], batch size: 35, lr: 3.95e-03, grad_scale: 32.0 2023-04-26 16:00:10,760 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=23217.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:00:15,563 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.275e+02 1.925e+02 2.300e+02 2.815e+02 4.609e+02, threshold=4.600e+02, percent-clipped=3.0 2023-04-26 16:00:18,604 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=23229.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:00:35,018 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.31 vs. limit=5.0 2023-04-26 16:00:55,705 INFO [finetune.py:976] (4/7) Epoch 5, batch 350, loss[loss=0.2298, simple_loss=0.2851, pruned_loss=0.08729, over 4915.00 frames. ], tot_loss[loss=0.2184, simple_loss=0.2785, pruned_loss=0.07919, over 792517.47 frames. ], batch size: 38, lr: 3.95e-03, grad_scale: 32.0 2023-04-26 16:01:07,642 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1552, 2.7846, 2.0794, 2.0686, 1.6573, 1.7160, 2.1304, 1.6031], device='cuda:4'), covar=tensor([0.1791, 0.1479, 0.1847, 0.1993, 0.2736, 0.2380, 0.1371, 0.2375], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0222, 0.0181, 0.0210, 0.0220, 0.0190, 0.0174, 0.0197], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 16:01:18,183 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=23278.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:01:31,693 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=23290.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:01:40,995 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=23295.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:01:55,883 INFO [finetune.py:976] (4/7) Epoch 5, batch 400, loss[loss=0.1876, simple_loss=0.2656, pruned_loss=0.05484, over 4916.00 frames. ], tot_loss[loss=0.219, simple_loss=0.2793, pruned_loss=0.0793, over 828621.15 frames. ], batch size: 38, lr: 3.95e-03, grad_scale: 32.0 2023-04-26 16:02:05,328 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.110e+02 1.686e+02 2.112e+02 2.575e+02 6.256e+02, threshold=4.223e+02, percent-clipped=1.0 2023-04-26 16:02:18,848 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=23343.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:02:21,896 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.4840, 3.5014, 2.6297, 4.1150, 3.5038, 3.4438, 1.7724, 3.4516], device='cuda:4'), covar=tensor([0.1628, 0.1231, 0.3915, 0.1517, 0.2642, 0.1874, 0.5089, 0.2224], device='cuda:4'), in_proj_covar=tensor([0.0248, 0.0222, 0.0261, 0.0312, 0.0306, 0.0258, 0.0275, 0.0276], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 16:02:29,821 INFO [finetune.py:976] (4/7) Epoch 5, batch 450, loss[loss=0.1977, simple_loss=0.264, pruned_loss=0.06571, over 4821.00 frames. ], tot_loss[loss=0.2169, simple_loss=0.2772, pruned_loss=0.07826, over 854782.15 frames. ], batch size: 39, lr: 3.95e-03, grad_scale: 32.0 2023-04-26 16:02:31,883 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-26 16:03:14,686 INFO [finetune.py:976] (4/7) Epoch 5, batch 500, loss[loss=0.2163, simple_loss=0.2649, pruned_loss=0.08389, over 4916.00 frames. ], tot_loss[loss=0.2158, simple_loss=0.2751, pruned_loss=0.07825, over 875786.77 frames. ], batch size: 37, lr: 3.95e-03, grad_scale: 32.0 2023-04-26 16:03:29,843 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.283e+02 1.686e+02 2.056e+02 2.772e+02 5.396e+02, threshold=4.112e+02, percent-clipped=3.0 2023-04-26 16:03:47,509 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=23450.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:03:54,137 INFO [finetune.py:976] (4/7) Epoch 5, batch 550, loss[loss=0.1674, simple_loss=0.2349, pruned_loss=0.04993, over 4775.00 frames. ], tot_loss[loss=0.2141, simple_loss=0.2729, pruned_loss=0.07768, over 893718.18 frames. ], batch size: 27, lr: 3.95e-03, grad_scale: 32.0 2023-04-26 16:03:56,064 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=23464.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:04:07,405 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=23481.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:04:12,216 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=23488.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:04:19,705 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=23498.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:04:27,505 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.56 vs. limit=2.0 2023-04-26 16:04:27,578 INFO [finetune.py:976] (4/7) Epoch 5, batch 600, loss[loss=0.225, simple_loss=0.2836, pruned_loss=0.08324, over 4901.00 frames. ], tot_loss[loss=0.2148, simple_loss=0.2733, pruned_loss=0.07819, over 908672.13 frames. ], batch size: 43, lr: 3.94e-03, grad_scale: 32.0 2023-04-26 16:04:36,131 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.343e+02 1.963e+02 2.277e+02 2.707e+02 6.010e+02, threshold=4.553e+02, percent-clipped=1.0 2023-04-26 16:04:39,605 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=23529.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:04:53,878 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=23549.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 16:05:01,024 INFO [finetune.py:976] (4/7) Epoch 5, batch 650, loss[loss=0.2374, simple_loss=0.3042, pruned_loss=0.08531, over 4738.00 frames. ], tot_loss[loss=0.2181, simple_loss=0.2772, pruned_loss=0.07953, over 919419.56 frames. ], batch size: 59, lr: 3.94e-03, grad_scale: 32.0 2023-04-26 16:05:08,396 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=23573.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:05:16,709 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=23585.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:05:34,426 INFO [finetune.py:976] (4/7) Epoch 5, batch 700, loss[loss=0.2874, simple_loss=0.3412, pruned_loss=0.1168, over 4808.00 frames. ], tot_loss[loss=0.2201, simple_loss=0.2793, pruned_loss=0.08045, over 926503.07 frames. ], batch size: 39, lr: 3.94e-03, grad_scale: 32.0 2023-04-26 16:05:36,665 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.90 vs. limit=2.0 2023-04-26 16:05:42,879 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.112e+02 1.928e+02 2.468e+02 2.939e+02 6.493e+02, threshold=4.936e+02, percent-clipped=4.0 2023-04-26 16:06:19,711 INFO [finetune.py:976] (4/7) Epoch 5, batch 750, loss[loss=0.1826, simple_loss=0.2327, pruned_loss=0.06624, over 4278.00 frames. ], tot_loss[loss=0.2211, simple_loss=0.2801, pruned_loss=0.08099, over 932704.09 frames. ], batch size: 18, lr: 3.94e-03, grad_scale: 32.0 2023-04-26 16:07:26,347 INFO [finetune.py:976] (4/7) Epoch 5, batch 800, loss[loss=0.2176, simple_loss=0.2588, pruned_loss=0.08816, over 4271.00 frames. ], tot_loss[loss=0.2181, simple_loss=0.2779, pruned_loss=0.07914, over 935982.17 frames. ], batch size: 18, lr: 3.94e-03, grad_scale: 32.0 2023-04-26 16:07:34,822 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.218e+02 1.744e+02 2.077e+02 2.568e+02 5.488e+02, threshold=4.154e+02, percent-clipped=2.0 2023-04-26 16:07:55,042 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4895, 0.9572, 1.3338, 1.8610, 1.6293, 1.4211, 1.4634, 1.4934], device='cuda:4'), covar=tensor([1.0104, 1.4109, 1.4478, 1.5796, 1.1611, 1.6444, 1.5380, 1.2598], device='cuda:4'), in_proj_covar=tensor([0.0425, 0.0462, 0.0550, 0.0568, 0.0455, 0.0480, 0.0491, 0.0492], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 16:08:00,117 INFO [finetune.py:976] (4/7) Epoch 5, batch 850, loss[loss=0.1755, simple_loss=0.2281, pruned_loss=0.06139, over 4834.00 frames. ], tot_loss[loss=0.2166, simple_loss=0.2764, pruned_loss=0.07839, over 941111.55 frames. ], batch size: 33, lr: 3.94e-03, grad_scale: 32.0 2023-04-26 16:08:02,024 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=23764.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:08:17,393 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.85 vs. limit=2.0 2023-04-26 16:08:35,443 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-26 16:08:39,276 INFO [finetune.py:976] (4/7) Epoch 5, batch 900, loss[loss=0.2117, simple_loss=0.2581, pruned_loss=0.0827, over 4813.00 frames. ], tot_loss[loss=0.2148, simple_loss=0.2742, pruned_loss=0.07769, over 946319.80 frames. ], batch size: 25, lr: 3.94e-03, grad_scale: 32.0 2023-04-26 16:08:40,390 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=23812.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:08:49,865 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7641, 2.1822, 1.1575, 1.5771, 2.2607, 1.7456, 1.6645, 1.7856], device='cuda:4'), covar=tensor([0.0530, 0.0385, 0.0349, 0.0567, 0.0240, 0.0556, 0.0536, 0.0631], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0027, 0.0024, 0.0031, 0.0022, 0.0030, 0.0030, 0.0031], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0045, 0.0038, 0.0049, 0.0038, 0.0048, 0.0047, 0.0050], device='cuda:4') 2023-04-26 16:08:54,019 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.136e+02 1.680e+02 2.075e+02 2.492e+02 8.869e+02, threshold=4.150e+02, percent-clipped=5.0 2023-04-26 16:09:13,203 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6314, 2.6112, 2.0990, 3.0313, 2.6505, 2.6147, 1.2044, 2.5915], device='cuda:4'), covar=tensor([0.2070, 0.1754, 0.4148, 0.3193, 0.3027, 0.2227, 0.4650, 0.2736], device='cuda:4'), in_proj_covar=tensor([0.0249, 0.0223, 0.0262, 0.0316, 0.0310, 0.0260, 0.0278, 0.0279], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 16:09:23,842 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=23844.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 16:09:46,929 INFO [finetune.py:976] (4/7) Epoch 5, batch 950, loss[loss=0.1682, simple_loss=0.2361, pruned_loss=0.05021, over 4750.00 frames. ], tot_loss[loss=0.2133, simple_loss=0.272, pruned_loss=0.07732, over 947212.08 frames. ], batch size: 27, lr: 3.94e-03, grad_scale: 32.0 2023-04-26 16:09:56,932 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=23873.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:10:14,840 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=23885.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:10:43,224 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2349, 1.5509, 1.3966, 1.9974, 1.7709, 1.9246, 1.4799, 4.2066], device='cuda:4'), covar=tensor([0.0677, 0.0825, 0.0821, 0.1178, 0.0653, 0.0702, 0.0785, 0.0119], device='cuda:4'), in_proj_covar=tensor([0.0040, 0.0040, 0.0042, 0.0046, 0.0041, 0.0041, 0.0040, 0.0062], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 16:10:44,367 INFO [finetune.py:976] (4/7) Epoch 5, batch 1000, loss[loss=0.2565, simple_loss=0.3027, pruned_loss=0.1051, over 4156.00 frames. ], tot_loss[loss=0.2144, simple_loss=0.274, pruned_loss=0.07747, over 949902.59 frames. ], batch size: 17, lr: 3.94e-03, grad_scale: 32.0 2023-04-26 16:10:45,107 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.3369, 1.3449, 1.3842, 0.9915, 1.3085, 1.0502, 1.7519, 1.2565], device='cuda:4'), covar=tensor([0.3893, 0.1713, 0.5356, 0.2830, 0.1756, 0.2368, 0.1711, 0.5071], device='cuda:4'), in_proj_covar=tensor([0.0353, 0.0358, 0.0441, 0.0374, 0.0402, 0.0386, 0.0401, 0.0423], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 16:10:52,028 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=23921.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:10:54,329 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.295e+02 1.932e+02 2.275e+02 2.774e+02 5.327e+02, threshold=4.551e+02, percent-clipped=3.0 2023-04-26 16:10:59,243 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=23933.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:11:17,667 INFO [finetune.py:976] (4/7) Epoch 5, batch 1050, loss[loss=0.2408, simple_loss=0.2934, pruned_loss=0.09413, over 4805.00 frames. ], tot_loss[loss=0.2167, simple_loss=0.2769, pruned_loss=0.07825, over 951724.49 frames. ], batch size: 25, lr: 3.94e-03, grad_scale: 32.0 2023-04-26 16:11:50,374 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8822, 2.5537, 0.9293, 1.2115, 1.8099, 1.2040, 3.3941, 1.5877], device='cuda:4'), covar=tensor([0.0769, 0.0893, 0.0991, 0.1389, 0.0611, 0.1101, 0.0238, 0.0760], device='cuda:4'), in_proj_covar=tensor([0.0055, 0.0071, 0.0053, 0.0050, 0.0054, 0.0055, 0.0083, 0.0053], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0008, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 16:12:31,172 INFO [finetune.py:976] (4/7) Epoch 5, batch 1100, loss[loss=0.2259, simple_loss=0.2818, pruned_loss=0.085, over 4736.00 frames. ], tot_loss[loss=0.2163, simple_loss=0.2769, pruned_loss=0.07786, over 952917.76 frames. ], batch size: 54, lr: 3.94e-03, grad_scale: 32.0 2023-04-26 16:12:45,514 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.175e+02 1.787e+02 2.210e+02 2.620e+02 5.269e+02, threshold=4.419e+02, percent-clipped=3.0 2023-04-26 16:12:53,353 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=24037.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:12:56,172 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.98 vs. limit=5.0 2023-04-26 16:13:08,673 INFO [finetune.py:976] (4/7) Epoch 5, batch 1150, loss[loss=0.2465, simple_loss=0.3112, pruned_loss=0.09093, over 4881.00 frames. ], tot_loss[loss=0.2182, simple_loss=0.279, pruned_loss=0.0787, over 953597.92 frames. ], batch size: 35, lr: 3.94e-03, grad_scale: 32.0 2023-04-26 16:13:28,118 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1653, 1.6900, 2.0527, 2.4757, 1.9244, 1.4959, 1.2602, 1.7841], device='cuda:4'), covar=tensor([0.4612, 0.4884, 0.2456, 0.3846, 0.4297, 0.3996, 0.6279, 0.3791], device='cuda:4'), in_proj_covar=tensor([0.0272, 0.0261, 0.0220, 0.0331, 0.0221, 0.0229, 0.0247, 0.0197], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 16:13:33,502 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=24098.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:13:42,166 INFO [finetune.py:976] (4/7) Epoch 5, batch 1200, loss[loss=0.2496, simple_loss=0.3057, pruned_loss=0.09681, over 4892.00 frames. ], tot_loss[loss=0.2175, simple_loss=0.2779, pruned_loss=0.07858, over 953301.72 frames. ], batch size: 35, lr: 3.94e-03, grad_scale: 32.0 2023-04-26 16:13:52,170 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.100e+02 1.779e+02 2.090e+02 2.399e+02 4.332e+02, threshold=4.179e+02, percent-clipped=0.0 2023-04-26 16:14:01,089 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6599, 1.9423, 1.0818, 1.4129, 2.1908, 1.6380, 1.5534, 1.5715], device='cuda:4'), covar=tensor([0.0534, 0.0391, 0.0368, 0.0583, 0.0279, 0.0555, 0.0521, 0.0602], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0027, 0.0024, 0.0030, 0.0022, 0.0030, 0.0029, 0.0031], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0049, 0.0038, 0.0048, 0.0047, 0.0050], device='cuda:4') 2023-04-26 16:14:01,725 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2705, 1.5797, 1.5157, 1.6688, 1.5822, 1.7203, 1.6715, 1.6239], device='cuda:4'), covar=tensor([0.7358, 1.2766, 1.0356, 0.9699, 1.1318, 1.6983, 1.2413, 1.1147], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0396, 0.0316, 0.0326, 0.0347, 0.0411, 0.0381, 0.0336], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 16:14:04,722 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=24144.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 16:14:12,579 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.00 vs. limit=5.0 2023-04-26 16:14:14,331 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5396, 1.3409, 1.8216, 1.7384, 1.4280, 1.1671, 1.4959, 0.9608], device='cuda:4'), covar=tensor([0.0727, 0.0904, 0.0535, 0.0807, 0.0905, 0.1528, 0.0821, 0.1021], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0078, 0.0076, 0.0070, 0.0081, 0.0097, 0.0083, 0.0079], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 16:14:15,789 INFO [finetune.py:976] (4/7) Epoch 5, batch 1250, loss[loss=0.1995, simple_loss=0.2604, pruned_loss=0.06925, over 4765.00 frames. ], tot_loss[loss=0.2145, simple_loss=0.2748, pruned_loss=0.0771, over 954757.00 frames. ], batch size: 23, lr: 3.94e-03, grad_scale: 32.0 2023-04-26 16:14:20,204 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-04-26 16:14:42,785 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=24192.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:15:00,605 INFO [finetune.py:976] (4/7) Epoch 5, batch 1300, loss[loss=0.2148, simple_loss=0.2781, pruned_loss=0.07576, over 4854.00 frames. ], tot_loss[loss=0.2121, simple_loss=0.2718, pruned_loss=0.07625, over 954332.86 frames. ], batch size: 44, lr: 3.94e-03, grad_scale: 32.0 2023-04-26 16:15:10,677 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.317e+02 1.846e+02 2.149e+02 2.713e+02 6.103e+02, threshold=4.299e+02, percent-clipped=1.0 2023-04-26 16:15:49,983 INFO [finetune.py:976] (4/7) Epoch 5, batch 1350, loss[loss=0.2211, simple_loss=0.2763, pruned_loss=0.08294, over 4820.00 frames. ], tot_loss[loss=0.2111, simple_loss=0.2707, pruned_loss=0.07579, over 956764.85 frames. ], batch size: 30, lr: 3.94e-03, grad_scale: 32.0 2023-04-26 16:16:26,099 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6861, 2.2072, 0.9936, 1.0898, 1.5203, 1.0611, 2.4603, 1.2213], device='cuda:4'), covar=tensor([0.0791, 0.0674, 0.0707, 0.1380, 0.0495, 0.1125, 0.0341, 0.0836], device='cuda:4'), in_proj_covar=tensor([0.0054, 0.0071, 0.0053, 0.0050, 0.0054, 0.0055, 0.0083, 0.0053], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0008, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 16:16:29,536 INFO [finetune.py:976] (4/7) Epoch 5, batch 1400, loss[loss=0.2555, simple_loss=0.3276, pruned_loss=0.09173, over 4925.00 frames. ], tot_loss[loss=0.2163, simple_loss=0.2757, pruned_loss=0.07845, over 956196.48 frames. ], batch size: 42, lr: 3.94e-03, grad_scale: 32.0 2023-04-26 16:16:38,504 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.222e+02 1.836e+02 2.129e+02 2.470e+02 6.262e+02, threshold=4.259e+02, percent-clipped=1.0 2023-04-26 16:16:38,993 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-26 16:17:19,155 INFO [finetune.py:976] (4/7) Epoch 5, batch 1450, loss[loss=0.1567, simple_loss=0.2128, pruned_loss=0.05031, over 4076.00 frames. ], tot_loss[loss=0.2179, simple_loss=0.2779, pruned_loss=0.07895, over 954763.00 frames. ], batch size: 17, lr: 3.94e-03, grad_scale: 32.0 2023-04-26 16:17:28,196 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6040, 2.0310, 1.7172, 1.9757, 1.5925, 1.6412, 1.7513, 1.4896], device='cuda:4'), covar=tensor([0.2136, 0.1277, 0.0960, 0.1265, 0.3490, 0.1342, 0.1818, 0.2576], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0333, 0.0242, 0.0306, 0.0326, 0.0287, 0.0275, 0.0299], device='cuda:4'), out_proj_covar=tensor([1.2776e-04, 1.3531e-04, 9.8408e-05, 1.2304e-04, 1.3432e-04, 1.1596e-04, 1.1334e-04, 1.2069e-04], device='cuda:4') 2023-04-26 16:18:00,685 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=24393.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:18:08,920 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9743, 2.5585, 2.0407, 1.8746, 1.5413, 1.6279, 2.1043, 1.5334], device='cuda:4'), covar=tensor([0.1966, 0.1594, 0.1952, 0.2190, 0.2963, 0.2427, 0.1364, 0.2490], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0222, 0.0181, 0.0210, 0.0219, 0.0190, 0.0173, 0.0197], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 16:18:16,237 INFO [finetune.py:976] (4/7) Epoch 5, batch 1500, loss[loss=0.2413, simple_loss=0.3019, pruned_loss=0.09032, over 4811.00 frames. ], tot_loss[loss=0.2194, simple_loss=0.2797, pruned_loss=0.07959, over 956107.89 frames. ], batch size: 40, lr: 3.94e-03, grad_scale: 32.0 2023-04-26 16:18:17,039 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.85 vs. limit=5.0 2023-04-26 16:18:25,733 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.120e+02 1.847e+02 2.193e+02 2.526e+02 4.286e+02, threshold=4.386e+02, percent-clipped=1.0 2023-04-26 16:18:49,533 INFO [finetune.py:976] (4/7) Epoch 5, batch 1550, loss[loss=0.2358, simple_loss=0.2957, pruned_loss=0.08794, over 4828.00 frames. ], tot_loss[loss=0.2201, simple_loss=0.2806, pruned_loss=0.07979, over 955667.69 frames. ], batch size: 33, lr: 3.94e-03, grad_scale: 32.0 2023-04-26 16:19:22,931 INFO [finetune.py:976] (4/7) Epoch 5, batch 1600, loss[loss=0.2181, simple_loss=0.2637, pruned_loss=0.08624, over 4831.00 frames. ], tot_loss[loss=0.2182, simple_loss=0.2778, pruned_loss=0.0793, over 955379.29 frames. ], batch size: 30, lr: 3.94e-03, grad_scale: 32.0 2023-04-26 16:19:32,014 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.098e+02 1.836e+02 2.156e+02 2.625e+02 3.904e+02, threshold=4.311e+02, percent-clipped=0.0 2023-04-26 16:19:32,145 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1021, 1.4147, 1.2136, 1.6061, 1.3990, 1.7240, 1.2470, 3.3245], device='cuda:4'), covar=tensor([0.0856, 0.1051, 0.1067, 0.1503, 0.0878, 0.0669, 0.1003, 0.0227], device='cuda:4'), in_proj_covar=tensor([0.0040, 0.0040, 0.0041, 0.0046, 0.0041, 0.0041, 0.0040, 0.0062], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 16:19:40,022 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6307, 3.6689, 0.8529, 1.9037, 2.0898, 2.5418, 2.0677, 1.1490], device='cuda:4'), covar=tensor([0.1332, 0.0923, 0.2182, 0.1364, 0.1058, 0.1078, 0.1465, 0.1887], device='cuda:4'), in_proj_covar=tensor([0.0122, 0.0262, 0.0145, 0.0128, 0.0138, 0.0160, 0.0125, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 16:19:45,368 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4284, 1.3027, 0.5567, 1.2382, 1.1612, 1.2877, 1.2894, 1.2824], device='cuda:4'), covar=tensor([0.0642, 0.0383, 0.0443, 0.0649, 0.0349, 0.0707, 0.0665, 0.0654], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0027, 0.0024, 0.0031, 0.0022, 0.0030, 0.0030, 0.0031], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0045, 0.0038, 0.0049, 0.0038, 0.0048, 0.0047, 0.0050], device='cuda:4') 2023-04-26 16:19:48,188 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=24547.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:19:56,588 INFO [finetune.py:976] (4/7) Epoch 5, batch 1650, loss[loss=0.223, simple_loss=0.2648, pruned_loss=0.09063, over 4328.00 frames. ], tot_loss[loss=0.2151, simple_loss=0.2744, pruned_loss=0.07783, over 956180.13 frames. ], batch size: 19, lr: 3.94e-03, grad_scale: 64.0 2023-04-26 16:20:01,606 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=24569.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:20:13,589 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1759, 2.0186, 2.2579, 2.3886, 2.4594, 1.9894, 1.6508, 2.0679], device='cuda:4'), covar=tensor([0.0858, 0.0978, 0.0534, 0.0609, 0.0599, 0.0921, 0.0931, 0.0656], device='cuda:4'), in_proj_covar=tensor([0.0205, 0.0207, 0.0183, 0.0180, 0.0181, 0.0195, 0.0167, 0.0192], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 16:20:25,486 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.31 vs. limit=5.0 2023-04-26 16:20:28,446 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=24608.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:20:30,147 INFO [finetune.py:976] (4/7) Epoch 5, batch 1700, loss[loss=0.226, simple_loss=0.2846, pruned_loss=0.08374, over 4899.00 frames. ], tot_loss[loss=0.2122, simple_loss=0.2711, pruned_loss=0.07668, over 954943.57 frames. ], batch size: 32, lr: 3.94e-03, grad_scale: 64.0 2023-04-26 16:20:38,586 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.174e+02 1.758e+02 2.083e+02 2.406e+02 5.348e+02, threshold=4.165e+02, percent-clipped=1.0 2023-04-26 16:20:38,701 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4336, 3.4745, 0.9717, 1.8233, 1.9292, 2.4238, 1.9856, 1.0586], device='cuda:4'), covar=tensor([0.1445, 0.0963, 0.2032, 0.1383, 0.1074, 0.1068, 0.1509, 0.2146], device='cuda:4'), in_proj_covar=tensor([0.0122, 0.0262, 0.0145, 0.0128, 0.0137, 0.0160, 0.0125, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 16:20:42,279 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=24630.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:20:50,347 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.57 vs. limit=2.0 2023-04-26 16:21:03,220 INFO [finetune.py:976] (4/7) Epoch 5, batch 1750, loss[loss=0.1697, simple_loss=0.2277, pruned_loss=0.05584, over 4764.00 frames. ], tot_loss[loss=0.2141, simple_loss=0.2727, pruned_loss=0.07777, over 954746.02 frames. ], batch size: 26, lr: 3.94e-03, grad_scale: 64.0 2023-04-26 16:21:06,394 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4484, 1.2046, 0.5646, 1.1672, 1.4719, 1.3522, 1.2520, 1.2740], device='cuda:4'), covar=tensor([0.0571, 0.0445, 0.0452, 0.0607, 0.0333, 0.0537, 0.0543, 0.0646], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0027, 0.0025, 0.0031, 0.0022, 0.0030, 0.0030, 0.0031], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0045, 0.0038, 0.0049, 0.0038, 0.0048, 0.0047, 0.0050], device='cuda:4') 2023-04-26 16:21:35,459 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=24693.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:21:48,678 INFO [finetune.py:976] (4/7) Epoch 5, batch 1800, loss[loss=0.2367, simple_loss=0.2936, pruned_loss=0.08991, over 4796.00 frames. ], tot_loss[loss=0.216, simple_loss=0.2755, pruned_loss=0.07826, over 954295.71 frames. ], batch size: 45, lr: 3.94e-03, grad_scale: 64.0 2023-04-26 16:21:54,924 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2936, 3.2691, 0.8622, 1.7888, 1.7051, 2.3750, 1.9226, 0.9287], device='cuda:4'), covar=tensor([0.1549, 0.1126, 0.2105, 0.1424, 0.1147, 0.1038, 0.1571, 0.2065], device='cuda:4'), in_proj_covar=tensor([0.0122, 0.0262, 0.0145, 0.0128, 0.0137, 0.0159, 0.0125, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 16:21:57,282 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.363e+02 2.030e+02 2.431e+02 2.911e+02 5.485e+02, threshold=4.863e+02, percent-clipped=5.0 2023-04-26 16:22:08,098 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=24741.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:22:27,731 INFO [finetune.py:976] (4/7) Epoch 5, batch 1850, loss[loss=0.2387, simple_loss=0.2994, pruned_loss=0.08897, over 4897.00 frames. ], tot_loss[loss=0.2171, simple_loss=0.2771, pruned_loss=0.07852, over 953084.59 frames. ], batch size: 37, lr: 3.94e-03, grad_scale: 64.0 2023-04-26 16:22:30,298 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=24765.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:23:01,618 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3158, 2.9231, 0.9753, 1.4331, 2.2859, 1.3174, 3.9601, 1.7948], device='cuda:4'), covar=tensor([0.0658, 0.0957, 0.0868, 0.1272, 0.0491, 0.1011, 0.0201, 0.0640], device='cuda:4'), in_proj_covar=tensor([0.0054, 0.0071, 0.0053, 0.0050, 0.0054, 0.0055, 0.0083, 0.0053], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0008, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 16:23:29,629 INFO [finetune.py:976] (4/7) Epoch 5, batch 1900, loss[loss=0.2449, simple_loss=0.3065, pruned_loss=0.09169, over 4903.00 frames. ], tot_loss[loss=0.2173, simple_loss=0.2777, pruned_loss=0.07843, over 954352.72 frames. ], batch size: 37, lr: 3.94e-03, grad_scale: 64.0 2023-04-26 16:23:44,192 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.223e+02 1.736e+02 2.141e+02 2.516e+02 4.890e+02, threshold=4.282e+02, percent-clipped=1.0 2023-04-26 16:23:50,607 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=24826.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:24:16,052 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=24848.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:24:29,176 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.77 vs. limit=2.0 2023-04-26 16:24:31,094 INFO [finetune.py:976] (4/7) Epoch 5, batch 1950, loss[loss=0.2055, simple_loss=0.269, pruned_loss=0.07097, over 4815.00 frames. ], tot_loss[loss=0.2153, simple_loss=0.2759, pruned_loss=0.0774, over 954295.73 frames. ], batch size: 25, lr: 3.94e-03, grad_scale: 64.0 2023-04-26 16:24:35,911 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.8490, 2.8485, 2.2048, 3.3054, 2.8915, 2.8592, 1.2040, 2.8183], device='cuda:4'), covar=tensor([0.2016, 0.1562, 0.3227, 0.2774, 0.3202, 0.2163, 0.5370, 0.2804], device='cuda:4'), in_proj_covar=tensor([0.0251, 0.0224, 0.0262, 0.0318, 0.0310, 0.0261, 0.0278, 0.0280], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 16:24:44,493 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4966, 1.1360, 0.5177, 1.1874, 1.2098, 1.3633, 1.2840, 1.2690], device='cuda:4'), covar=tensor([0.0669, 0.0416, 0.0457, 0.0671, 0.0336, 0.0746, 0.0642, 0.0717], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0027, 0.0025, 0.0031, 0.0022, 0.0030, 0.0030, 0.0031], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0045, 0.0038, 0.0049, 0.0038, 0.0048, 0.0048, 0.0050], device='cuda:4') 2023-04-26 16:24:57,880 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-04-26 16:24:58,639 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=24903.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:25:03,331 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=24909.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:25:04,918 INFO [finetune.py:976] (4/7) Epoch 5, batch 2000, loss[loss=0.2155, simple_loss=0.2727, pruned_loss=0.07918, over 4814.00 frames. ], tot_loss[loss=0.2139, simple_loss=0.2738, pruned_loss=0.07704, over 954943.03 frames. ], batch size: 41, lr: 3.94e-03, grad_scale: 64.0 2023-04-26 16:25:13,838 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.128e+02 1.797e+02 2.078e+02 2.506e+02 3.857e+02, threshold=4.156e+02, percent-clipped=0.0 2023-04-26 16:25:13,929 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=24925.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:25:37,718 INFO [finetune.py:976] (4/7) Epoch 5, batch 2050, loss[loss=0.2385, simple_loss=0.2884, pruned_loss=0.09427, over 4875.00 frames. ], tot_loss[loss=0.2106, simple_loss=0.2701, pruned_loss=0.07552, over 955555.45 frames. ], batch size: 34, lr: 3.94e-03, grad_scale: 64.0 2023-04-26 16:25:59,897 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=24995.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:26:10,616 INFO [finetune.py:976] (4/7) Epoch 5, batch 2100, loss[loss=0.2449, simple_loss=0.2774, pruned_loss=0.1062, over 4191.00 frames. ], tot_loss[loss=0.2105, simple_loss=0.2699, pruned_loss=0.07557, over 957248.68 frames. ], batch size: 18, lr: 3.94e-03, grad_scale: 64.0 2023-04-26 16:26:13,591 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2240, 1.4264, 1.4504, 1.5828, 1.5170, 1.6997, 1.5804, 1.5703], device='cuda:4'), covar=tensor([0.7879, 1.0758, 0.9737, 0.9402, 1.0426, 1.5357, 1.1168, 1.0022], device='cuda:4'), in_proj_covar=tensor([0.0314, 0.0399, 0.0318, 0.0328, 0.0349, 0.0414, 0.0381, 0.0337], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 16:26:21,077 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.078e+02 1.851e+02 2.091e+02 2.608e+02 4.715e+02, threshold=4.183e+02, percent-clipped=2.0 2023-04-26 16:26:25,563 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.66 vs. limit=2.0 2023-04-26 16:26:40,933 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=25056.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 16:26:43,814 INFO [finetune.py:976] (4/7) Epoch 5, batch 2150, loss[loss=0.1839, simple_loss=0.2629, pruned_loss=0.05249, over 4904.00 frames. ], tot_loss[loss=0.2127, simple_loss=0.273, pruned_loss=0.07621, over 956369.58 frames. ], batch size: 35, lr: 3.94e-03, grad_scale: 64.0 2023-04-26 16:27:17,133 INFO [finetune.py:976] (4/7) Epoch 5, batch 2200, loss[loss=0.2398, simple_loss=0.3124, pruned_loss=0.08362, over 4805.00 frames. ], tot_loss[loss=0.2149, simple_loss=0.2758, pruned_loss=0.07701, over 955179.19 frames. ], batch size: 40, lr: 3.94e-03, grad_scale: 64.0 2023-04-26 16:27:19,111 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6393, 2.4279, 1.5848, 1.5621, 1.2976, 1.3295, 1.7168, 1.2367], device='cuda:4'), covar=tensor([0.2075, 0.1676, 0.2028, 0.2360, 0.3085, 0.2479, 0.1360, 0.2411], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0223, 0.0180, 0.0210, 0.0218, 0.0190, 0.0173, 0.0197], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 16:27:24,269 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=25121.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:27:27,081 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.188e+02 1.776e+02 2.171e+02 2.634e+02 4.750e+02, threshold=4.343e+02, percent-clipped=2.0 2023-04-26 16:27:38,391 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-26 16:27:50,269 INFO [finetune.py:976] (4/7) Epoch 5, batch 2250, loss[loss=0.2031, simple_loss=0.2786, pruned_loss=0.06385, over 4800.00 frames. ], tot_loss[loss=0.2183, simple_loss=0.2787, pruned_loss=0.079, over 954452.52 frames. ], batch size: 51, lr: 3.94e-03, grad_scale: 64.0 2023-04-26 16:28:07,838 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-26 16:28:26,329 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=25203.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:28:26,893 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=25204.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:28:29,068 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.51 vs. limit=2.0 2023-04-26 16:28:32,137 INFO [finetune.py:976] (4/7) Epoch 5, batch 2300, loss[loss=0.225, simple_loss=0.2819, pruned_loss=0.08402, over 4109.00 frames. ], tot_loss[loss=0.2171, simple_loss=0.2782, pruned_loss=0.07802, over 953324.49 frames. ], batch size: 65, lr: 3.94e-03, grad_scale: 64.0 2023-04-26 16:28:52,194 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.169e+02 1.823e+02 2.210e+02 2.598e+02 6.851e+02, threshold=4.421e+02, percent-clipped=2.0 2023-04-26 16:28:52,298 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=25225.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:29:05,739 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9007, 2.6073, 1.8380, 1.7360, 1.3810, 1.4951, 1.8959, 1.3701], device='cuda:4'), covar=tensor([0.1916, 0.1688, 0.1971, 0.2238, 0.2971, 0.2388, 0.1404, 0.2509], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0223, 0.0181, 0.0210, 0.0219, 0.0191, 0.0173, 0.0198], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 16:29:26,418 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=25251.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:29:38,270 INFO [finetune.py:976] (4/7) Epoch 5, batch 2350, loss[loss=0.1925, simple_loss=0.2458, pruned_loss=0.06963, over 4795.00 frames. ], tot_loss[loss=0.2142, simple_loss=0.275, pruned_loss=0.07669, over 953470.41 frames. ], batch size: 51, lr: 3.94e-03, grad_scale: 64.0 2023-04-26 16:29:56,790 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6989, 1.8506, 0.7797, 1.4123, 1.8208, 1.4949, 1.4589, 1.5141], device='cuda:4'), covar=tensor([0.0647, 0.0361, 0.0429, 0.0675, 0.0299, 0.0740, 0.0724, 0.0727], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0027, 0.0024, 0.0030, 0.0022, 0.0030, 0.0030, 0.0031], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0049, 0.0038, 0.0048, 0.0047, 0.0050], device='cuda:4') 2023-04-26 16:29:57,357 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=25273.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:30:01,087 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.77 vs. limit=2.0 2023-04-26 16:30:39,274 INFO [finetune.py:976] (4/7) Epoch 5, batch 2400, loss[loss=0.1899, simple_loss=0.2449, pruned_loss=0.06743, over 4701.00 frames. ], tot_loss[loss=0.2123, simple_loss=0.2722, pruned_loss=0.07618, over 953389.16 frames. ], batch size: 23, lr: 3.94e-03, grad_scale: 64.0 2023-04-26 16:30:47,873 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5383, 1.3750, 4.3419, 4.0428, 3.8484, 4.1253, 4.0981, 3.7564], device='cuda:4'), covar=tensor([0.6886, 0.5820, 0.0970, 0.1708, 0.1004, 0.1317, 0.1197, 0.1636], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0303, 0.0414, 0.0422, 0.0356, 0.0409, 0.0320, 0.0377], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-26 16:30:48,374 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.145e+02 1.795e+02 2.098e+02 2.522e+02 5.509e+02, threshold=4.195e+02, percent-clipped=3.0 2023-04-26 16:31:12,381 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=25351.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 16:31:18,440 INFO [finetune.py:976] (4/7) Epoch 5, batch 2450, loss[loss=0.2038, simple_loss=0.2549, pruned_loss=0.07636, over 4865.00 frames. ], tot_loss[loss=0.2101, simple_loss=0.2694, pruned_loss=0.07542, over 953889.11 frames. ], batch size: 49, lr: 3.94e-03, grad_scale: 64.0 2023-04-26 16:32:08,176 INFO [finetune.py:976] (4/7) Epoch 5, batch 2500, loss[loss=0.1984, simple_loss=0.2626, pruned_loss=0.06714, over 4817.00 frames. ], tot_loss[loss=0.2106, simple_loss=0.2698, pruned_loss=0.07568, over 951158.80 frames. ], batch size: 40, lr: 3.94e-03, grad_scale: 64.0 2023-04-26 16:32:10,195 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.51 vs. limit=2.0 2023-04-26 16:32:15,245 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=25421.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:32:17,551 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.133e+02 1.862e+02 2.343e+02 2.826e+02 5.817e+02, threshold=4.685e+02, percent-clipped=1.0 2023-04-26 16:32:19,511 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=25428.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 16:32:41,938 INFO [finetune.py:976] (4/7) Epoch 5, batch 2550, loss[loss=0.259, simple_loss=0.3164, pruned_loss=0.1007, over 4807.00 frames. ], tot_loss[loss=0.2163, simple_loss=0.2755, pruned_loss=0.07849, over 951769.05 frames. ], batch size: 41, lr: 3.94e-03, grad_scale: 32.0 2023-04-26 16:32:47,380 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=25469.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:33:01,431 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4246, 3.0152, 0.8397, 1.7581, 1.7698, 2.1618, 1.7829, 0.9645], device='cuda:4'), covar=tensor([0.1346, 0.0963, 0.1960, 0.1301, 0.0989, 0.0992, 0.1503, 0.1787], device='cuda:4'), in_proj_covar=tensor([0.0122, 0.0263, 0.0146, 0.0128, 0.0138, 0.0160, 0.0125, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 16:33:01,477 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=25489.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 16:33:11,417 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=25504.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:33:15,644 INFO [finetune.py:976] (4/7) Epoch 5, batch 2600, loss[loss=0.2329, simple_loss=0.2933, pruned_loss=0.08626, over 4850.00 frames. ], tot_loss[loss=0.2167, simple_loss=0.2764, pruned_loss=0.07851, over 951424.22 frames. ], batch size: 31, lr: 3.94e-03, grad_scale: 32.0 2023-04-26 16:33:20,702 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6461, 1.3863, 1.8759, 1.8941, 1.4716, 1.1753, 1.5341, 0.9999], device='cuda:4'), covar=tensor([0.0678, 0.1033, 0.0592, 0.1027, 0.0971, 0.1527, 0.0988, 0.1160], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0077, 0.0075, 0.0070, 0.0081, 0.0096, 0.0083, 0.0079], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 16:33:25,302 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.303e+02 1.830e+02 2.137e+02 2.517e+02 4.934e+02, threshold=4.274e+02, percent-clipped=1.0 2023-04-26 16:33:43,926 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=25552.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:33:49,359 INFO [finetune.py:976] (4/7) Epoch 5, batch 2650, loss[loss=0.2121, simple_loss=0.2651, pruned_loss=0.07953, over 4782.00 frames. ], tot_loss[loss=0.2191, simple_loss=0.2787, pruned_loss=0.07973, over 950414.51 frames. ], batch size: 29, lr: 3.94e-03, grad_scale: 32.0 2023-04-26 16:33:53,365 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-26 16:34:02,999 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-26 16:34:28,767 INFO [finetune.py:976] (4/7) Epoch 5, batch 2700, loss[loss=0.1864, simple_loss=0.2499, pruned_loss=0.0615, over 4889.00 frames. ], tot_loss[loss=0.2159, simple_loss=0.2765, pruned_loss=0.0776, over 951385.93 frames. ], batch size: 35, lr: 3.94e-03, grad_scale: 32.0 2023-04-26 16:34:28,915 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4474, 1.7630, 2.2021, 2.8795, 2.1277, 1.6850, 1.5703, 2.0518], device='cuda:4'), covar=tensor([0.3951, 0.4587, 0.2095, 0.3580, 0.4153, 0.3586, 0.5520, 0.3763], device='cuda:4'), in_proj_covar=tensor([0.0272, 0.0259, 0.0219, 0.0328, 0.0219, 0.0228, 0.0244, 0.0195], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 16:34:48,614 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.304e+02 1.757e+02 2.118e+02 2.619e+02 4.731e+02, threshold=4.237e+02, percent-clipped=2.0 2023-04-26 16:35:20,266 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=25651.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 16:35:37,124 INFO [finetune.py:976] (4/7) Epoch 5, batch 2750, loss[loss=0.1977, simple_loss=0.2524, pruned_loss=0.0715, over 4866.00 frames. ], tot_loss[loss=0.2133, simple_loss=0.2731, pruned_loss=0.07672, over 949521.77 frames. ], batch size: 31, lr: 3.94e-03, grad_scale: 32.0 2023-04-26 16:36:19,572 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.8729, 2.7869, 2.2132, 3.2969, 2.8682, 2.8768, 1.1146, 2.8117], device='cuda:4'), covar=tensor([0.1911, 0.1558, 0.3154, 0.2791, 0.2813, 0.2162, 0.5660, 0.2742], device='cuda:4'), in_proj_covar=tensor([0.0247, 0.0222, 0.0255, 0.0310, 0.0303, 0.0255, 0.0273, 0.0275], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 16:36:24,102 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=25699.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:36:32,334 INFO [finetune.py:976] (4/7) Epoch 5, batch 2800, loss[loss=0.1752, simple_loss=0.236, pruned_loss=0.05717, over 4763.00 frames. ], tot_loss[loss=0.21, simple_loss=0.2696, pruned_loss=0.07524, over 951720.27 frames. ], batch size: 27, lr: 3.93e-03, grad_scale: 32.0 2023-04-26 16:36:47,384 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.154e+02 1.821e+02 2.084e+02 2.532e+02 5.696e+02, threshold=4.167e+02, percent-clipped=4.0 2023-04-26 16:37:19,756 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-26 16:37:32,911 INFO [finetune.py:976] (4/7) Epoch 5, batch 2850, loss[loss=0.1484, simple_loss=0.2184, pruned_loss=0.03925, over 4715.00 frames. ], tot_loss[loss=0.209, simple_loss=0.2682, pruned_loss=0.07487, over 950452.75 frames. ], batch size: 23, lr: 3.93e-03, grad_scale: 32.0 2023-04-26 16:37:47,000 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=25784.0, num_to_drop=1, layers_to_drop={3} 2023-04-26 16:38:06,695 INFO [finetune.py:976] (4/7) Epoch 5, batch 2900, loss[loss=0.233, simple_loss=0.3015, pruned_loss=0.08224, over 4839.00 frames. ], tot_loss[loss=0.2106, simple_loss=0.27, pruned_loss=0.07563, over 949952.39 frames. ], batch size: 49, lr: 3.93e-03, grad_scale: 32.0 2023-04-26 16:38:15,790 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.320e+02 1.753e+02 2.187e+02 2.584e+02 6.163e+02, threshold=4.374e+02, percent-clipped=2.0 2023-04-26 16:38:30,896 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8012, 2.2026, 0.8991, 1.1271, 1.5111, 1.1167, 2.4280, 1.2943], device='cuda:4'), covar=tensor([0.0775, 0.0624, 0.0710, 0.1287, 0.0487, 0.1049, 0.0306, 0.0770], device='cuda:4'), in_proj_covar=tensor([0.0054, 0.0071, 0.0053, 0.0049, 0.0054, 0.0055, 0.0083, 0.0053], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 16:38:37,781 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=25858.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 16:38:39,945 INFO [finetune.py:976] (4/7) Epoch 5, batch 2950, loss[loss=0.2359, simple_loss=0.2977, pruned_loss=0.08699, over 4906.00 frames. ], tot_loss[loss=0.214, simple_loss=0.2742, pruned_loss=0.07688, over 950206.38 frames. ], batch size: 37, lr: 3.93e-03, grad_scale: 32.0 2023-04-26 16:39:12,519 INFO [finetune.py:976] (4/7) Epoch 5, batch 3000, loss[loss=0.2186, simple_loss=0.2829, pruned_loss=0.07709, over 4869.00 frames. ], tot_loss[loss=0.2159, simple_loss=0.2762, pruned_loss=0.07781, over 950299.26 frames. ], batch size: 31, lr: 3.93e-03, grad_scale: 32.0 2023-04-26 16:39:12,519 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-26 16:39:29,056 INFO [finetune.py:1010] (4/7) Epoch 5, validation: loss=0.1595, simple_loss=0.233, pruned_loss=0.04303, over 2265189.00 frames. 2023-04-26 16:39:29,057 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6259MB 2023-04-26 16:39:41,225 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=25919.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 16:39:51,735 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.259e+02 1.795e+02 2.151e+02 2.692e+02 4.010e+02, threshold=4.303e+02, percent-clipped=0.0 2023-04-26 16:39:52,580 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.60 vs. limit=2.0 2023-04-26 16:40:35,905 INFO [finetune.py:976] (4/7) Epoch 5, batch 3050, loss[loss=0.2405, simple_loss=0.301, pruned_loss=0.08999, over 4880.00 frames. ], tot_loss[loss=0.2151, simple_loss=0.276, pruned_loss=0.07712, over 950233.25 frames. ], batch size: 32, lr: 3.93e-03, grad_scale: 32.0 2023-04-26 16:41:08,651 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-26 16:41:32,850 INFO [finetune.py:976] (4/7) Epoch 5, batch 3100, loss[loss=0.1955, simple_loss=0.2507, pruned_loss=0.07016, over 4887.00 frames. ], tot_loss[loss=0.2129, simple_loss=0.2739, pruned_loss=0.07595, over 951036.19 frames. ], batch size: 32, lr: 3.93e-03, grad_scale: 32.0 2023-04-26 16:41:43,947 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.102e+02 1.748e+02 1.976e+02 2.307e+02 6.292e+02, threshold=3.952e+02, percent-clipped=1.0 2023-04-26 16:42:06,248 INFO [finetune.py:976] (4/7) Epoch 5, batch 3150, loss[loss=0.24, simple_loss=0.2826, pruned_loss=0.0987, over 4830.00 frames. ], tot_loss[loss=0.2098, simple_loss=0.27, pruned_loss=0.07479, over 950604.37 frames. ], batch size: 25, lr: 3.93e-03, grad_scale: 32.0 2023-04-26 16:42:18,188 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8371, 1.2595, 1.6306, 1.9136, 1.6537, 1.2839, 0.8945, 1.3394], device='cuda:4'), covar=tensor([0.4464, 0.5066, 0.2393, 0.3221, 0.4203, 0.3423, 0.5698, 0.3447], device='cuda:4'), in_proj_covar=tensor([0.0274, 0.0260, 0.0220, 0.0328, 0.0220, 0.0229, 0.0244, 0.0195], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 16:42:27,118 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=26084.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 16:42:46,455 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=26102.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:42:54,365 INFO [finetune.py:976] (4/7) Epoch 5, batch 3200, loss[loss=0.2213, simple_loss=0.2781, pruned_loss=0.08227, over 4855.00 frames. ], tot_loss[loss=0.2064, simple_loss=0.2663, pruned_loss=0.07321, over 952268.83 frames. ], batch size: 44, lr: 3.93e-03, grad_scale: 32.0 2023-04-26 16:43:09,264 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.183e+02 1.750e+02 2.081e+02 2.590e+02 4.901e+02, threshold=4.161e+02, percent-clipped=2.0 2023-04-26 16:43:14,415 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=26132.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 16:43:31,948 INFO [finetune.py:976] (4/7) Epoch 5, batch 3250, loss[loss=0.1935, simple_loss=0.2595, pruned_loss=0.06374, over 4761.00 frames. ], tot_loss[loss=0.2072, simple_loss=0.2668, pruned_loss=0.07375, over 950117.03 frames. ], batch size: 27, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 16:43:33,286 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=26163.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:43:37,344 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=26168.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:43:42,248 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1120, 1.3889, 1.3507, 1.4780, 1.4253, 1.6006, 1.4927, 1.4682], device='cuda:4'), covar=tensor([0.6710, 0.9178, 0.8563, 0.7798, 0.9264, 1.4148, 1.0059, 0.9236], device='cuda:4'), in_proj_covar=tensor([0.0314, 0.0395, 0.0316, 0.0326, 0.0346, 0.0410, 0.0378, 0.0335], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 16:43:58,987 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.11 vs. limit=5.0 2023-04-26 16:44:05,485 INFO [finetune.py:976] (4/7) Epoch 5, batch 3300, loss[loss=0.2282, simple_loss=0.2744, pruned_loss=0.09106, over 4232.00 frames. ], tot_loss[loss=0.2114, simple_loss=0.2714, pruned_loss=0.07568, over 950269.24 frames. ], batch size: 18, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 16:44:07,398 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=26214.0, num_to_drop=1, layers_to_drop={3} 2023-04-26 16:44:16,103 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.343e+02 1.793e+02 2.255e+02 2.677e+02 5.857e+02, threshold=4.510e+02, percent-clipped=2.0 2023-04-26 16:44:17,431 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=26229.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:44:32,048 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=26250.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:44:38,662 INFO [finetune.py:976] (4/7) Epoch 5, batch 3350, loss[loss=0.2448, simple_loss=0.3066, pruned_loss=0.09146, over 4843.00 frames. ], tot_loss[loss=0.2142, simple_loss=0.2743, pruned_loss=0.07699, over 949343.76 frames. ], batch size: 47, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 16:44:44,944 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.98 vs. limit=5.0 2023-04-26 16:45:50,830 INFO [finetune.py:976] (4/7) Epoch 5, batch 3400, loss[loss=0.1636, simple_loss=0.242, pruned_loss=0.04255, over 4803.00 frames. ], tot_loss[loss=0.2141, simple_loss=0.2748, pruned_loss=0.07671, over 949379.95 frames. ], batch size: 51, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 16:45:50,977 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=26311.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:46:13,227 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.088e+02 1.738e+02 2.082e+02 2.437e+02 3.720e+02, threshold=4.164e+02, percent-clipped=0.0 2023-04-26 16:46:56,946 INFO [finetune.py:976] (4/7) Epoch 5, batch 3450, loss[loss=0.2568, simple_loss=0.2998, pruned_loss=0.1069, over 4917.00 frames. ], tot_loss[loss=0.2133, simple_loss=0.2747, pruned_loss=0.07596, over 950697.31 frames. ], batch size: 38, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 16:48:03,178 INFO [finetune.py:976] (4/7) Epoch 5, batch 3500, loss[loss=0.2254, simple_loss=0.2847, pruned_loss=0.08308, over 4828.00 frames. ], tot_loss[loss=0.2116, simple_loss=0.2725, pruned_loss=0.07533, over 950546.88 frames. ], batch size: 38, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 16:48:05,087 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=26414.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:48:25,050 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.307e+02 1.713e+02 2.095e+02 2.561e+02 6.592e+02, threshold=4.191e+02, percent-clipped=2.0 2023-04-26 16:48:39,277 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6516, 1.9802, 1.6183, 1.8702, 1.6013, 1.5758, 1.6687, 1.4222], device='cuda:4'), covar=tensor([0.2492, 0.1889, 0.1148, 0.1640, 0.3954, 0.1710, 0.2146, 0.2678], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0327, 0.0235, 0.0298, 0.0317, 0.0281, 0.0269, 0.0291], device='cuda:4'), out_proj_covar=tensor([1.2527e-04, 1.3291e-04, 9.5617e-05, 1.1990e-04, 1.3060e-04, 1.1366e-04, 1.1049e-04, 1.1760e-04], device='cuda:4') 2023-04-26 16:48:49,904 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.56 vs. limit=2.0 2023-04-26 16:48:50,383 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.3303, 3.3045, 2.4081, 3.9090, 3.3619, 3.3538, 1.1843, 3.3312], device='cuda:4'), covar=tensor([0.1849, 0.1298, 0.3012, 0.2043, 0.2875, 0.2041, 0.6474, 0.2480], device='cuda:4'), in_proj_covar=tensor([0.0249, 0.0224, 0.0258, 0.0314, 0.0308, 0.0259, 0.0278, 0.0282], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 16:49:08,574 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=26458.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:49:10,352 INFO [finetune.py:976] (4/7) Epoch 5, batch 3550, loss[loss=0.1473, simple_loss=0.2171, pruned_loss=0.03878, over 4900.00 frames. ], tot_loss[loss=0.2081, simple_loss=0.2688, pruned_loss=0.07368, over 950257.70 frames. ], batch size: 36, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 16:49:24,647 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=26475.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:49:49,836 INFO [finetune.py:976] (4/7) Epoch 5, batch 3600, loss[loss=0.216, simple_loss=0.2617, pruned_loss=0.08513, over 4835.00 frames. ], tot_loss[loss=0.2052, simple_loss=0.2655, pruned_loss=0.07242, over 951003.42 frames. ], batch size: 33, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 16:49:51,771 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=26514.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 16:49:57,837 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=26524.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:49:59,594 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.194e+02 1.718e+02 2.043e+02 2.547e+02 4.174e+02, threshold=4.086e+02, percent-clipped=0.0 2023-04-26 16:50:06,240 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0251, 2.4858, 1.0236, 1.3109, 1.9575, 1.2253, 3.3026, 1.6972], device='cuda:4'), covar=tensor([0.0679, 0.0748, 0.0848, 0.1261, 0.0521, 0.1049, 0.0234, 0.0658], device='cuda:4'), in_proj_covar=tensor([0.0054, 0.0070, 0.0052, 0.0049, 0.0054, 0.0055, 0.0083, 0.0053], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 16:50:11,955 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-04-26 16:50:23,342 INFO [finetune.py:976] (4/7) Epoch 5, batch 3650, loss[loss=0.2467, simple_loss=0.3159, pruned_loss=0.08878, over 4840.00 frames. ], tot_loss[loss=0.2082, simple_loss=0.2685, pruned_loss=0.07398, over 952645.21 frames. ], batch size: 49, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 16:50:24,020 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=26562.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 16:50:32,497 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5745, 3.2238, 1.3943, 1.9785, 1.9460, 2.4846, 1.9738, 1.3164], device='cuda:4'), covar=tensor([0.1178, 0.0760, 0.1569, 0.1104, 0.0970, 0.0884, 0.1293, 0.1995], device='cuda:4'), in_proj_covar=tensor([0.0122, 0.0261, 0.0146, 0.0128, 0.0138, 0.0160, 0.0125, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 16:50:53,161 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=26606.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:50:57,092 INFO [finetune.py:976] (4/7) Epoch 5, batch 3700, loss[loss=0.2143, simple_loss=0.2873, pruned_loss=0.0706, over 4831.00 frames. ], tot_loss[loss=0.2143, simple_loss=0.2746, pruned_loss=0.07704, over 952469.40 frames. ], batch size: 47, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 16:51:06,781 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.079e+02 1.827e+02 2.264e+02 2.659e+02 6.887e+02, threshold=4.529e+02, percent-clipped=2.0 2023-04-26 16:51:17,521 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.9947, 5.0971, 3.5641, 5.7722, 5.0528, 5.0923, 2.8452, 4.9799], device='cuda:4'), covar=tensor([0.1395, 0.0845, 0.2246, 0.0711, 0.2030, 0.1544, 0.4511, 0.1878], device='cuda:4'), in_proj_covar=tensor([0.0247, 0.0221, 0.0254, 0.0312, 0.0305, 0.0255, 0.0275, 0.0278], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 16:51:29,840 INFO [finetune.py:976] (4/7) Epoch 5, batch 3750, loss[loss=0.2072, simple_loss=0.2713, pruned_loss=0.07154, over 4819.00 frames. ], tot_loss[loss=0.2167, simple_loss=0.277, pruned_loss=0.0782, over 952869.03 frames. ], batch size: 30, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 16:52:20,490 INFO [finetune.py:976] (4/7) Epoch 5, batch 3800, loss[loss=0.2282, simple_loss=0.3066, pruned_loss=0.07492, over 4809.00 frames. ], tot_loss[loss=0.2165, simple_loss=0.2777, pruned_loss=0.07766, over 952250.29 frames. ], batch size: 38, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 16:52:21,235 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1859, 1.3799, 1.4681, 1.6177, 1.5399, 1.7366, 1.5738, 1.5967], device='cuda:4'), covar=tensor([0.7684, 1.0634, 0.9603, 0.8916, 1.0336, 1.4658, 1.0996, 1.0027], device='cuda:4'), in_proj_covar=tensor([0.0315, 0.0396, 0.0319, 0.0327, 0.0348, 0.0412, 0.0379, 0.0336], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 16:52:31,669 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.252e+02 1.856e+02 2.180e+02 2.581e+02 5.516e+02, threshold=4.361e+02, percent-clipped=1.0 2023-04-26 16:52:52,597 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=26758.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:52:54,307 INFO [finetune.py:976] (4/7) Epoch 5, batch 3850, loss[loss=0.2195, simple_loss=0.2723, pruned_loss=0.0833, over 4722.00 frames. ], tot_loss[loss=0.2143, simple_loss=0.2759, pruned_loss=0.07634, over 953067.24 frames. ], batch size: 59, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 16:53:00,825 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=26770.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:53:41,437 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=26806.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:53:44,435 INFO [finetune.py:976] (4/7) Epoch 5, batch 3900, loss[loss=0.1847, simple_loss=0.2358, pruned_loss=0.06687, over 4758.00 frames. ], tot_loss[loss=0.2111, simple_loss=0.2721, pruned_loss=0.07503, over 954626.14 frames. ], batch size: 59, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 16:54:03,739 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=26824.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:54:05,466 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.093e+02 1.758e+02 2.108e+02 2.612e+02 6.373e+02, threshold=4.215e+02, percent-clipped=3.0 2023-04-26 16:54:12,381 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7200, 1.2537, 4.5656, 4.2484, 4.0855, 4.3249, 4.2702, 4.0055], device='cuda:4'), covar=tensor([0.6592, 0.6170, 0.0963, 0.1620, 0.1047, 0.1653, 0.1244, 0.1545], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0306, 0.0415, 0.0419, 0.0354, 0.0407, 0.0321, 0.0376], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-26 16:54:50,051 INFO [finetune.py:976] (4/7) Epoch 5, batch 3950, loss[loss=0.1969, simple_loss=0.2527, pruned_loss=0.07058, over 4766.00 frames. ], tot_loss[loss=0.2092, simple_loss=0.2695, pruned_loss=0.07447, over 955867.49 frames. ], batch size: 54, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 16:54:58,571 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.57 vs. limit=5.0 2023-04-26 16:55:08,487 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=26872.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:55:20,212 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0593, 2.6317, 0.9962, 1.2679, 2.0141, 1.2192, 3.6804, 1.7217], device='cuda:4'), covar=tensor([0.0755, 0.0784, 0.0920, 0.1496, 0.0591, 0.1143, 0.0333, 0.0754], device='cuda:4'), in_proj_covar=tensor([0.0054, 0.0070, 0.0052, 0.0049, 0.0054, 0.0055, 0.0082, 0.0053], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 16:55:40,405 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7196, 1.9642, 1.5918, 1.3693, 1.2996, 1.3144, 1.6258, 1.2142], device='cuda:4'), covar=tensor([0.2118, 0.1813, 0.1956, 0.2314, 0.3048, 0.2466, 0.1466, 0.2546], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0220, 0.0178, 0.0208, 0.0214, 0.0188, 0.0170, 0.0195], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 16:55:47,039 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=26904.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 16:55:48,208 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=26906.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:55:51,740 INFO [finetune.py:976] (4/7) Epoch 5, batch 4000, loss[loss=0.2404, simple_loss=0.313, pruned_loss=0.0839, over 4865.00 frames. ], tot_loss[loss=0.2094, simple_loss=0.2689, pruned_loss=0.07491, over 953344.61 frames. ], batch size: 44, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 16:56:02,926 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.248e+02 1.675e+02 2.014e+02 2.454e+02 6.687e+02, threshold=4.029e+02, percent-clipped=1.0 2023-04-26 16:56:17,513 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=26950.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:56:19,908 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=26954.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:56:19,988 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7409, 1.2621, 1.3224, 1.3619, 1.9745, 1.5537, 1.2177, 1.2830], device='cuda:4'), covar=tensor([0.1535, 0.1632, 0.2194, 0.1467, 0.0863, 0.1700, 0.2135, 0.1977], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0332, 0.0349, 0.0306, 0.0343, 0.0339, 0.0307, 0.0349], device='cuda:4'), out_proj_covar=tensor([6.6843e-05, 7.1064e-05, 7.5588e-05, 6.4042e-05, 7.2591e-05, 7.3724e-05, 6.6641e-05, 7.5246e-05], device='cuda:4') 2023-04-26 16:56:25,066 INFO [finetune.py:976] (4/7) Epoch 5, batch 4050, loss[loss=0.2156, simple_loss=0.2704, pruned_loss=0.08045, over 4854.00 frames. ], tot_loss[loss=0.2128, simple_loss=0.2727, pruned_loss=0.07645, over 953455.70 frames. ], batch size: 31, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 16:56:28,127 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=26965.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 16:56:29,378 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.6591, 1.7471, 1.5338, 1.3614, 1.7882, 1.4851, 2.1296, 1.3197], device='cuda:4'), covar=tensor([0.4085, 0.1512, 0.5311, 0.2820, 0.1770, 0.2398, 0.1664, 0.4738], device='cuda:4'), in_proj_covar=tensor([0.0353, 0.0356, 0.0440, 0.0371, 0.0400, 0.0388, 0.0395, 0.0424], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 16:56:58,984 INFO [finetune.py:976] (4/7) Epoch 5, batch 4100, loss[loss=0.189, simple_loss=0.2598, pruned_loss=0.0591, over 4921.00 frames. ], tot_loss[loss=0.2152, simple_loss=0.2758, pruned_loss=0.07734, over 952960.33 frames. ], batch size: 38, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 16:56:59,092 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=27011.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:57:20,560 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.204e+02 1.859e+02 2.211e+02 2.694e+02 5.784e+02, threshold=4.421e+02, percent-clipped=3.0 2023-04-26 16:57:41,500 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5091, 3.7119, 0.9687, 1.9008, 2.0015, 2.4219, 2.0002, 1.0680], device='cuda:4'), covar=tensor([0.1502, 0.1036, 0.2246, 0.1431, 0.1158, 0.1218, 0.1624, 0.2120], device='cuda:4'), in_proj_covar=tensor([0.0120, 0.0257, 0.0144, 0.0127, 0.0137, 0.0158, 0.0123, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 16:57:50,516 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9633, 1.7169, 2.1650, 2.3067, 1.7265, 1.3306, 1.9325, 1.1343], device='cuda:4'), covar=tensor([0.0792, 0.1088, 0.0764, 0.0921, 0.1089, 0.1651, 0.1050, 0.1264], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0076, 0.0075, 0.0069, 0.0079, 0.0096, 0.0082, 0.0078], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 16:58:04,488 INFO [finetune.py:976] (4/7) Epoch 5, batch 4150, loss[loss=0.2204, simple_loss=0.2825, pruned_loss=0.07916, over 4907.00 frames. ], tot_loss[loss=0.217, simple_loss=0.2776, pruned_loss=0.07821, over 953431.31 frames. ], batch size: 36, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 16:58:16,080 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=27070.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:58:47,720 INFO [finetune.py:976] (4/7) Epoch 5, batch 4200, loss[loss=0.2059, simple_loss=0.2733, pruned_loss=0.0693, over 4897.00 frames. ], tot_loss[loss=0.2155, simple_loss=0.2769, pruned_loss=0.07708, over 954720.66 frames. ], batch size: 37, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 16:58:53,014 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=27118.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:58:58,902 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.181e+01 1.721e+02 2.038e+02 2.503e+02 4.432e+02, threshold=4.077e+02, percent-clipped=1.0 2023-04-26 16:59:13,620 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=27148.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:59:17,271 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5247, 1.4736, 4.1202, 3.8662, 3.6989, 3.8520, 3.8438, 3.6074], device='cuda:4'), covar=tensor([0.6352, 0.5386, 0.0981, 0.1576, 0.1061, 0.1430, 0.1537, 0.1481], device='cuda:4'), in_proj_covar=tensor([0.0315, 0.0309, 0.0419, 0.0423, 0.0357, 0.0410, 0.0322, 0.0379], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-26 16:59:21,419 INFO [finetune.py:976] (4/7) Epoch 5, batch 4250, loss[loss=0.151, simple_loss=0.2209, pruned_loss=0.04052, over 4778.00 frames. ], tot_loss[loss=0.214, simple_loss=0.2749, pruned_loss=0.07651, over 953690.01 frames. ], batch size: 28, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 16:59:29,613 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.49 vs. limit=2.0 2023-04-26 16:59:54,273 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=27209.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 16:59:55,343 INFO [finetune.py:976] (4/7) Epoch 5, batch 4300, loss[loss=0.1998, simple_loss=0.2709, pruned_loss=0.0643, over 4829.00 frames. ], tot_loss[loss=0.2097, simple_loss=0.2707, pruned_loss=0.07435, over 955152.92 frames. ], batch size: 40, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 17:00:17,881 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.057e+02 1.798e+02 2.267e+02 2.763e+02 5.468e+02, threshold=4.535e+02, percent-clipped=5.0 2023-04-26 17:01:04,150 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=27260.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 17:01:04,678 INFO [finetune.py:976] (4/7) Epoch 5, batch 4350, loss[loss=0.2339, simple_loss=0.2815, pruned_loss=0.09318, over 4908.00 frames. ], tot_loss[loss=0.2073, simple_loss=0.2678, pruned_loss=0.07344, over 955800.30 frames. ], batch size: 35, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 17:01:11,504 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.7784, 3.7369, 2.6974, 4.3883, 3.8717, 3.7698, 1.6820, 3.7405], device='cuda:4'), covar=tensor([0.1890, 0.1292, 0.3302, 0.1872, 0.3244, 0.2118, 0.6136, 0.2346], device='cuda:4'), in_proj_covar=tensor([0.0250, 0.0223, 0.0256, 0.0314, 0.0306, 0.0256, 0.0277, 0.0279], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 17:01:46,638 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=27306.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:01:49,558 INFO [finetune.py:976] (4/7) Epoch 5, batch 4400, loss[loss=0.1655, simple_loss=0.2363, pruned_loss=0.04735, over 4745.00 frames. ], tot_loss[loss=0.2087, simple_loss=0.2687, pruned_loss=0.07436, over 956627.35 frames. ], batch size: 27, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 17:02:00,134 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.042e+02 1.703e+02 2.157e+02 2.468e+02 4.465e+02, threshold=4.314e+02, percent-clipped=0.0 2023-04-26 17:02:23,094 INFO [finetune.py:976] (4/7) Epoch 5, batch 4450, loss[loss=0.1968, simple_loss=0.27, pruned_loss=0.06178, over 4722.00 frames. ], tot_loss[loss=0.2127, simple_loss=0.2731, pruned_loss=0.07612, over 956266.80 frames. ], batch size: 59, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 17:02:28,762 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2540, 2.9680, 1.0137, 1.5479, 2.1572, 1.6114, 4.1516, 2.0656], device='cuda:4'), covar=tensor([0.0702, 0.0879, 0.0914, 0.1383, 0.0572, 0.0974, 0.0291, 0.0637], device='cuda:4'), in_proj_covar=tensor([0.0053, 0.0070, 0.0052, 0.0049, 0.0053, 0.0054, 0.0082, 0.0053], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 17:02:56,797 INFO [finetune.py:976] (4/7) Epoch 5, batch 4500, loss[loss=0.2287, simple_loss=0.2911, pruned_loss=0.08314, over 4897.00 frames. ], tot_loss[loss=0.2126, simple_loss=0.2737, pruned_loss=0.0757, over 957071.84 frames. ], batch size: 36, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 17:03:17,441 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.259e+02 1.778e+02 2.173e+02 2.607e+02 4.915e+02, threshold=4.346e+02, percent-clipped=1.0 2023-04-26 17:04:02,850 INFO [finetune.py:976] (4/7) Epoch 5, batch 4550, loss[loss=0.224, simple_loss=0.2778, pruned_loss=0.08505, over 4737.00 frames. ], tot_loss[loss=0.2152, simple_loss=0.2763, pruned_loss=0.07703, over 956797.50 frames. ], batch size: 23, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 17:04:31,478 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=27504.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:04:34,458 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3938, 3.0819, 1.0072, 1.5295, 2.1325, 1.4623, 4.1221, 2.0154], device='cuda:4'), covar=tensor([0.0670, 0.0791, 0.0915, 0.1327, 0.0583, 0.1037, 0.0210, 0.0633], device='cuda:4'), in_proj_covar=tensor([0.0054, 0.0070, 0.0052, 0.0049, 0.0054, 0.0055, 0.0082, 0.0053], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 17:04:36,154 INFO [finetune.py:976] (4/7) Epoch 5, batch 4600, loss[loss=0.1995, simple_loss=0.2714, pruned_loss=0.06379, over 4913.00 frames. ], tot_loss[loss=0.2144, simple_loss=0.2756, pruned_loss=0.07663, over 957344.91 frames. ], batch size: 37, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 17:04:46,279 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.146e+02 1.803e+02 2.105e+02 2.477e+02 5.679e+02, threshold=4.210e+02, percent-clipped=3.0 2023-04-26 17:05:08,954 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=27560.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 17:05:09,029 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-26 17:05:09,449 INFO [finetune.py:976] (4/7) Epoch 5, batch 4650, loss[loss=0.1935, simple_loss=0.2569, pruned_loss=0.065, over 4864.00 frames. ], tot_loss[loss=0.2112, simple_loss=0.2718, pruned_loss=0.07535, over 956752.94 frames. ], batch size: 31, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 17:05:26,678 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-26 17:05:40,079 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=27606.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:05:41,302 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=27608.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 17:05:43,542 INFO [finetune.py:976] (4/7) Epoch 5, batch 4700, loss[loss=0.1775, simple_loss=0.219, pruned_loss=0.06799, over 4163.00 frames. ], tot_loss[loss=0.2074, simple_loss=0.2677, pruned_loss=0.07354, over 956749.60 frames. ], batch size: 18, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 17:05:54,152 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.191e+02 1.680e+02 2.058e+02 2.502e+02 5.893e+02, threshold=4.117e+02, percent-clipped=4.0 2023-04-26 17:06:23,333 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=27654.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:06:34,204 INFO [finetune.py:976] (4/7) Epoch 5, batch 4750, loss[loss=0.2426, simple_loss=0.2984, pruned_loss=0.09336, over 4854.00 frames. ], tot_loss[loss=0.2073, simple_loss=0.2669, pruned_loss=0.07382, over 956129.87 frames. ], batch size: 49, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 17:06:43,373 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4256, 1.7090, 1.4974, 1.9093, 1.7461, 2.0528, 1.5621, 3.1797], device='cuda:4'), covar=tensor([0.0602, 0.0637, 0.0682, 0.0937, 0.0528, 0.0637, 0.0652, 0.0174], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0040, 0.0041, 0.0045, 0.0041, 0.0041, 0.0040, 0.0061], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 17:06:55,908 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5447, 1.7156, 1.3504, 1.0196, 1.2361, 1.2002, 1.3057, 1.1462], device='cuda:4'), covar=tensor([0.1990, 0.1533, 0.1943, 0.2273, 0.2835, 0.2199, 0.1439, 0.2356], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0219, 0.0178, 0.0208, 0.0214, 0.0187, 0.0170, 0.0194], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 17:06:57,114 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4429, 1.8438, 1.5719, 1.7518, 1.3435, 1.5457, 1.6161, 1.3227], device='cuda:4'), covar=tensor([0.2254, 0.1607, 0.1317, 0.1706, 0.3516, 0.1486, 0.1961, 0.2581], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0324, 0.0235, 0.0299, 0.0319, 0.0279, 0.0267, 0.0291], device='cuda:4'), out_proj_covar=tensor([1.2485e-04, 1.3170e-04, 9.5839e-05, 1.2031e-04, 1.3119e-04, 1.1277e-04, 1.0994e-04, 1.1759e-04], device='cuda:4') 2023-04-26 17:07:06,420 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0523, 0.6449, 0.9389, 0.7078, 1.2404, 0.8914, 0.7669, 0.9575], device='cuda:4'), covar=tensor([0.1588, 0.1602, 0.1989, 0.1487, 0.0965, 0.1455, 0.1616, 0.1902], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0331, 0.0349, 0.0305, 0.0341, 0.0337, 0.0306, 0.0349], device='cuda:4'), out_proj_covar=tensor([6.6531e-05, 7.0886e-05, 7.5435e-05, 6.3757e-05, 7.2149e-05, 7.3106e-05, 6.6441e-05, 7.5128e-05], device='cuda:4') 2023-04-26 17:07:40,901 INFO [finetune.py:976] (4/7) Epoch 5, batch 4800, loss[loss=0.2888, simple_loss=0.3455, pruned_loss=0.1161, over 4866.00 frames. ], tot_loss[loss=0.2095, simple_loss=0.2696, pruned_loss=0.07471, over 956786.56 frames. ], batch size: 44, lr: 3.93e-03, grad_scale: 16.0 2023-04-26 17:07:48,589 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=27714.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 17:08:02,772 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.360e+02 1.880e+02 2.179e+02 2.503e+02 4.628e+02, threshold=4.358e+02, percent-clipped=2.0 2023-04-26 17:08:33,584 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4971, 1.6848, 0.6914, 1.2428, 1.8143, 1.3632, 1.2907, 1.3316], device='cuda:4'), covar=tensor([0.0572, 0.0407, 0.0454, 0.0629, 0.0303, 0.0620, 0.0583, 0.0659], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0026, 0.0024, 0.0030, 0.0021, 0.0030, 0.0029, 0.0031], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0049, 0.0038, 0.0048, 0.0047, 0.0050], device='cuda:4') 2023-04-26 17:08:41,288 INFO [finetune.py:976] (4/7) Epoch 5, batch 4850, loss[loss=0.2275, simple_loss=0.3001, pruned_loss=0.0774, over 4869.00 frames. ], tot_loss[loss=0.2105, simple_loss=0.2716, pruned_loss=0.07467, over 957223.80 frames. ], batch size: 34, lr: 3.92e-03, grad_scale: 16.0 2023-04-26 17:08:41,970 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.9671, 3.9506, 2.8875, 4.6442, 4.0565, 4.0496, 1.9544, 3.9446], device='cuda:4'), covar=tensor([0.1570, 0.1005, 0.3043, 0.1247, 0.3131, 0.1641, 0.5454, 0.2359], device='cuda:4'), in_proj_covar=tensor([0.0248, 0.0223, 0.0256, 0.0315, 0.0306, 0.0256, 0.0277, 0.0278], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 17:08:50,936 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=27775.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 17:09:01,190 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8812, 2.3484, 1.9789, 2.3147, 1.8317, 2.0805, 2.1000, 1.5867], device='cuda:4'), covar=tensor([0.2419, 0.1461, 0.1048, 0.1383, 0.3058, 0.1328, 0.2044, 0.3018], device='cuda:4'), in_proj_covar=tensor([0.0306, 0.0325, 0.0235, 0.0299, 0.0318, 0.0279, 0.0268, 0.0290], device='cuda:4'), out_proj_covar=tensor([1.2453e-04, 1.3185e-04, 9.5839e-05, 1.2020e-04, 1.3086e-04, 1.1297e-04, 1.1022e-04, 1.1734e-04], device='cuda:4') 2023-04-26 17:09:14,393 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=27804.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:09:18,625 INFO [finetune.py:976] (4/7) Epoch 5, batch 4900, loss[loss=0.1736, simple_loss=0.2404, pruned_loss=0.05337, over 4672.00 frames. ], tot_loss[loss=0.2124, simple_loss=0.2737, pruned_loss=0.07556, over 957845.57 frames. ], batch size: 23, lr: 3.92e-03, grad_scale: 16.0 2023-04-26 17:09:23,908 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.7609, 3.5583, 2.7638, 4.2741, 3.5549, 3.7246, 1.7386, 3.7151], device='cuda:4'), covar=tensor([0.1693, 0.1273, 0.3478, 0.1311, 0.2659, 0.1665, 0.5054, 0.2010], device='cuda:4'), in_proj_covar=tensor([0.0249, 0.0223, 0.0256, 0.0316, 0.0307, 0.0258, 0.0279, 0.0279], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 17:09:30,227 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.253e+02 1.856e+02 2.309e+02 2.750e+02 8.138e+02, threshold=4.618e+02, percent-clipped=6.0 2023-04-26 17:09:35,316 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=5.32 vs. limit=5.0 2023-04-26 17:09:46,105 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=27852.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:09:52,541 INFO [finetune.py:976] (4/7) Epoch 5, batch 4950, loss[loss=0.2442, simple_loss=0.3002, pruned_loss=0.09409, over 4147.00 frames. ], tot_loss[loss=0.2123, simple_loss=0.2741, pruned_loss=0.07523, over 956458.50 frames. ], batch size: 65, lr: 3.92e-03, grad_scale: 16.0 2023-04-26 17:09:55,141 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.0314, 2.2135, 2.0095, 2.2381, 1.9891, 2.3320, 2.1369, 2.1242], device='cuda:4'), covar=tensor([0.6044, 1.1064, 1.0155, 0.7688, 0.9759, 1.2054, 1.2206, 1.0634], device='cuda:4'), in_proj_covar=tensor([0.0315, 0.0397, 0.0319, 0.0327, 0.0348, 0.0414, 0.0379, 0.0334], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 17:10:31,676 INFO [finetune.py:976] (4/7) Epoch 5, batch 5000, loss[loss=0.2126, simple_loss=0.2817, pruned_loss=0.07179, over 4840.00 frames. ], tot_loss[loss=0.2098, simple_loss=0.2715, pruned_loss=0.07406, over 953544.16 frames. ], batch size: 47, lr: 3.92e-03, grad_scale: 16.0 2023-04-26 17:10:42,356 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.076e+02 1.656e+02 1.985e+02 2.433e+02 4.076e+02, threshold=3.970e+02, percent-clipped=0.0 2023-04-26 17:10:54,668 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-04-26 17:11:04,342 INFO [finetune.py:976] (4/7) Epoch 5, batch 5050, loss[loss=0.2409, simple_loss=0.2945, pruned_loss=0.0936, over 4843.00 frames. ], tot_loss[loss=0.2082, simple_loss=0.2693, pruned_loss=0.07348, over 954465.22 frames. ], batch size: 47, lr: 3.92e-03, grad_scale: 16.0 2023-04-26 17:12:05,265 INFO [finetune.py:976] (4/7) Epoch 5, batch 5100, loss[loss=0.1672, simple_loss=0.2249, pruned_loss=0.05474, over 4772.00 frames. ], tot_loss[loss=0.2074, simple_loss=0.2679, pruned_loss=0.07343, over 954781.93 frames. ], batch size: 26, lr: 3.92e-03, grad_scale: 16.0 2023-04-26 17:12:27,654 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 1.777e+02 2.079e+02 2.348e+02 6.496e+02, threshold=4.158e+02, percent-clipped=5.0 2023-04-26 17:13:12,919 INFO [finetune.py:976] (4/7) Epoch 5, batch 5150, loss[loss=0.2358, simple_loss=0.2966, pruned_loss=0.08745, over 4210.00 frames. ], tot_loss[loss=0.2076, simple_loss=0.2679, pruned_loss=0.07369, over 954741.50 frames. ], batch size: 65, lr: 3.92e-03, grad_scale: 16.0 2023-04-26 17:13:31,489 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=28070.0, num_to_drop=1, layers_to_drop={3} 2023-04-26 17:13:44,368 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.6754, 4.7134, 3.3002, 5.4602, 4.8671, 4.7218, 2.2603, 4.8194], device='cuda:4'), covar=tensor([0.1714, 0.1144, 0.3148, 0.0977, 0.3908, 0.1777, 0.5580, 0.1975], device='cuda:4'), in_proj_covar=tensor([0.0249, 0.0223, 0.0257, 0.0316, 0.0307, 0.0259, 0.0279, 0.0279], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 17:13:54,543 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.36 vs. limit=5.0 2023-04-26 17:14:09,046 INFO [finetune.py:976] (4/7) Epoch 5, batch 5200, loss[loss=0.2211, simple_loss=0.2927, pruned_loss=0.07475, over 4806.00 frames. ], tot_loss[loss=0.2105, simple_loss=0.2714, pruned_loss=0.07483, over 952896.95 frames. ], batch size: 45, lr: 3.92e-03, grad_scale: 16.0 2023-04-26 17:14:12,144 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9313, 1.7123, 2.2756, 2.3399, 1.8437, 1.4723, 2.0631, 1.1250], device='cuda:4'), covar=tensor([0.0770, 0.0975, 0.0695, 0.0939, 0.0959, 0.1617, 0.0961, 0.1262], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0076, 0.0075, 0.0069, 0.0080, 0.0097, 0.0082, 0.0078], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 17:14:19,886 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.281e+02 1.900e+02 2.150e+02 2.501e+02 5.102e+02, threshold=4.301e+02, percent-clipped=1.0 2023-04-26 17:14:21,783 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-26 17:14:40,618 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1848, 3.1903, 0.9330, 1.6249, 1.6744, 2.1986, 1.8398, 1.0192], device='cuda:4'), covar=tensor([0.1755, 0.1413, 0.2248, 0.1856, 0.1343, 0.1440, 0.1675, 0.2297], device='cuda:4'), in_proj_covar=tensor([0.0122, 0.0263, 0.0146, 0.0129, 0.0139, 0.0161, 0.0124, 0.0128], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 17:14:42,336 INFO [finetune.py:976] (4/7) Epoch 5, batch 5250, loss[loss=0.2189, simple_loss=0.2917, pruned_loss=0.07307, over 4812.00 frames. ], tot_loss[loss=0.2137, simple_loss=0.275, pruned_loss=0.07626, over 953448.23 frames. ], batch size: 30, lr: 3.92e-03, grad_scale: 32.0 2023-04-26 17:15:16,326 INFO [finetune.py:976] (4/7) Epoch 5, batch 5300, loss[loss=0.1735, simple_loss=0.2523, pruned_loss=0.04734, over 4818.00 frames. ], tot_loss[loss=0.2149, simple_loss=0.276, pruned_loss=0.0769, over 954251.04 frames. ], batch size: 39, lr: 3.92e-03, grad_scale: 32.0 2023-04-26 17:15:27,005 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.419e+02 1.781e+02 2.085e+02 2.383e+02 5.799e+02, threshold=4.171e+02, percent-clipped=2.0 2023-04-26 17:15:28,333 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=28229.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:15:46,208 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=28255.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 17:15:49,729 INFO [finetune.py:976] (4/7) Epoch 5, batch 5350, loss[loss=0.2035, simple_loss=0.2609, pruned_loss=0.07306, over 4878.00 frames. ], tot_loss[loss=0.2145, simple_loss=0.2759, pruned_loss=0.07658, over 954404.69 frames. ], batch size: 31, lr: 3.92e-03, grad_scale: 32.0 2023-04-26 17:16:07,726 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=28287.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:16:09,994 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=28290.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:16:23,681 INFO [finetune.py:976] (4/7) Epoch 5, batch 5400, loss[loss=0.1808, simple_loss=0.2496, pruned_loss=0.05605, over 4926.00 frames. ], tot_loss[loss=0.2103, simple_loss=0.272, pruned_loss=0.07435, over 954453.42 frames. ], batch size: 33, lr: 3.92e-03, grad_scale: 32.0 2023-04-26 17:16:27,321 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=28316.0, num_to_drop=1, layers_to_drop={3} 2023-04-26 17:16:34,348 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.191e+02 1.689e+02 2.005e+02 2.420e+02 4.760e+02, threshold=4.009e+02, percent-clipped=1.0 2023-04-26 17:16:49,199 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=28348.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:16:55,843 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-04-26 17:16:57,452 INFO [finetune.py:976] (4/7) Epoch 5, batch 5450, loss[loss=0.2114, simple_loss=0.2684, pruned_loss=0.07723, over 4861.00 frames. ], tot_loss[loss=0.2095, simple_loss=0.2695, pruned_loss=0.07472, over 955063.84 frames. ], batch size: 31, lr: 3.92e-03, grad_scale: 32.0 2023-04-26 17:17:01,821 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3438, 1.4971, 3.8724, 3.6412, 3.4588, 3.7309, 3.7288, 3.4346], device='cuda:4'), covar=tensor([0.6809, 0.5387, 0.1150, 0.1527, 0.1118, 0.1757, 0.1277, 0.1521], device='cuda:4'), in_proj_covar=tensor([0.0316, 0.0309, 0.0419, 0.0423, 0.0358, 0.0413, 0.0321, 0.0379], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-26 17:17:03,557 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=28370.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 17:17:04,195 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5172, 3.0101, 0.9694, 1.6533, 2.1628, 1.4612, 4.3882, 2.1404], device='cuda:4'), covar=tensor([0.0651, 0.0994, 0.0919, 0.1378, 0.0598, 0.1034, 0.0219, 0.0640], device='cuda:4'), in_proj_covar=tensor([0.0054, 0.0070, 0.0052, 0.0049, 0.0054, 0.0054, 0.0082, 0.0053], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 17:17:22,347 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.45 vs. limit=5.0 2023-04-26 17:17:37,439 INFO [finetune.py:976] (4/7) Epoch 5, batch 5500, loss[loss=0.2136, simple_loss=0.2778, pruned_loss=0.0747, over 4804.00 frames. ], tot_loss[loss=0.206, simple_loss=0.2659, pruned_loss=0.07302, over 955972.98 frames. ], batch size: 51, lr: 3.92e-03, grad_scale: 32.0 2023-04-26 17:17:47,037 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=28418.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 17:17:47,059 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0658, 2.4401, 0.8588, 1.2825, 1.4356, 1.7617, 1.5886, 0.8223], device='cuda:4'), covar=tensor([0.1859, 0.1799, 0.2174, 0.1995, 0.1470, 0.1361, 0.1887, 0.1823], device='cuda:4'), in_proj_covar=tensor([0.0124, 0.0267, 0.0148, 0.0130, 0.0141, 0.0164, 0.0126, 0.0130], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 17:17:59,136 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.103e+02 1.711e+02 2.116e+02 2.529e+02 5.388e+02, threshold=4.231e+02, percent-clipped=4.0 2023-04-26 17:18:08,682 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5314, 1.2884, 1.6231, 1.9400, 1.6891, 1.5040, 1.5843, 1.6196], device='cuda:4'), covar=tensor([0.9651, 1.2138, 1.2684, 1.3906, 1.0314, 1.4874, 1.4757, 1.2590], device='cuda:4'), in_proj_covar=tensor([0.0421, 0.0454, 0.0538, 0.0560, 0.0450, 0.0472, 0.0485, 0.0486], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 17:18:11,212 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.74 vs. limit=2.0 2023-04-26 17:18:49,577 INFO [finetune.py:976] (4/7) Epoch 5, batch 5550, loss[loss=0.2842, simple_loss=0.3358, pruned_loss=0.1163, over 4740.00 frames. ], tot_loss[loss=0.2088, simple_loss=0.2688, pruned_loss=0.07442, over 956530.67 frames. ], batch size: 54, lr: 3.92e-03, grad_scale: 32.0 2023-04-26 17:19:09,789 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=28486.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:19:14,463 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=28492.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:19:25,623 INFO [finetune.py:976] (4/7) Epoch 5, batch 5600, loss[loss=0.2672, simple_loss=0.3161, pruned_loss=0.1091, over 4794.00 frames. ], tot_loss[loss=0.2117, simple_loss=0.2724, pruned_loss=0.07547, over 954863.35 frames. ], batch size: 25, lr: 3.92e-03, grad_scale: 32.0 2023-04-26 17:19:34,908 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.156e+02 1.757e+02 2.198e+02 2.682e+02 4.335e+02, threshold=4.395e+02, percent-clipped=1.0 2023-04-26 17:19:47,087 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=28547.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:19:51,014 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=28553.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:19:55,615 INFO [finetune.py:976] (4/7) Epoch 5, batch 5650, loss[loss=0.2521, simple_loss=0.3103, pruned_loss=0.09701, over 4729.00 frames. ], tot_loss[loss=0.2139, simple_loss=0.2752, pruned_loss=0.07634, over 954137.84 frames. ], batch size: 59, lr: 3.92e-03, grad_scale: 32.0 2023-04-26 17:19:55,689 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0156, 2.4797, 1.0861, 1.3399, 2.0457, 1.1393, 3.3271, 1.6610], device='cuda:4'), covar=tensor([0.0723, 0.0662, 0.0873, 0.1404, 0.0510, 0.1124, 0.0236, 0.0715], device='cuda:4'), in_proj_covar=tensor([0.0054, 0.0070, 0.0052, 0.0049, 0.0054, 0.0054, 0.0082, 0.0053], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 17:20:09,689 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=28585.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:20:25,139 INFO [finetune.py:976] (4/7) Epoch 5, batch 5700, loss[loss=0.1632, simple_loss=0.2234, pruned_loss=0.05153, over 4250.00 frames. ], tot_loss[loss=0.2105, simple_loss=0.2709, pruned_loss=0.07503, over 940268.06 frames. ], batch size: 18, lr: 3.92e-03, grad_scale: 32.0 2023-04-26 17:20:25,177 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=28611.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 17:20:34,603 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.220e+02 1.668e+02 2.021e+02 2.528e+02 3.624e+02, threshold=4.042e+02, percent-clipped=0.0 2023-04-26 17:20:58,935 INFO [finetune.py:976] (4/7) Epoch 6, batch 0, loss[loss=0.176, simple_loss=0.2455, pruned_loss=0.05321, over 4748.00 frames. ], tot_loss[loss=0.176, simple_loss=0.2455, pruned_loss=0.05321, over 4748.00 frames. ], batch size: 27, lr: 3.92e-03, grad_scale: 32.0 2023-04-26 17:20:58,935 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-26 17:21:02,329 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2904, 1.7121, 1.4987, 1.8259, 1.6415, 1.9028, 1.5019, 2.8815], device='cuda:4'), covar=tensor([0.0642, 0.0684, 0.0657, 0.1045, 0.0567, 0.0461, 0.0647, 0.0213], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0040, 0.0041, 0.0045, 0.0041, 0.0040, 0.0040, 0.0061], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 17:21:13,392 INFO [finetune.py:1010] (4/7) Epoch 6, validation: loss=0.1605, simple_loss=0.2337, pruned_loss=0.04366, over 2265189.00 frames. 2023-04-26 17:21:13,392 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6259MB 2023-04-26 17:21:19,203 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=28643.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:21:19,249 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6613, 1.6282, 1.7934, 1.9916, 2.0593, 1.6021, 1.1637, 1.7698], device='cuda:4'), covar=tensor([0.0924, 0.1120, 0.0690, 0.0653, 0.0563, 0.0954, 0.1074, 0.0634], device='cuda:4'), in_proj_covar=tensor([0.0202, 0.0204, 0.0181, 0.0177, 0.0177, 0.0192, 0.0165, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 17:21:21,107 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0004, 1.8762, 2.1355, 2.3577, 2.3649, 1.8485, 1.4414, 2.0333], device='cuda:4'), covar=tensor([0.0945, 0.1067, 0.0640, 0.0661, 0.0622, 0.0977, 0.1038, 0.0648], device='cuda:4'), in_proj_covar=tensor([0.0202, 0.0204, 0.0181, 0.0177, 0.0177, 0.0192, 0.0165, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 17:21:39,898 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5584, 1.6086, 0.8126, 1.2520, 1.6844, 1.4582, 1.3935, 1.4209], device='cuda:4'), covar=tensor([0.0567, 0.0412, 0.0415, 0.0603, 0.0308, 0.0597, 0.0538, 0.0636], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0026, 0.0024, 0.0030, 0.0021, 0.0030, 0.0029, 0.0031], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0049, 0.0037, 0.0048, 0.0047, 0.0050], device='cuda:4') 2023-04-26 17:21:59,786 INFO [finetune.py:976] (4/7) Epoch 6, batch 50, loss[loss=0.2287, simple_loss=0.2919, pruned_loss=0.08272, over 4905.00 frames. ], tot_loss[loss=0.2159, simple_loss=0.2784, pruned_loss=0.07669, over 217362.07 frames. ], batch size: 46, lr: 3.92e-03, grad_scale: 32.0 2023-04-26 17:22:11,071 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1020, 1.5925, 1.3680, 1.7287, 1.5301, 2.3134, 1.3083, 3.6362], device='cuda:4'), covar=tensor([0.0705, 0.0810, 0.0844, 0.1214, 0.0714, 0.0434, 0.0779, 0.0133], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0039, 0.0041, 0.0045, 0.0041, 0.0040, 0.0040, 0.0061], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 17:22:24,839 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.026e+02 1.869e+02 2.241e+02 2.636e+02 5.025e+02, threshold=4.483e+02, percent-clipped=7.0 2023-04-26 17:22:33,641 INFO [finetune.py:976] (4/7) Epoch 6, batch 100, loss[loss=0.2223, simple_loss=0.2847, pruned_loss=0.07996, over 4848.00 frames. ], tot_loss[loss=0.2065, simple_loss=0.269, pruned_loss=0.07205, over 382523.45 frames. ], batch size: 44, lr: 3.92e-03, grad_scale: 32.0 2023-04-26 17:22:43,757 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1486, 1.6345, 1.4411, 1.7490, 1.5459, 2.2272, 1.3602, 3.5690], device='cuda:4'), covar=tensor([0.0704, 0.0753, 0.0773, 0.1159, 0.0649, 0.0468, 0.0739, 0.0133], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0039, 0.0041, 0.0045, 0.0041, 0.0040, 0.0040, 0.0061], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 17:22:58,423 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.45 vs. limit=2.0 2023-04-26 17:23:06,831 INFO [finetune.py:976] (4/7) Epoch 6, batch 150, loss[loss=0.1934, simple_loss=0.2598, pruned_loss=0.06349, over 4905.00 frames. ], tot_loss[loss=0.2058, simple_loss=0.2658, pruned_loss=0.0729, over 509449.48 frames. ], batch size: 37, lr: 3.92e-03, grad_scale: 32.0 2023-04-26 17:23:07,011 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.92 vs. limit=5.0 2023-04-26 17:23:36,889 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.113e+02 1.735e+02 2.162e+02 2.514e+02 4.413e+02, threshold=4.325e+02, percent-clipped=0.0 2023-04-26 17:23:57,267 INFO [finetune.py:976] (4/7) Epoch 6, batch 200, loss[loss=0.2068, simple_loss=0.265, pruned_loss=0.07432, over 4827.00 frames. ], tot_loss[loss=0.209, simple_loss=0.2672, pruned_loss=0.07534, over 608010.82 frames. ], batch size: 38, lr: 3.92e-03, grad_scale: 32.0 2023-04-26 17:23:59,203 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=28842.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:24:08,147 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=28848.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:24:22,285 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=28861.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:24:55,323 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=28885.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:25:03,860 INFO [finetune.py:976] (4/7) Epoch 6, batch 250, loss[loss=0.2376, simple_loss=0.3009, pruned_loss=0.08722, over 4808.00 frames. ], tot_loss[loss=0.2088, simple_loss=0.2683, pruned_loss=0.07465, over 684594.01 frames. ], batch size: 39, lr: 3.92e-03, grad_scale: 16.0 2023-04-26 17:25:29,050 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=28911.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 17:25:40,825 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=28922.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:25:44,146 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2291, 2.9651, 0.9373, 1.5833, 1.6192, 2.1650, 1.7227, 1.0187], device='cuda:4'), covar=tensor([0.1455, 0.1007, 0.1916, 0.1409, 0.1158, 0.1038, 0.1510, 0.1871], device='cuda:4'), in_proj_covar=tensor([0.0121, 0.0261, 0.0146, 0.0128, 0.0138, 0.0160, 0.0124, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 17:25:46,466 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.358e+02 1.791e+02 2.229e+02 2.957e+02 8.157e+02, threshold=4.458e+02, percent-clipped=7.0 2023-04-26 17:25:54,903 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=28933.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:25:59,504 INFO [finetune.py:976] (4/7) Epoch 6, batch 300, loss[loss=0.2137, simple_loss=0.2852, pruned_loss=0.07112, over 4825.00 frames. ], tot_loss[loss=0.2131, simple_loss=0.2735, pruned_loss=0.07635, over 746392.80 frames. ], batch size: 39, lr: 3.92e-03, grad_scale: 16.0 2023-04-26 17:26:08,045 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=28943.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:26:28,685 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=28959.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 17:27:05,111 INFO [finetune.py:976] (4/7) Epoch 6, batch 350, loss[loss=0.2382, simple_loss=0.3089, pruned_loss=0.08376, over 4814.00 frames. ], tot_loss[loss=0.2137, simple_loss=0.2749, pruned_loss=0.07624, over 793804.81 frames. ], batch size: 39, lr: 3.92e-03, grad_scale: 16.0 2023-04-26 17:27:11,879 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=28991.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:27:24,955 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=29002.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:27:53,124 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.162e+02 1.828e+02 2.192e+02 2.525e+02 3.973e+02, threshold=4.385e+02, percent-clipped=0.0 2023-04-26 17:27:59,802 INFO [finetune.py:976] (4/7) Epoch 6, batch 400, loss[loss=0.172, simple_loss=0.2433, pruned_loss=0.05039, over 4789.00 frames. ], tot_loss[loss=0.2142, simple_loss=0.2759, pruned_loss=0.07631, over 830397.55 frames. ], batch size: 29, lr: 3.92e-03, grad_scale: 16.0 2023-04-26 17:28:01,770 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.79 vs. limit=2.0 2023-04-26 17:28:16,909 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=29063.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:28:32,185 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.81 vs. limit=5.0 2023-04-26 17:28:33,172 INFO [finetune.py:976] (4/7) Epoch 6, batch 450, loss[loss=0.224, simple_loss=0.2859, pruned_loss=0.08103, over 4821.00 frames. ], tot_loss[loss=0.212, simple_loss=0.2737, pruned_loss=0.0751, over 858368.41 frames. ], batch size: 33, lr: 3.92e-03, grad_scale: 16.0 2023-04-26 17:28:59,674 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.119e+02 1.667e+02 1.998e+02 2.325e+02 3.966e+02, threshold=3.996e+02, percent-clipped=0.0 2023-04-26 17:29:06,386 INFO [finetune.py:976] (4/7) Epoch 6, batch 500, loss[loss=0.2329, simple_loss=0.2944, pruned_loss=0.08568, over 4829.00 frames. ], tot_loss[loss=0.2088, simple_loss=0.2701, pruned_loss=0.07376, over 879146.13 frames. ], batch size: 39, lr: 3.92e-03, grad_scale: 16.0 2023-04-26 17:29:08,313 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=29142.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:29:13,407 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=29148.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:29:13,804 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.75 vs. limit=2.0 2023-04-26 17:29:37,378 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6207, 1.2438, 1.3516, 1.3191, 1.8704, 1.4687, 1.2113, 1.2573], device='cuda:4'), covar=tensor([0.1459, 0.1369, 0.1643, 0.1346, 0.0685, 0.1375, 0.1931, 0.1849], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0328, 0.0347, 0.0304, 0.0338, 0.0333, 0.0304, 0.0347], device='cuda:4'), out_proj_covar=tensor([6.6162e-05, 6.9969e-05, 7.4965e-05, 6.3463e-05, 7.1518e-05, 7.2190e-05, 6.5961e-05, 7.4830e-05], device='cuda:4') 2023-04-26 17:29:39,681 INFO [finetune.py:976] (4/7) Epoch 6, batch 550, loss[loss=0.1627, simple_loss=0.2239, pruned_loss=0.05077, over 4823.00 frames. ], tot_loss[loss=0.206, simple_loss=0.2667, pruned_loss=0.07265, over 897648.75 frames. ], batch size: 25, lr: 3.92e-03, grad_scale: 16.0 2023-04-26 17:29:40,346 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=29190.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:29:44,457 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=29196.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:29:52,487 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-26 17:29:59,122 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=29217.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:30:06,230 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.252e+02 1.833e+02 2.160e+02 2.601e+02 5.201e+02, threshold=4.320e+02, percent-clipped=1.0 2023-04-26 17:30:11,157 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=29236.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:30:12,890 INFO [finetune.py:976] (4/7) Epoch 6, batch 600, loss[loss=0.2096, simple_loss=0.2611, pruned_loss=0.07908, over 4707.00 frames. ], tot_loss[loss=0.2073, simple_loss=0.268, pruned_loss=0.07332, over 909262.98 frames. ], batch size: 23, lr: 3.92e-03, grad_scale: 16.0 2023-04-26 17:30:30,677 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5845, 1.8110, 0.9979, 1.3858, 2.0243, 1.5237, 1.4508, 1.5177], device='cuda:4'), covar=tensor([0.0531, 0.0384, 0.0368, 0.0564, 0.0288, 0.0547, 0.0527, 0.0582], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0026, 0.0024, 0.0030, 0.0021, 0.0030, 0.0029, 0.0031], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0045, 0.0038, 0.0049, 0.0037, 0.0048, 0.0047, 0.0050], device='cuda:4') 2023-04-26 17:30:31,761 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.8446, 4.7425, 3.3121, 5.5859, 4.8618, 4.9484, 2.1250, 4.7857], device='cuda:4'), covar=tensor([0.1289, 0.0837, 0.2748, 0.0875, 0.3919, 0.1527, 0.5720, 0.2127], device='cuda:4'), in_proj_covar=tensor([0.0246, 0.0219, 0.0254, 0.0311, 0.0305, 0.0256, 0.0275, 0.0276], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 17:30:47,032 INFO [finetune.py:976] (4/7) Epoch 6, batch 650, loss[loss=0.2008, simple_loss=0.2678, pruned_loss=0.06688, over 4834.00 frames. ], tot_loss[loss=0.2099, simple_loss=0.2712, pruned_loss=0.07424, over 920040.93 frames. ], batch size: 33, lr: 3.92e-03, grad_scale: 16.0 2023-04-26 17:30:57,626 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=29297.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:31:42,433 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.237e+02 1.829e+02 2.148e+02 2.649e+02 3.979e+02, threshold=4.296e+02, percent-clipped=0.0 2023-04-26 17:32:01,093 INFO [finetune.py:976] (4/7) Epoch 6, batch 700, loss[loss=0.2082, simple_loss=0.2739, pruned_loss=0.07123, over 4913.00 frames. ], tot_loss[loss=0.2104, simple_loss=0.2724, pruned_loss=0.0742, over 929152.18 frames. ], batch size: 42, lr: 3.92e-03, grad_scale: 16.0 2023-04-26 17:32:25,507 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=29358.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:33:07,831 INFO [finetune.py:976] (4/7) Epoch 6, batch 750, loss[loss=0.2527, simple_loss=0.3018, pruned_loss=0.1018, over 4885.00 frames. ], tot_loss[loss=0.2112, simple_loss=0.273, pruned_loss=0.07468, over 934754.31 frames. ], batch size: 32, lr: 3.92e-03, grad_scale: 16.0 2023-04-26 17:33:11,532 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5810, 1.9103, 0.8429, 1.4429, 2.1204, 1.5868, 1.5735, 1.5158], device='cuda:4'), covar=tensor([0.0543, 0.0398, 0.0390, 0.0583, 0.0270, 0.0595, 0.0539, 0.0631], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0026, 0.0024, 0.0030, 0.0021, 0.0030, 0.0029, 0.0031], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0049, 0.0037, 0.0048, 0.0047, 0.0050], device='cuda:4') 2023-04-26 17:33:12,186 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5264, 0.9311, 1.2040, 1.0485, 1.7181, 1.2951, 1.0030, 1.2226], device='cuda:4'), covar=tensor([0.1416, 0.1714, 0.2034, 0.1463, 0.0822, 0.1605, 0.2087, 0.1928], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0331, 0.0351, 0.0307, 0.0341, 0.0336, 0.0306, 0.0351], device='cuda:4'), out_proj_covar=tensor([6.6674e-05, 7.0771e-05, 7.5908e-05, 6.4122e-05, 7.2096e-05, 7.2793e-05, 6.6338e-05, 7.5628e-05], device='cuda:4') 2023-04-26 17:33:42,224 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.8672, 1.9166, 1.7889, 1.4987, 2.0458, 1.6675, 2.6132, 1.4989], device='cuda:4'), covar=tensor([0.4065, 0.1803, 0.5543, 0.3507, 0.1913, 0.2389, 0.1384, 0.5060], device='cuda:4'), in_proj_covar=tensor([0.0352, 0.0359, 0.0440, 0.0371, 0.0399, 0.0388, 0.0393, 0.0424], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 17:34:03,539 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.229e+02 1.738e+02 2.111e+02 2.599e+02 4.246e+02, threshold=4.221e+02, percent-clipped=0.0 2023-04-26 17:34:14,521 INFO [finetune.py:976] (4/7) Epoch 6, batch 800, loss[loss=0.1839, simple_loss=0.2499, pruned_loss=0.05902, over 4793.00 frames. ], tot_loss[loss=0.2109, simple_loss=0.2732, pruned_loss=0.07432, over 941275.44 frames. ], batch size: 29, lr: 3.92e-03, grad_scale: 16.0 2023-04-26 17:34:48,530 INFO [finetune.py:976] (4/7) Epoch 6, batch 850, loss[loss=0.1866, simple_loss=0.2574, pruned_loss=0.05794, over 4785.00 frames. ], tot_loss[loss=0.2078, simple_loss=0.2701, pruned_loss=0.07274, over 943441.86 frames. ], batch size: 28, lr: 3.92e-03, grad_scale: 16.0 2023-04-26 17:35:08,009 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=29517.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:35:14,515 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.189e+02 1.686e+02 1.967e+02 2.412e+02 4.596e+02, threshold=3.934e+02, percent-clipped=4.0 2023-04-26 17:35:22,137 INFO [finetune.py:976] (4/7) Epoch 6, batch 900, loss[loss=0.153, simple_loss=0.22, pruned_loss=0.043, over 4827.00 frames. ], tot_loss[loss=0.2055, simple_loss=0.2671, pruned_loss=0.07194, over 946629.10 frames. ], batch size: 30, lr: 3.91e-03, grad_scale: 16.0 2023-04-26 17:35:30,015 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=29551.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:35:38,986 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=29565.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:35:56,022 INFO [finetune.py:976] (4/7) Epoch 6, batch 950, loss[loss=0.2033, simple_loss=0.2639, pruned_loss=0.0714, over 4905.00 frames. ], tot_loss[loss=0.2041, simple_loss=0.2653, pruned_loss=0.07147, over 948432.99 frames. ], batch size: 35, lr: 3.91e-03, grad_scale: 16.0 2023-04-26 17:35:57,878 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=29592.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:36:10,621 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=29612.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:36:15,252 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.24 vs. limit=5.0 2023-04-26 17:36:27,225 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.239e+02 1.864e+02 2.273e+02 2.659e+02 4.416e+02, threshold=4.547e+02, percent-clipped=3.0 2023-04-26 17:36:39,505 INFO [finetune.py:976] (4/7) Epoch 6, batch 1000, loss[loss=0.2491, simple_loss=0.3229, pruned_loss=0.08769, over 4847.00 frames. ], tot_loss[loss=0.2069, simple_loss=0.2683, pruned_loss=0.07273, over 949978.52 frames. ], batch size: 49, lr: 3.91e-03, grad_scale: 16.0 2023-04-26 17:36:46,066 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-26 17:37:09,273 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=29658.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:37:11,732 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4197, 1.1248, 0.3208, 1.1132, 1.1633, 1.2713, 1.1594, 1.1588], device='cuda:4'), covar=tensor([0.0615, 0.0464, 0.0535, 0.0665, 0.0356, 0.0642, 0.0631, 0.0712], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0026, 0.0024, 0.0030, 0.0021, 0.0030, 0.0029, 0.0031], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0045, 0.0038, 0.0049, 0.0038, 0.0048, 0.0048, 0.0050], device='cuda:4') 2023-04-26 17:37:11,743 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=29662.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:37:12,963 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1318, 1.5939, 1.3549, 1.6915, 1.4954, 1.7530, 1.3522, 3.3873], device='cuda:4'), covar=tensor([0.0666, 0.0759, 0.0795, 0.1184, 0.0654, 0.0641, 0.0784, 0.0159], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0040, 0.0041, 0.0045, 0.0041, 0.0040, 0.0040, 0.0061], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 17:37:52,136 INFO [finetune.py:976] (4/7) Epoch 6, batch 1050, loss[loss=0.2107, simple_loss=0.2805, pruned_loss=0.07046, over 4803.00 frames. ], tot_loss[loss=0.2089, simple_loss=0.271, pruned_loss=0.07338, over 953076.24 frames. ], batch size: 45, lr: 3.91e-03, grad_scale: 16.0 2023-04-26 17:37:55,322 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5888, 1.1660, 1.3600, 1.2547, 1.7881, 1.4534, 1.1915, 1.3376], device='cuda:4'), covar=tensor([0.1702, 0.1526, 0.1950, 0.1683, 0.1030, 0.1630, 0.2318, 0.2013], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0333, 0.0352, 0.0308, 0.0343, 0.0337, 0.0307, 0.0353], device='cuda:4'), out_proj_covar=tensor([6.6993e-05, 7.1131e-05, 7.6340e-05, 6.4350e-05, 7.2431e-05, 7.3096e-05, 6.6587e-05, 7.6092e-05], device='cuda:4') 2023-04-26 17:38:04,013 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.6791, 3.6028, 2.7097, 4.2541, 3.6638, 3.7552, 1.6970, 3.6320], device='cuda:4'), covar=tensor([0.1816, 0.1180, 0.3113, 0.1842, 0.3161, 0.1862, 0.5568, 0.2518], device='cuda:4'), in_proj_covar=tensor([0.0245, 0.0218, 0.0254, 0.0311, 0.0303, 0.0255, 0.0274, 0.0275], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 17:38:06,513 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2022, 1.3314, 1.2350, 1.5650, 1.4124, 1.5461, 1.2613, 2.4415], device='cuda:4'), covar=tensor([0.0613, 0.0855, 0.0812, 0.1294, 0.0677, 0.0547, 0.0806, 0.0236], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0040, 0.0041, 0.0045, 0.0041, 0.0040, 0.0040, 0.0061], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 17:38:13,535 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=29706.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:38:34,960 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=29723.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 17:38:39,285 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.169e+02 1.744e+02 2.054e+02 2.475e+02 7.840e+02, threshold=4.108e+02, percent-clipped=2.0 2023-04-26 17:38:57,682 INFO [finetune.py:976] (4/7) Epoch 6, batch 1100, loss[loss=0.2017, simple_loss=0.2808, pruned_loss=0.0613, over 4905.00 frames. ], tot_loss[loss=0.2102, simple_loss=0.2727, pruned_loss=0.07381, over 954775.32 frames. ], batch size: 36, lr: 3.91e-03, grad_scale: 16.0 2023-04-26 17:39:58,705 INFO [finetune.py:976] (4/7) Epoch 6, batch 1150, loss[loss=0.2107, simple_loss=0.2736, pruned_loss=0.07385, over 4840.00 frames. ], tot_loss[loss=0.2096, simple_loss=0.2725, pruned_loss=0.07342, over 953867.16 frames. ], batch size: 30, lr: 3.91e-03, grad_scale: 16.0 2023-04-26 17:40:07,751 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=29802.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:40:24,363 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.056e+02 1.784e+02 2.105e+02 2.513e+02 8.942e+02, threshold=4.210e+02, percent-clipped=4.0 2023-04-26 17:40:31,970 INFO [finetune.py:976] (4/7) Epoch 6, batch 1200, loss[loss=0.1899, simple_loss=0.2555, pruned_loss=0.06208, over 4795.00 frames. ], tot_loss[loss=0.207, simple_loss=0.2696, pruned_loss=0.07224, over 954748.41 frames. ], batch size: 45, lr: 3.91e-03, grad_scale: 16.0 2023-04-26 17:40:48,154 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=29863.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:41:05,314 INFO [finetune.py:976] (4/7) Epoch 6, batch 1250, loss[loss=0.1849, simple_loss=0.2554, pruned_loss=0.05717, over 4821.00 frames. ], tot_loss[loss=0.2052, simple_loss=0.2668, pruned_loss=0.0718, over 954380.14 frames. ], batch size: 41, lr: 3.91e-03, grad_scale: 16.0 2023-04-26 17:41:07,189 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=29892.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:41:17,267 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=29907.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:41:19,234 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-04-26 17:41:24,592 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-26 17:41:30,522 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.076e+02 1.701e+02 1.937e+02 2.233e+02 5.324e+02, threshold=3.875e+02, percent-clipped=2.0 2023-04-26 17:41:38,730 INFO [finetune.py:976] (4/7) Epoch 6, batch 1300, loss[loss=0.1898, simple_loss=0.2527, pruned_loss=0.06345, over 4747.00 frames. ], tot_loss[loss=0.2027, simple_loss=0.264, pruned_loss=0.07069, over 954671.00 frames. ], batch size: 27, lr: 3.91e-03, grad_scale: 16.0 2023-04-26 17:41:39,345 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=29940.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:42:12,109 INFO [finetune.py:976] (4/7) Epoch 6, batch 1350, loss[loss=0.2034, simple_loss=0.2604, pruned_loss=0.07319, over 4708.00 frames. ], tot_loss[loss=0.2027, simple_loss=0.2641, pruned_loss=0.07069, over 955892.30 frames. ], batch size: 23, lr: 3.91e-03, grad_scale: 16.0 2023-04-26 17:42:32,672 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=30018.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 17:42:39,142 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.331e+02 1.856e+02 2.103e+02 2.497e+02 6.080e+02, threshold=4.207e+02, percent-clipped=3.0 2023-04-26 17:42:52,098 INFO [finetune.py:976] (4/7) Epoch 6, batch 1400, loss[loss=0.1939, simple_loss=0.2727, pruned_loss=0.05751, over 4903.00 frames. ], tot_loss[loss=0.2051, simple_loss=0.2669, pruned_loss=0.07159, over 953181.02 frames. ], batch size: 37, lr: 3.91e-03, grad_scale: 16.0 2023-04-26 17:43:32,886 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-26 17:43:56,757 INFO [finetune.py:976] (4/7) Epoch 6, batch 1450, loss[loss=0.1684, simple_loss=0.2364, pruned_loss=0.0502, over 4790.00 frames. ], tot_loss[loss=0.2067, simple_loss=0.2691, pruned_loss=0.07217, over 954972.82 frames. ], batch size: 26, lr: 3.91e-03, grad_scale: 16.0 2023-04-26 17:44:14,547 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.59 vs. limit=5.0 2023-04-26 17:44:42,748 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=30125.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:44:44,462 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.421e+02 1.906e+02 2.234e+02 2.642e+02 4.905e+02, threshold=4.468e+02, percent-clipped=1.0 2023-04-26 17:44:57,147 INFO [finetune.py:976] (4/7) Epoch 6, batch 1500, loss[loss=0.2399, simple_loss=0.291, pruned_loss=0.09441, over 4910.00 frames. ], tot_loss[loss=0.2098, simple_loss=0.2727, pruned_loss=0.07348, over 956389.28 frames. ], batch size: 37, lr: 3.91e-03, grad_scale: 16.0 2023-04-26 17:45:25,797 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=30158.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:45:59,231 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=30186.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:46:00,954 INFO [finetune.py:976] (4/7) Epoch 6, batch 1550, loss[loss=0.2039, simple_loss=0.2577, pruned_loss=0.07508, over 4863.00 frames. ], tot_loss[loss=0.2106, simple_loss=0.2735, pruned_loss=0.07387, over 954924.81 frames. ], batch size: 34, lr: 3.91e-03, grad_scale: 16.0 2023-04-26 17:46:30,716 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=30207.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:46:31,941 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=30209.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:46:42,511 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5640, 1.6790, 0.7411, 1.3330, 1.8351, 1.4822, 1.4354, 1.4109], device='cuda:4'), covar=tensor([0.0510, 0.0399, 0.0433, 0.0572, 0.0316, 0.0549, 0.0518, 0.0596], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0026, 0.0024, 0.0030, 0.0021, 0.0030, 0.0029, 0.0031], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0045, 0.0038, 0.0049, 0.0037, 0.0048, 0.0047, 0.0050], device='cuda:4') 2023-04-26 17:46:45,645 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.90 vs. limit=5.0 2023-04-26 17:46:56,315 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.224e+02 1.842e+02 2.143e+02 2.537e+02 6.404e+02, threshold=4.287e+02, percent-clipped=2.0 2023-04-26 17:47:09,505 INFO [finetune.py:976] (4/7) Epoch 6, batch 1600, loss[loss=0.1621, simple_loss=0.2368, pruned_loss=0.04364, over 4782.00 frames. ], tot_loss[loss=0.2074, simple_loss=0.2699, pruned_loss=0.07247, over 955512.42 frames. ], batch size: 29, lr: 3.91e-03, grad_scale: 16.0 2023-04-26 17:47:31,235 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=30255.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:47:41,445 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=30270.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:47:53,335 INFO [finetune.py:976] (4/7) Epoch 6, batch 1650, loss[loss=0.2017, simple_loss=0.2639, pruned_loss=0.06972, over 4896.00 frames. ], tot_loss[loss=0.2071, simple_loss=0.2687, pruned_loss=0.07276, over 955394.17 frames. ], batch size: 46, lr: 3.91e-03, grad_scale: 16.0 2023-04-26 17:48:13,366 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=30318.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 17:48:19,319 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.157e+02 1.688e+02 2.054e+02 2.477e+02 3.898e+02, threshold=4.108e+02, percent-clipped=0.0 2023-04-26 17:48:26,040 INFO [finetune.py:976] (4/7) Epoch 6, batch 1700, loss[loss=0.168, simple_loss=0.2428, pruned_loss=0.0466, over 4781.00 frames. ], tot_loss[loss=0.2052, simple_loss=0.266, pruned_loss=0.07217, over 955405.72 frames. ], batch size: 29, lr: 3.91e-03, grad_scale: 16.0 2023-04-26 17:48:51,272 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=30360.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:49:00,652 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=30366.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:49:26,051 INFO [finetune.py:976] (4/7) Epoch 6, batch 1750, loss[loss=0.2306, simple_loss=0.2981, pruned_loss=0.08152, over 4825.00 frames. ], tot_loss[loss=0.2073, simple_loss=0.2682, pruned_loss=0.07316, over 956808.03 frames. ], batch size: 39, lr: 3.91e-03, grad_scale: 16.0 2023-04-26 17:50:03,363 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=30421.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:50:08,045 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.357e+02 1.827e+02 2.242e+02 2.631e+02 5.410e+02, threshold=4.483e+02, percent-clipped=4.0 2023-04-26 17:50:20,188 INFO [finetune.py:976] (4/7) Epoch 6, batch 1800, loss[loss=0.2028, simple_loss=0.2702, pruned_loss=0.06764, over 4859.00 frames. ], tot_loss[loss=0.2094, simple_loss=0.271, pruned_loss=0.07386, over 956823.37 frames. ], batch size: 31, lr: 3.91e-03, grad_scale: 16.0 2023-04-26 17:50:43,034 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=30458.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:51:01,090 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.91 vs. limit=5.0 2023-04-26 17:51:09,062 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=30481.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:51:13,898 INFO [finetune.py:976] (4/7) Epoch 6, batch 1850, loss[loss=0.2185, simple_loss=0.2734, pruned_loss=0.08185, over 4796.00 frames. ], tot_loss[loss=0.2107, simple_loss=0.2722, pruned_loss=0.07456, over 954553.81 frames. ], batch size: 26, lr: 3.91e-03, grad_scale: 16.0 2023-04-26 17:51:24,865 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=30506.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:51:40,049 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.288e+02 1.791e+02 2.106e+02 2.463e+02 4.581e+02, threshold=4.213e+02, percent-clipped=1.0 2023-04-26 17:51:47,205 INFO [finetune.py:976] (4/7) Epoch 6, batch 1900, loss[loss=0.1918, simple_loss=0.2707, pruned_loss=0.05646, over 4799.00 frames. ], tot_loss[loss=0.21, simple_loss=0.272, pruned_loss=0.07396, over 953226.73 frames. ], batch size: 41, lr: 3.91e-03, grad_scale: 16.0 2023-04-26 17:51:53,524 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.85 vs. limit=2.0 2023-04-26 17:51:54,636 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0256, 2.4525, 0.9993, 1.2905, 1.8411, 1.2772, 3.3544, 1.6628], device='cuda:4'), covar=tensor([0.0749, 0.0694, 0.0878, 0.1442, 0.0543, 0.1072, 0.0379, 0.0707], device='cuda:4'), in_proj_covar=tensor([0.0054, 0.0070, 0.0052, 0.0049, 0.0053, 0.0054, 0.0081, 0.0052], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 17:52:06,948 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3724, 2.8926, 1.0534, 1.6316, 1.6820, 2.1705, 1.8295, 1.1026], device='cuda:4'), covar=tensor([0.1345, 0.1273, 0.1739, 0.1283, 0.1073, 0.1017, 0.1408, 0.1585], device='cuda:4'), in_proj_covar=tensor([0.0121, 0.0259, 0.0146, 0.0127, 0.0137, 0.0159, 0.0122, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 17:52:17,585 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=30565.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:52:34,610 INFO [finetune.py:976] (4/7) Epoch 6, batch 1950, loss[loss=0.1772, simple_loss=0.2438, pruned_loss=0.05531, over 4870.00 frames. ], tot_loss[loss=0.2062, simple_loss=0.2684, pruned_loss=0.07199, over 952746.68 frames. ], batch size: 31, lr: 3.91e-03, grad_scale: 16.0 2023-04-26 17:52:41,530 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.9194, 3.8886, 2.8169, 4.5561, 4.0398, 3.9554, 1.9025, 3.9085], device='cuda:4'), covar=tensor([0.1737, 0.1335, 0.3272, 0.1664, 0.2842, 0.2026, 0.5682, 0.2422], device='cuda:4'), in_proj_covar=tensor([0.0245, 0.0220, 0.0256, 0.0313, 0.0304, 0.0256, 0.0276, 0.0277], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 17:52:55,790 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7868, 2.3700, 1.9760, 2.2796, 1.7460, 1.9807, 2.0666, 1.6960], device='cuda:4'), covar=tensor([0.2562, 0.1919, 0.1063, 0.1662, 0.3406, 0.1478, 0.2060, 0.3047], device='cuda:4'), in_proj_covar=tensor([0.0305, 0.0325, 0.0236, 0.0299, 0.0318, 0.0280, 0.0266, 0.0291], device='cuda:4'), out_proj_covar=tensor([1.2414e-04, 1.3220e-04, 9.5765e-05, 1.1990e-04, 1.3091e-04, 1.1324e-04, 1.0910e-04, 1.1744e-04], device='cuda:4') 2023-04-26 17:53:00,334 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.015e+02 1.660e+02 1.977e+02 2.493e+02 4.247e+02, threshold=3.954e+02, percent-clipped=1.0 2023-04-26 17:53:08,042 INFO [finetune.py:976] (4/7) Epoch 6, batch 2000, loss[loss=0.1901, simple_loss=0.2453, pruned_loss=0.06749, over 4865.00 frames. ], tot_loss[loss=0.2055, simple_loss=0.267, pruned_loss=0.07198, over 952642.16 frames. ], batch size: 31, lr: 3.91e-03, grad_scale: 16.0 2023-04-26 17:53:17,878 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=30655.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:53:30,701 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1976, 1.6537, 1.4294, 1.8672, 1.6519, 2.1080, 1.4080, 3.5636], device='cuda:4'), covar=tensor([0.0697, 0.0804, 0.0798, 0.1182, 0.0681, 0.0501, 0.0780, 0.0171], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0039, 0.0041, 0.0045, 0.0041, 0.0040, 0.0040, 0.0061], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 17:53:41,352 INFO [finetune.py:976] (4/7) Epoch 6, batch 2050, loss[loss=0.1969, simple_loss=0.2528, pruned_loss=0.07051, over 4913.00 frames. ], tot_loss[loss=0.2011, simple_loss=0.2629, pruned_loss=0.0697, over 955318.40 frames. ], batch size: 37, lr: 3.91e-03, grad_scale: 16.0 2023-04-26 17:53:41,466 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6113, 1.5659, 0.5987, 1.3599, 1.7674, 1.5241, 1.4707, 1.4008], device='cuda:4'), covar=tensor([0.0563, 0.0434, 0.0450, 0.0602, 0.0287, 0.0579, 0.0568, 0.0657], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0027, 0.0024, 0.0030, 0.0021, 0.0030, 0.0029, 0.0031], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0045, 0.0038, 0.0049, 0.0037, 0.0048, 0.0047, 0.0050], device='cuda:4') 2023-04-26 17:53:58,937 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=30716.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:53:58,994 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=30716.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 17:54:12,372 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.098e+02 1.673e+02 1.948e+02 2.190e+02 5.524e+02, threshold=3.896e+02, percent-clipped=3.0 2023-04-26 17:54:24,211 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0156, 1.2860, 1.1132, 1.5702, 1.3257, 1.5155, 1.1924, 2.4339], device='cuda:4'), covar=tensor([0.0686, 0.0886, 0.0914, 0.1321, 0.0741, 0.0535, 0.0862, 0.0264], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0039, 0.0041, 0.0045, 0.0041, 0.0040, 0.0040, 0.0061], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 17:54:25,817 INFO [finetune.py:976] (4/7) Epoch 6, batch 2100, loss[loss=0.2157, simple_loss=0.2913, pruned_loss=0.07007, over 4817.00 frames. ], tot_loss[loss=0.2028, simple_loss=0.2646, pruned_loss=0.07045, over 956379.44 frames. ], batch size: 41, lr: 3.91e-03, grad_scale: 16.0 2023-04-26 17:54:54,265 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=30781.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:54:59,005 INFO [finetune.py:976] (4/7) Epoch 6, batch 2150, loss[loss=0.2371, simple_loss=0.2979, pruned_loss=0.08814, over 4814.00 frames. ], tot_loss[loss=0.2046, simple_loss=0.267, pruned_loss=0.07111, over 956800.49 frames. ], batch size: 51, lr: 3.91e-03, grad_scale: 16.0 2023-04-26 17:55:00,230 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5447, 1.3109, 1.8950, 1.7754, 1.3924, 1.1663, 1.4792, 1.0417], device='cuda:4'), covar=tensor([0.0662, 0.0990, 0.0479, 0.0765, 0.0897, 0.1407, 0.0754, 0.0903], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0075, 0.0074, 0.0068, 0.0078, 0.0095, 0.0081, 0.0077], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 17:55:21,841 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.93 vs. limit=2.0 2023-04-26 17:55:40,195 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 1.783e+02 2.191e+02 2.615e+02 4.573e+02, threshold=4.381e+02, percent-clipped=3.0 2023-04-26 17:55:40,854 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=30829.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:55:54,381 INFO [finetune.py:976] (4/7) Epoch 6, batch 2200, loss[loss=0.213, simple_loss=0.286, pruned_loss=0.07004, over 4890.00 frames. ], tot_loss[loss=0.2083, simple_loss=0.2709, pruned_loss=0.07289, over 957485.72 frames. ], batch size: 35, lr: 3.91e-03, grad_scale: 16.0 2023-04-26 17:56:14,050 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.34 vs. limit=5.0 2023-04-26 17:56:23,794 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=30865.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:56:51,897 INFO [finetune.py:976] (4/7) Epoch 6, batch 2250, loss[loss=0.1772, simple_loss=0.2388, pruned_loss=0.05781, over 4714.00 frames. ], tot_loss[loss=0.209, simple_loss=0.2713, pruned_loss=0.07333, over 956594.28 frames. ], batch size: 23, lr: 3.91e-03, grad_scale: 32.0 2023-04-26 17:57:11,902 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=30903.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:57:22,864 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=30913.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:57:40,804 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=30925.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:57:42,531 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.062e+02 1.673e+02 1.940e+02 2.404e+02 3.842e+02, threshold=3.880e+02, percent-clipped=0.0 2023-04-26 17:57:46,804 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=30934.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:57:50,255 INFO [finetune.py:976] (4/7) Epoch 6, batch 2300, loss[loss=0.1775, simple_loss=0.2323, pruned_loss=0.06132, over 4738.00 frames. ], tot_loss[loss=0.2083, simple_loss=0.2708, pruned_loss=0.07289, over 956790.24 frames. ], batch size: 23, lr: 3.91e-03, grad_scale: 32.0 2023-04-26 17:58:07,541 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=30964.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:58:21,369 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=30986.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:58:23,529 INFO [finetune.py:976] (4/7) Epoch 6, batch 2350, loss[loss=0.2198, simple_loss=0.2884, pruned_loss=0.07567, over 4380.00 frames. ], tot_loss[loss=0.2058, simple_loss=0.2683, pruned_loss=0.07163, over 956677.84 frames. ], batch size: 66, lr: 3.91e-03, grad_scale: 32.0 2023-04-26 17:58:28,711 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=30995.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:58:39,469 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=31011.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 17:58:39,497 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=31011.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 17:58:42,545 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=31016.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:58:49,764 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.071e+02 1.679e+02 2.129e+02 2.680e+02 4.388e+02, threshold=4.258e+02, percent-clipped=1.0 2023-04-26 17:58:56,922 INFO [finetune.py:976] (4/7) Epoch 6, batch 2400, loss[loss=0.1898, simple_loss=0.2477, pruned_loss=0.06599, over 4857.00 frames. ], tot_loss[loss=0.2044, simple_loss=0.2662, pruned_loss=0.07136, over 958730.39 frames. ], batch size: 49, lr: 3.91e-03, grad_scale: 32.0 2023-04-26 17:59:14,975 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=31064.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 17:59:19,940 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=31072.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 17:59:30,875 INFO [finetune.py:976] (4/7) Epoch 6, batch 2450, loss[loss=0.1654, simple_loss=0.226, pruned_loss=0.05242, over 4782.00 frames. ], tot_loss[loss=0.2016, simple_loss=0.2628, pruned_loss=0.0702, over 955565.23 frames. ], batch size: 26, lr: 3.91e-03, grad_scale: 32.0 2023-04-26 17:59:57,657 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.941e+02 2.240e+02 2.729e+02 4.178e+02, threshold=4.480e+02, percent-clipped=0.0 2023-04-26 18:00:04,430 INFO [finetune.py:976] (4/7) Epoch 6, batch 2500, loss[loss=0.2104, simple_loss=0.2852, pruned_loss=0.0678, over 4721.00 frames. ], tot_loss[loss=0.2057, simple_loss=0.2667, pruned_loss=0.07232, over 957248.10 frames. ], batch size: 54, lr: 3.91e-03, grad_scale: 32.0 2023-04-26 18:00:37,629 INFO [finetune.py:976] (4/7) Epoch 6, batch 2550, loss[loss=0.2621, simple_loss=0.325, pruned_loss=0.09956, over 4867.00 frames. ], tot_loss[loss=0.209, simple_loss=0.2707, pruned_loss=0.07363, over 957486.89 frames. ], batch size: 34, lr: 3.91e-03, grad_scale: 32.0 2023-04-26 18:00:37,704 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.8294, 4.7112, 3.2385, 5.5529, 4.9178, 4.8799, 2.5052, 4.8500], device='cuda:4'), covar=tensor([0.1220, 0.0937, 0.2677, 0.0850, 0.3337, 0.1682, 0.4838, 0.1808], device='cuda:4'), in_proj_covar=tensor([0.0243, 0.0219, 0.0253, 0.0312, 0.0301, 0.0254, 0.0273, 0.0273], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 18:01:09,725 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.313e+02 1.772e+02 2.143e+02 2.617e+02 5.206e+02, threshold=4.285e+02, percent-clipped=3.0 2023-04-26 18:01:22,513 INFO [finetune.py:976] (4/7) Epoch 6, batch 2600, loss[loss=0.2163, simple_loss=0.2779, pruned_loss=0.07731, over 4825.00 frames. ], tot_loss[loss=0.2093, simple_loss=0.2713, pruned_loss=0.07364, over 956663.21 frames. ], batch size: 38, lr: 3.91e-03, grad_scale: 32.0 2023-04-26 18:01:51,863 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=31259.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:02:10,442 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5510, 3.8456, 0.6711, 2.2595, 2.0566, 2.5826, 2.2840, 0.8916], device='cuda:4'), covar=tensor([0.1464, 0.0789, 0.2380, 0.1276, 0.1154, 0.1174, 0.1476, 0.2513], device='cuda:4'), in_proj_covar=tensor([0.0123, 0.0262, 0.0147, 0.0127, 0.0138, 0.0161, 0.0123, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 18:02:14,134 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=31281.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:02:24,319 INFO [finetune.py:976] (4/7) Epoch 6, batch 2650, loss[loss=0.1762, simple_loss=0.2405, pruned_loss=0.05591, over 4873.00 frames. ], tot_loss[loss=0.2098, simple_loss=0.2721, pruned_loss=0.07376, over 957058.21 frames. ], batch size: 34, lr: 3.91e-03, grad_scale: 32.0 2023-04-26 18:02:24,976 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=31290.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:02:52,939 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5066, 0.9745, 4.1007, 3.5342, 3.6349, 3.8714, 3.7468, 3.3980], device='cuda:4'), covar=tensor([0.9723, 0.9535, 0.1929, 0.3718, 0.2364, 0.2882, 0.2691, 0.3717], device='cuda:4'), in_proj_covar=tensor([0.0314, 0.0307, 0.0415, 0.0421, 0.0356, 0.0411, 0.0319, 0.0376], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-26 18:02:55,883 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=31311.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:03:18,341 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.052e+02 1.837e+02 2.164e+02 2.605e+02 4.419e+02, threshold=4.327e+02, percent-clipped=1.0 2023-04-26 18:03:30,924 INFO [finetune.py:976] (4/7) Epoch 6, batch 2700, loss[loss=0.1536, simple_loss=0.219, pruned_loss=0.04411, over 4794.00 frames. ], tot_loss[loss=0.208, simple_loss=0.2708, pruned_loss=0.07261, over 957855.02 frames. ], batch size: 51, lr: 3.90e-03, grad_scale: 32.0 2023-04-26 18:03:51,466 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3624, 1.1584, 1.5641, 1.5021, 1.2261, 1.0800, 1.1911, 0.7898], device='cuda:4'), covar=tensor([0.0657, 0.0862, 0.0586, 0.0715, 0.0927, 0.1454, 0.0777, 0.0933], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0074, 0.0073, 0.0067, 0.0078, 0.0094, 0.0080, 0.0076], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 18:04:01,833 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=31359.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:04:14,066 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=31367.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 18:04:14,100 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0915, 1.7655, 1.9485, 2.4688, 2.4036, 2.0278, 1.5727, 2.0344], device='cuda:4'), covar=tensor([0.0942, 0.1301, 0.0872, 0.0651, 0.0652, 0.0917, 0.1099, 0.0762], device='cuda:4'), in_proj_covar=tensor([0.0203, 0.0208, 0.0182, 0.0179, 0.0180, 0.0196, 0.0167, 0.0190], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 18:04:14,686 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9773, 1.5125, 1.4839, 1.7671, 1.5456, 2.0438, 1.3662, 3.4375], device='cuda:4'), covar=tensor([0.0699, 0.0730, 0.0760, 0.1151, 0.0629, 0.0547, 0.0762, 0.0155], device='cuda:4'), in_proj_covar=tensor([0.0040, 0.0040, 0.0041, 0.0045, 0.0041, 0.0040, 0.0040, 0.0061], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 18:04:44,537 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4216, 1.5935, 1.7165, 1.8785, 1.7399, 1.8859, 1.8002, 1.8198], device='cuda:4'), covar=tensor([0.6553, 1.0570, 0.7939, 0.7810, 0.9061, 1.3273, 1.0471, 0.9048], device='cuda:4'), in_proj_covar=tensor([0.0317, 0.0393, 0.0316, 0.0326, 0.0345, 0.0411, 0.0375, 0.0331], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 18:04:45,000 INFO [finetune.py:976] (4/7) Epoch 6, batch 2750, loss[loss=0.1726, simple_loss=0.2479, pruned_loss=0.04872, over 4754.00 frames. ], tot_loss[loss=0.2051, simple_loss=0.2676, pruned_loss=0.07131, over 955954.50 frames. ], batch size: 26, lr: 3.90e-03, grad_scale: 32.0 2023-04-26 18:05:10,768 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=31409.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:05:28,795 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.050e+02 1.757e+02 2.143e+02 2.485e+02 4.889e+02, threshold=4.286e+02, percent-clipped=1.0 2023-04-26 18:05:35,955 INFO [finetune.py:976] (4/7) Epoch 6, batch 2800, loss[loss=0.1654, simple_loss=0.2416, pruned_loss=0.04462, over 4905.00 frames. ], tot_loss[loss=0.2009, simple_loss=0.2632, pruned_loss=0.06933, over 957767.00 frames. ], batch size: 32, lr: 3.90e-03, grad_scale: 32.0 2023-04-26 18:05:37,912 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5280, 1.0308, 1.2877, 1.2363, 1.7071, 1.3734, 1.0640, 1.2879], device='cuda:4'), covar=tensor([0.1561, 0.1487, 0.1976, 0.1256, 0.0814, 0.1316, 0.1833, 0.1829], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0335, 0.0354, 0.0307, 0.0344, 0.0338, 0.0311, 0.0353], device='cuda:4'), out_proj_covar=tensor([6.6845e-05, 7.1524e-05, 7.6737e-05, 6.4063e-05, 7.2549e-05, 7.3058e-05, 6.7400e-05, 7.5995e-05], device='cuda:4') 2023-04-26 18:05:57,393 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=31470.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:06:09,971 INFO [finetune.py:976] (4/7) Epoch 6, batch 2850, loss[loss=0.2253, simple_loss=0.2837, pruned_loss=0.08345, over 4821.00 frames. ], tot_loss[loss=0.2001, simple_loss=0.2623, pruned_loss=0.06891, over 958771.48 frames. ], batch size: 38, lr: 3.90e-03, grad_scale: 32.0 2023-04-26 18:06:11,333 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=31491.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 18:06:15,673 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6877, 1.2738, 1.2246, 1.4202, 1.9195, 1.5914, 1.1852, 1.1789], device='cuda:4'), covar=tensor([0.1460, 0.1520, 0.2219, 0.1351, 0.0723, 0.1671, 0.2150, 0.2095], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0333, 0.0353, 0.0306, 0.0343, 0.0337, 0.0310, 0.0352], device='cuda:4'), out_proj_covar=tensor([6.6665e-05, 7.1261e-05, 7.6551e-05, 6.3813e-05, 7.2349e-05, 7.2878e-05, 6.7223e-05, 7.5830e-05], device='cuda:4') 2023-04-26 18:06:36,454 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.094e+02 1.825e+02 2.131e+02 2.409e+02 4.261e+02, threshold=4.261e+02, percent-clipped=0.0 2023-04-26 18:06:40,030 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-26 18:06:43,783 INFO [finetune.py:976] (4/7) Epoch 6, batch 2900, loss[loss=0.24, simple_loss=0.3129, pruned_loss=0.08354, over 4942.00 frames. ], tot_loss[loss=0.2016, simple_loss=0.2641, pruned_loss=0.06953, over 953608.78 frames. ], batch size: 39, lr: 3.90e-03, grad_scale: 32.0 2023-04-26 18:06:51,763 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=31552.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 18:06:55,991 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=31559.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:07:11,689 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=31581.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:07:16,956 INFO [finetune.py:976] (4/7) Epoch 6, batch 2950, loss[loss=0.2125, simple_loss=0.2748, pruned_loss=0.07512, over 4921.00 frames. ], tot_loss[loss=0.2048, simple_loss=0.2682, pruned_loss=0.07076, over 954617.76 frames. ], batch size: 42, lr: 3.90e-03, grad_scale: 32.0 2023-04-26 18:07:17,656 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=31590.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:07:28,066 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=31607.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:07:42,550 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.159e+02 1.797e+02 2.161e+02 2.728e+02 5.785e+02, threshold=4.321e+02, percent-clipped=2.0 2023-04-26 18:07:43,218 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=31629.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:07:49,139 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=31638.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:07:49,695 INFO [finetune.py:976] (4/7) Epoch 6, batch 3000, loss[loss=0.1335, simple_loss=0.2065, pruned_loss=0.0302, over 4768.00 frames. ], tot_loss[loss=0.2067, simple_loss=0.2698, pruned_loss=0.07179, over 954455.70 frames. ], batch size: 26, lr: 3.90e-03, grad_scale: 32.0 2023-04-26 18:07:49,695 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-26 18:07:55,419 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1987, 2.5152, 1.0382, 1.3362, 1.9046, 1.3716, 3.1120, 1.8281], device='cuda:4'), covar=tensor([0.0606, 0.0555, 0.0771, 0.1339, 0.0480, 0.0950, 0.0277, 0.0586], device='cuda:4'), in_proj_covar=tensor([0.0054, 0.0070, 0.0052, 0.0049, 0.0053, 0.0054, 0.0081, 0.0052], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 18:08:00,215 INFO [finetune.py:1010] (4/7) Epoch 6, validation: loss=0.1565, simple_loss=0.2301, pruned_loss=0.04144, over 2265189.00 frames. 2023-04-26 18:08:00,215 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6259MB 2023-04-26 18:08:17,853 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=31667.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 18:08:35,823 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.6469, 1.8096, 1.6956, 1.3806, 1.8082, 1.5040, 2.2926, 1.4857], device='cuda:4'), covar=tensor([0.3545, 0.1435, 0.4314, 0.2601, 0.1439, 0.2120, 0.1346, 0.4095], device='cuda:4'), in_proj_covar=tensor([0.0349, 0.0355, 0.0438, 0.0367, 0.0392, 0.0384, 0.0387, 0.0418], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 18:08:48,665 INFO [finetune.py:976] (4/7) Epoch 6, batch 3050, loss[loss=0.1933, simple_loss=0.2693, pruned_loss=0.05866, over 4796.00 frames. ], tot_loss[loss=0.2059, simple_loss=0.2697, pruned_loss=0.07105, over 955251.45 frames. ], batch size: 29, lr: 3.90e-03, grad_scale: 32.0 2023-04-26 18:09:10,821 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=31704.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:09:23,336 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=31715.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 18:09:32,702 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.99 vs. limit=5.0 2023-04-26 18:09:34,560 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.72 vs. limit=2.0 2023-04-26 18:09:40,854 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.151e+02 1.692e+02 2.096e+02 2.514e+02 4.783e+02, threshold=4.192e+02, percent-clipped=1.0 2023-04-26 18:09:53,205 INFO [finetune.py:976] (4/7) Epoch 6, batch 3100, loss[loss=0.1625, simple_loss=0.2285, pruned_loss=0.04831, over 4806.00 frames. ], tot_loss[loss=0.2045, simple_loss=0.2678, pruned_loss=0.07053, over 957110.06 frames. ], batch size: 40, lr: 3.90e-03, grad_scale: 32.0 2023-04-26 18:10:25,857 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=31765.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:10:25,889 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8540, 2.3870, 1.1812, 1.5321, 2.2980, 1.7053, 1.6573, 1.8570], device='cuda:4'), covar=tensor([0.0507, 0.0332, 0.0365, 0.0576, 0.0247, 0.0579, 0.0535, 0.0578], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0026, 0.0024, 0.0030, 0.0021, 0.0030, 0.0029, 0.0031], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0049, 0.0037, 0.0048, 0.0047, 0.0050], device='cuda:4') 2023-04-26 18:10:25,914 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=31765.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:10:56,985 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=31788.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:10:57,966 INFO [finetune.py:976] (4/7) Epoch 6, batch 3150, loss[loss=0.1658, simple_loss=0.2282, pruned_loss=0.05172, over 4852.00 frames. ], tot_loss[loss=0.2025, simple_loss=0.2654, pruned_loss=0.0698, over 956020.24 frames. ], batch size: 44, lr: 3.90e-03, grad_scale: 32.0 2023-04-26 18:11:42,879 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=31823.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:11:51,413 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.949e+01 1.610e+02 1.947e+02 2.415e+02 6.715e+02, threshold=3.894e+02, percent-clipped=1.0 2023-04-26 18:12:02,482 INFO [finetune.py:976] (4/7) Epoch 6, batch 3200, loss[loss=0.1969, simple_loss=0.2601, pruned_loss=0.06684, over 4838.00 frames. ], tot_loss[loss=0.2007, simple_loss=0.2626, pruned_loss=0.06941, over 956414.65 frames. ], batch size: 47, lr: 3.90e-03, grad_scale: 32.0 2023-04-26 18:12:14,986 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=31847.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 18:12:21,991 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=31849.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:12:57,057 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=31878.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:13:06,260 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=31884.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:13:09,142 INFO [finetune.py:976] (4/7) Epoch 6, batch 3250, loss[loss=0.2287, simple_loss=0.2936, pruned_loss=0.08189, over 4738.00 frames. ], tot_loss[loss=0.1999, simple_loss=0.2621, pruned_loss=0.06881, over 955906.63 frames. ], batch size: 59, lr: 3.90e-03, grad_scale: 32.0 2023-04-26 18:13:38,301 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0041, 1.4367, 1.3941, 2.0114, 2.1834, 1.7722, 1.7165, 1.4194], device='cuda:4'), covar=tensor([0.2336, 0.2131, 0.2120, 0.1597, 0.1266, 0.2344, 0.2539, 0.1949], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0335, 0.0354, 0.0308, 0.0344, 0.0338, 0.0311, 0.0355], device='cuda:4'), out_proj_covar=tensor([6.7068e-05, 7.1651e-05, 7.6838e-05, 6.4096e-05, 7.2461e-05, 7.3062e-05, 6.7538e-05, 7.6530e-05], device='cuda:4') 2023-04-26 18:13:39,527 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=31915.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:13:57,725 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.120e+02 1.762e+02 2.242e+02 2.827e+02 6.116e+02, threshold=4.483e+02, percent-clipped=8.0 2023-04-26 18:14:00,889 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2691, 1.5750, 1.4427, 1.8374, 1.6857, 2.1850, 1.5086, 3.4817], device='cuda:4'), covar=tensor([0.0655, 0.0773, 0.0796, 0.1149, 0.0614, 0.0428, 0.0754, 0.0184], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0039, 0.0041, 0.0045, 0.0041, 0.0040, 0.0040, 0.0061], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 18:14:10,189 INFO [finetune.py:976] (4/7) Epoch 6, batch 3300, loss[loss=0.1712, simple_loss=0.2434, pruned_loss=0.0495, over 4894.00 frames. ], tot_loss[loss=0.2011, simple_loss=0.2642, pruned_loss=0.06897, over 954862.76 frames. ], batch size: 35, lr: 3.90e-03, grad_scale: 32.0 2023-04-26 18:14:10,319 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=31939.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:14:17,112 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3645, 1.2838, 3.7291, 3.4525, 3.3311, 3.5667, 3.5174, 3.3114], device='cuda:4'), covar=tensor([0.6965, 0.5541, 0.1142, 0.1803, 0.1089, 0.1722, 0.2648, 0.1692], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0303, 0.0411, 0.0414, 0.0350, 0.0405, 0.0313, 0.0370], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-26 18:14:18,852 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=31950.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:14:20,728 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.13 vs. limit=5.0 2023-04-26 18:14:35,919 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=31976.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:14:43,708 INFO [finetune.py:976] (4/7) Epoch 6, batch 3350, loss[loss=0.2066, simple_loss=0.2608, pruned_loss=0.07623, over 4154.00 frames. ], tot_loss[loss=0.2031, simple_loss=0.2664, pruned_loss=0.06993, over 953003.61 frames. ], batch size: 65, lr: 3.90e-03, grad_scale: 32.0 2023-04-26 18:14:46,097 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.51 vs. limit=2.0 2023-04-26 18:14:54,438 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0904, 1.2294, 1.3891, 1.5904, 1.5187, 1.6635, 1.5010, 1.5228], device='cuda:4'), covar=tensor([0.5741, 0.9032, 0.7703, 0.6911, 0.8088, 1.1669, 0.8945, 0.8250], device='cuda:4'), in_proj_covar=tensor([0.0316, 0.0393, 0.0317, 0.0325, 0.0344, 0.0408, 0.0373, 0.0331], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 18:15:00,878 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=32011.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:15:12,029 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.180e+02 1.734e+02 2.033e+02 2.508e+02 4.432e+02, threshold=4.066e+02, percent-clipped=0.0 2023-04-26 18:15:24,409 INFO [finetune.py:976] (4/7) Epoch 6, batch 3400, loss[loss=0.1664, simple_loss=0.2242, pruned_loss=0.05432, over 4437.00 frames. ], tot_loss[loss=0.2048, simple_loss=0.2681, pruned_loss=0.07074, over 954213.59 frames. ], batch size: 19, lr: 3.90e-03, grad_scale: 32.0 2023-04-26 18:15:55,466 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=32060.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:16:04,715 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=32065.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:16:20,375 INFO [finetune.py:976] (4/7) Epoch 6, batch 3450, loss[loss=0.2621, simple_loss=0.3133, pruned_loss=0.1054, over 4732.00 frames. ], tot_loss[loss=0.2054, simple_loss=0.2691, pruned_loss=0.07081, over 953131.68 frames. ], batch size: 54, lr: 3.90e-03, grad_scale: 32.0 2023-04-26 18:16:20,523 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8195, 1.0397, 1.3383, 1.5116, 1.4851, 1.5880, 1.3801, 1.4134], device='cuda:4'), covar=tensor([0.6118, 0.8540, 0.7423, 0.6743, 0.8662, 1.3032, 0.8980, 0.7643], device='cuda:4'), in_proj_covar=tensor([0.0315, 0.0392, 0.0316, 0.0324, 0.0343, 0.0407, 0.0373, 0.0329], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 18:16:37,418 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=32113.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:16:47,433 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.129e+02 1.699e+02 2.063e+02 2.455e+02 6.391e+02, threshold=4.126e+02, percent-clipped=3.0 2023-04-26 18:16:54,107 INFO [finetune.py:976] (4/7) Epoch 6, batch 3500, loss[loss=0.2012, simple_loss=0.2588, pruned_loss=0.07174, over 4906.00 frames. ], tot_loss[loss=0.2028, simple_loss=0.266, pruned_loss=0.0698, over 954282.61 frames. ], batch size: 36, lr: 3.90e-03, grad_scale: 32.0 2023-04-26 18:16:57,312 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=32144.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:16:59,147 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=32147.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 18:17:01,682 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.14 vs. limit=2.0 2023-04-26 18:17:21,483 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=32179.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:17:33,409 INFO [finetune.py:976] (4/7) Epoch 6, batch 3550, loss[loss=0.2389, simple_loss=0.2896, pruned_loss=0.09409, over 4805.00 frames. ], tot_loss[loss=0.2006, simple_loss=0.2633, pruned_loss=0.0689, over 956725.64 frames. ], batch size: 25, lr: 3.90e-03, grad_scale: 32.0 2023-04-26 18:17:43,433 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=32195.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 18:17:45,318 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6664, 3.6499, 0.9182, 1.9220, 2.0911, 2.4337, 2.0911, 1.0438], device='cuda:4'), covar=tensor([0.1304, 0.0878, 0.2125, 0.1435, 0.1042, 0.1155, 0.1465, 0.2028], device='cuda:4'), in_proj_covar=tensor([0.0122, 0.0261, 0.0147, 0.0128, 0.0138, 0.0161, 0.0124, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 18:18:29,258 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.147e+02 1.692e+02 2.038e+02 2.540e+02 1.815e+03, threshold=4.076e+02, percent-clipped=3.0 2023-04-26 18:18:39,221 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=32234.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:18:47,840 INFO [finetune.py:976] (4/7) Epoch 6, batch 3600, loss[loss=0.1966, simple_loss=0.257, pruned_loss=0.0681, over 4927.00 frames. ], tot_loss[loss=0.2011, simple_loss=0.2633, pruned_loss=0.06947, over 957076.23 frames. ], batch size: 33, lr: 3.90e-03, grad_scale: 32.0 2023-04-26 18:18:59,926 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2872, 2.9760, 0.9257, 1.7011, 1.7050, 2.1761, 1.8077, 1.0133], device='cuda:4'), covar=tensor([0.1414, 0.1079, 0.2022, 0.1342, 0.1180, 0.1027, 0.1655, 0.1890], device='cuda:4'), in_proj_covar=tensor([0.0122, 0.0261, 0.0147, 0.0128, 0.0138, 0.0161, 0.0124, 0.0127], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 18:19:31,473 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=32271.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:19:55,461 INFO [finetune.py:976] (4/7) Epoch 6, batch 3650, loss[loss=0.1797, simple_loss=0.2488, pruned_loss=0.05534, over 4927.00 frames. ], tot_loss[loss=0.2016, simple_loss=0.2641, pruned_loss=0.06957, over 953762.98 frames. ], batch size: 33, lr: 3.90e-03, grad_scale: 32.0 2023-04-26 18:20:06,033 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.73 vs. limit=5.0 2023-04-26 18:20:16,844 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=32306.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:20:43,394 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.015e+02 1.808e+02 2.117e+02 2.519e+02 3.732e+02, threshold=4.235e+02, percent-clipped=0.0 2023-04-26 18:21:00,979 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.8681, 1.8889, 1.6145, 1.4143, 1.9840, 1.5266, 2.3671, 1.3824], device='cuda:4'), covar=tensor([0.3948, 0.1621, 0.4892, 0.3279, 0.1737, 0.2518, 0.1306, 0.4634], device='cuda:4'), in_proj_covar=tensor([0.0347, 0.0356, 0.0434, 0.0365, 0.0392, 0.0383, 0.0387, 0.0416], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 18:21:03,308 INFO [finetune.py:976] (4/7) Epoch 6, batch 3700, loss[loss=0.1394, simple_loss=0.2089, pruned_loss=0.0349, over 4713.00 frames. ], tot_loss[loss=0.2053, simple_loss=0.2684, pruned_loss=0.07109, over 954341.47 frames. ], batch size: 23, lr: 3.90e-03, grad_scale: 32.0 2023-04-26 18:21:24,296 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.6252, 4.4529, 3.1338, 5.3047, 4.6699, 4.5942, 2.1866, 4.5793], device='cuda:4'), covar=tensor([0.1507, 0.1031, 0.3157, 0.0987, 0.3073, 0.1703, 0.5398, 0.1964], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0220, 0.0256, 0.0312, 0.0302, 0.0255, 0.0277, 0.0275], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 18:21:26,215 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9023, 1.4201, 1.5190, 1.6801, 2.1180, 1.7918, 1.4396, 1.4096], device='cuda:4'), covar=tensor([0.1696, 0.1686, 0.1950, 0.1407, 0.0917, 0.1552, 0.2534, 0.1863], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0336, 0.0353, 0.0306, 0.0343, 0.0338, 0.0311, 0.0355], device='cuda:4'), out_proj_covar=tensor([6.6829e-05, 7.1662e-05, 7.6618e-05, 6.3731e-05, 7.2373e-05, 7.3055e-05, 6.7355e-05, 7.6434e-05], device='cuda:4') 2023-04-26 18:21:26,775 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=32360.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:22:04,011 INFO [finetune.py:976] (4/7) Epoch 6, batch 3750, loss[loss=0.1794, simple_loss=0.249, pruned_loss=0.05486, over 4692.00 frames. ], tot_loss[loss=0.2063, simple_loss=0.2696, pruned_loss=0.07153, over 954508.61 frames. ], batch size: 59, lr: 3.90e-03, grad_scale: 32.0 2023-04-26 18:22:21,625 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=32408.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:22:35,218 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.756e+02 2.047e+02 2.329e+02 4.527e+02, threshold=4.095e+02, percent-clipped=1.0 2023-04-26 18:22:40,983 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.48 vs. limit=2.0 2023-04-26 18:22:43,808 INFO [finetune.py:976] (4/7) Epoch 6, batch 3800, loss[loss=0.1833, simple_loss=0.2566, pruned_loss=0.05501, over 4760.00 frames. ], tot_loss[loss=0.2076, simple_loss=0.2713, pruned_loss=0.0719, over 955820.81 frames. ], batch size: 28, lr: 3.90e-03, grad_scale: 32.0 2023-04-26 18:22:53,231 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=32444.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:23:08,414 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.10 vs. limit=5.0 2023-04-26 18:23:21,584 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=32479.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:23:33,783 INFO [finetune.py:976] (4/7) Epoch 6, batch 3850, loss[loss=0.1862, simple_loss=0.2531, pruned_loss=0.05959, over 4903.00 frames. ], tot_loss[loss=0.207, simple_loss=0.2706, pruned_loss=0.07173, over 957732.75 frames. ], batch size: 43, lr: 3.90e-03, grad_scale: 32.0 2023-04-26 18:23:41,606 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=32492.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:23:41,653 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0099, 1.3970, 1.2379, 1.6143, 1.4550, 1.6487, 1.2772, 2.4480], device='cuda:4'), covar=tensor([0.0668, 0.0805, 0.0845, 0.1250, 0.0686, 0.0473, 0.0789, 0.0240], device='cuda:4'), in_proj_covar=tensor([0.0040, 0.0040, 0.0042, 0.0046, 0.0041, 0.0040, 0.0040, 0.0062], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 18:24:04,918 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2906, 1.3070, 1.4039, 1.6324, 1.5707, 1.2378, 0.9116, 1.4209], device='cuda:4'), covar=tensor([0.0903, 0.1268, 0.0825, 0.0656, 0.0684, 0.0948, 0.1019, 0.0683], device='cuda:4'), in_proj_covar=tensor([0.0202, 0.0207, 0.0181, 0.0179, 0.0180, 0.0195, 0.0166, 0.0189], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 18:24:18,831 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=32527.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:24:19,366 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.258e+02 1.697e+02 2.014e+02 2.514e+02 4.116e+02, threshold=4.029e+02, percent-clipped=1.0 2023-04-26 18:24:24,612 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=32534.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:24:26,368 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6172, 1.3750, 1.8049, 1.8416, 1.4179, 1.2760, 1.5088, 0.9132], device='cuda:4'), covar=tensor([0.0661, 0.1066, 0.0635, 0.0801, 0.1008, 0.1439, 0.0871, 0.1029], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0074, 0.0073, 0.0067, 0.0077, 0.0094, 0.0080, 0.0076], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 18:24:28,027 INFO [finetune.py:976] (4/7) Epoch 6, batch 3900, loss[loss=0.1743, simple_loss=0.2425, pruned_loss=0.05305, over 4835.00 frames. ], tot_loss[loss=0.206, simple_loss=0.269, pruned_loss=0.07154, over 956865.87 frames. ], batch size: 47, lr: 3.90e-03, grad_scale: 32.0 2023-04-26 18:24:37,980 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1700, 2.6270, 1.1039, 1.4068, 2.0030, 1.2542, 3.6492, 1.9767], device='cuda:4'), covar=tensor([0.0697, 0.0659, 0.0877, 0.1368, 0.0568, 0.1100, 0.0293, 0.0666], device='cuda:4'), in_proj_covar=tensor([0.0053, 0.0070, 0.0052, 0.0048, 0.0053, 0.0054, 0.0081, 0.0052], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 18:25:09,813 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=32571.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:25:22,278 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=32582.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:25:33,447 INFO [finetune.py:976] (4/7) Epoch 6, batch 3950, loss[loss=0.221, simple_loss=0.2795, pruned_loss=0.08124, over 4829.00 frames. ], tot_loss[loss=0.2031, simple_loss=0.2651, pruned_loss=0.07054, over 957286.19 frames. ], batch size: 30, lr: 3.90e-03, grad_scale: 16.0 2023-04-26 18:25:50,974 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=32606.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:25:58,795 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=32619.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:26:10,336 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.148e+02 1.669e+02 1.908e+02 2.385e+02 4.552e+02, threshold=3.816e+02, percent-clipped=3.0 2023-04-26 18:26:22,977 INFO [finetune.py:976] (4/7) Epoch 6, batch 4000, loss[loss=0.2417, simple_loss=0.2955, pruned_loss=0.09395, over 4820.00 frames. ], tot_loss[loss=0.2024, simple_loss=0.2638, pruned_loss=0.07048, over 957123.25 frames. ], batch size: 38, lr: 3.90e-03, grad_scale: 16.0 2023-04-26 18:26:51,941 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=32654.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:26:52,667 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1124, 1.6434, 1.9946, 2.0466, 1.9263, 1.6239, 1.0682, 1.6349], device='cuda:4'), covar=tensor([0.3800, 0.4171, 0.1886, 0.3173, 0.3444, 0.3332, 0.5388, 0.3009], device='cuda:4'), in_proj_covar=tensor([0.0278, 0.0256, 0.0220, 0.0327, 0.0217, 0.0229, 0.0242, 0.0192], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 18:27:01,490 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2020, 1.2245, 1.3592, 1.5715, 1.5427, 1.2268, 0.8948, 1.3573], device='cuda:4'), covar=tensor([0.0971, 0.1308, 0.0858, 0.0678, 0.0731, 0.0995, 0.1067, 0.0698], device='cuda:4'), in_proj_covar=tensor([0.0201, 0.0205, 0.0179, 0.0177, 0.0179, 0.0193, 0.0165, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 18:27:18,612 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.76 vs. limit=2.0 2023-04-26 18:27:19,404 INFO [finetune.py:976] (4/7) Epoch 6, batch 4050, loss[loss=0.1927, simple_loss=0.2661, pruned_loss=0.0596, over 4826.00 frames. ], tot_loss[loss=0.2059, simple_loss=0.2676, pruned_loss=0.07211, over 955768.06 frames. ], batch size: 47, lr: 3.90e-03, grad_scale: 16.0 2023-04-26 18:27:23,532 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6088, 3.6005, 2.9792, 3.2172, 2.4911, 2.9187, 2.9001, 2.7638], device='cuda:4'), covar=tensor([0.2288, 0.1298, 0.0814, 0.1208, 0.2848, 0.1366, 0.2147, 0.2380], device='cuda:4'), in_proj_covar=tensor([0.0305, 0.0326, 0.0236, 0.0298, 0.0320, 0.0279, 0.0266, 0.0289], device='cuda:4'), out_proj_covar=tensor([1.2370e-04, 1.3229e-04, 9.5733e-05, 1.1987e-04, 1.3160e-04, 1.1282e-04, 1.0936e-04, 1.1650e-04], device='cuda:4') 2023-04-26 18:27:30,526 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9587, 1.9992, 1.6801, 1.6093, 2.2235, 1.6962, 2.5749, 1.4256], device='cuda:4'), covar=tensor([0.3784, 0.1714, 0.4583, 0.3287, 0.1604, 0.2395, 0.1584, 0.4707], device='cuda:4'), in_proj_covar=tensor([0.0348, 0.0355, 0.0435, 0.0365, 0.0394, 0.0384, 0.0385, 0.0417], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 18:27:45,870 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3989, 2.8067, 0.9924, 1.5310, 2.0769, 1.4337, 4.2390, 2.0835], device='cuda:4'), covar=tensor([0.0646, 0.1050, 0.0931, 0.1308, 0.0555, 0.1039, 0.0188, 0.0615], device='cuda:4'), in_proj_covar=tensor([0.0053, 0.0070, 0.0052, 0.0048, 0.0053, 0.0054, 0.0081, 0.0052], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 18:27:46,365 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.068e+02 1.872e+02 2.266e+02 2.687e+02 3.966e+02, threshold=4.532e+02, percent-clipped=2.0 2023-04-26 18:27:52,927 INFO [finetune.py:976] (4/7) Epoch 6, batch 4100, loss[loss=0.2245, simple_loss=0.2861, pruned_loss=0.08148, over 4912.00 frames. ], tot_loss[loss=0.2063, simple_loss=0.2682, pruned_loss=0.07217, over 954640.48 frames. ], batch size: 38, lr: 3.90e-03, grad_scale: 16.0 2023-04-26 18:28:15,850 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-26 18:28:26,742 INFO [finetune.py:976] (4/7) Epoch 6, batch 4150, loss[loss=0.2778, simple_loss=0.3328, pruned_loss=0.1114, over 4807.00 frames. ], tot_loss[loss=0.2083, simple_loss=0.2706, pruned_loss=0.073, over 953693.05 frames. ], batch size: 38, lr: 3.90e-03, grad_scale: 16.0 2023-04-26 18:29:05,515 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.362e+02 1.802e+02 2.275e+02 2.602e+02 5.008e+02, threshold=4.549e+02, percent-clipped=2.0 2023-04-26 18:29:10,552 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7537, 1.4837, 2.0218, 2.1122, 1.5238, 1.3068, 1.6943, 1.0839], device='cuda:4'), covar=tensor([0.0835, 0.1090, 0.0580, 0.0806, 0.1041, 0.1441, 0.0877, 0.1045], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0076, 0.0074, 0.0068, 0.0078, 0.0096, 0.0081, 0.0078], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 18:29:12,123 INFO [finetune.py:976] (4/7) Epoch 6, batch 4200, loss[loss=0.2141, simple_loss=0.2748, pruned_loss=0.07665, over 4829.00 frames. ], tot_loss[loss=0.2074, simple_loss=0.2705, pruned_loss=0.07219, over 954532.11 frames. ], batch size: 44, lr: 3.90e-03, grad_scale: 16.0 2023-04-26 18:29:12,831 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2830, 1.6231, 1.4566, 2.0978, 1.6844, 1.9782, 1.5700, 4.4956], device='cuda:4'), covar=tensor([0.0683, 0.0900, 0.0936, 0.1249, 0.0739, 0.0609, 0.0831, 0.0125], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0040, 0.0041, 0.0045, 0.0041, 0.0040, 0.0040, 0.0061], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 18:29:33,632 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-04-26 18:29:45,429 INFO [finetune.py:976] (4/7) Epoch 6, batch 4250, loss[loss=0.1908, simple_loss=0.26, pruned_loss=0.06086, over 4850.00 frames. ], tot_loss[loss=0.2046, simple_loss=0.2677, pruned_loss=0.07074, over 953606.01 frames. ], batch size: 47, lr: 3.90e-03, grad_scale: 16.0 2023-04-26 18:29:50,254 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-26 18:30:13,035 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.159e+01 1.639e+02 1.929e+02 2.275e+02 6.095e+02, threshold=3.858e+02, percent-clipped=1.0 2023-04-26 18:30:19,136 INFO [finetune.py:976] (4/7) Epoch 6, batch 4300, loss[loss=0.1738, simple_loss=0.2405, pruned_loss=0.05353, over 4933.00 frames. ], tot_loss[loss=0.2034, simple_loss=0.2654, pruned_loss=0.07067, over 955301.81 frames. ], batch size: 38, lr: 3.90e-03, grad_scale: 16.0 2023-04-26 18:30:27,919 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.45 vs. limit=5.0 2023-04-26 18:30:47,973 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=32966.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:31:19,600 INFO [finetune.py:976] (4/7) Epoch 6, batch 4350, loss[loss=0.2288, simple_loss=0.2858, pruned_loss=0.08585, over 4803.00 frames. ], tot_loss[loss=0.202, simple_loss=0.2635, pruned_loss=0.07021, over 956120.53 frames. ], batch size: 45, lr: 3.90e-03, grad_scale: 16.0 2023-04-26 18:32:12,345 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2456, 1.1430, 1.7227, 1.5682, 1.1767, 1.0154, 1.3667, 1.0081], device='cuda:4'), covar=tensor([0.0803, 0.1069, 0.0550, 0.0687, 0.1016, 0.1549, 0.0852, 0.0925], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0076, 0.0074, 0.0068, 0.0079, 0.0096, 0.0082, 0.0078], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 18:32:12,352 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=33027.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:32:13,431 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.910e+01 1.721e+02 1.970e+02 2.465e+02 4.001e+02, threshold=3.941e+02, percent-clipped=3.0 2023-04-26 18:32:13,998 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-04-26 18:32:24,806 INFO [finetune.py:976] (4/7) Epoch 6, batch 4400, loss[loss=0.1844, simple_loss=0.236, pruned_loss=0.06644, over 4307.00 frames. ], tot_loss[loss=0.2044, simple_loss=0.2662, pruned_loss=0.0713, over 956208.20 frames. ], batch size: 18, lr: 3.90e-03, grad_scale: 16.0 2023-04-26 18:32:27,405 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-26 18:32:45,510 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0474, 2.7360, 2.1674, 2.4257, 1.8718, 2.2732, 2.4089, 1.9215], device='cuda:4'), covar=tensor([0.2333, 0.1075, 0.0890, 0.1304, 0.2795, 0.1118, 0.1894, 0.2348], device='cuda:4'), in_proj_covar=tensor([0.0302, 0.0322, 0.0234, 0.0296, 0.0317, 0.0276, 0.0264, 0.0287], device='cuda:4'), out_proj_covar=tensor([1.2259e-04, 1.3057e-04, 9.4818e-05, 1.1896e-04, 1.3055e-04, 1.1138e-04, 1.0834e-04, 1.1563e-04], device='cuda:4') 2023-04-26 18:33:06,921 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=33069.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:33:11,028 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5563, 3.8916, 0.7351, 2.1590, 2.0539, 2.6267, 2.2675, 0.9101], device='cuda:4'), covar=tensor([0.1368, 0.0971, 0.2158, 0.1293, 0.1113, 0.1122, 0.1455, 0.2321], device='cuda:4'), in_proj_covar=tensor([0.0122, 0.0260, 0.0145, 0.0127, 0.0138, 0.0160, 0.0123, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 18:33:17,892 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1068, 1.5576, 1.3110, 2.0358, 1.6964, 2.0578, 1.5191, 4.1591], device='cuda:4'), covar=tensor([0.0678, 0.0806, 0.0891, 0.1199, 0.0681, 0.0685, 0.0815, 0.0156], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0039, 0.0041, 0.0045, 0.0041, 0.0040, 0.0040, 0.0061], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 18:33:20,947 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1327, 1.4135, 1.2092, 1.6582, 1.4390, 2.0514, 1.3142, 3.5434], device='cuda:4'), covar=tensor([0.0658, 0.0812, 0.0889, 0.1271, 0.0713, 0.0555, 0.0819, 0.0144], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0039, 0.0041, 0.0045, 0.0041, 0.0040, 0.0040, 0.0061], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 18:33:29,651 INFO [finetune.py:976] (4/7) Epoch 6, batch 4450, loss[loss=0.2297, simple_loss=0.2921, pruned_loss=0.08365, over 4810.00 frames. ], tot_loss[loss=0.2059, simple_loss=0.2685, pruned_loss=0.07169, over 954962.85 frames. ], batch size: 40, lr: 3.90e-03, grad_scale: 16.0 2023-04-26 18:33:43,443 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0754, 1.3746, 1.2226, 1.6398, 1.3813, 1.8290, 1.2772, 3.0321], device='cuda:4'), covar=tensor([0.0665, 0.0781, 0.0840, 0.1207, 0.0686, 0.0549, 0.0812, 0.0199], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0039, 0.0041, 0.0045, 0.0041, 0.0040, 0.0040, 0.0061], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 18:33:54,494 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8747, 2.5422, 1.8445, 1.8583, 1.3955, 1.3949, 1.9670, 1.2841], device='cuda:4'), covar=tensor([0.1825, 0.1771, 0.1835, 0.2223, 0.2753, 0.2180, 0.1279, 0.2409], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0218, 0.0174, 0.0204, 0.0209, 0.0185, 0.0166, 0.0191], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-26 18:34:08,078 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3468, 3.6524, 1.0170, 1.9631, 1.9553, 2.6277, 2.0468, 1.0470], device='cuda:4'), covar=tensor([0.1475, 0.0931, 0.1977, 0.1306, 0.1146, 0.1014, 0.1548, 0.2054], device='cuda:4'), in_proj_covar=tensor([0.0121, 0.0259, 0.0144, 0.0127, 0.0138, 0.0159, 0.0123, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 18:34:12,571 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.224e+02 1.742e+02 2.062e+02 2.633e+02 5.031e+02, threshold=4.124e+02, percent-clipped=5.0 2023-04-26 18:34:13,337 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=33130.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:34:18,682 INFO [finetune.py:976] (4/7) Epoch 6, batch 4500, loss[loss=0.2239, simple_loss=0.2801, pruned_loss=0.08383, over 4917.00 frames. ], tot_loss[loss=0.2076, simple_loss=0.2704, pruned_loss=0.07237, over 956650.17 frames. ], batch size: 33, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:34:29,042 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=33155.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:34:49,185 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-04-26 18:34:52,375 INFO [finetune.py:976] (4/7) Epoch 6, batch 4550, loss[loss=0.1783, simple_loss=0.2436, pruned_loss=0.05654, over 4829.00 frames. ], tot_loss[loss=0.2078, simple_loss=0.271, pruned_loss=0.07227, over 954349.77 frames. ], batch size: 30, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:34:54,322 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0728, 2.7698, 2.1972, 2.4709, 1.9330, 2.2562, 2.4685, 1.8818], device='cuda:4'), covar=tensor([0.2507, 0.1272, 0.0922, 0.1476, 0.2864, 0.1277, 0.1861, 0.2715], device='cuda:4'), in_proj_covar=tensor([0.0303, 0.0323, 0.0234, 0.0297, 0.0318, 0.0276, 0.0264, 0.0288], device='cuda:4'), out_proj_covar=tensor([1.2322e-04, 1.3087e-04, 9.4978e-05, 1.1919e-04, 1.3076e-04, 1.1165e-04, 1.0853e-04, 1.1593e-04], device='cuda:4') 2023-04-26 18:34:56,774 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9751, 2.5326, 2.0762, 2.3448, 1.7463, 2.1211, 2.0901, 1.6822], device='cuda:4'), covar=tensor([0.2330, 0.1393, 0.1004, 0.1331, 0.3396, 0.1203, 0.2050, 0.2948], device='cuda:4'), in_proj_covar=tensor([0.0303, 0.0323, 0.0234, 0.0297, 0.0318, 0.0276, 0.0264, 0.0288], device='cuda:4'), out_proj_covar=tensor([1.2321e-04, 1.3086e-04, 9.4997e-05, 1.1919e-04, 1.3079e-04, 1.1165e-04, 1.0854e-04, 1.1593e-04], device='cuda:4') 2023-04-26 18:35:09,512 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=33216.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:35:18,796 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.085e+02 1.680e+02 1.969e+02 2.447e+02 4.378e+02, threshold=3.937e+02, percent-clipped=2.0 2023-04-26 18:35:25,830 INFO [finetune.py:976] (4/7) Epoch 6, batch 4600, loss[loss=0.1728, simple_loss=0.2412, pruned_loss=0.05219, over 4792.00 frames. ], tot_loss[loss=0.2064, simple_loss=0.2702, pruned_loss=0.07127, over 954596.11 frames. ], batch size: 45, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:35:36,851 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0283, 2.0007, 1.7470, 1.6652, 2.1895, 1.7150, 2.6061, 1.5901], device='cuda:4'), covar=tensor([0.3857, 0.2014, 0.5254, 0.3293, 0.1782, 0.2667, 0.1487, 0.4506], device='cuda:4'), in_proj_covar=tensor([0.0352, 0.0358, 0.0441, 0.0368, 0.0397, 0.0387, 0.0390, 0.0423], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 18:35:59,783 INFO [finetune.py:976] (4/7) Epoch 6, batch 4650, loss[loss=0.1863, simple_loss=0.2414, pruned_loss=0.06558, over 4708.00 frames. ], tot_loss[loss=0.2031, simple_loss=0.2664, pruned_loss=0.06984, over 954175.30 frames. ], batch size: 23, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:36:01,691 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.7126, 4.5258, 3.1695, 5.3680, 4.6858, 4.6943, 2.0351, 4.6447], device='cuda:4'), covar=tensor([0.1405, 0.1015, 0.3115, 0.0884, 0.2670, 0.1670, 0.5658, 0.1973], device='cuda:4'), in_proj_covar=tensor([0.0245, 0.0221, 0.0256, 0.0314, 0.0305, 0.0257, 0.0277, 0.0276], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 18:36:06,628 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5575, 1.2224, 1.6663, 1.9781, 1.7096, 1.5648, 1.6027, 1.6712], device='cuda:4'), covar=tensor([0.8443, 1.0796, 1.1378, 1.2646, 0.9686, 1.2853, 1.3367, 1.1249], device='cuda:4'), in_proj_covar=tensor([0.0414, 0.0440, 0.0525, 0.0545, 0.0443, 0.0462, 0.0474, 0.0473], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 18:36:20,281 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=33322.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:36:22,855 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.68 vs. limit=5.0 2023-04-26 18:36:25,472 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.206e+02 1.668e+02 2.007e+02 2.366e+02 4.187e+02, threshold=4.014e+02, percent-clipped=2.0 2023-04-26 18:36:32,593 INFO [finetune.py:976] (4/7) Epoch 6, batch 4700, loss[loss=0.1516, simple_loss=0.201, pruned_loss=0.05107, over 4304.00 frames. ], tot_loss[loss=0.2003, simple_loss=0.2629, pruned_loss=0.06882, over 954727.51 frames. ], batch size: 65, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:37:20,899 INFO [finetune.py:976] (4/7) Epoch 6, batch 4750, loss[loss=0.2081, simple_loss=0.2601, pruned_loss=0.07809, over 4386.00 frames. ], tot_loss[loss=0.1985, simple_loss=0.261, pruned_loss=0.06805, over 955051.17 frames. ], batch size: 19, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:38:06,057 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=33425.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:38:14,206 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.077e+02 1.777e+02 2.013e+02 2.553e+02 6.681e+02, threshold=4.026e+02, percent-clipped=2.0 2023-04-26 18:38:27,282 INFO [finetune.py:976] (4/7) Epoch 6, batch 4800, loss[loss=0.214, simple_loss=0.2903, pruned_loss=0.06889, over 4900.00 frames. ], tot_loss[loss=0.2015, simple_loss=0.2635, pruned_loss=0.06976, over 955317.79 frames. ], batch size: 36, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:39:01,134 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=33472.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:39:12,306 INFO [finetune.py:976] (4/7) Epoch 6, batch 4850, loss[loss=0.1965, simple_loss=0.2593, pruned_loss=0.06682, over 4897.00 frames. ], tot_loss[loss=0.2042, simple_loss=0.2668, pruned_loss=0.0708, over 955742.24 frames. ], batch size: 32, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:39:13,545 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0221, 1.3859, 1.9330, 2.1889, 1.7858, 1.4085, 1.0636, 1.5738], device='cuda:4'), covar=tensor([0.4005, 0.4410, 0.1932, 0.3353, 0.3515, 0.3500, 0.5544, 0.3080], device='cuda:4'), in_proj_covar=tensor([0.0278, 0.0255, 0.0219, 0.0326, 0.0215, 0.0228, 0.0240, 0.0191], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 18:39:19,949 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9893, 2.4433, 1.0318, 1.2310, 1.8644, 1.2338, 3.2232, 1.5673], device='cuda:4'), covar=tensor([0.0700, 0.0677, 0.0881, 0.1347, 0.0571, 0.1041, 0.0253, 0.0700], device='cuda:4'), in_proj_covar=tensor([0.0053, 0.0069, 0.0052, 0.0048, 0.0053, 0.0054, 0.0081, 0.0052], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 18:39:27,146 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=33511.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:39:38,471 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.113e+02 1.796e+02 2.126e+02 2.490e+02 5.374e+02, threshold=4.252e+02, percent-clipped=2.0 2023-04-26 18:39:41,034 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=33533.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:39:44,984 INFO [finetune.py:976] (4/7) Epoch 6, batch 4900, loss[loss=0.1986, simple_loss=0.2666, pruned_loss=0.0653, over 4920.00 frames. ], tot_loss[loss=0.2057, simple_loss=0.2685, pruned_loss=0.07148, over 956232.12 frames. ], batch size: 41, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:40:11,225 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7298, 4.2827, 0.8382, 2.3513, 2.4643, 2.8095, 2.5444, 0.9603], device='cuda:4'), covar=tensor([0.1397, 0.0846, 0.2221, 0.1225, 0.0953, 0.1136, 0.1538, 0.2184], device='cuda:4'), in_proj_covar=tensor([0.0121, 0.0259, 0.0144, 0.0126, 0.0137, 0.0159, 0.0122, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 18:40:18,324 INFO [finetune.py:976] (4/7) Epoch 6, batch 4950, loss[loss=0.2267, simple_loss=0.2916, pruned_loss=0.08083, over 4796.00 frames. ], tot_loss[loss=0.2081, simple_loss=0.2715, pruned_loss=0.07236, over 955104.25 frames. ], batch size: 45, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:40:27,257 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.43 vs. limit=2.0 2023-04-26 18:40:27,388 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.7533, 3.7080, 2.8049, 4.3602, 3.8087, 3.7539, 1.7348, 3.6518], device='cuda:4'), covar=tensor([0.1694, 0.1342, 0.3222, 0.1617, 0.2250, 0.1776, 0.5313, 0.2246], device='cuda:4'), in_proj_covar=tensor([0.0243, 0.0218, 0.0253, 0.0311, 0.0301, 0.0254, 0.0274, 0.0273], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 18:40:40,487 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=33622.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:40:44,643 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.256e+02 1.697e+02 1.943e+02 2.276e+02 3.620e+02, threshold=3.887e+02, percent-clipped=0.0 2023-04-26 18:40:51,674 INFO [finetune.py:976] (4/7) Epoch 6, batch 5000, loss[loss=0.1704, simple_loss=0.2379, pruned_loss=0.05146, over 4844.00 frames. ], tot_loss[loss=0.2055, simple_loss=0.2691, pruned_loss=0.07095, over 956284.98 frames. ], batch size: 31, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:41:24,493 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=33670.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:41:28,777 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=33677.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:41:32,991 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-26 18:41:36,417 INFO [finetune.py:976] (4/7) Epoch 6, batch 5050, loss[loss=0.1911, simple_loss=0.2528, pruned_loss=0.06472, over 4869.00 frames. ], tot_loss[loss=0.2037, simple_loss=0.2665, pruned_loss=0.07048, over 956575.31 frames. ], batch size: 34, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:42:01,201 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=33725.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:42:03,508 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.184e+02 1.719e+02 2.034e+02 2.333e+02 5.936e+02, threshold=4.068e+02, percent-clipped=3.0 2023-04-26 18:42:14,953 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=33738.0, num_to_drop=1, layers_to_drop={3} 2023-04-26 18:42:15,449 INFO [finetune.py:976] (4/7) Epoch 6, batch 5100, loss[loss=0.1839, simple_loss=0.2619, pruned_loss=0.05297, over 4789.00 frames. ], tot_loss[loss=0.199, simple_loss=0.2623, pruned_loss=0.06785, over 957164.96 frames. ], batch size: 28, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:42:59,574 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.56 vs. limit=2.0 2023-04-26 18:43:00,672 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=33773.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:43:21,825 INFO [finetune.py:976] (4/7) Epoch 6, batch 5150, loss[loss=0.2115, simple_loss=0.2749, pruned_loss=0.07405, over 4901.00 frames. ], tot_loss[loss=0.1992, simple_loss=0.262, pruned_loss=0.06821, over 957503.64 frames. ], batch size: 43, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:43:33,826 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9074, 1.8466, 2.2114, 2.2749, 1.7309, 1.4685, 2.0720, 0.9550], device='cuda:4'), covar=tensor([0.0832, 0.1172, 0.0717, 0.1304, 0.1185, 0.1798, 0.0990, 0.1453], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0075, 0.0074, 0.0068, 0.0079, 0.0096, 0.0081, 0.0078], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 18:43:40,822 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-04-26 18:43:42,912 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=33811.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:43:44,067 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5852, 1.2341, 1.6238, 1.9838, 1.7091, 1.5013, 1.5114, 1.5924], device='cuda:4'), covar=tensor([0.8119, 1.1899, 1.2064, 1.1362, 0.9456, 1.3805, 1.4340, 1.2310], device='cuda:4'), in_proj_covar=tensor([0.0411, 0.0437, 0.0521, 0.0541, 0.0439, 0.0459, 0.0473, 0.0471], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 18:43:54,702 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=33828.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:43:55,207 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.192e+02 1.912e+02 2.137e+02 2.570e+02 6.415e+02, threshold=4.274e+02, percent-clipped=3.0 2023-04-26 18:44:07,683 INFO [finetune.py:976] (4/7) Epoch 6, batch 5200, loss[loss=0.2077, simple_loss=0.2754, pruned_loss=0.07001, over 4821.00 frames. ], tot_loss[loss=0.2021, simple_loss=0.266, pruned_loss=0.06913, over 958334.79 frames. ], batch size: 30, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:44:27,122 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6105, 1.4414, 4.1012, 3.8445, 3.6180, 3.9000, 3.7607, 3.5936], device='cuda:4'), covar=tensor([0.6865, 0.5512, 0.1044, 0.1710, 0.1026, 0.1866, 0.2143, 0.1364], device='cuda:4'), in_proj_covar=tensor([0.0314, 0.0308, 0.0414, 0.0417, 0.0354, 0.0408, 0.0316, 0.0370], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-26 18:44:38,048 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=33859.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:44:49,967 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1978, 3.2340, 0.7623, 1.6988, 1.6323, 2.1414, 1.8121, 1.0449], device='cuda:4'), covar=tensor([0.1879, 0.1548, 0.2578, 0.1815, 0.1392, 0.1494, 0.1849, 0.2189], device='cuda:4'), in_proj_covar=tensor([0.0121, 0.0260, 0.0145, 0.0127, 0.0138, 0.0159, 0.0123, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 18:45:14,785 INFO [finetune.py:976] (4/7) Epoch 6, batch 5250, loss[loss=0.2364, simple_loss=0.2964, pruned_loss=0.08819, over 4847.00 frames. ], tot_loss[loss=0.2047, simple_loss=0.2689, pruned_loss=0.07025, over 957816.51 frames. ], batch size: 44, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:45:58,334 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.252e+02 1.714e+02 2.064e+02 2.470e+02 4.995e+02, threshold=4.129e+02, percent-clipped=2.0 2023-04-26 18:46:04,412 INFO [finetune.py:976] (4/7) Epoch 6, batch 5300, loss[loss=0.2567, simple_loss=0.3088, pruned_loss=0.1023, over 4828.00 frames. ], tot_loss[loss=0.2071, simple_loss=0.2709, pruned_loss=0.07166, over 956756.68 frames. ], batch size: 30, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:46:04,528 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=33939.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:46:10,183 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-04-26 18:46:21,713 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.9124, 3.7843, 2.6203, 4.4755, 3.8231, 3.9299, 1.8738, 3.8603], device='cuda:4'), covar=tensor([0.1854, 0.1183, 0.3485, 0.1486, 0.2091, 0.1840, 0.5353, 0.2425], device='cuda:4'), in_proj_covar=tensor([0.0243, 0.0219, 0.0253, 0.0310, 0.0302, 0.0254, 0.0274, 0.0273], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 18:46:38,100 INFO [finetune.py:976] (4/7) Epoch 6, batch 5350, loss[loss=0.1896, simple_loss=0.2657, pruned_loss=0.05672, over 4894.00 frames. ], tot_loss[loss=0.2075, simple_loss=0.2714, pruned_loss=0.07185, over 957528.51 frames. ], batch size: 32, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:46:46,591 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=34000.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:47:06,985 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.964e+01 1.645e+02 1.895e+02 2.308e+02 4.781e+02, threshold=3.789e+02, percent-clipped=2.0 2023-04-26 18:47:09,494 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=34033.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 18:47:13,057 INFO [finetune.py:976] (4/7) Epoch 6, batch 5400, loss[loss=0.201, simple_loss=0.2532, pruned_loss=0.07436, over 4769.00 frames. ], tot_loss[loss=0.2056, simple_loss=0.2693, pruned_loss=0.07099, over 958402.53 frames. ], batch size: 26, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:47:16,164 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=34044.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:47:20,911 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-26 18:47:25,101 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6995, 1.6436, 1.9651, 2.0551, 1.5982, 1.3001, 1.7421, 1.0313], device='cuda:4'), covar=tensor([0.0963, 0.0934, 0.0694, 0.1031, 0.1069, 0.1241, 0.0867, 0.1072], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0075, 0.0074, 0.0068, 0.0079, 0.0096, 0.0081, 0.0078], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 18:47:46,754 INFO [finetune.py:976] (4/7) Epoch 6, batch 5450, loss[loss=0.1887, simple_loss=0.2551, pruned_loss=0.06115, over 4827.00 frames. ], tot_loss[loss=0.2023, simple_loss=0.2652, pruned_loss=0.06969, over 957839.32 frames. ], batch size: 33, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:47:56,394 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9428, 1.4119, 4.6651, 4.3504, 4.1046, 4.3374, 4.1895, 4.0475], device='cuda:4'), covar=tensor([0.6504, 0.5603, 0.0889, 0.1688, 0.1014, 0.1473, 0.2097, 0.1642], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0306, 0.0411, 0.0415, 0.0352, 0.0406, 0.0315, 0.0369], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-26 18:47:57,050 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=34105.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:48:11,377 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.9638, 2.0329, 2.1827, 2.8237, 2.8989, 2.6644, 2.5506, 2.3222], device='cuda:4'), covar=tensor([0.1586, 0.1626, 0.1552, 0.1192, 0.0903, 0.1249, 0.1944, 0.1706], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0329, 0.0348, 0.0301, 0.0338, 0.0329, 0.0307, 0.0351], device='cuda:4'), out_proj_covar=tensor([6.5485e-05, 7.0328e-05, 7.5346e-05, 6.2537e-05, 7.1191e-05, 7.1011e-05, 6.6328e-05, 7.5749e-05], device='cuda:4') 2023-04-26 18:48:12,447 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=34128.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:48:12,948 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.072e+02 1.622e+02 2.010e+02 2.382e+02 4.007e+02, threshold=4.020e+02, percent-clipped=1.0 2023-04-26 18:48:24,303 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5307, 1.7863, 2.2417, 2.8419, 2.2657, 1.6916, 1.5972, 2.0343], device='cuda:4'), covar=tensor([0.3957, 0.4293, 0.1995, 0.3822, 0.3953, 0.3425, 0.5166, 0.3288], device='cuda:4'), in_proj_covar=tensor([0.0277, 0.0254, 0.0218, 0.0324, 0.0215, 0.0228, 0.0239, 0.0190], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 18:48:25,395 INFO [finetune.py:976] (4/7) Epoch 6, batch 5500, loss[loss=0.1746, simple_loss=0.234, pruned_loss=0.05757, over 4801.00 frames. ], tot_loss[loss=0.1993, simple_loss=0.2616, pruned_loss=0.06848, over 957367.64 frames. ], batch size: 26, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:48:29,204 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=34145.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:49:00,089 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.63 vs. limit=2.0 2023-04-26 18:49:10,838 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=34176.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:49:30,902 INFO [finetune.py:976] (4/7) Epoch 6, batch 5550, loss[loss=0.2165, simple_loss=0.2791, pruned_loss=0.07696, over 4904.00 frames. ], tot_loss[loss=0.2029, simple_loss=0.2648, pruned_loss=0.07053, over 955535.55 frames. ], batch size: 32, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:49:52,184 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=34206.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:50:24,793 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.272e+02 1.890e+02 2.190e+02 2.677e+02 6.143e+02, threshold=4.380e+02, percent-clipped=3.0 2023-04-26 18:50:36,983 INFO [finetune.py:976] (4/7) Epoch 6, batch 5600, loss[loss=0.2059, simple_loss=0.2842, pruned_loss=0.06379, over 4800.00 frames. ], tot_loss[loss=0.2054, simple_loss=0.2683, pruned_loss=0.07121, over 955217.12 frames. ], batch size: 29, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:51:28,269 INFO [finetune.py:976] (4/7) Epoch 6, batch 5650, loss[loss=0.2573, simple_loss=0.3195, pruned_loss=0.0976, over 4811.00 frames. ], tot_loss[loss=0.2076, simple_loss=0.2713, pruned_loss=0.07196, over 956006.77 frames. ], batch size: 45, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:51:31,888 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=34295.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:52:02,999 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.004e+02 1.702e+02 2.093e+02 2.545e+02 5.557e+02, threshold=4.186e+02, percent-clipped=1.0 2023-04-26 18:52:05,462 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=34333.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:52:08,281 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-04-26 18:52:08,526 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9561, 1.4587, 1.6155, 1.7109, 2.1363, 1.8438, 1.5176, 1.4771], device='cuda:4'), covar=tensor([0.2076, 0.1871, 0.2058, 0.1720, 0.0983, 0.1358, 0.2569, 0.2085], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0331, 0.0351, 0.0303, 0.0340, 0.0330, 0.0309, 0.0354], device='cuda:4'), out_proj_covar=tensor([6.6215e-05, 7.0631e-05, 7.6161e-05, 6.3011e-05, 7.1675e-05, 7.1329e-05, 6.6821e-05, 7.6220e-05], device='cuda:4') 2023-04-26 18:52:09,024 INFO [finetune.py:976] (4/7) Epoch 6, batch 5700, loss[loss=0.21, simple_loss=0.271, pruned_loss=0.07457, over 4080.00 frames. ], tot_loss[loss=0.2045, simple_loss=0.2666, pruned_loss=0.07121, over 935740.41 frames. ], batch size: 18, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:52:39,657 INFO [finetune.py:976] (4/7) Epoch 7, batch 0, loss[loss=0.234, simple_loss=0.2915, pruned_loss=0.08821, over 4894.00 frames. ], tot_loss[loss=0.234, simple_loss=0.2915, pruned_loss=0.08821, over 4894.00 frames. ], batch size: 36, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:52:39,657 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-26 18:52:50,219 INFO [finetune.py:1010] (4/7) Epoch 7, validation: loss=0.1579, simple_loss=0.2317, pruned_loss=0.04207, over 2265189.00 frames. 2023-04-26 18:52:50,220 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6259MB 2023-04-26 18:52:59,340 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=34381.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:53:11,519 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=34400.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:53:23,181 INFO [finetune.py:976] (4/7) Epoch 7, batch 50, loss[loss=0.1466, simple_loss=0.2199, pruned_loss=0.03671, over 4734.00 frames. ], tot_loss[loss=0.2115, simple_loss=0.2742, pruned_loss=0.07437, over 217238.81 frames. ], batch size: 27, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:53:23,287 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0143, 1.4030, 1.2189, 1.6098, 1.4388, 1.4746, 1.3103, 2.4202], device='cuda:4'), covar=tensor([0.0634, 0.0701, 0.0766, 0.1114, 0.0575, 0.0604, 0.0694, 0.0261], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0039, 0.0041, 0.0045, 0.0041, 0.0040, 0.0039, 0.0061], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 18:53:32,046 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.224e+02 1.800e+02 2.094e+02 2.566e+02 4.468e+02, threshold=4.189e+02, percent-clipped=1.0 2023-04-26 18:53:34,634 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-26 18:53:56,441 INFO [finetune.py:976] (4/7) Epoch 7, batch 100, loss[loss=0.1861, simple_loss=0.2579, pruned_loss=0.05711, over 4789.00 frames. ], tot_loss[loss=0.2019, simple_loss=0.265, pruned_loss=0.06941, over 382057.20 frames. ], batch size: 51, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:54:12,172 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3762, 1.8544, 1.6125, 2.1440, 1.8606, 2.1402, 1.5879, 4.4824], device='cuda:4'), covar=tensor([0.0583, 0.0738, 0.0792, 0.1124, 0.0623, 0.0515, 0.0734, 0.0110], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0039, 0.0041, 0.0044, 0.0040, 0.0040, 0.0039, 0.0060], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 18:54:19,344 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=34501.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:54:29,756 INFO [finetune.py:976] (4/7) Epoch 7, batch 150, loss[loss=0.2455, simple_loss=0.3011, pruned_loss=0.09497, over 4718.00 frames. ], tot_loss[loss=0.196, simple_loss=0.2583, pruned_loss=0.06683, over 509083.35 frames. ], batch size: 59, lr: 3.89e-03, grad_scale: 16.0 2023-04-26 18:54:30,513 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4635, 1.3754, 1.7706, 1.7923, 1.3994, 1.1290, 1.4913, 0.9809], device='cuda:4'), covar=tensor([0.0718, 0.0904, 0.0562, 0.0842, 0.0966, 0.1374, 0.0818, 0.0880], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0075, 0.0073, 0.0067, 0.0078, 0.0095, 0.0080, 0.0077], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 18:54:39,165 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.220e+02 1.717e+02 2.069e+02 2.417e+02 7.374e+02, threshold=4.138e+02, percent-clipped=4.0 2023-04-26 18:54:40,571 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7708, 1.0567, 1.3627, 1.4987, 1.5067, 1.6738, 1.4543, 1.4557], device='cuda:4'), covar=tensor([0.6059, 0.8241, 0.7565, 0.7910, 0.8269, 1.2358, 0.8532, 0.7243], device='cuda:4'), in_proj_covar=tensor([0.0320, 0.0393, 0.0318, 0.0329, 0.0344, 0.0410, 0.0374, 0.0330], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 18:55:03,722 INFO [finetune.py:976] (4/7) Epoch 7, batch 200, loss[loss=0.1882, simple_loss=0.2514, pruned_loss=0.06245, over 4901.00 frames. ], tot_loss[loss=0.198, simple_loss=0.2594, pruned_loss=0.06823, over 610661.29 frames. ], batch size: 43, lr: 3.89e-03, grad_scale: 32.0 2023-04-26 18:55:27,261 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=34593.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:55:33,682 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=34595.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:56:04,706 INFO [finetune.py:976] (4/7) Epoch 7, batch 250, loss[loss=0.1803, simple_loss=0.2391, pruned_loss=0.06077, over 4769.00 frames. ], tot_loss[loss=0.2044, simple_loss=0.2659, pruned_loss=0.07142, over 686638.55 frames. ], batch size: 26, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 18:56:09,038 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=34622.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:56:17,340 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=34626.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:56:19,025 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.076e+02 1.883e+02 2.183e+02 2.769e+02 4.729e+02, threshold=4.366e+02, percent-clipped=3.0 2023-04-26 18:56:39,246 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=34643.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:56:51,385 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=34654.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:57:04,728 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-26 18:57:11,714 INFO [finetune.py:976] (4/7) Epoch 7, batch 300, loss[loss=0.1834, simple_loss=0.2583, pruned_loss=0.05422, over 4897.00 frames. ], tot_loss[loss=0.2059, simple_loss=0.269, pruned_loss=0.07137, over 749559.28 frames. ], batch size: 32, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 18:57:35,515 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=34683.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 18:57:37,962 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=34687.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:57:51,404 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=34700.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:58:03,716 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=34710.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 18:58:12,742 INFO [finetune.py:976] (4/7) Epoch 7, batch 350, loss[loss=0.2415, simple_loss=0.3075, pruned_loss=0.08772, over 4820.00 frames. ], tot_loss[loss=0.2091, simple_loss=0.2723, pruned_loss=0.07296, over 795129.94 frames. ], batch size: 33, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 18:58:28,029 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.287e+02 1.776e+02 2.211e+02 2.674e+02 4.769e+02, threshold=4.422e+02, percent-clipped=1.0 2023-04-26 18:58:50,331 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=34748.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:59:02,248 INFO [finetune.py:976] (4/7) Epoch 7, batch 400, loss[loss=0.2252, simple_loss=0.2824, pruned_loss=0.08401, over 4785.00 frames. ], tot_loss[loss=0.2072, simple_loss=0.2715, pruned_loss=0.07145, over 830567.29 frames. ], batch size: 26, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 18:59:04,061 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.4909, 4.4116, 3.0245, 5.1648, 4.5592, 4.4569, 1.9577, 4.2883], device='cuda:4'), covar=tensor([0.1533, 0.0934, 0.3302, 0.0944, 0.2351, 0.1637, 0.5564, 0.2183], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0219, 0.0254, 0.0310, 0.0302, 0.0254, 0.0274, 0.0276], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 18:59:05,925 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=34771.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 18:59:26,226 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=34801.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 18:59:36,189 INFO [finetune.py:976] (4/7) Epoch 7, batch 450, loss[loss=0.1708, simple_loss=0.2416, pruned_loss=0.05003, over 4896.00 frames. ], tot_loss[loss=0.2065, simple_loss=0.2701, pruned_loss=0.07141, over 858049.74 frames. ], batch size: 37, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 18:59:37,490 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1212, 2.9926, 0.8568, 1.2167, 1.9555, 1.2884, 3.8215, 1.6211], device='cuda:4'), covar=tensor([0.0862, 0.0879, 0.1191, 0.1694, 0.0729, 0.1372, 0.0316, 0.0935], device='cuda:4'), in_proj_covar=tensor([0.0053, 0.0069, 0.0051, 0.0048, 0.0053, 0.0054, 0.0081, 0.0052], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 18:59:45,544 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.159e+02 1.735e+02 2.065e+02 2.383e+02 5.194e+02, threshold=4.130e+02, percent-clipped=2.0 2023-04-26 18:59:58,681 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=34849.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:00:09,438 INFO [finetune.py:976] (4/7) Epoch 7, batch 500, loss[loss=0.2346, simple_loss=0.2829, pruned_loss=0.09315, over 4899.00 frames. ], tot_loss[loss=0.2026, simple_loss=0.2661, pruned_loss=0.06951, over 880993.87 frames. ], batch size: 35, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 19:00:42,229 INFO [finetune.py:976] (4/7) Epoch 7, batch 550, loss[loss=0.154, simple_loss=0.221, pruned_loss=0.04351, over 4742.00 frames. ], tot_loss[loss=0.1989, simple_loss=0.2621, pruned_loss=0.06787, over 898779.24 frames. ], batch size: 27, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 19:00:51,094 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.195e+02 1.669e+02 2.030e+02 2.418e+02 4.759e+02, threshold=4.059e+02, percent-clipped=2.0 2023-04-26 19:01:10,925 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=34949.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:01:12,018 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-26 19:01:26,569 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.48 vs. limit=2.0 2023-04-26 19:01:32,182 INFO [finetune.py:976] (4/7) Epoch 7, batch 600, loss[loss=0.2466, simple_loss=0.3018, pruned_loss=0.09573, over 4931.00 frames. ], tot_loss[loss=0.1999, simple_loss=0.2626, pruned_loss=0.06855, over 910535.45 frames. ], batch size: 33, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 19:01:45,331 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=34978.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 19:01:54,262 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=34982.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:01:58,045 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.58 vs. limit=5.0 2023-04-26 19:02:08,845 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-04-26 19:02:27,944 INFO [finetune.py:976] (4/7) Epoch 7, batch 650, loss[loss=0.1745, simple_loss=0.244, pruned_loss=0.0525, over 4866.00 frames. ], tot_loss[loss=0.2026, simple_loss=0.266, pruned_loss=0.06958, over 921538.04 frames. ], batch size: 34, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 19:02:42,342 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.097e+02 1.760e+02 2.002e+02 2.405e+02 4.301e+02, threshold=4.004e+02, percent-clipped=1.0 2023-04-26 19:03:19,048 INFO [finetune.py:976] (4/7) Epoch 7, batch 700, loss[loss=0.2063, simple_loss=0.2691, pruned_loss=0.07173, over 4907.00 frames. ], tot_loss[loss=0.2038, simple_loss=0.2676, pruned_loss=0.06995, over 925907.66 frames. ], batch size: 32, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 19:03:19,119 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=35066.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 19:03:47,925 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=35092.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:04:19,897 INFO [finetune.py:976] (4/7) Epoch 7, batch 750, loss[loss=0.2522, simple_loss=0.3047, pruned_loss=0.09982, over 4229.00 frames. ], tot_loss[loss=0.2069, simple_loss=0.2708, pruned_loss=0.07148, over 932713.71 frames. ], batch size: 66, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 19:04:33,726 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.243e+02 1.705e+02 2.087e+02 2.408e+02 7.580e+02, threshold=4.175e+02, percent-clipped=5.0 2023-04-26 19:05:05,733 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5783, 3.7614, 0.8679, 1.9341, 1.8756, 2.5959, 1.9670, 0.9908], device='cuda:4'), covar=tensor([0.1413, 0.0892, 0.2075, 0.1335, 0.1166, 0.1028, 0.1497, 0.2020], device='cuda:4'), in_proj_covar=tensor([0.0120, 0.0257, 0.0144, 0.0126, 0.0137, 0.0158, 0.0122, 0.0126], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 19:05:05,772 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=35153.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:05:25,182 INFO [finetune.py:976] (4/7) Epoch 7, batch 800, loss[loss=0.1894, simple_loss=0.262, pruned_loss=0.05835, over 4898.00 frames. ], tot_loss[loss=0.2056, simple_loss=0.2698, pruned_loss=0.0707, over 938290.32 frames. ], batch size: 35, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 19:06:12,640 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.64 vs. limit=2.0 2023-04-26 19:06:20,321 INFO [finetune.py:976] (4/7) Epoch 7, batch 850, loss[loss=0.1785, simple_loss=0.243, pruned_loss=0.05698, over 4922.00 frames. ], tot_loss[loss=0.2036, simple_loss=0.267, pruned_loss=0.07007, over 943454.50 frames. ], batch size: 38, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 19:06:32,702 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.117e+02 1.732e+02 2.031e+02 2.393e+02 4.995e+02, threshold=4.061e+02, percent-clipped=2.0 2023-04-26 19:06:58,280 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=35249.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:07:20,079 INFO [finetune.py:976] (4/7) Epoch 7, batch 900, loss[loss=0.1744, simple_loss=0.2359, pruned_loss=0.05647, over 4894.00 frames. ], tot_loss[loss=0.2019, simple_loss=0.2649, pruned_loss=0.06947, over 947382.87 frames. ], batch size: 43, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 19:07:20,665 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8412, 2.2977, 2.0581, 2.1944, 1.5818, 1.9384, 1.9239, 1.6540], device='cuda:4'), covar=tensor([0.1902, 0.1214, 0.0725, 0.1069, 0.2933, 0.1146, 0.1616, 0.2051], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0326, 0.0238, 0.0298, 0.0319, 0.0278, 0.0267, 0.0290], device='cuda:4'), out_proj_covar=tensor([1.2532e-04, 1.3224e-04, 9.6261e-05, 1.1984e-04, 1.3137e-04, 1.1242e-04, 1.0927e-04, 1.1653e-04], device='cuda:4') 2023-04-26 19:07:38,488 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=35278.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 19:07:40,977 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=35282.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:07:54,292 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.81 vs. limit=2.0 2023-04-26 19:08:02,084 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=35297.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:08:14,477 INFO [finetune.py:976] (4/7) Epoch 7, batch 950, loss[loss=0.1616, simple_loss=0.2329, pruned_loss=0.04514, over 4812.00 frames. ], tot_loss[loss=0.2012, simple_loss=0.2637, pruned_loss=0.06934, over 950945.29 frames. ], batch size: 25, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 19:08:27,253 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=35326.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:08:29,016 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.023e+02 1.623e+02 1.972e+02 2.345e+02 3.927e+02, threshold=3.944e+02, percent-clipped=0.0 2023-04-26 19:08:29,702 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=35330.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:08:38,481 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6374, 1.3567, 1.7154, 2.0505, 1.7849, 1.6180, 1.6646, 1.6692], device='cuda:4'), covar=tensor([0.7819, 1.0604, 1.1044, 1.1124, 0.9113, 1.2495, 1.3319, 1.1402], device='cuda:4'), in_proj_covar=tensor([0.0413, 0.0436, 0.0521, 0.0541, 0.0440, 0.0461, 0.0473, 0.0472], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 19:08:50,433 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=35345.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 19:09:01,890 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.77 vs. limit=5.0 2023-04-26 19:09:20,446 INFO [finetune.py:976] (4/7) Epoch 7, batch 1000, loss[loss=0.1835, simple_loss=0.2503, pruned_loss=0.05836, over 4745.00 frames. ], tot_loss[loss=0.2024, simple_loss=0.2655, pruned_loss=0.06965, over 951552.25 frames. ], batch size: 26, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 19:09:20,553 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=35366.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 19:09:29,988 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=35374.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:10:06,697 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=35406.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 19:10:18,216 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=35414.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 19:10:25,353 INFO [finetune.py:976] (4/7) Epoch 7, batch 1050, loss[loss=0.1974, simple_loss=0.2598, pruned_loss=0.06745, over 4912.00 frames. ], tot_loss[loss=0.2059, simple_loss=0.27, pruned_loss=0.07096, over 953860.23 frames. ], batch size: 36, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 19:10:39,105 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.246e+02 1.810e+02 2.204e+02 2.753e+02 5.526e+02, threshold=4.408e+02, percent-clipped=1.0 2023-04-26 19:10:48,434 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=35435.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:11:06,921 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=35448.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:11:31,122 INFO [finetune.py:976] (4/7) Epoch 7, batch 1100, loss[loss=0.2229, simple_loss=0.291, pruned_loss=0.07743, over 4893.00 frames. ], tot_loss[loss=0.2079, simple_loss=0.2715, pruned_loss=0.07212, over 955336.94 frames. ], batch size: 43, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 19:11:44,636 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6033, 1.4655, 0.6226, 1.2800, 1.5891, 1.4483, 1.3442, 1.3727], device='cuda:4'), covar=tensor([0.0602, 0.0431, 0.0469, 0.0619, 0.0325, 0.0653, 0.0585, 0.0700], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0026, 0.0024, 0.0031, 0.0021, 0.0030, 0.0029, 0.0031], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0045, 0.0038, 0.0050, 0.0038, 0.0048, 0.0048, 0.0050], device='cuda:4') 2023-04-26 19:11:51,479 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.27 vs. limit=5.0 2023-04-26 19:11:57,258 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7625, 2.2694, 1.7121, 1.4468, 1.2835, 1.3232, 1.8632, 1.2325], device='cuda:4'), covar=tensor([0.1877, 0.1570, 0.1713, 0.2207, 0.2687, 0.2114, 0.1214, 0.2243], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0219, 0.0175, 0.0206, 0.0210, 0.0187, 0.0167, 0.0191], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 19:12:09,916 INFO [finetune.py:976] (4/7) Epoch 7, batch 1150, loss[loss=0.1886, simple_loss=0.2599, pruned_loss=0.0587, over 4888.00 frames. ], tot_loss[loss=0.21, simple_loss=0.2736, pruned_loss=0.07316, over 957257.33 frames. ], batch size: 43, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 19:12:18,334 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.291e+02 1.835e+02 2.101e+02 2.448e+02 6.183e+02, threshold=4.202e+02, percent-clipped=2.0 2023-04-26 19:12:19,088 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=35530.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:12:43,336 INFO [finetune.py:976] (4/7) Epoch 7, batch 1200, loss[loss=0.191, simple_loss=0.2638, pruned_loss=0.05914, over 4740.00 frames. ], tot_loss[loss=0.2079, simple_loss=0.2712, pruned_loss=0.07234, over 955943.90 frames. ], batch size: 59, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 19:13:00,092 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=35591.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:13:17,334 INFO [finetune.py:976] (4/7) Epoch 7, batch 1250, loss[loss=0.1999, simple_loss=0.271, pruned_loss=0.06434, over 4816.00 frames. ], tot_loss[loss=0.2053, simple_loss=0.268, pruned_loss=0.0713, over 957605.86 frames. ], batch size: 41, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 19:13:26,240 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.030e+02 1.763e+02 2.030e+02 2.428e+02 4.549e+02, threshold=4.060e+02, percent-clipped=1.0 2023-04-26 19:13:27,678 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.23 vs. limit=5.0 2023-04-26 19:13:35,170 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.83 vs. limit=5.0 2023-04-26 19:13:44,941 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4139, 1.1767, 1.5408, 1.5137, 1.2676, 1.1296, 1.2614, 0.7341], device='cuda:4'), covar=tensor([0.0710, 0.1207, 0.0770, 0.0809, 0.0961, 0.1373, 0.0834, 0.1081], device='cuda:4'), in_proj_covar=tensor([0.0066, 0.0075, 0.0072, 0.0067, 0.0077, 0.0095, 0.0080, 0.0076], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 19:13:51,308 INFO [finetune.py:976] (4/7) Epoch 7, batch 1300, loss[loss=0.2137, simple_loss=0.2853, pruned_loss=0.07111, over 4820.00 frames. ], tot_loss[loss=0.202, simple_loss=0.2643, pruned_loss=0.06985, over 956488.48 frames. ], batch size: 38, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 19:14:04,831 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.99 vs. limit=2.0 2023-04-26 19:14:14,093 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=35701.0, num_to_drop=1, layers_to_drop={3} 2023-04-26 19:14:25,119 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.8720, 2.8432, 2.3084, 3.3195, 2.8604, 2.9192, 1.3692, 2.7964], device='cuda:4'), covar=tensor([0.2189, 0.1732, 0.3186, 0.2591, 0.3480, 0.2199, 0.5713, 0.2803], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0220, 0.0254, 0.0311, 0.0304, 0.0255, 0.0275, 0.0274], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 19:14:35,193 INFO [finetune.py:976] (4/7) Epoch 7, batch 1350, loss[loss=0.1781, simple_loss=0.241, pruned_loss=0.05754, over 4826.00 frames. ], tot_loss[loss=0.2014, simple_loss=0.2638, pruned_loss=0.06951, over 957785.76 frames. ], batch size: 30, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 19:14:54,298 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.038e+02 1.639e+02 1.977e+02 2.381e+02 4.004e+02, threshold=3.953e+02, percent-clipped=0.0 2023-04-26 19:14:54,987 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=35730.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:15:17,740 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=35748.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:15:40,833 INFO [finetune.py:976] (4/7) Epoch 7, batch 1400, loss[loss=0.2025, simple_loss=0.2806, pruned_loss=0.06224, over 4840.00 frames. ], tot_loss[loss=0.2045, simple_loss=0.2679, pruned_loss=0.07055, over 956466.24 frames. ], batch size: 40, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 19:15:44,442 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=35771.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:16:06,188 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=35796.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:16:26,257 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=35811.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:16:30,134 INFO [finetune.py:976] (4/7) Epoch 7, batch 1450, loss[loss=0.16, simple_loss=0.223, pruned_loss=0.04854, over 4776.00 frames. ], tot_loss[loss=0.2054, simple_loss=0.2692, pruned_loss=0.07079, over 957135.25 frames. ], batch size: 26, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 19:16:50,309 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.278e+02 1.770e+02 2.190e+02 2.562e+02 4.335e+02, threshold=4.381e+02, percent-clipped=3.0 2023-04-26 19:16:58,484 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=35832.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:17:41,154 INFO [finetune.py:976] (4/7) Epoch 7, batch 1500, loss[loss=0.2084, simple_loss=0.2716, pruned_loss=0.07266, over 4917.00 frames. ], tot_loss[loss=0.2063, simple_loss=0.27, pruned_loss=0.07128, over 958570.92 frames. ], batch size: 33, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 19:17:43,870 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-04-26 19:17:46,521 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=35872.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:17:55,867 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=35886.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:18:31,248 INFO [finetune.py:976] (4/7) Epoch 7, batch 1550, loss[loss=0.1744, simple_loss=0.2461, pruned_loss=0.0513, over 4888.00 frames. ], tot_loss[loss=0.2063, simple_loss=0.27, pruned_loss=0.07127, over 958167.27 frames. ], batch size: 35, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 19:18:51,455 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.224e+02 1.799e+02 2.113e+02 2.516e+02 6.961e+02, threshold=4.225e+02, percent-clipped=2.0 2023-04-26 19:19:14,916 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-26 19:19:38,739 INFO [finetune.py:976] (4/7) Epoch 7, batch 1600, loss[loss=0.1755, simple_loss=0.244, pruned_loss=0.05353, over 4844.00 frames. ], tot_loss[loss=0.2038, simple_loss=0.2669, pruned_loss=0.0704, over 957272.61 frames. ], batch size: 47, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 19:19:57,969 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.16 vs. limit=5.0 2023-04-26 19:20:30,770 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=36001.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 19:20:45,229 INFO [finetune.py:976] (4/7) Epoch 7, batch 1650, loss[loss=0.1683, simple_loss=0.2397, pruned_loss=0.04849, over 4775.00 frames. ], tot_loss[loss=0.2011, simple_loss=0.2638, pruned_loss=0.06921, over 957949.24 frames. ], batch size: 28, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 19:20:59,361 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.12 vs. limit=5.0 2023-04-26 19:20:59,791 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.070e+02 1.778e+02 2.122e+02 2.492e+02 4.114e+02, threshold=4.243e+02, percent-clipped=0.0 2023-04-26 19:21:00,969 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=36030.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:21:26,277 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=36049.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 19:21:48,195 INFO [finetune.py:976] (4/7) Epoch 7, batch 1700, loss[loss=0.1746, simple_loss=0.2278, pruned_loss=0.06069, over 4812.00 frames. ], tot_loss[loss=0.201, simple_loss=0.263, pruned_loss=0.06954, over 959452.54 frames. ], batch size: 25, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 19:21:56,086 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=36078.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:22:10,251 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.7515, 1.7526, 1.7640, 1.4979, 1.9766, 1.6526, 2.5255, 1.5098], device='cuda:4'), covar=tensor([0.4019, 0.1902, 0.4954, 0.3172, 0.1811, 0.2481, 0.1419, 0.5017], device='cuda:4'), in_proj_covar=tensor([0.0348, 0.0354, 0.0436, 0.0364, 0.0392, 0.0385, 0.0388, 0.0421], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 19:22:21,089 INFO [finetune.py:976] (4/7) Epoch 7, batch 1750, loss[loss=0.1856, simple_loss=0.2586, pruned_loss=0.05624, over 4761.00 frames. ], tot_loss[loss=0.2036, simple_loss=0.2658, pruned_loss=0.07069, over 959545.79 frames. ], batch size: 26, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 19:22:23,538 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-26 19:22:28,810 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=36127.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:22:29,945 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.331e+02 1.866e+02 2.133e+02 2.694e+02 4.644e+02, threshold=4.265e+02, percent-clipped=2.0 2023-04-26 19:22:54,680 INFO [finetune.py:976] (4/7) Epoch 7, batch 1800, loss[loss=0.2238, simple_loss=0.2871, pruned_loss=0.08029, over 4858.00 frames. ], tot_loss[loss=0.2066, simple_loss=0.2695, pruned_loss=0.07186, over 956992.48 frames. ], batch size: 34, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 19:22:55,383 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=36167.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:23:05,440 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=36181.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:23:08,950 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=36186.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:23:44,253 INFO [finetune.py:976] (4/7) Epoch 7, batch 1850, loss[loss=0.184, simple_loss=0.2606, pruned_loss=0.05367, over 4761.00 frames. ], tot_loss[loss=0.2078, simple_loss=0.2705, pruned_loss=0.07254, over 957649.35 frames. ], batch size: 28, lr: 3.88e-03, grad_scale: 32.0 2023-04-26 19:24:03,617 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.013e+02 1.720e+02 2.075e+02 2.567e+02 3.756e+02, threshold=4.150e+02, percent-clipped=0.0 2023-04-26 19:24:07,670 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=36234.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:24:19,055 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=36242.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:24:41,757 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.3701, 3.2954, 2.6439, 3.8940, 3.2892, 3.4337, 1.4651, 3.3330], device='cuda:4'), covar=tensor([0.1976, 0.1269, 0.2985, 0.1972, 0.3111, 0.1813, 0.5935, 0.2380], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0219, 0.0253, 0.0313, 0.0304, 0.0254, 0.0275, 0.0274], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 19:24:45,934 INFO [finetune.py:976] (4/7) Epoch 7, batch 1900, loss[loss=0.1747, simple_loss=0.243, pruned_loss=0.05319, over 4781.00 frames. ], tot_loss[loss=0.2081, simple_loss=0.2714, pruned_loss=0.07244, over 958039.39 frames. ], batch size: 25, lr: 3.87e-03, grad_scale: 32.0 2023-04-26 19:24:47,282 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2241, 2.6262, 1.0070, 1.3607, 1.8574, 1.2687, 3.1151, 1.6640], device='cuda:4'), covar=tensor([0.0670, 0.0572, 0.0770, 0.1322, 0.0521, 0.1015, 0.0309, 0.0682], device='cuda:4'), in_proj_covar=tensor([0.0053, 0.0069, 0.0051, 0.0048, 0.0053, 0.0053, 0.0080, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0011, 0.0007], device='cuda:4') 2023-04-26 19:25:13,713 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4337, 3.3186, 0.9554, 1.7484, 1.8853, 2.4411, 1.9249, 1.0765], device='cuda:4'), covar=tensor([0.1507, 0.1068, 0.2181, 0.1449, 0.1130, 0.1080, 0.1500, 0.1928], device='cuda:4'), in_proj_covar=tensor([0.0120, 0.0257, 0.0145, 0.0127, 0.0137, 0.0158, 0.0122, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 19:25:19,583 INFO [finetune.py:976] (4/7) Epoch 7, batch 1950, loss[loss=0.1733, simple_loss=0.2463, pruned_loss=0.05016, over 4928.00 frames. ], tot_loss[loss=0.2066, simple_loss=0.2698, pruned_loss=0.07169, over 958483.38 frames. ], batch size: 33, lr: 3.87e-03, grad_scale: 32.0 2023-04-26 19:25:27,368 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.259e+02 1.677e+02 1.937e+02 2.378e+02 4.833e+02, threshold=3.873e+02, percent-clipped=2.0 2023-04-26 19:25:58,965 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.71 vs. limit=2.0 2023-04-26 19:26:09,641 INFO [finetune.py:976] (4/7) Epoch 7, batch 2000, loss[loss=0.175, simple_loss=0.2411, pruned_loss=0.05443, over 4394.00 frames. ], tot_loss[loss=0.204, simple_loss=0.2668, pruned_loss=0.07063, over 958467.75 frames. ], batch size: 19, lr: 3.87e-03, grad_scale: 32.0 2023-04-26 19:26:50,718 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7617, 2.0899, 1.0796, 1.4609, 2.1281, 1.6582, 1.6033, 1.6600], device='cuda:4'), covar=tensor([0.0528, 0.0369, 0.0349, 0.0596, 0.0252, 0.0520, 0.0533, 0.0594], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0026, 0.0024, 0.0030, 0.0021, 0.0030, 0.0029, 0.0030], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0050, 0.0037, 0.0048, 0.0048, 0.0050], device='cuda:4') 2023-04-26 19:27:00,247 INFO [finetune.py:976] (4/7) Epoch 7, batch 2050, loss[loss=0.2, simple_loss=0.251, pruned_loss=0.07451, over 4927.00 frames. ], tot_loss[loss=0.201, simple_loss=0.2632, pruned_loss=0.06937, over 959865.38 frames. ], batch size: 46, lr: 3.87e-03, grad_scale: 32.0 2023-04-26 19:27:07,117 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=36427.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:27:08,258 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.074e+02 1.550e+02 1.908e+02 2.415e+02 5.206e+02, threshold=3.816e+02, percent-clipped=3.0 2023-04-26 19:27:43,883 INFO [finetune.py:976] (4/7) Epoch 7, batch 2100, loss[loss=0.1455, simple_loss=0.2243, pruned_loss=0.03331, over 4895.00 frames. ], tot_loss[loss=0.1998, simple_loss=0.2624, pruned_loss=0.06859, over 959555.34 frames. ], batch size: 32, lr: 3.87e-03, grad_scale: 32.0 2023-04-26 19:27:45,093 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=36467.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:27:49,910 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=36475.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:27:56,128 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7404, 2.4255, 1.8146, 1.7054, 1.2941, 1.3202, 1.8963, 1.2439], device='cuda:4'), covar=tensor([0.1809, 0.1571, 0.1583, 0.1947, 0.2489, 0.2069, 0.1133, 0.2178], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0218, 0.0173, 0.0204, 0.0209, 0.0186, 0.0165, 0.0190], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-26 19:28:17,167 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=36515.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:28:17,704 INFO [finetune.py:976] (4/7) Epoch 7, batch 2150, loss[loss=0.2329, simple_loss=0.2932, pruned_loss=0.08632, over 4918.00 frames. ], tot_loss[loss=0.2023, simple_loss=0.2652, pruned_loss=0.06964, over 959943.32 frames. ], batch size: 42, lr: 3.87e-03, grad_scale: 32.0 2023-04-26 19:28:25,936 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.260e+02 1.822e+02 2.196e+02 2.668e+02 7.231e+02, threshold=4.392e+02, percent-clipped=1.0 2023-04-26 19:28:30,879 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=36537.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:28:40,078 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-26 19:28:51,287 INFO [finetune.py:976] (4/7) Epoch 7, batch 2200, loss[loss=0.2003, simple_loss=0.2606, pruned_loss=0.07002, over 4797.00 frames. ], tot_loss[loss=0.2045, simple_loss=0.2677, pruned_loss=0.0706, over 958656.58 frames. ], batch size: 25, lr: 3.87e-03, grad_scale: 64.0 2023-04-26 19:29:04,002 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4080, 1.6111, 1.6720, 2.1508, 1.7605, 2.1380, 1.5838, 4.6025], device='cuda:4'), covar=tensor([0.0667, 0.0798, 0.0811, 0.1157, 0.0690, 0.0587, 0.0747, 0.0136], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0039, 0.0041, 0.0044, 0.0041, 0.0040, 0.0039, 0.0060], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 19:29:28,457 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=36605.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:29:30,352 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=36608.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:29:38,711 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=36612.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:29:41,093 INFO [finetune.py:976] (4/7) Epoch 7, batch 2250, loss[loss=0.2047, simple_loss=0.2568, pruned_loss=0.07632, over 4925.00 frames. ], tot_loss[loss=0.2062, simple_loss=0.2695, pruned_loss=0.07144, over 958966.00 frames. ], batch size: 28, lr: 3.87e-03, grad_scale: 64.0 2023-04-26 19:30:00,405 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.228e+02 1.830e+02 2.114e+02 2.598e+02 4.357e+02, threshold=4.229e+02, percent-clipped=0.0 2023-04-26 19:30:12,633 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4617, 1.2697, 4.1269, 3.8620, 3.6280, 3.8776, 3.8330, 3.6402], device='cuda:4'), covar=tensor([0.7242, 0.5901, 0.0942, 0.1518, 0.1177, 0.1631, 0.2239, 0.1535], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0305, 0.0411, 0.0414, 0.0353, 0.0409, 0.0318, 0.0367], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-26 19:30:24,826 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1799, 1.5978, 2.0273, 2.4937, 1.9022, 1.5782, 1.2582, 1.8383], device='cuda:4'), covar=tensor([0.4389, 0.4764, 0.2258, 0.3371, 0.4059, 0.3742, 0.5543, 0.3140], device='cuda:4'), in_proj_covar=tensor([0.0278, 0.0253, 0.0217, 0.0321, 0.0214, 0.0228, 0.0238, 0.0189], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-26 19:30:40,443 INFO [finetune.py:976] (4/7) Epoch 7, batch 2300, loss[loss=0.1835, simple_loss=0.2511, pruned_loss=0.05792, over 4880.00 frames. ], tot_loss[loss=0.2067, simple_loss=0.2703, pruned_loss=0.07151, over 959549.07 frames. ], batch size: 32, lr: 3.87e-03, grad_scale: 64.0 2023-04-26 19:30:40,578 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=36666.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:30:42,379 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=36669.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:30:45,355 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=36673.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:31:02,276 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=36700.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 19:31:23,676 INFO [finetune.py:976] (4/7) Epoch 7, batch 2350, loss[loss=0.1576, simple_loss=0.2219, pruned_loss=0.04659, over 4756.00 frames. ], tot_loss[loss=0.2048, simple_loss=0.2677, pruned_loss=0.07093, over 956564.04 frames. ], batch size: 27, lr: 3.87e-03, grad_scale: 64.0 2023-04-26 19:31:24,892 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4424, 1.7485, 1.7796, 2.1719, 1.8765, 2.2257, 1.5101, 4.3644], device='cuda:4'), covar=tensor([0.0614, 0.0782, 0.0759, 0.1182, 0.0650, 0.0517, 0.0768, 0.0144], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0039, 0.0041, 0.0045, 0.0041, 0.0040, 0.0039, 0.0061], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 19:31:38,221 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.607e+01 1.814e+02 2.081e+02 2.499e+02 5.543e+02, threshold=4.163e+02, percent-clipped=2.0 2023-04-26 19:31:38,359 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3753, 2.3399, 1.9137, 2.1751, 2.4333, 1.9037, 3.1756, 1.8135], device='cuda:4'), covar=tensor([0.4987, 0.2286, 0.4905, 0.3687, 0.2331, 0.2967, 0.2186, 0.4322], device='cuda:4'), in_proj_covar=tensor([0.0349, 0.0356, 0.0438, 0.0367, 0.0392, 0.0387, 0.0389, 0.0423], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 19:32:19,385 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-04-26 19:32:21,174 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=36761.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 19:32:30,354 INFO [finetune.py:976] (4/7) Epoch 7, batch 2400, loss[loss=0.1852, simple_loss=0.2459, pruned_loss=0.06228, over 4868.00 frames. ], tot_loss[loss=0.2025, simple_loss=0.2649, pruned_loss=0.07002, over 956199.13 frames. ], batch size: 31, lr: 3.87e-03, grad_scale: 64.0 2023-04-26 19:32:33,413 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8204, 1.7272, 1.9638, 2.3698, 2.2191, 1.7654, 1.4833, 1.9001], device='cuda:4'), covar=tensor([0.0898, 0.1125, 0.0671, 0.0569, 0.0618, 0.0920, 0.0870, 0.0599], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0205, 0.0180, 0.0176, 0.0179, 0.0193, 0.0162, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 19:32:56,367 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=36805.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 19:33:03,828 INFO [finetune.py:976] (4/7) Epoch 7, batch 2450, loss[loss=0.1731, simple_loss=0.2434, pruned_loss=0.05134, over 4759.00 frames. ], tot_loss[loss=0.2009, simple_loss=0.2626, pruned_loss=0.06957, over 953689.39 frames. ], batch size: 27, lr: 3.87e-03, grad_scale: 64.0 2023-04-26 19:33:12,660 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.003e+02 1.628e+02 1.856e+02 2.205e+02 3.828e+02, threshold=3.712e+02, percent-clipped=0.0 2023-04-26 19:33:18,529 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=36837.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:33:37,049 INFO [finetune.py:976] (4/7) Epoch 7, batch 2500, loss[loss=0.1907, simple_loss=0.2608, pruned_loss=0.06028, over 4822.00 frames. ], tot_loss[loss=0.2023, simple_loss=0.2642, pruned_loss=0.07015, over 954871.06 frames. ], batch size: 39, lr: 3.87e-03, grad_scale: 64.0 2023-04-26 19:33:37,687 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=36866.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 19:33:51,063 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=36885.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:34:07,403 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=36911.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:34:10,363 INFO [finetune.py:976] (4/7) Epoch 7, batch 2550, loss[loss=0.1883, simple_loss=0.2604, pruned_loss=0.05812, over 4924.00 frames. ], tot_loss[loss=0.2031, simple_loss=0.2664, pruned_loss=0.06993, over 955150.40 frames. ], batch size: 38, lr: 3.87e-03, grad_scale: 64.0 2023-04-26 19:34:12,639 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5842, 1.1092, 1.3520, 1.2443, 1.7480, 1.3678, 1.1010, 1.3508], device='cuda:4'), covar=tensor([0.1676, 0.1586, 0.2196, 0.1636, 0.0961, 0.1615, 0.2244, 0.1940], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0331, 0.0357, 0.0306, 0.0343, 0.0332, 0.0310, 0.0356], device='cuda:4'), out_proj_covar=tensor([6.6110e-05, 7.0474e-05, 7.7324e-05, 6.3652e-05, 7.2266e-05, 7.1617e-05, 6.7107e-05, 7.6457e-05], device='cuda:4') 2023-04-26 19:34:20,132 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.245e+02 1.780e+02 2.108e+02 2.520e+02 3.900e+02, threshold=4.217e+02, percent-clipped=2.0 2023-04-26 19:34:45,096 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=36958.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:34:52,520 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=36961.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:34:54,336 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=36964.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:34:55,486 INFO [finetune.py:976] (4/7) Epoch 7, batch 2600, loss[loss=0.2451, simple_loss=0.2998, pruned_loss=0.09524, over 4884.00 frames. ], tot_loss[loss=0.2062, simple_loss=0.2699, pruned_loss=0.07132, over 954853.92 frames. ], batch size: 32, lr: 3.87e-03, grad_scale: 64.0 2023-04-26 19:35:00,963 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=36968.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:35:04,503 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=36972.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:35:48,154 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=37004.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:36:01,127 INFO [finetune.py:976] (4/7) Epoch 7, batch 2650, loss[loss=0.2087, simple_loss=0.2623, pruned_loss=0.07758, over 4832.00 frames. ], tot_loss[loss=0.2058, simple_loss=0.2697, pruned_loss=0.07096, over 953939.87 frames. ], batch size: 30, lr: 3.87e-03, grad_scale: 64.0 2023-04-26 19:36:03,050 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=37019.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:36:09,956 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.140e+02 1.731e+02 2.035e+02 2.586e+02 4.493e+02, threshold=4.069e+02, percent-clipped=1.0 2023-04-26 19:36:28,800 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=37056.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 19:36:34,289 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=37065.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:36:34,794 INFO [finetune.py:976] (4/7) Epoch 7, batch 2700, loss[loss=0.1728, simple_loss=0.2457, pruned_loss=0.04998, over 4831.00 frames. ], tot_loss[loss=0.2046, simple_loss=0.2684, pruned_loss=0.07042, over 954288.06 frames. ], batch size: 38, lr: 3.87e-03, grad_scale: 64.0 2023-04-26 19:36:49,133 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5054, 1.4897, 4.4039, 4.1151, 3.8653, 4.1371, 4.0361, 3.8160], device='cuda:4'), covar=tensor([0.7211, 0.5365, 0.0957, 0.1495, 0.1028, 0.1559, 0.1549, 0.1481], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0307, 0.0410, 0.0415, 0.0352, 0.0409, 0.0319, 0.0369], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-26 19:36:53,244 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5034, 1.2124, 0.4523, 1.2663, 1.2142, 1.4116, 1.3120, 1.3014], device='cuda:4'), covar=tensor([0.0586, 0.0428, 0.0489, 0.0622, 0.0336, 0.0590, 0.0525, 0.0678], device='cuda:4'), in_proj_covar=tensor([0.0031, 0.0026, 0.0024, 0.0030, 0.0021, 0.0030, 0.0030, 0.0030], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0050, 0.0038, 0.0048, 0.0048, 0.0050], device='cuda:4') 2023-04-26 19:37:05,908 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8623, 1.3814, 1.6575, 1.5701, 1.6340, 1.3522, 0.6939, 1.3042], device='cuda:4'), covar=tensor([0.4151, 0.4284, 0.2133, 0.2943, 0.3248, 0.3239, 0.4988, 0.2948], device='cuda:4'), in_proj_covar=tensor([0.0279, 0.0253, 0.0217, 0.0321, 0.0214, 0.0228, 0.0237, 0.0189], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-26 19:37:19,208 INFO [finetune.py:976] (4/7) Epoch 7, batch 2750, loss[loss=0.1966, simple_loss=0.2661, pruned_loss=0.06354, over 4829.00 frames. ], tot_loss[loss=0.2037, simple_loss=0.2664, pruned_loss=0.07047, over 955336.63 frames. ], batch size: 47, lr: 3.87e-03, grad_scale: 64.0 2023-04-26 19:37:32,275 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.030e+02 1.627e+02 1.918e+02 2.292e+02 3.296e+02, threshold=3.835e+02, percent-clipped=0.0 2023-04-26 19:38:17,424 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=37161.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 19:38:20,451 INFO [finetune.py:976] (4/7) Epoch 7, batch 2800, loss[loss=0.191, simple_loss=0.2576, pruned_loss=0.06226, over 4826.00 frames. ], tot_loss[loss=0.1999, simple_loss=0.2625, pruned_loss=0.06869, over 955837.76 frames. ], batch size: 30, lr: 3.87e-03, grad_scale: 64.0 2023-04-26 19:38:41,283 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.72 vs. limit=2.0 2023-04-26 19:39:00,021 INFO [finetune.py:976] (4/7) Epoch 7, batch 2850, loss[loss=0.1864, simple_loss=0.2511, pruned_loss=0.06085, over 4935.00 frames. ], tot_loss[loss=0.1986, simple_loss=0.2612, pruned_loss=0.06797, over 956870.45 frames. ], batch size: 38, lr: 3.87e-03, grad_scale: 32.0 2023-04-26 19:39:08,538 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.213e+02 1.711e+02 1.920e+02 2.384e+02 5.364e+02, threshold=3.840e+02, percent-clipped=2.0 2023-04-26 19:39:21,972 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5450, 1.0868, 1.5888, 1.9139, 1.6215, 1.4902, 1.5112, 1.5820], device='cuda:4'), covar=tensor([0.7341, 1.0456, 1.0119, 1.1102, 0.9278, 1.2288, 1.3214, 1.0797], device='cuda:4'), in_proj_covar=tensor([0.0413, 0.0435, 0.0518, 0.0540, 0.0441, 0.0461, 0.0473, 0.0472], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 19:39:30,387 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.96 vs. limit=2.0 2023-04-26 19:39:30,817 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=37261.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:39:32,635 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=37264.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:39:33,743 INFO [finetune.py:976] (4/7) Epoch 7, batch 2900, loss[loss=0.2631, simple_loss=0.3276, pruned_loss=0.09931, over 4905.00 frames. ], tot_loss[loss=0.2017, simple_loss=0.2645, pruned_loss=0.06949, over 957203.87 frames. ], batch size: 36, lr: 3.87e-03, grad_scale: 32.0 2023-04-26 19:39:34,409 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=37267.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:39:35,024 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=37268.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:39:35,046 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=37268.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:40:02,585 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=37309.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:40:10,641 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=37312.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:40:11,880 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=37314.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:40:13,065 INFO [finetune.py:976] (4/7) Epoch 7, batch 2950, loss[loss=0.1762, simple_loss=0.2369, pruned_loss=0.05775, over 4377.00 frames. ], tot_loss[loss=0.2036, simple_loss=0.267, pruned_loss=0.0701, over 956601.28 frames. ], batch size: 19, lr: 3.87e-03, grad_scale: 32.0 2023-04-26 19:40:13,124 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=37316.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:40:14,403 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7608, 3.7356, 1.0311, 2.1327, 2.2447, 2.5940, 2.2722, 1.1274], device='cuda:4'), covar=tensor([0.1280, 0.1008, 0.1973, 0.1214, 0.0931, 0.1036, 0.1387, 0.1951], device='cuda:4'), in_proj_covar=tensor([0.0120, 0.0257, 0.0144, 0.0126, 0.0137, 0.0157, 0.0122, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 19:40:14,502 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.75 vs. limit=2.0 2023-04-26 19:40:24,039 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2225, 1.5108, 1.4871, 1.7921, 1.5966, 1.9540, 1.3534, 3.6419], device='cuda:4'), covar=tensor([0.0721, 0.0876, 0.0855, 0.1308, 0.0710, 0.0500, 0.0793, 0.0167], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0039, 0.0041, 0.0045, 0.0041, 0.0040, 0.0039, 0.0060], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 19:40:24,695 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4114, 1.3715, 1.7262, 1.7383, 1.4280, 1.2005, 1.4975, 1.0412], device='cuda:4'), covar=tensor([0.0898, 0.0739, 0.0608, 0.0768, 0.0866, 0.1280, 0.0703, 0.0925], device='cuda:4'), in_proj_covar=tensor([0.0066, 0.0075, 0.0073, 0.0067, 0.0077, 0.0096, 0.0081, 0.0077], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 19:40:32,670 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=37329.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:40:33,131 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.057e+02 1.889e+02 2.237e+02 2.673e+02 8.962e+02, threshold=4.474e+02, percent-clipped=7.0 2023-04-26 19:41:05,685 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=37356.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 19:41:08,605 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=37360.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:41:17,431 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=37364.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:41:18,548 INFO [finetune.py:976] (4/7) Epoch 7, batch 3000, loss[loss=0.2547, simple_loss=0.3173, pruned_loss=0.09604, over 4926.00 frames. ], tot_loss[loss=0.206, simple_loss=0.2694, pruned_loss=0.07127, over 958140.13 frames. ], batch size: 33, lr: 3.87e-03, grad_scale: 32.0 2023-04-26 19:41:18,548 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-26 19:41:40,584 INFO [finetune.py:1010] (4/7) Epoch 7, validation: loss=0.1559, simple_loss=0.2289, pruned_loss=0.04148, over 2265189.00 frames. 2023-04-26 19:41:40,599 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6259MB 2023-04-26 19:41:52,028 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9390, 1.7767, 2.0270, 2.3285, 2.3186, 1.9429, 1.5364, 1.8773], device='cuda:4'), covar=tensor([0.0864, 0.1042, 0.0628, 0.0552, 0.0580, 0.0862, 0.0947, 0.0666], device='cuda:4'), in_proj_covar=tensor([0.0202, 0.0207, 0.0183, 0.0179, 0.0181, 0.0196, 0.0165, 0.0190], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 19:42:14,831 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4491, 3.0396, 0.9349, 1.8739, 1.7841, 2.1890, 1.8514, 1.0030], device='cuda:4'), covar=tensor([0.1339, 0.1190, 0.1877, 0.1275, 0.1085, 0.1080, 0.1497, 0.1781], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0257, 0.0144, 0.0126, 0.0137, 0.0157, 0.0122, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 19:42:15,401 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=37404.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 19:42:22,081 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=5.06 vs. limit=5.0 2023-04-26 19:42:23,695 INFO [finetune.py:976] (4/7) Epoch 7, batch 3050, loss[loss=0.2077, simple_loss=0.2745, pruned_loss=0.07042, over 4887.00 frames. ], tot_loss[loss=0.2055, simple_loss=0.2695, pruned_loss=0.07074, over 957926.70 frames. ], batch size: 36, lr: 3.87e-03, grad_scale: 32.0 2023-04-26 19:42:24,507 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.64 vs. limit=2.0 2023-04-26 19:42:30,236 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=37425.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:42:34,155 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.257e+01 1.766e+02 2.163e+02 2.758e+02 4.245e+02, threshold=4.325e+02, percent-clipped=0.0 2023-04-26 19:42:54,536 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=37461.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 19:42:57,505 INFO [finetune.py:976] (4/7) Epoch 7, batch 3100, loss[loss=0.227, simple_loss=0.2834, pruned_loss=0.08533, over 4843.00 frames. ], tot_loss[loss=0.2034, simple_loss=0.2672, pruned_loss=0.06984, over 958013.26 frames. ], batch size: 49, lr: 3.87e-03, grad_scale: 32.0 2023-04-26 19:43:11,702 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1074, 2.5423, 1.0373, 1.3326, 1.8565, 1.2315, 3.2941, 1.5642], device='cuda:4'), covar=tensor([0.0646, 0.0841, 0.0899, 0.1184, 0.0573, 0.0975, 0.0243, 0.0678], device='cuda:4'), in_proj_covar=tensor([0.0052, 0.0068, 0.0050, 0.0048, 0.0052, 0.0053, 0.0079, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-26 19:43:22,636 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4451, 1.0962, 0.3899, 1.1637, 1.1392, 1.3359, 1.2471, 1.2197], device='cuda:4'), covar=tensor([0.0539, 0.0438, 0.0491, 0.0593, 0.0337, 0.0575, 0.0542, 0.0630], device='cuda:4'), in_proj_covar=tensor([0.0030, 0.0026, 0.0024, 0.0030, 0.0021, 0.0030, 0.0029, 0.0030], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0050, 0.0038, 0.0048, 0.0048, 0.0049], device='cuda:4') 2023-04-26 19:43:26,082 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=37509.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 19:43:31,233 INFO [finetune.py:976] (4/7) Epoch 7, batch 3150, loss[loss=0.1637, simple_loss=0.2311, pruned_loss=0.04815, over 4907.00 frames. ], tot_loss[loss=0.2009, simple_loss=0.2641, pruned_loss=0.06887, over 958504.04 frames. ], batch size: 37, lr: 3.87e-03, grad_scale: 32.0 2023-04-26 19:43:51,673 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.5810, 3.5449, 2.7631, 4.2083, 3.6373, 3.7037, 1.8646, 3.5315], device='cuda:4'), covar=tensor([0.1925, 0.1564, 0.3238, 0.1869, 0.3251, 0.2092, 0.5613, 0.2523], device='cuda:4'), in_proj_covar=tensor([0.0246, 0.0221, 0.0253, 0.0312, 0.0306, 0.0254, 0.0276, 0.0275], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 19:43:52,228 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.195e+02 1.638e+02 2.016e+02 2.443e+02 5.326e+02, threshold=4.032e+02, percent-clipped=1.0 2023-04-26 19:44:03,350 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0268, 1.9472, 2.2758, 2.4901, 1.8911, 1.6391, 1.8926, 1.1114], device='cuda:4'), covar=tensor([0.0729, 0.1022, 0.0730, 0.1006, 0.1034, 0.1468, 0.0964, 0.1178], device='cuda:4'), in_proj_covar=tensor([0.0065, 0.0074, 0.0073, 0.0067, 0.0077, 0.0095, 0.0080, 0.0076], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 19:44:13,945 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.7165, 3.6693, 2.7524, 4.3460, 3.8593, 3.7771, 1.6564, 3.6915], device='cuda:4'), covar=tensor([0.1809, 0.1394, 0.3058, 0.1750, 0.3252, 0.2056, 0.5978, 0.2352], device='cuda:4'), in_proj_covar=tensor([0.0246, 0.0221, 0.0253, 0.0312, 0.0306, 0.0255, 0.0276, 0.0275], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 19:44:36,862 INFO [finetune.py:976] (4/7) Epoch 7, batch 3200, loss[loss=0.1591, simple_loss=0.2242, pruned_loss=0.04705, over 4810.00 frames. ], tot_loss[loss=0.1976, simple_loss=0.2606, pruned_loss=0.06727, over 959393.44 frames. ], batch size: 45, lr: 3.87e-03, grad_scale: 32.0 2023-04-26 19:44:37,573 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=37567.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:45:42,732 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=37614.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:45:43,284 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=37615.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:45:43,828 INFO [finetune.py:976] (4/7) Epoch 7, batch 3250, loss[loss=0.2092, simple_loss=0.2864, pruned_loss=0.06602, over 4833.00 frames. ], tot_loss[loss=0.1976, simple_loss=0.2611, pruned_loss=0.06709, over 957560.88 frames. ], batch size: 39, lr: 3.87e-03, grad_scale: 32.0 2023-04-26 19:45:55,399 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=37624.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:46:03,765 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.639e+01 1.714e+02 1.975e+02 2.453e+02 9.656e+02, threshold=3.950e+02, percent-clipped=3.0 2023-04-26 19:46:41,391 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=37660.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:46:42,586 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=37662.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:46:45,455 INFO [finetune.py:976] (4/7) Epoch 7, batch 3300, loss[loss=0.2335, simple_loss=0.2966, pruned_loss=0.08519, over 4898.00 frames. ], tot_loss[loss=0.2006, simple_loss=0.2645, pruned_loss=0.06837, over 956212.64 frames. ], batch size: 43, lr: 3.87e-03, grad_scale: 32.0 2023-04-26 19:47:48,147 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=37708.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:47:57,371 INFO [finetune.py:976] (4/7) Epoch 7, batch 3350, loss[loss=0.23, simple_loss=0.2905, pruned_loss=0.08476, over 4833.00 frames. ], tot_loss[loss=0.2027, simple_loss=0.2666, pruned_loss=0.06941, over 956990.81 frames. ], batch size: 49, lr: 3.87e-03, grad_scale: 32.0 2023-04-26 19:47:58,714 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4823, 2.3938, 2.0326, 2.2501, 2.5785, 2.0044, 3.3720, 1.8546], device='cuda:4'), covar=tensor([0.4144, 0.2449, 0.4676, 0.3654, 0.2216, 0.2995, 0.1838, 0.4581], device='cuda:4'), in_proj_covar=tensor([0.0348, 0.0354, 0.0434, 0.0365, 0.0390, 0.0384, 0.0386, 0.0420], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 19:48:00,372 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=37720.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:48:12,888 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.265e+02 1.789e+02 2.091e+02 2.525e+02 4.436e+02, threshold=4.182e+02, percent-clipped=1.0 2023-04-26 19:48:36,820 INFO [finetune.py:976] (4/7) Epoch 7, batch 3400, loss[loss=0.2352, simple_loss=0.3074, pruned_loss=0.08146, over 4907.00 frames. ], tot_loss[loss=0.2037, simple_loss=0.2686, pruned_loss=0.06944, over 957335.26 frames. ], batch size: 37, lr: 3.87e-03, grad_scale: 32.0 2023-04-26 19:49:02,170 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4556, 3.4778, 0.9765, 1.7964, 1.9072, 2.4846, 1.9258, 0.9434], device='cuda:4'), covar=tensor([0.1436, 0.0872, 0.1931, 0.1421, 0.1166, 0.1039, 0.1615, 0.2368], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0254, 0.0143, 0.0125, 0.0135, 0.0155, 0.0121, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 19:49:02,425 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.75 vs. limit=5.0 2023-04-26 19:49:36,094 INFO [finetune.py:976] (4/7) Epoch 7, batch 3450, loss[loss=0.2041, simple_loss=0.2731, pruned_loss=0.06759, over 4786.00 frames. ], tot_loss[loss=0.2022, simple_loss=0.267, pruned_loss=0.06874, over 954616.84 frames. ], batch size: 29, lr: 3.87e-03, grad_scale: 32.0 2023-04-26 19:49:55,704 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.199e+01 1.617e+02 2.034e+02 2.503e+02 3.981e+02, threshold=4.069e+02, percent-clipped=0.0 2023-04-26 19:50:37,107 INFO [finetune.py:976] (4/7) Epoch 7, batch 3500, loss[loss=0.2156, simple_loss=0.2645, pruned_loss=0.08329, over 4915.00 frames. ], tot_loss[loss=0.2019, simple_loss=0.2659, pruned_loss=0.06892, over 955837.72 frames. ], batch size: 46, lr: 3.86e-03, grad_scale: 32.0 2023-04-26 19:50:44,470 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1801, 1.4588, 1.4735, 1.7142, 1.6095, 1.8432, 1.4129, 3.3283], device='cuda:4'), covar=tensor([0.0726, 0.0888, 0.0833, 0.1317, 0.0682, 0.0550, 0.0840, 0.0185], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0039, 0.0041, 0.0045, 0.0041, 0.0040, 0.0039, 0.0060], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 19:51:04,186 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.13 vs. limit=2.0 2023-04-26 19:51:14,033 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1895, 1.5550, 4.8685, 4.2643, 4.3274, 4.5425, 4.2766, 4.1668], device='cuda:4'), covar=tensor([0.7993, 0.8103, 0.1382, 0.2724, 0.1958, 0.2660, 0.2700, 0.2238], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0305, 0.0409, 0.0413, 0.0350, 0.0407, 0.0317, 0.0366], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-26 19:51:17,011 INFO [finetune.py:976] (4/7) Epoch 7, batch 3550, loss[loss=0.1587, simple_loss=0.2302, pruned_loss=0.04356, over 4763.00 frames. ], tot_loss[loss=0.2, simple_loss=0.2636, pruned_loss=0.06825, over 957606.58 frames. ], batch size: 28, lr: 3.86e-03, grad_scale: 32.0 2023-04-26 19:51:23,288 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7832, 1.2787, 1.5632, 1.6077, 1.5209, 1.2336, 0.7397, 1.2738], device='cuda:4'), covar=tensor([0.3832, 0.4235, 0.2119, 0.2815, 0.3399, 0.3215, 0.5276, 0.2968], device='cuda:4'), in_proj_covar=tensor([0.0281, 0.0255, 0.0218, 0.0324, 0.0217, 0.0230, 0.0239, 0.0190], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 19:51:27,502 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=37924.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:51:36,395 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.005e+02 1.609e+02 1.946e+02 2.387e+02 4.801e+02, threshold=3.892e+02, percent-clipped=2.0 2023-04-26 19:51:40,072 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4609, 3.5744, 0.9673, 1.8033, 1.9762, 2.4701, 1.9957, 1.1089], device='cuda:4'), covar=tensor([0.1483, 0.0959, 0.2128, 0.1443, 0.1195, 0.1084, 0.1575, 0.2090], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0255, 0.0143, 0.0125, 0.0136, 0.0155, 0.0120, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 19:51:58,391 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4651, 3.3273, 0.8601, 1.6863, 1.8203, 2.3026, 1.8948, 1.0761], device='cuda:4'), covar=tensor([0.1405, 0.1017, 0.2091, 0.1379, 0.1121, 0.1062, 0.1473, 0.1935], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0254, 0.0143, 0.0125, 0.0136, 0.0155, 0.0120, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 19:52:01,352 INFO [finetune.py:976] (4/7) Epoch 7, batch 3600, loss[loss=0.1868, simple_loss=0.2546, pruned_loss=0.05948, over 4866.00 frames. ], tot_loss[loss=0.1982, simple_loss=0.261, pruned_loss=0.06772, over 955532.22 frames. ], batch size: 44, lr: 3.86e-03, grad_scale: 32.0 2023-04-26 19:52:05,027 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=37972.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:52:09,400 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=37979.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:52:35,872 INFO [finetune.py:976] (4/7) Epoch 7, batch 3650, loss[loss=0.2586, simple_loss=0.3178, pruned_loss=0.0997, over 4810.00 frames. ], tot_loss[loss=0.2014, simple_loss=0.2638, pruned_loss=0.0695, over 956038.09 frames. ], batch size: 41, lr: 3.86e-03, grad_scale: 32.0 2023-04-26 19:52:38,385 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=38020.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:52:43,304 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4012, 1.3594, 1.6814, 1.6799, 1.3494, 1.1310, 1.4693, 0.9033], device='cuda:4'), covar=tensor([0.0731, 0.0782, 0.0510, 0.0642, 0.0876, 0.1271, 0.0781, 0.0962], device='cuda:4'), in_proj_covar=tensor([0.0065, 0.0074, 0.0072, 0.0067, 0.0076, 0.0095, 0.0080, 0.0076], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 19:52:44,356 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.213e+02 1.731e+02 1.989e+02 2.580e+02 1.075e+03, threshold=3.977e+02, percent-clipped=3.0 2023-04-26 19:52:51,021 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=38040.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:53:08,983 INFO [finetune.py:976] (4/7) Epoch 7, batch 3700, loss[loss=0.1571, simple_loss=0.225, pruned_loss=0.04457, over 4771.00 frames. ], tot_loss[loss=0.2031, simple_loss=0.2659, pruned_loss=0.07015, over 952146.50 frames. ], batch size: 26, lr: 3.86e-03, grad_scale: 32.0 2023-04-26 19:53:10,247 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=38068.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:53:17,044 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6140, 1.3823, 1.2934, 1.4269, 1.9232, 1.5397, 1.2551, 1.2660], device='cuda:4'), covar=tensor([0.1893, 0.1652, 0.2241, 0.1547, 0.1015, 0.1923, 0.2529, 0.2646], device='cuda:4'), in_proj_covar=tensor([0.0306, 0.0326, 0.0353, 0.0301, 0.0338, 0.0324, 0.0304, 0.0351], device='cuda:4'), out_proj_covar=tensor([6.5293e-05, 6.9385e-05, 7.6426e-05, 6.2463e-05, 7.1202e-05, 6.9946e-05, 6.5685e-05, 7.5535e-05], device='cuda:4') 2023-04-26 19:53:41,669 INFO [finetune.py:976] (4/7) Epoch 7, batch 3750, loss[loss=0.2342, simple_loss=0.2893, pruned_loss=0.08953, over 4864.00 frames. ], tot_loss[loss=0.2054, simple_loss=0.2685, pruned_loss=0.07118, over 952181.78 frames. ], batch size: 31, lr: 3.86e-03, grad_scale: 32.0 2023-04-26 19:53:50,668 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.250e+02 1.891e+02 2.141e+02 2.716e+02 4.079e+02, threshold=4.281e+02, percent-clipped=1.0 2023-04-26 19:54:31,493 INFO [finetune.py:976] (4/7) Epoch 7, batch 3800, loss[loss=0.2313, simple_loss=0.3003, pruned_loss=0.08117, over 4909.00 frames. ], tot_loss[loss=0.208, simple_loss=0.2709, pruned_loss=0.07254, over 952473.90 frames. ], batch size: 37, lr: 3.86e-03, grad_scale: 32.0 2023-04-26 19:54:43,669 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.81 vs. limit=2.0 2023-04-26 19:55:29,004 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-04-26 19:55:36,932 INFO [finetune.py:976] (4/7) Epoch 7, batch 3850, loss[loss=0.1585, simple_loss=0.2263, pruned_loss=0.04537, over 4773.00 frames. ], tot_loss[loss=0.2051, simple_loss=0.2686, pruned_loss=0.0708, over 951902.07 frames. ], batch size: 29, lr: 3.86e-03, grad_scale: 32.0 2023-04-26 19:55:56,088 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.094e+02 1.647e+02 1.925e+02 2.333e+02 3.999e+02, threshold=3.850e+02, percent-clipped=0.0 2023-04-26 19:56:18,615 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.3708, 1.3891, 1.4492, 1.1447, 1.4592, 1.0997, 1.8419, 1.3119], device='cuda:4'), covar=tensor([0.3525, 0.1544, 0.4579, 0.2531, 0.1547, 0.2095, 0.1680, 0.4565], device='cuda:4'), in_proj_covar=tensor([0.0349, 0.0355, 0.0436, 0.0365, 0.0392, 0.0385, 0.0386, 0.0422], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 19:56:41,861 INFO [finetune.py:976] (4/7) Epoch 7, batch 3900, loss[loss=0.1578, simple_loss=0.2052, pruned_loss=0.05523, over 2784.00 frames. ], tot_loss[loss=0.2034, simple_loss=0.2663, pruned_loss=0.07022, over 952341.63 frames. ], batch size: 11, lr: 3.86e-03, grad_scale: 32.0 2023-04-26 19:57:13,806 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.76 vs. limit=2.0 2023-04-26 19:57:13,993 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-04-26 19:57:47,850 INFO [finetune.py:976] (4/7) Epoch 7, batch 3950, loss[loss=0.1723, simple_loss=0.2281, pruned_loss=0.05824, over 4827.00 frames. ], tot_loss[loss=0.2011, simple_loss=0.2632, pruned_loss=0.06949, over 952032.10 frames. ], batch size: 25, lr: 3.86e-03, grad_scale: 32.0 2023-04-26 19:57:58,931 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-26 19:58:06,964 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8652, 2.4269, 2.1743, 2.2330, 1.7841, 1.9597, 2.0159, 1.6800], device='cuda:4'), covar=tensor([0.2011, 0.1406, 0.0883, 0.1206, 0.3225, 0.1332, 0.2133, 0.2621], device='cuda:4'), in_proj_covar=tensor([0.0302, 0.0321, 0.0233, 0.0293, 0.0313, 0.0273, 0.0262, 0.0285], device='cuda:4'), out_proj_covar=tensor([1.2272e-04, 1.2973e-04, 9.4343e-05, 1.1747e-04, 1.2889e-04, 1.1037e-04, 1.0733e-04, 1.1467e-04], device='cuda:4') 2023-04-26 19:58:09,819 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.007e+02 1.650e+02 1.940e+02 2.337e+02 4.288e+02, threshold=3.879e+02, percent-clipped=1.0 2023-04-26 19:58:18,665 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=38335.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:58:31,229 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4273, 1.4191, 1.6491, 1.7232, 1.4173, 1.2087, 1.5723, 1.1397], device='cuda:4'), covar=tensor([0.0726, 0.0561, 0.0560, 0.0572, 0.0780, 0.1065, 0.0553, 0.0793], device='cuda:4'), in_proj_covar=tensor([0.0065, 0.0074, 0.0072, 0.0067, 0.0076, 0.0095, 0.0080, 0.0076], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 19:58:41,911 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9176, 1.3570, 1.7669, 1.9178, 1.7486, 1.3978, 0.9819, 1.4900], device='cuda:4'), covar=tensor([0.3793, 0.4122, 0.1924, 0.3060, 0.3148, 0.3079, 0.5275, 0.2666], device='cuda:4'), in_proj_covar=tensor([0.0280, 0.0254, 0.0219, 0.0324, 0.0217, 0.0231, 0.0239, 0.0190], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 19:58:43,708 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=38360.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 19:58:53,667 INFO [finetune.py:976] (4/7) Epoch 7, batch 4000, loss[loss=0.2245, simple_loss=0.2842, pruned_loss=0.08237, over 4871.00 frames. ], tot_loss[loss=0.1995, simple_loss=0.2614, pruned_loss=0.06875, over 952686.01 frames. ], batch size: 34, lr: 3.86e-03, grad_scale: 32.0 2023-04-26 19:58:55,154 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-26 19:59:57,910 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.2960, 3.3290, 2.4817, 3.8494, 3.3928, 3.3431, 1.3977, 3.3422], device='cuda:4'), covar=tensor([0.1980, 0.1418, 0.3207, 0.2240, 0.3683, 0.2128, 0.6182, 0.2541], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0219, 0.0252, 0.0310, 0.0304, 0.0255, 0.0274, 0.0273], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 20:00:00,679 INFO [finetune.py:976] (4/7) Epoch 7, batch 4050, loss[loss=0.2425, simple_loss=0.2995, pruned_loss=0.09276, over 4817.00 frames. ], tot_loss[loss=0.2046, simple_loss=0.2664, pruned_loss=0.07136, over 952574.33 frames. ], batch size: 39, lr: 3.86e-03, grad_scale: 32.0 2023-04-26 20:00:10,090 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=38421.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:00:21,444 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.926e+01 1.887e+02 2.250e+02 2.854e+02 6.950e+02, threshold=4.500e+02, percent-clipped=7.0 2023-04-26 20:01:01,407 INFO [finetune.py:976] (4/7) Epoch 7, batch 4100, loss[loss=0.1739, simple_loss=0.2422, pruned_loss=0.05276, over 4862.00 frames. ], tot_loss[loss=0.2061, simple_loss=0.2691, pruned_loss=0.07155, over 954098.93 frames. ], batch size: 31, lr: 3.86e-03, grad_scale: 32.0 2023-04-26 20:01:40,046 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-26 20:02:13,352 INFO [finetune.py:976] (4/7) Epoch 7, batch 4150, loss[loss=0.2065, simple_loss=0.2764, pruned_loss=0.06834, over 4750.00 frames. ], tot_loss[loss=0.2062, simple_loss=0.2697, pruned_loss=0.0714, over 956125.04 frames. ], batch size: 59, lr: 3.86e-03, grad_scale: 32.0 2023-04-26 20:02:14,155 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.59 vs. limit=2.0 2023-04-26 20:02:28,364 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.215e+02 1.719e+02 1.998e+02 2.310e+02 4.665e+02, threshold=3.995e+02, percent-clipped=1.0 2023-04-26 20:02:57,367 INFO [finetune.py:976] (4/7) Epoch 7, batch 4200, loss[loss=0.2288, simple_loss=0.2883, pruned_loss=0.08461, over 4869.00 frames. ], tot_loss[loss=0.2056, simple_loss=0.2692, pruned_loss=0.07098, over 953173.57 frames. ], batch size: 34, lr: 3.86e-03, grad_scale: 32.0 2023-04-26 20:03:07,035 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8202, 2.4903, 2.1779, 2.2929, 1.6742, 2.1597, 2.0891, 1.7833], device='cuda:4'), covar=tensor([0.2630, 0.1239, 0.0737, 0.1330, 0.3263, 0.1150, 0.1948, 0.2876], device='cuda:4'), in_proj_covar=tensor([0.0305, 0.0323, 0.0234, 0.0295, 0.0316, 0.0275, 0.0263, 0.0288], device='cuda:4'), out_proj_covar=tensor([1.2382e-04, 1.3063e-04, 9.4891e-05, 1.1824e-04, 1.3009e-04, 1.1109e-04, 1.0777e-04, 1.1573e-04], device='cuda:4') 2023-04-26 20:03:14,410 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.45 vs. limit=5.0 2023-04-26 20:03:29,903 INFO [finetune.py:976] (4/7) Epoch 7, batch 4250, loss[loss=0.2082, simple_loss=0.2703, pruned_loss=0.07308, over 4830.00 frames. ], tot_loss[loss=0.2028, simple_loss=0.2662, pruned_loss=0.0697, over 953869.69 frames. ], batch size: 33, lr: 3.86e-03, grad_scale: 32.0 2023-04-26 20:03:34,155 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.72 vs. limit=5.0 2023-04-26 20:03:40,413 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.148e+02 1.629e+02 2.024e+02 2.405e+02 4.304e+02, threshold=4.048e+02, percent-clipped=1.0 2023-04-26 20:03:44,059 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=38635.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:04:03,668 INFO [finetune.py:976] (4/7) Epoch 7, batch 4300, loss[loss=0.2186, simple_loss=0.2811, pruned_loss=0.07808, over 4914.00 frames. ], tot_loss[loss=0.2017, simple_loss=0.2648, pruned_loss=0.06929, over 955528.59 frames. ], batch size: 36, lr: 3.86e-03, grad_scale: 32.0 2023-04-26 20:04:05,638 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8512, 2.5491, 2.1373, 2.3103, 1.7460, 2.0842, 2.1921, 1.6806], device='cuda:4'), covar=tensor([0.2523, 0.1246, 0.0855, 0.1216, 0.3148, 0.1287, 0.1919, 0.2922], device='cuda:4'), in_proj_covar=tensor([0.0303, 0.0321, 0.0233, 0.0294, 0.0315, 0.0273, 0.0262, 0.0286], device='cuda:4'), out_proj_covar=tensor([1.2299e-04, 1.2982e-04, 9.4420e-05, 1.1781e-04, 1.2962e-04, 1.1045e-04, 1.0733e-04, 1.1511e-04], device='cuda:4') 2023-04-26 20:04:06,658 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3219, 3.3823, 0.8475, 1.7216, 1.7939, 2.4007, 1.9128, 1.0288], device='cuda:4'), covar=tensor([0.1484, 0.0968, 0.2141, 0.1359, 0.1138, 0.1007, 0.1387, 0.1921], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0253, 0.0142, 0.0125, 0.0135, 0.0155, 0.0120, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 20:04:26,428 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=38683.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:05:02,267 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-26 20:05:03,901 INFO [finetune.py:976] (4/7) Epoch 7, batch 4350, loss[loss=0.1931, simple_loss=0.25, pruned_loss=0.06815, over 4934.00 frames. ], tot_loss[loss=0.1991, simple_loss=0.2619, pruned_loss=0.06812, over 955793.57 frames. ], batch size: 33, lr: 3.86e-03, grad_scale: 32.0 2023-04-26 20:05:08,718 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=38716.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:05:12,694 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-26 20:05:18,152 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.164e+02 1.707e+02 2.023e+02 2.548e+02 4.531e+02, threshold=4.045e+02, percent-clipped=2.0 2023-04-26 20:05:52,729 INFO [finetune.py:976] (4/7) Epoch 7, batch 4400, loss[loss=0.2273, simple_loss=0.2827, pruned_loss=0.08596, over 4906.00 frames. ], tot_loss[loss=0.2002, simple_loss=0.2631, pruned_loss=0.0687, over 954578.72 frames. ], batch size: 36, lr: 3.86e-03, grad_scale: 32.0 2023-04-26 20:06:26,655 INFO [finetune.py:976] (4/7) Epoch 7, batch 4450, loss[loss=0.1933, simple_loss=0.2747, pruned_loss=0.05597, over 4781.00 frames. ], tot_loss[loss=0.202, simple_loss=0.2654, pruned_loss=0.06924, over 953918.93 frames. ], batch size: 29, lr: 3.86e-03, grad_scale: 32.0 2023-04-26 20:06:36,497 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=38823.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:06:45,522 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.120e+02 1.756e+02 2.089e+02 2.730e+02 5.109e+02, threshold=4.178e+02, percent-clipped=5.0 2023-04-26 20:07:37,018 INFO [finetune.py:976] (4/7) Epoch 7, batch 4500, loss[loss=0.205, simple_loss=0.2716, pruned_loss=0.06921, over 4835.00 frames. ], tot_loss[loss=0.2034, simple_loss=0.2667, pruned_loss=0.07007, over 953782.08 frames. ], batch size: 49, lr: 3.86e-03, grad_scale: 32.0 2023-04-26 20:08:00,055 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=38884.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:08:34,527 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-26 20:08:44,165 INFO [finetune.py:976] (4/7) Epoch 7, batch 4550, loss[loss=0.2084, simple_loss=0.2769, pruned_loss=0.06996, over 4727.00 frames. ], tot_loss[loss=0.2051, simple_loss=0.2685, pruned_loss=0.07089, over 952362.70 frames. ], batch size: 59, lr: 3.86e-03, grad_scale: 32.0 2023-04-26 20:08:58,069 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.107e+02 1.848e+02 2.064e+02 2.542e+02 5.076e+02, threshold=4.127e+02, percent-clipped=2.0 2023-04-26 20:09:19,266 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.95 vs. limit=2.0 2023-04-26 20:09:50,305 INFO [finetune.py:976] (4/7) Epoch 7, batch 4600, loss[loss=0.1534, simple_loss=0.2287, pruned_loss=0.03905, over 4812.00 frames. ], tot_loss[loss=0.2034, simple_loss=0.2673, pruned_loss=0.06977, over 953682.02 frames. ], batch size: 38, lr: 3.86e-03, grad_scale: 32.0 2023-04-26 20:09:58,905 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3958, 1.8261, 1.8331, 2.1570, 2.1338, 1.9390, 1.6799, 4.4820], device='cuda:4'), covar=tensor([0.0599, 0.0726, 0.0709, 0.1122, 0.0561, 0.0611, 0.0706, 0.0119], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0039, 0.0041, 0.0044, 0.0040, 0.0040, 0.0039, 0.0060], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 20:10:56,081 INFO [finetune.py:976] (4/7) Epoch 7, batch 4650, loss[loss=0.2199, simple_loss=0.2767, pruned_loss=0.08152, over 4791.00 frames. ], tot_loss[loss=0.2002, simple_loss=0.2638, pruned_loss=0.0683, over 954225.75 frames. ], batch size: 29, lr: 3.86e-03, grad_scale: 32.0 2023-04-26 20:10:56,168 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=39016.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:11:16,119 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.165e+02 1.581e+02 1.907e+02 2.292e+02 5.308e+02, threshold=3.814e+02, percent-clipped=1.0 2023-04-26 20:12:02,799 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=39064.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:12:03,961 INFO [finetune.py:976] (4/7) Epoch 7, batch 4700, loss[loss=0.1903, simple_loss=0.2444, pruned_loss=0.06807, over 4760.00 frames. ], tot_loss[loss=0.1975, simple_loss=0.2608, pruned_loss=0.06709, over 954450.23 frames. ], batch size: 54, lr: 3.86e-03, grad_scale: 32.0 2023-04-26 20:12:24,603 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8141, 3.6147, 0.9532, 1.9331, 2.2209, 2.4957, 2.1897, 1.0107], device='cuda:4'), covar=tensor([0.1264, 0.1018, 0.2062, 0.1350, 0.0978, 0.1173, 0.1444, 0.2103], device='cuda:4'), in_proj_covar=tensor([0.0120, 0.0258, 0.0145, 0.0126, 0.0137, 0.0157, 0.0122, 0.0125], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 20:13:04,274 INFO [finetune.py:976] (4/7) Epoch 7, batch 4750, loss[loss=0.1623, simple_loss=0.2461, pruned_loss=0.03927, over 4849.00 frames. ], tot_loss[loss=0.1968, simple_loss=0.2596, pruned_loss=0.06701, over 953921.05 frames. ], batch size: 44, lr: 3.86e-03, grad_scale: 32.0 2023-04-26 20:13:24,488 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.169e+02 1.720e+02 2.202e+02 2.598e+02 5.909e+02, threshold=4.403e+02, percent-clipped=7.0 2023-04-26 20:13:25,911 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3098, 1.4831, 1.6724, 1.8002, 1.7152, 1.8690, 1.7485, 1.7317], device='cuda:4'), covar=tensor([0.5400, 0.7675, 0.6136, 0.6121, 0.6616, 0.9998, 0.7615, 0.6795], device='cuda:4'), in_proj_covar=tensor([0.0323, 0.0393, 0.0316, 0.0328, 0.0343, 0.0407, 0.0370, 0.0329], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 20:14:16,832 INFO [finetune.py:976] (4/7) Epoch 7, batch 4800, loss[loss=0.1922, simple_loss=0.2632, pruned_loss=0.06063, over 4788.00 frames. ], tot_loss[loss=0.1999, simple_loss=0.2624, pruned_loss=0.06871, over 952927.91 frames. ], batch size: 29, lr: 3.86e-03, grad_scale: 32.0 2023-04-26 20:14:31,376 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=39179.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:14:40,493 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=39185.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:14:51,294 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.97 vs. limit=2.0 2023-04-26 20:15:12,274 INFO [finetune.py:976] (4/7) Epoch 7, batch 4850, loss[loss=0.2082, simple_loss=0.2696, pruned_loss=0.07342, over 4819.00 frames. ], tot_loss[loss=0.2039, simple_loss=0.2671, pruned_loss=0.07034, over 953978.73 frames. ], batch size: 30, lr: 3.86e-03, grad_scale: 64.0 2023-04-26 20:15:21,785 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.088e+02 1.823e+02 2.117e+02 2.667e+02 5.725e+02, threshold=4.234e+02, percent-clipped=1.0 2023-04-26 20:15:31,647 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=39246.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:15:45,478 INFO [finetune.py:976] (4/7) Epoch 7, batch 4900, loss[loss=0.2219, simple_loss=0.2725, pruned_loss=0.0856, over 4112.00 frames. ], tot_loss[loss=0.2043, simple_loss=0.2677, pruned_loss=0.07042, over 951410.94 frames. ], batch size: 66, lr: 3.86e-03, grad_scale: 64.0 2023-04-26 20:15:46,030 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.63 vs. limit=5.0 2023-04-26 20:16:09,151 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=39284.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:16:19,567 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2727, 1.6874, 1.5658, 1.9915, 2.0405, 1.9423, 1.5788, 4.3129], device='cuda:4'), covar=tensor([0.0635, 0.0762, 0.0805, 0.1194, 0.0610, 0.0576, 0.0727, 0.0141], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0039, 0.0041, 0.0044, 0.0040, 0.0040, 0.0039, 0.0060], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 20:16:50,960 INFO [finetune.py:976] (4/7) Epoch 7, batch 4950, loss[loss=0.1965, simple_loss=0.2604, pruned_loss=0.06623, over 4814.00 frames. ], tot_loss[loss=0.2041, simple_loss=0.268, pruned_loss=0.07012, over 950804.66 frames. ], batch size: 33, lr: 3.86e-03, grad_scale: 64.0 2023-04-26 20:17:01,408 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4174, 1.7198, 1.5019, 1.9717, 1.9946, 1.9836, 1.5449, 4.4082], device='cuda:4'), covar=tensor([0.0614, 0.0804, 0.0834, 0.1242, 0.0620, 0.0571, 0.0785, 0.0114], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0039, 0.0041, 0.0044, 0.0040, 0.0040, 0.0039, 0.0060], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 20:17:06,091 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.056e+02 1.715e+02 2.035e+02 2.513e+02 5.677e+02, threshold=4.070e+02, percent-clipped=1.0 2023-04-26 20:17:25,347 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=39345.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:17:49,760 INFO [finetune.py:976] (4/7) Epoch 7, batch 5000, loss[loss=0.1588, simple_loss=0.226, pruned_loss=0.04576, over 4899.00 frames. ], tot_loss[loss=0.2007, simple_loss=0.265, pruned_loss=0.06819, over 951877.68 frames. ], batch size: 46, lr: 3.86e-03, grad_scale: 64.0 2023-04-26 20:18:06,526 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9155, 4.1000, 0.6874, 2.0595, 2.3812, 2.8273, 2.4850, 0.9499], device='cuda:4'), covar=tensor([0.1405, 0.0922, 0.2430, 0.1513, 0.1100, 0.1155, 0.1424, 0.2327], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0255, 0.0144, 0.0125, 0.0136, 0.0156, 0.0121, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 20:18:23,014 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.47 vs. limit=2.0 2023-04-26 20:18:23,435 INFO [finetune.py:976] (4/7) Epoch 7, batch 5050, loss[loss=0.1646, simple_loss=0.2356, pruned_loss=0.04681, over 4770.00 frames. ], tot_loss[loss=0.1987, simple_loss=0.2623, pruned_loss=0.0676, over 952226.61 frames. ], batch size: 27, lr: 3.85e-03, grad_scale: 64.0 2023-04-26 20:18:33,401 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.120e+02 1.684e+02 2.008e+02 2.415e+02 4.173e+02, threshold=4.016e+02, percent-clipped=2.0 2023-04-26 20:18:33,622 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-04-26 20:18:47,760 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0919, 2.6044, 1.0788, 1.4185, 1.9947, 1.2023, 3.4634, 1.6747], device='cuda:4'), covar=tensor([0.0689, 0.0626, 0.0811, 0.1268, 0.0537, 0.1075, 0.0224, 0.0671], device='cuda:4'), in_proj_covar=tensor([0.0052, 0.0068, 0.0050, 0.0048, 0.0052, 0.0053, 0.0079, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-26 20:18:54,473 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=39463.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:18:56,692 INFO [finetune.py:976] (4/7) Epoch 7, batch 5100, loss[loss=0.1849, simple_loss=0.2461, pruned_loss=0.06183, over 4908.00 frames. ], tot_loss[loss=0.1963, simple_loss=0.2597, pruned_loss=0.06647, over 953110.98 frames. ], batch size: 43, lr: 3.85e-03, grad_scale: 64.0 2023-04-26 20:19:05,633 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=39479.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:19:29,889 INFO [finetune.py:976] (4/7) Epoch 7, batch 5150, loss[loss=0.1812, simple_loss=0.2546, pruned_loss=0.05387, over 4736.00 frames. ], tot_loss[loss=0.1982, simple_loss=0.2611, pruned_loss=0.0676, over 953937.30 frames. ], batch size: 59, lr: 3.85e-03, grad_scale: 32.0 2023-04-26 20:19:35,824 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=39524.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:19:38,070 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=39527.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:19:40,455 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.309e+02 1.798e+02 2.123e+02 2.599e+02 5.486e+02, threshold=4.247e+02, percent-clipped=4.0 2023-04-26 20:19:47,588 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=39541.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:20:03,309 INFO [finetune.py:976] (4/7) Epoch 7, batch 5200, loss[loss=0.2202, simple_loss=0.2878, pruned_loss=0.07628, over 4925.00 frames. ], tot_loss[loss=0.2016, simple_loss=0.2654, pruned_loss=0.06891, over 953642.64 frames. ], batch size: 38, lr: 3.85e-03, grad_scale: 32.0 2023-04-26 20:20:04,128 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=5.14 vs. limit=5.0 2023-04-26 20:20:12,792 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6002, 1.8449, 2.3936, 3.0624, 2.3050, 1.9175, 1.5915, 2.2283], device='cuda:4'), covar=tensor([0.3603, 0.4013, 0.1752, 0.3112, 0.3713, 0.3199, 0.4888, 0.2973], device='cuda:4'), in_proj_covar=tensor([0.0279, 0.0253, 0.0218, 0.0322, 0.0214, 0.0229, 0.0237, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-26 20:20:30,012 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1223, 1.3002, 1.4986, 1.6645, 1.5973, 1.7426, 1.5899, 1.5611], device='cuda:4'), covar=tensor([0.5536, 0.7604, 0.6245, 0.5991, 0.7772, 1.1051, 0.7605, 0.7291], device='cuda:4'), in_proj_covar=tensor([0.0322, 0.0390, 0.0316, 0.0325, 0.0342, 0.0406, 0.0368, 0.0328], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 20:20:37,216 INFO [finetune.py:976] (4/7) Epoch 7, batch 5250, loss[loss=0.2321, simple_loss=0.2978, pruned_loss=0.08316, over 4733.00 frames. ], tot_loss[loss=0.2035, simple_loss=0.2675, pruned_loss=0.06974, over 952181.63 frames. ], batch size: 54, lr: 3.85e-03, grad_scale: 32.0 2023-04-26 20:20:37,315 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1373, 2.5209, 1.0962, 1.3528, 2.0417, 1.2683, 3.3948, 1.6406], device='cuda:4'), covar=tensor([0.0709, 0.0587, 0.0783, 0.1390, 0.0513, 0.1072, 0.0376, 0.0714], device='cuda:4'), in_proj_covar=tensor([0.0053, 0.0068, 0.0051, 0.0048, 0.0053, 0.0053, 0.0079, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0011, 0.0007], device='cuda:4') 2023-04-26 20:20:47,795 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.203e+02 1.776e+02 2.030e+02 2.603e+02 8.469e+02, threshold=4.060e+02, percent-clipped=1.0 2023-04-26 20:20:53,190 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=39639.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 20:20:53,231 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1325, 1.5283, 1.9817, 2.3849, 1.9126, 1.5435, 1.1844, 1.6942], device='cuda:4'), covar=tensor([0.3986, 0.4295, 0.1976, 0.2996, 0.3409, 0.3482, 0.5449, 0.2973], device='cuda:4'), in_proj_covar=tensor([0.0280, 0.0255, 0.0219, 0.0324, 0.0215, 0.0230, 0.0238, 0.0189], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-26 20:20:53,792 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=39640.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:21:03,839 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1552, 1.5595, 1.4145, 1.8951, 1.7878, 1.8459, 1.4565, 4.1100], device='cuda:4'), covar=tensor([0.0748, 0.0931, 0.0943, 0.1352, 0.0719, 0.0713, 0.0886, 0.0183], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0039, 0.0041, 0.0044, 0.0040, 0.0040, 0.0039, 0.0060], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-26 20:21:10,367 INFO [finetune.py:976] (4/7) Epoch 7, batch 5300, loss[loss=0.2136, simple_loss=0.2868, pruned_loss=0.07023, over 4894.00 frames. ], tot_loss[loss=0.2028, simple_loss=0.2675, pruned_loss=0.06909, over 952759.98 frames. ], batch size: 36, lr: 3.85e-03, grad_scale: 32.0 2023-04-26 20:21:50,876 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=39700.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 20:21:57,873 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.03 vs. limit=5.0 2023-04-26 20:22:11,715 INFO [finetune.py:976] (4/7) Epoch 7, batch 5350, loss[loss=0.1528, simple_loss=0.2332, pruned_loss=0.03614, over 4896.00 frames. ], tot_loss[loss=0.2029, simple_loss=0.2676, pruned_loss=0.06913, over 953525.67 frames. ], batch size: 36, lr: 3.85e-03, grad_scale: 32.0 2023-04-26 20:22:31,524 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.006e+02 1.675e+02 2.001e+02 2.311e+02 3.893e+02, threshold=4.002e+02, percent-clipped=0.0 2023-04-26 20:23:17,038 INFO [finetune.py:976] (4/7) Epoch 7, batch 5400, loss[loss=0.1689, simple_loss=0.2362, pruned_loss=0.05082, over 4770.00 frames. ], tot_loss[loss=0.1991, simple_loss=0.2636, pruned_loss=0.0673, over 954820.50 frames. ], batch size: 28, lr: 3.85e-03, grad_scale: 32.0 2023-04-26 20:23:17,766 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=39767.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:24:18,122 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=39815.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:24:18,614 INFO [finetune.py:976] (4/7) Epoch 7, batch 5450, loss[loss=0.2184, simple_loss=0.2714, pruned_loss=0.0827, over 4839.00 frames. ], tot_loss[loss=0.1977, simple_loss=0.2615, pruned_loss=0.06699, over 955389.93 frames. ], batch size: 25, lr: 3.85e-03, grad_scale: 32.0 2023-04-26 20:24:20,496 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=39819.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:24:31,607 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=39828.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:24:33,271 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.150e+02 1.624e+02 1.910e+02 2.215e+02 3.764e+02, threshold=3.819e+02, percent-clipped=0.0 2023-04-26 20:24:40,440 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=39841.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:24:43,726 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.42 vs. limit=5.0 2023-04-26 20:24:57,926 INFO [finetune.py:976] (4/7) Epoch 7, batch 5500, loss[loss=0.2066, simple_loss=0.2681, pruned_loss=0.0726, over 4814.00 frames. ], tot_loss[loss=0.1946, simple_loss=0.2582, pruned_loss=0.06554, over 956229.61 frames. ], batch size: 38, lr: 3.85e-03, grad_scale: 32.0 2023-04-26 20:25:04,140 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=39876.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:25:12,054 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=39889.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:25:31,741 INFO [finetune.py:976] (4/7) Epoch 7, batch 5550, loss[loss=0.1538, simple_loss=0.2216, pruned_loss=0.04304, over 4761.00 frames. ], tot_loss[loss=0.1974, simple_loss=0.2606, pruned_loss=0.06711, over 954192.03 frames. ], batch size: 28, lr: 3.85e-03, grad_scale: 32.0 2023-04-26 20:25:31,879 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7241, 2.4514, 1.8070, 1.6883, 1.3246, 1.3569, 1.9527, 1.2983], device='cuda:4'), covar=tensor([0.1801, 0.1593, 0.1649, 0.2034, 0.2606, 0.2133, 0.1050, 0.2166], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0219, 0.0174, 0.0206, 0.0208, 0.0187, 0.0164, 0.0189], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-26 20:25:40,907 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.156e+02 1.695e+02 2.043e+02 2.631e+02 5.173e+02, threshold=4.085e+02, percent-clipped=3.0 2023-04-26 20:25:46,470 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=39940.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:25:49,037 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.75 vs. limit=2.0 2023-04-26 20:26:02,970 INFO [finetune.py:976] (4/7) Epoch 7, batch 5600, loss[loss=0.1995, simple_loss=0.256, pruned_loss=0.07153, over 4759.00 frames. ], tot_loss[loss=0.201, simple_loss=0.2648, pruned_loss=0.06863, over 952327.48 frames. ], batch size: 27, lr: 3.85e-03, grad_scale: 32.0 2023-04-26 20:26:15,009 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.93 vs. limit=5.0 2023-04-26 20:26:15,865 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=39988.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:26:19,990 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=39995.0, num_to_drop=1, layers_to_drop={3} 2023-04-26 20:26:38,932 INFO [finetune.py:976] (4/7) Epoch 7, batch 5650, loss[loss=0.1865, simple_loss=0.2595, pruned_loss=0.05674, over 4817.00 frames. ], tot_loss[loss=0.203, simple_loss=0.2684, pruned_loss=0.06882, over 954822.40 frames. ], batch size: 33, lr: 3.85e-03, grad_scale: 32.0 2023-04-26 20:26:53,974 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.232e+02 1.674e+02 2.052e+02 2.439e+02 4.352e+02, threshold=4.105e+02, percent-clipped=2.0 2023-04-26 20:27:26,568 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2409, 2.7602, 1.6589, 2.1727, 2.7863, 2.1794, 2.1259, 2.2937], device='cuda:4'), covar=tensor([0.0458, 0.0307, 0.0294, 0.0512, 0.0207, 0.0492, 0.0492, 0.0511], device='cuda:4'), in_proj_covar=tensor([0.0030, 0.0026, 0.0024, 0.0030, 0.0021, 0.0030, 0.0029, 0.0030], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0050, 0.0037, 0.0048, 0.0048, 0.0050], device='cuda:4') 2023-04-26 20:27:27,069 INFO [finetune.py:976] (4/7) Epoch 7, batch 5700, loss[loss=0.1431, simple_loss=0.2079, pruned_loss=0.03915, over 4177.00 frames. ], tot_loss[loss=0.2004, simple_loss=0.2645, pruned_loss=0.06811, over 939796.94 frames. ], batch size: 18, lr: 3.85e-03, grad_scale: 32.0 2023-04-26 20:27:47,428 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.9450, 3.0159, 3.3380, 3.6317, 3.2414, 2.8479, 2.4205, 3.0045], device='cuda:4'), covar=tensor([0.0764, 0.0749, 0.0404, 0.0528, 0.0531, 0.0861, 0.0761, 0.0588], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0205, 0.0181, 0.0178, 0.0179, 0.0194, 0.0162, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 20:28:14,055 INFO [finetune.py:976] (4/7) Epoch 8, batch 0, loss[loss=0.2027, simple_loss=0.2652, pruned_loss=0.0701, over 4691.00 frames. ], tot_loss[loss=0.2027, simple_loss=0.2652, pruned_loss=0.0701, over 4691.00 frames. ], batch size: 59, lr: 3.85e-03, grad_scale: 32.0 2023-04-26 20:28:14,055 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-26 20:28:25,756 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4105, 1.2370, 1.6421, 1.5603, 1.3078, 1.1428, 1.3070, 0.8255], device='cuda:4'), covar=tensor([0.0657, 0.0742, 0.0514, 0.0661, 0.0814, 0.1362, 0.0692, 0.0838], device='cuda:4'), in_proj_covar=tensor([0.0066, 0.0074, 0.0073, 0.0067, 0.0077, 0.0096, 0.0080, 0.0075], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 20:28:30,561 INFO [finetune.py:1010] (4/7) Epoch 8, validation: loss=0.1574, simple_loss=0.2299, pruned_loss=0.04247, over 2265189.00 frames. 2023-04-26 20:28:30,562 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6336MB 2023-04-26 20:29:03,032 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=40119.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:29:05,429 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=40123.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:29:10,786 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.231e+02 1.752e+02 2.089e+02 2.536e+02 4.447e+02, threshold=4.178e+02, percent-clipped=1.0 2023-04-26 20:29:20,981 INFO [finetune.py:976] (4/7) Epoch 8, batch 50, loss[loss=0.193, simple_loss=0.2597, pruned_loss=0.06319, over 4904.00 frames. ], tot_loss[loss=0.2044, simple_loss=0.2682, pruned_loss=0.07033, over 215184.47 frames. ], batch size: 36, lr: 3.85e-03, grad_scale: 32.0 2023-04-26 20:29:35,500 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2543, 1.6034, 1.3690, 1.8428, 1.6597, 1.9244, 1.4318, 3.6123], device='cuda:4'), covar=tensor([0.0634, 0.0784, 0.0833, 0.1153, 0.0643, 0.0531, 0.0764, 0.0133], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0044, 0.0040, 0.0040, 0.0039, 0.0059], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-26 20:29:36,058 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=40167.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:29:38,550 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=40171.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:29:43,526 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-26 20:29:48,549 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7424, 1.1296, 1.2793, 1.4899, 1.8905, 1.5131, 1.3616, 1.2239], device='cuda:4'), covar=tensor([0.1712, 0.1887, 0.2376, 0.1526, 0.0941, 0.1750, 0.2217, 0.2371], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0328, 0.0357, 0.0302, 0.0340, 0.0328, 0.0308, 0.0353], device='cuda:4'), out_proj_covar=tensor([6.5448e-05, 6.9659e-05, 7.7380e-05, 6.2724e-05, 7.1663e-05, 7.0582e-05, 6.6661e-05, 7.5886e-05], device='cuda:4') 2023-04-26 20:29:54,378 INFO [finetune.py:976] (4/7) Epoch 8, batch 100, loss[loss=0.2161, simple_loss=0.2885, pruned_loss=0.07189, over 4927.00 frames. ], tot_loss[loss=0.2014, simple_loss=0.2643, pruned_loss=0.06923, over 380512.11 frames. ], batch size: 33, lr: 3.85e-03, grad_scale: 32.0 2023-04-26 20:29:55,611 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=40195.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:30:03,083 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8226, 1.8951, 1.0367, 1.5847, 1.7991, 1.6660, 1.6199, 1.6968], device='cuda:4'), covar=tensor([0.0552, 0.0305, 0.0374, 0.0554, 0.0295, 0.0603, 0.0590, 0.0578], device='cuda:4'), in_proj_covar=tensor([0.0030, 0.0026, 0.0024, 0.0030, 0.0021, 0.0030, 0.0029, 0.0030], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0050, 0.0038, 0.0048, 0.0048, 0.0050], device='cuda:4') 2023-04-26 20:30:10,502 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=40218.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:30:18,313 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.006e+02 1.606e+02 1.925e+02 2.421e+02 4.111e+02, threshold=3.850e+02, percent-clipped=0.0 2023-04-26 20:30:24,701 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9235, 1.3610, 1.5039, 1.5439, 2.0902, 1.6639, 1.4195, 1.3974], device='cuda:4'), covar=tensor([0.1442, 0.1750, 0.1844, 0.1417, 0.1045, 0.1777, 0.2493, 0.2104], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0327, 0.0355, 0.0302, 0.0340, 0.0326, 0.0308, 0.0353], device='cuda:4'), out_proj_covar=tensor([6.5269e-05, 6.9397e-05, 7.7042e-05, 6.2644e-05, 7.1687e-05, 7.0282e-05, 6.6512e-05, 7.5698e-05], device='cuda:4') 2023-04-26 20:30:28,096 INFO [finetune.py:976] (4/7) Epoch 8, batch 150, loss[loss=0.1668, simple_loss=0.2379, pruned_loss=0.04782, over 4767.00 frames. ], tot_loss[loss=0.1947, simple_loss=0.2573, pruned_loss=0.06607, over 507254.17 frames. ], batch size: 26, lr: 3.85e-03, grad_scale: 32.0 2023-04-26 20:30:36,427 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=40256.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:30:41,942 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3777, 1.0549, 1.1810, 1.1579, 1.5842, 1.3225, 1.1495, 1.1729], device='cuda:4'), covar=tensor([0.1178, 0.1158, 0.1317, 0.1372, 0.0638, 0.1132, 0.1405, 0.1530], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0327, 0.0355, 0.0302, 0.0341, 0.0326, 0.0308, 0.0353], device='cuda:4'), out_proj_covar=tensor([6.5316e-05, 6.9444e-05, 7.7040e-05, 6.2704e-05, 7.1784e-05, 7.0352e-05, 6.6608e-05, 7.5773e-05], device='cuda:4') 2023-04-26 20:30:46,200 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.3448, 1.3402, 1.3988, 1.1321, 1.4567, 1.1060, 1.7085, 1.3560], device='cuda:4'), covar=tensor([0.3720, 0.1614, 0.5182, 0.2434, 0.1557, 0.2086, 0.1729, 0.4506], device='cuda:4'), in_proj_covar=tensor([0.0345, 0.0350, 0.0432, 0.0361, 0.0386, 0.0380, 0.0382, 0.0417], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 20:30:46,202 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=40272.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:30:50,433 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=40279.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:31:01,727 INFO [finetune.py:976] (4/7) Epoch 8, batch 200, loss[loss=0.1747, simple_loss=0.2422, pruned_loss=0.05361, over 4897.00 frames. ], tot_loss[loss=0.1957, simple_loss=0.2577, pruned_loss=0.06681, over 605790.90 frames. ], batch size: 43, lr: 3.85e-03, grad_scale: 32.0 2023-04-26 20:31:02,443 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=40295.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 20:31:24,761 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=40330.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:31:25,242 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.151e+02 1.700e+02 1.993e+02 2.528e+02 5.564e+02, threshold=3.985e+02, percent-clipped=2.0 2023-04-26 20:31:26,598 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=40333.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:31:34,055 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=40343.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 20:31:35,038 INFO [finetune.py:976] (4/7) Epoch 8, batch 250, loss[loss=0.2263, simple_loss=0.2797, pruned_loss=0.08641, over 4816.00 frames. ], tot_loss[loss=0.1978, simple_loss=0.2608, pruned_loss=0.06741, over 680891.70 frames. ], batch size: 30, lr: 3.85e-03, grad_scale: 32.0 2023-04-26 20:32:05,769 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=40391.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:32:07,912 INFO [finetune.py:976] (4/7) Epoch 8, batch 300, loss[loss=0.2524, simple_loss=0.2995, pruned_loss=0.1026, over 4758.00 frames. ], tot_loss[loss=0.2011, simple_loss=0.2652, pruned_loss=0.06856, over 741365.74 frames. ], batch size: 27, lr: 3.85e-03, grad_scale: 32.0 2023-04-26 20:32:27,219 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=40423.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:32:31,999 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.115e+02 1.648e+02 2.026e+02 2.422e+02 3.959e+02, threshold=4.051e+02, percent-clipped=0.0 2023-04-26 20:32:32,321 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.55 vs. limit=2.0 2023-04-26 20:32:45,857 INFO [finetune.py:976] (4/7) Epoch 8, batch 350, loss[loss=0.1883, simple_loss=0.2659, pruned_loss=0.05541, over 4862.00 frames. ], tot_loss[loss=0.2022, simple_loss=0.2667, pruned_loss=0.0689, over 790844.92 frames. ], batch size: 34, lr: 3.85e-03, grad_scale: 32.0 2023-04-26 20:33:27,707 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=40471.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:33:27,754 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=40471.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:33:52,008 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=40492.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:33:53,134 INFO [finetune.py:976] (4/7) Epoch 8, batch 400, loss[loss=0.2101, simple_loss=0.2751, pruned_loss=0.07258, over 4740.00 frames. ], tot_loss[loss=0.2041, simple_loss=0.2687, pruned_loss=0.06977, over 826457.80 frames. ], batch size: 54, lr: 3.85e-03, grad_scale: 32.0 2023-04-26 20:34:24,185 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4336, 0.5993, 1.2320, 1.8035, 1.5192, 1.3366, 1.3082, 1.4001], device='cuda:4'), covar=tensor([0.7031, 0.9726, 0.9816, 1.0369, 0.8523, 1.1246, 1.1676, 1.0105], device='cuda:4'), in_proj_covar=tensor([0.0408, 0.0426, 0.0511, 0.0531, 0.0439, 0.0458, 0.0469, 0.0468], device='cuda:4'), out_proj_covar=tensor([9.9245e-05, 1.0549e-04, 1.1527e-04, 1.2600e-04, 1.0674e-04, 1.1094e-04, 1.1291e-04, 1.1375e-04], device='cuda:4') 2023-04-26 20:34:32,787 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=40519.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:34:45,209 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.151e+02 1.623e+02 1.962e+02 2.378e+02 5.703e+02, threshold=3.923e+02, percent-clipped=1.0 2023-04-26 20:34:54,477 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2418, 3.0204, 0.8120, 1.6749, 1.7585, 2.1626, 1.7779, 0.9211], device='cuda:4'), covar=tensor([0.1526, 0.0975, 0.1979, 0.1384, 0.1088, 0.1027, 0.1532, 0.1950], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0255, 0.0143, 0.0125, 0.0136, 0.0155, 0.0122, 0.0124], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 20:35:05,023 INFO [finetune.py:976] (4/7) Epoch 8, batch 450, loss[loss=0.2332, simple_loss=0.2825, pruned_loss=0.09198, over 4822.00 frames. ], tot_loss[loss=0.2029, simple_loss=0.2673, pruned_loss=0.06926, over 854433.80 frames. ], batch size: 33, lr: 3.85e-03, grad_scale: 32.0 2023-04-26 20:35:09,657 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6314, 1.1918, 1.3269, 1.3486, 1.8650, 1.5353, 1.2279, 1.3086], device='cuda:4'), covar=tensor([0.1812, 0.1813, 0.2208, 0.1556, 0.1017, 0.1590, 0.2243, 0.2410], device='cuda:4'), in_proj_covar=tensor([0.0306, 0.0325, 0.0355, 0.0300, 0.0339, 0.0326, 0.0308, 0.0352], device='cuda:4'), out_proj_covar=tensor([6.5105e-05, 6.9146e-05, 7.6860e-05, 6.2256e-05, 7.1366e-05, 7.0150e-05, 6.6501e-05, 7.5435e-05], device='cuda:4') 2023-04-26 20:35:17,314 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=40551.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:35:18,619 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=40553.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:35:38,797 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=40574.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:35:56,423 INFO [finetune.py:976] (4/7) Epoch 8, batch 500, loss[loss=0.1737, simple_loss=0.2293, pruned_loss=0.05907, over 4489.00 frames. ], tot_loss[loss=0.2004, simple_loss=0.2639, pruned_loss=0.06842, over 876096.62 frames. ], batch size: 19, lr: 3.85e-03, grad_scale: 32.0 2023-04-26 20:36:37,615 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=40628.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:36:39,370 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.145e+02 1.684e+02 2.034e+02 2.425e+02 5.158e+02, threshold=4.068e+02, percent-clipped=3.0 2023-04-26 20:36:52,501 INFO [finetune.py:976] (4/7) Epoch 8, batch 550, loss[loss=0.1993, simple_loss=0.2437, pruned_loss=0.07748, over 4816.00 frames. ], tot_loss[loss=0.1964, simple_loss=0.2597, pruned_loss=0.06652, over 895011.97 frames. ], batch size: 25, lr: 3.85e-03, grad_scale: 32.0 2023-04-26 20:36:56,990 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=40651.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:37:47,171 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=40686.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:37:52,033 INFO [finetune.py:976] (4/7) Epoch 8, batch 600, loss[loss=0.236, simple_loss=0.2915, pruned_loss=0.09026, over 4760.00 frames. ], tot_loss[loss=0.1972, simple_loss=0.2607, pruned_loss=0.0669, over 907293.18 frames. ], batch size: 54, lr: 3.84e-03, grad_scale: 32.0 2023-04-26 20:38:21,432 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=40712.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:38:44,483 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.285e+02 1.754e+02 2.110e+02 2.532e+02 4.405e+02, threshold=4.220e+02, percent-clipped=2.0 2023-04-26 20:38:45,238 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0479, 2.1621, 1.8212, 1.8633, 2.2841, 1.7780, 2.8593, 1.6065], device='cuda:4'), covar=tensor([0.4527, 0.2169, 0.5259, 0.3893, 0.2083, 0.2924, 0.1483, 0.4959], device='cuda:4'), in_proj_covar=tensor([0.0347, 0.0351, 0.0433, 0.0363, 0.0388, 0.0382, 0.0384, 0.0420], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 20:39:02,721 INFO [finetune.py:976] (4/7) Epoch 8, batch 650, loss[loss=0.1818, simple_loss=0.2781, pruned_loss=0.04281, over 4817.00 frames. ], tot_loss[loss=0.2006, simple_loss=0.2652, pruned_loss=0.06802, over 918237.09 frames. ], batch size: 39, lr: 3.84e-03, grad_scale: 32.0 2023-04-26 20:39:50,532 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-04-26 20:40:08,308 INFO [finetune.py:976] (4/7) Epoch 8, batch 700, loss[loss=0.193, simple_loss=0.2666, pruned_loss=0.05973, over 4237.00 frames. ], tot_loss[loss=0.2012, simple_loss=0.2665, pruned_loss=0.06792, over 925396.48 frames. ], batch size: 65, lr: 3.84e-03, grad_scale: 32.0 2023-04-26 20:40:12,216 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2725, 1.4772, 1.6513, 1.7780, 1.6438, 1.7483, 1.7761, 1.7121], device='cuda:4'), covar=tensor([0.5487, 0.7662, 0.6464, 0.6207, 0.7080, 1.0176, 0.7131, 0.6844], device='cuda:4'), in_proj_covar=tensor([0.0320, 0.0386, 0.0314, 0.0325, 0.0340, 0.0404, 0.0365, 0.0327], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 20:40:55,214 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.166e+02 1.764e+02 2.066e+02 2.587e+02 3.741e+02, threshold=4.132e+02, percent-clipped=0.0 2023-04-26 20:41:14,348 INFO [finetune.py:976] (4/7) Epoch 8, batch 750, loss[loss=0.2596, simple_loss=0.317, pruned_loss=0.1011, over 4759.00 frames. ], tot_loss[loss=0.2028, simple_loss=0.2681, pruned_loss=0.06882, over 931391.84 frames. ], batch size: 59, lr: 3.84e-03, grad_scale: 32.0 2023-04-26 20:41:16,854 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=40848.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:41:17,791 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.73 vs. limit=2.0 2023-04-26 20:41:24,326 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=40851.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:41:38,694 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=40865.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:41:44,673 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=40874.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:41:57,693 INFO [finetune.py:976] (4/7) Epoch 8, batch 800, loss[loss=0.2127, simple_loss=0.2647, pruned_loss=0.08036, over 4296.00 frames. ], tot_loss[loss=0.2015, simple_loss=0.2669, pruned_loss=0.06811, over 936228.51 frames. ], batch size: 66, lr: 3.84e-03, grad_scale: 32.0 2023-04-26 20:42:00,812 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=40899.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:42:01,462 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=40900.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:42:16,691 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=40922.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:42:19,184 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=40926.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:42:20,387 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=40928.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:42:22,146 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.615e+02 1.856e+02 2.266e+02 4.587e+02, threshold=3.712e+02, percent-clipped=2.0 2023-04-26 20:42:31,002 INFO [finetune.py:976] (4/7) Epoch 8, batch 850, loss[loss=0.1651, simple_loss=0.219, pruned_loss=0.05558, over 4734.00 frames. ], tot_loss[loss=0.1996, simple_loss=0.2646, pruned_loss=0.06734, over 940147.89 frames. ], batch size: 23, lr: 3.84e-03, grad_scale: 32.0 2023-04-26 20:42:41,361 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=40961.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 20:42:48,586 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-26 20:42:52,274 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=40976.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:42:58,798 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=40986.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:43:03,442 INFO [finetune.py:976] (4/7) Epoch 8, batch 900, loss[loss=0.2188, simple_loss=0.2719, pruned_loss=0.08289, over 4856.00 frames. ], tot_loss[loss=0.1989, simple_loss=0.2629, pruned_loss=0.0674, over 942603.53 frames. ], batch size: 44, lr: 3.84e-03, grad_scale: 32.0 2023-04-26 20:43:05,785 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=40997.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:43:12,071 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=41007.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:43:28,963 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.199e+02 1.671e+02 1.992e+02 2.459e+02 6.072e+02, threshold=3.985e+02, percent-clipped=2.0 2023-04-26 20:43:30,878 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=41034.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:43:35,072 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.64 vs. limit=2.0 2023-04-26 20:43:37,319 INFO [finetune.py:976] (4/7) Epoch 8, batch 950, loss[loss=0.1767, simple_loss=0.2408, pruned_loss=0.05626, over 4836.00 frames. ], tot_loss[loss=0.1977, simple_loss=0.2608, pruned_loss=0.0673, over 944129.81 frames. ], batch size: 30, lr: 3.84e-03, grad_scale: 32.0 2023-04-26 20:43:38,029 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=41045.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:43:46,387 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=41058.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:44:10,578 INFO [finetune.py:976] (4/7) Epoch 8, batch 1000, loss[loss=0.2502, simple_loss=0.3152, pruned_loss=0.0926, over 4755.00 frames. ], tot_loss[loss=0.2004, simple_loss=0.2638, pruned_loss=0.06847, over 946993.72 frames. ], batch size: 54, lr: 3.84e-03, grad_scale: 32.0 2023-04-26 20:44:16,011 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=41102.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:44:18,379 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=41106.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:44:35,362 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8061, 2.2847, 0.9001, 1.0446, 1.8133, 1.0689, 3.1883, 1.3913], device='cuda:4'), covar=tensor([0.0943, 0.0941, 0.0976, 0.2026, 0.0652, 0.1447, 0.0480, 0.1063], device='cuda:4'), in_proj_covar=tensor([0.0052, 0.0068, 0.0050, 0.0048, 0.0052, 0.0053, 0.0079, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0011, 0.0007], device='cuda:4') 2023-04-26 20:44:35,825 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.105e+02 1.783e+02 2.181e+02 2.591e+02 5.810e+02, threshold=4.362e+02, percent-clipped=3.0 2023-04-26 20:44:44,113 INFO [finetune.py:976] (4/7) Epoch 8, batch 1050, loss[loss=0.1659, simple_loss=0.2345, pruned_loss=0.04861, over 4689.00 frames. ], tot_loss[loss=0.2038, simple_loss=0.2678, pruned_loss=0.06988, over 950331.51 frames. ], batch size: 23, lr: 3.84e-03, grad_scale: 32.0 2023-04-26 20:44:46,608 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=41148.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:44:49,639 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3736, 3.1879, 2.5403, 2.5994, 2.3021, 2.6421, 2.5653, 2.0894], device='cuda:4'), covar=tensor([0.2617, 0.1600, 0.0916, 0.1563, 0.2862, 0.1302, 0.2381, 0.3116], device='cuda:4'), in_proj_covar=tensor([0.0303, 0.0323, 0.0234, 0.0296, 0.0320, 0.0275, 0.0263, 0.0287], device='cuda:4'), out_proj_covar=tensor([1.2275e-04, 1.3062e-04, 9.4610e-05, 1.1864e-04, 1.3140e-04, 1.1121e-04, 1.0799e-04, 1.1521e-04], device='cuda:4') 2023-04-26 20:44:54,942 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1127, 2.6031, 0.9107, 1.3055, 1.9707, 1.2686, 3.6887, 1.9318], device='cuda:4'), covar=tensor([0.0687, 0.0739, 0.0850, 0.1309, 0.0533, 0.0988, 0.0232, 0.0587], device='cuda:4'), in_proj_covar=tensor([0.0052, 0.0068, 0.0050, 0.0048, 0.0052, 0.0053, 0.0079, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0011, 0.0007], device='cuda:4') 2023-04-26 20:44:56,159 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=41163.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:44:56,738 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5861, 2.8284, 1.2003, 1.8259, 1.8813, 2.2455, 1.8708, 1.1383], device='cuda:4'), covar=tensor([0.1131, 0.0955, 0.1555, 0.1151, 0.0952, 0.0907, 0.1233, 0.1817], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0254, 0.0142, 0.0124, 0.0136, 0.0155, 0.0120, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 20:45:27,978 INFO [finetune.py:976] (4/7) Epoch 8, batch 1100, loss[loss=0.1799, simple_loss=0.2505, pruned_loss=0.05461, over 4888.00 frames. ], tot_loss[loss=0.2034, simple_loss=0.268, pruned_loss=0.0694, over 951853.73 frames. ], batch size: 32, lr: 3.84e-03, grad_scale: 32.0 2023-04-26 20:45:29,275 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=41196.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:45:31,994 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-26 20:45:45,864 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4342, 1.6883, 1.1717, 1.1060, 1.0904, 1.0835, 1.1385, 0.9789], device='cuda:4'), covar=tensor([0.2076, 0.1486, 0.2086, 0.2153, 0.2945, 0.2679, 0.1388, 0.2464], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0217, 0.0172, 0.0204, 0.0206, 0.0185, 0.0163, 0.0189], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-26 20:45:50,756 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=41221.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:45:53,187 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.9779, 2.8872, 2.2674, 3.3308, 2.8984, 2.9315, 1.1376, 2.8067], device='cuda:4'), covar=tensor([0.1788, 0.1518, 0.2933, 0.2620, 0.2951, 0.2032, 0.5617, 0.2754], device='cuda:4'), in_proj_covar=tensor([0.0246, 0.0220, 0.0253, 0.0311, 0.0304, 0.0255, 0.0275, 0.0274], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 20:45:58,236 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.055e+02 1.713e+02 2.125e+02 2.514e+02 5.728e+02, threshold=4.249e+02, percent-clipped=2.0 2023-04-26 20:46:17,816 INFO [finetune.py:976] (4/7) Epoch 8, batch 1150, loss[loss=0.1868, simple_loss=0.2601, pruned_loss=0.05677, over 4839.00 frames. ], tot_loss[loss=0.2041, simple_loss=0.2685, pruned_loss=0.06983, over 951502.40 frames. ], batch size: 30, lr: 3.84e-03, grad_scale: 32.0 2023-04-26 20:46:27,336 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.14 vs. limit=2.0 2023-04-26 20:46:36,331 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=41256.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 20:46:50,211 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-26 20:47:24,549 INFO [finetune.py:976] (4/7) Epoch 8, batch 1200, loss[loss=0.1709, simple_loss=0.2409, pruned_loss=0.05042, over 4764.00 frames. ], tot_loss[loss=0.2018, simple_loss=0.2663, pruned_loss=0.06865, over 953361.98 frames. ], batch size: 26, lr: 3.84e-03, grad_scale: 32.0 2023-04-26 20:47:44,084 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=41307.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:47:59,149 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.011e+02 1.672e+02 1.949e+02 2.366e+02 5.490e+02, threshold=3.899e+02, percent-clipped=2.0 2023-04-26 20:48:02,165 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.8151, 2.0443, 1.9618, 2.1928, 1.9500, 2.2118, 2.0583, 1.9989], device='cuda:4'), covar=tensor([0.5199, 0.9019, 0.7495, 0.6583, 0.8175, 1.0737, 0.9321, 0.8618], device='cuda:4'), in_proj_covar=tensor([0.0324, 0.0389, 0.0318, 0.0328, 0.0343, 0.0408, 0.0368, 0.0330], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 20:48:08,614 INFO [finetune.py:976] (4/7) Epoch 8, batch 1250, loss[loss=0.1858, simple_loss=0.243, pruned_loss=0.06432, over 4800.00 frames. ], tot_loss[loss=0.1986, simple_loss=0.2626, pruned_loss=0.06735, over 952618.48 frames. ], batch size: 51, lr: 3.84e-03, grad_scale: 32.0 2023-04-26 20:48:14,609 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=41353.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:48:15,343 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.71 vs. limit=2.0 2023-04-26 20:48:15,821 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=41355.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:48:29,027 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1840, 1.6379, 2.1335, 2.5882, 2.0360, 1.6305, 1.2769, 1.8385], device='cuda:4'), covar=tensor([0.4036, 0.3990, 0.1852, 0.2899, 0.3421, 0.3253, 0.5276, 0.2833], device='cuda:4'), in_proj_covar=tensor([0.0283, 0.0254, 0.0220, 0.0324, 0.0216, 0.0231, 0.0238, 0.0189], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-26 20:48:42,255 INFO [finetune.py:976] (4/7) Epoch 8, batch 1300, loss[loss=0.2173, simple_loss=0.2808, pruned_loss=0.0769, over 4832.00 frames. ], tot_loss[loss=0.1974, simple_loss=0.2605, pruned_loss=0.06713, over 953717.33 frames. ], batch size: 47, lr: 3.84e-03, grad_scale: 32.0 2023-04-26 20:48:46,732 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=41401.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:48:55,787 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4995, 2.9841, 0.9049, 1.6385, 2.1062, 1.6883, 4.4577, 2.3896], device='cuda:4'), covar=tensor([0.0655, 0.0944, 0.0986, 0.1483, 0.0608, 0.1022, 0.0266, 0.0608], device='cuda:4'), in_proj_covar=tensor([0.0052, 0.0069, 0.0051, 0.0048, 0.0052, 0.0053, 0.0080, 0.0052], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 20:49:05,847 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.125e+02 1.753e+02 2.019e+02 2.733e+02 6.111e+02, threshold=4.038e+02, percent-clipped=8.0 2023-04-26 20:49:15,122 INFO [finetune.py:976] (4/7) Epoch 8, batch 1350, loss[loss=0.2178, simple_loss=0.2782, pruned_loss=0.07876, over 4899.00 frames. ], tot_loss[loss=0.1986, simple_loss=0.2615, pruned_loss=0.0679, over 953793.58 frames. ], batch size: 32, lr: 3.84e-03, grad_scale: 32.0 2023-04-26 20:49:25,037 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=41458.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:49:35,646 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.96 vs. limit=5.0 2023-04-26 20:49:47,691 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6984, 1.9048, 1.8571, 2.0536, 1.8568, 2.0509, 1.9630, 1.8638], device='cuda:4'), covar=tensor([0.5979, 0.8554, 0.7255, 0.6221, 0.7627, 1.0469, 0.9052, 0.8237], device='cuda:4'), in_proj_covar=tensor([0.0323, 0.0387, 0.0318, 0.0328, 0.0342, 0.0407, 0.0367, 0.0330], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 20:49:48,137 INFO [finetune.py:976] (4/7) Epoch 8, batch 1400, loss[loss=0.2178, simple_loss=0.2769, pruned_loss=0.07937, over 4825.00 frames. ], tot_loss[loss=0.2024, simple_loss=0.2661, pruned_loss=0.0694, over 952403.21 frames. ], batch size: 39, lr: 3.84e-03, grad_scale: 64.0 2023-04-26 20:50:06,416 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=41521.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:50:12,930 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.248e+02 1.632e+02 2.112e+02 2.588e+02 6.346e+02, threshold=4.225e+02, percent-clipped=4.0 2023-04-26 20:50:21,306 INFO [finetune.py:976] (4/7) Epoch 8, batch 1450, loss[loss=0.2899, simple_loss=0.34, pruned_loss=0.1199, over 4796.00 frames. ], tot_loss[loss=0.2043, simple_loss=0.2684, pruned_loss=0.07011, over 953920.05 frames. ], batch size: 51, lr: 3.84e-03, grad_scale: 64.0 2023-04-26 20:50:30,558 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=41556.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 20:50:38,926 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=41569.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:50:54,400 INFO [finetune.py:976] (4/7) Epoch 8, batch 1500, loss[loss=0.1987, simple_loss=0.2691, pruned_loss=0.0642, over 4847.00 frames. ], tot_loss[loss=0.2044, simple_loss=0.2692, pruned_loss=0.06979, over 955444.43 frames. ], batch size: 44, lr: 3.84e-03, grad_scale: 64.0 2023-04-26 20:51:02,379 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=41604.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:51:12,499 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7587, 1.2862, 1.3614, 1.4283, 1.9007, 1.5289, 1.2005, 1.3840], device='cuda:4'), covar=tensor([0.1353, 0.1403, 0.2036, 0.1364, 0.0828, 0.1349, 0.1917, 0.1847], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0325, 0.0354, 0.0302, 0.0340, 0.0326, 0.0309, 0.0353], device='cuda:4'), out_proj_covar=tensor([6.5237e-05, 6.9132e-05, 7.6500e-05, 6.2644e-05, 7.1664e-05, 7.0296e-05, 6.6843e-05, 7.5885e-05], device='cuda:4') 2023-04-26 20:51:25,119 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.252e+02 1.649e+02 2.032e+02 2.388e+02 6.025e+02, threshold=4.063e+02, percent-clipped=2.0 2023-04-26 20:51:35,918 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9309, 1.7599, 1.9660, 2.2700, 2.2533, 1.8593, 1.5423, 1.9067], device='cuda:4'), covar=tensor([0.1046, 0.1134, 0.0702, 0.0624, 0.0625, 0.0967, 0.0879, 0.0728], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0206, 0.0183, 0.0179, 0.0179, 0.0192, 0.0162, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 20:51:38,883 INFO [finetune.py:976] (4/7) Epoch 8, batch 1550, loss[loss=0.2162, simple_loss=0.2826, pruned_loss=0.07489, over 4823.00 frames. ], tot_loss[loss=0.2039, simple_loss=0.269, pruned_loss=0.0694, over 955096.51 frames. ], batch size: 47, lr: 3.84e-03, grad_scale: 64.0 2023-04-26 20:51:48,300 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8618, 1.6712, 2.0473, 2.2052, 1.9264, 1.8187, 1.9442, 1.9608], device='cuda:4'), covar=tensor([0.6965, 0.9997, 1.1238, 1.0088, 0.8795, 1.2201, 1.2718, 1.1399], device='cuda:4'), in_proj_covar=tensor([0.0406, 0.0424, 0.0505, 0.0526, 0.0435, 0.0454, 0.0465, 0.0464], device='cuda:4'), out_proj_covar=tensor([9.8680e-05, 1.0509e-04, 1.1414e-04, 1.2488e-04, 1.0580e-04, 1.1004e-04, 1.1201e-04, 1.1269e-04], device='cuda:4') 2023-04-26 20:51:57,775 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=41653.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:51:58,450 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=41654.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:52:02,712 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-26 20:52:47,105 INFO [finetune.py:976] (4/7) Epoch 8, batch 1600, loss[loss=0.1688, simple_loss=0.2461, pruned_loss=0.04575, over 4760.00 frames. ], tot_loss[loss=0.2014, simple_loss=0.2657, pruned_loss=0.06855, over 955512.55 frames. ], batch size: 27, lr: 3.84e-03, grad_scale: 64.0 2023-04-26 20:52:56,991 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=41701.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:52:57,018 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=41701.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:53:12,207 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=41715.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 20:53:22,287 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.077e+02 1.620e+02 1.919e+02 2.261e+02 4.784e+02, threshold=3.838e+02, percent-clipped=1.0 2023-04-26 20:53:30,671 INFO [finetune.py:976] (4/7) Epoch 8, batch 1650, loss[loss=0.2191, simple_loss=0.2727, pruned_loss=0.08275, over 4914.00 frames. ], tot_loss[loss=0.1974, simple_loss=0.2615, pruned_loss=0.06667, over 952445.17 frames. ], batch size: 36, lr: 3.84e-03, grad_scale: 64.0 2023-04-26 20:53:33,800 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=41749.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:53:40,318 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=41758.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:53:48,897 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-04-26 20:54:03,919 INFO [finetune.py:976] (4/7) Epoch 8, batch 1700, loss[loss=0.2676, simple_loss=0.3206, pruned_loss=0.1073, over 4047.00 frames. ], tot_loss[loss=0.1953, simple_loss=0.2589, pruned_loss=0.06582, over 951586.64 frames. ], batch size: 65, lr: 3.84e-03, grad_scale: 64.0 2023-04-26 20:54:04,026 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1661, 1.5509, 1.4171, 1.7382, 1.6512, 1.9615, 1.4482, 3.4571], device='cuda:4'), covar=tensor([0.0672, 0.0816, 0.0808, 0.1205, 0.0675, 0.0549, 0.0787, 0.0156], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0039, 0.0040, 0.0044, 0.0040, 0.0040, 0.0039, 0.0059], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-26 20:54:11,555 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=41806.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:54:29,248 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.116e+02 1.826e+02 2.142e+02 2.567e+02 5.646e+02, threshold=4.284e+02, percent-clipped=4.0 2023-04-26 20:54:37,047 INFO [finetune.py:976] (4/7) Epoch 8, batch 1750, loss[loss=0.18, simple_loss=0.25, pruned_loss=0.05499, over 4901.00 frames. ], tot_loss[loss=0.1968, simple_loss=0.2607, pruned_loss=0.0664, over 954270.95 frames. ], batch size: 35, lr: 3.84e-03, grad_scale: 64.0 2023-04-26 20:54:39,555 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-26 20:55:01,325 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9872, 1.6287, 1.5464, 1.7016, 2.2215, 1.8277, 1.5309, 1.4558], device='cuda:4'), covar=tensor([0.1650, 0.1389, 0.2218, 0.1310, 0.0732, 0.1549, 0.2148, 0.2216], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0327, 0.0357, 0.0303, 0.0342, 0.0328, 0.0312, 0.0356], device='cuda:4'), out_proj_covar=tensor([6.5725e-05, 6.9510e-05, 7.7241e-05, 6.2810e-05, 7.1811e-05, 7.0717e-05, 6.7293e-05, 7.6504e-05], device='cuda:4') 2023-04-26 20:55:10,268 INFO [finetune.py:976] (4/7) Epoch 8, batch 1800, loss[loss=0.177, simple_loss=0.2627, pruned_loss=0.0456, over 4900.00 frames. ], tot_loss[loss=0.1984, simple_loss=0.263, pruned_loss=0.06691, over 953363.94 frames. ], batch size: 37, lr: 3.84e-03, grad_scale: 64.0 2023-04-26 20:55:18,155 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=41906.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:55:21,199 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3466, 1.4919, 1.3269, 1.7198, 1.5854, 1.9089, 1.4117, 3.5266], device='cuda:4'), covar=tensor([0.0629, 0.0783, 0.0808, 0.1187, 0.0672, 0.0563, 0.0778, 0.0136], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0039, 0.0040, 0.0044, 0.0041, 0.0040, 0.0039, 0.0059], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-26 20:55:34,483 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8772, 2.3999, 2.0598, 2.3026, 1.7507, 1.9788, 2.0231, 1.7013], device='cuda:4'), covar=tensor([0.2426, 0.1398, 0.0957, 0.1365, 0.3071, 0.1497, 0.2275, 0.2719], device='cuda:4'), in_proj_covar=tensor([0.0303, 0.0323, 0.0233, 0.0295, 0.0320, 0.0275, 0.0263, 0.0286], device='cuda:4'), out_proj_covar=tensor([1.2312e-04, 1.3052e-04, 9.4323e-05, 1.1815e-04, 1.3130e-04, 1.1110e-04, 1.0791e-04, 1.1511e-04], device='cuda:4') 2023-04-26 20:55:35,584 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.171e+02 1.841e+02 2.087e+02 2.578e+02 5.698e+02, threshold=4.173e+02, percent-clipped=3.0 2023-04-26 20:55:43,856 INFO [finetune.py:976] (4/7) Epoch 8, batch 1850, loss[loss=0.1681, simple_loss=0.244, pruned_loss=0.04612, over 4814.00 frames. ], tot_loss[loss=0.2023, simple_loss=0.2668, pruned_loss=0.06888, over 954457.38 frames. ], batch size: 39, lr: 3.84e-03, grad_scale: 64.0 2023-04-26 20:55:46,403 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=41948.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 20:55:58,246 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7605, 2.5867, 1.6298, 1.6768, 1.3250, 1.3322, 1.7412, 1.2933], device='cuda:4'), covar=tensor([0.1782, 0.1274, 0.1744, 0.1851, 0.2512, 0.2035, 0.1226, 0.2202], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0219, 0.0174, 0.0205, 0.0207, 0.0186, 0.0164, 0.0190], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-26 20:55:58,800 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=41967.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 20:56:02,291 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.14 vs. limit=2.0 2023-04-26 20:56:17,002 INFO [finetune.py:976] (4/7) Epoch 8, batch 1900, loss[loss=0.1805, simple_loss=0.2523, pruned_loss=0.05431, over 4906.00 frames. ], tot_loss[loss=0.2031, simple_loss=0.268, pruned_loss=0.06907, over 954790.75 frames. ], batch size: 36, lr: 3.84e-03, grad_scale: 32.0 2023-04-26 20:56:27,979 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=42009.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 20:56:28,517 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=42010.0, num_to_drop=1, layers_to_drop={3} 2023-04-26 20:56:37,133 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6577, 1.4167, 1.3373, 1.4777, 1.9590, 1.5840, 1.3253, 1.2752], device='cuda:4'), covar=tensor([0.1683, 0.1290, 0.2059, 0.1241, 0.0739, 0.1291, 0.2114, 0.1888], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0325, 0.0355, 0.0302, 0.0340, 0.0327, 0.0310, 0.0354], device='cuda:4'), out_proj_covar=tensor([6.5388e-05, 6.9058e-05, 7.6941e-05, 6.2551e-05, 7.1525e-05, 7.0379e-05, 6.7014e-05, 7.6084e-05], device='cuda:4') 2023-04-26 20:56:49,879 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.129e+02 1.688e+02 2.018e+02 2.426e+02 3.814e+02, threshold=4.036e+02, percent-clipped=0.0 2023-04-26 20:57:08,240 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6261, 1.2506, 1.7478, 2.0455, 1.7543, 1.6059, 1.6605, 1.6927], device='cuda:4'), covar=tensor([0.6548, 0.8913, 0.9061, 0.9158, 0.8001, 1.0325, 1.0639, 0.9249], device='cuda:4'), in_proj_covar=tensor([0.0409, 0.0426, 0.0507, 0.0530, 0.0438, 0.0456, 0.0469, 0.0467], device='cuda:4'), out_proj_covar=tensor([9.9458e-05, 1.0560e-04, 1.1470e-04, 1.2563e-04, 1.0646e-04, 1.1049e-04, 1.1277e-04, 1.1352e-04], device='cuda:4') 2023-04-26 20:57:08,688 INFO [finetune.py:976] (4/7) Epoch 8, batch 1950, loss[loss=0.1833, simple_loss=0.2463, pruned_loss=0.06008, over 4817.00 frames. ], tot_loss[loss=0.2001, simple_loss=0.2655, pruned_loss=0.06738, over 956115.52 frames. ], batch size: 38, lr: 3.84e-03, grad_scale: 32.0 2023-04-26 20:58:13,786 INFO [finetune.py:976] (4/7) Epoch 8, batch 2000, loss[loss=0.1972, simple_loss=0.2623, pruned_loss=0.06604, over 4900.00 frames. ], tot_loss[loss=0.1979, simple_loss=0.2623, pruned_loss=0.06676, over 955566.37 frames. ], batch size: 36, lr: 3.84e-03, grad_scale: 32.0 2023-04-26 20:58:22,126 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.68 vs. limit=5.0 2023-04-26 20:59:06,779 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.093e+02 1.619e+02 1.918e+02 2.281e+02 4.896e+02, threshold=3.837e+02, percent-clipped=1.0 2023-04-26 20:59:17,303 INFO [finetune.py:976] (4/7) Epoch 8, batch 2050, loss[loss=0.1688, simple_loss=0.2405, pruned_loss=0.04852, over 4851.00 frames. ], tot_loss[loss=0.1949, simple_loss=0.259, pruned_loss=0.06543, over 956994.25 frames. ], batch size: 44, lr: 3.84e-03, grad_scale: 32.0 2023-04-26 21:00:01,143 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.61 vs. limit=5.0 2023-04-26 21:00:05,392 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6301, 1.2919, 4.5822, 4.2399, 3.9325, 4.2925, 4.1839, 3.9743], device='cuda:4'), covar=tensor([0.7568, 0.6704, 0.0930, 0.1889, 0.1242, 0.2282, 0.1289, 0.1450], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0307, 0.0407, 0.0414, 0.0351, 0.0407, 0.0317, 0.0368], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-26 21:00:12,509 INFO [finetune.py:976] (4/7) Epoch 8, batch 2100, loss[loss=0.1496, simple_loss=0.2295, pruned_loss=0.03492, over 4807.00 frames. ], tot_loss[loss=0.1953, simple_loss=0.2589, pruned_loss=0.06584, over 958919.80 frames. ], batch size: 25, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:00:20,672 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5118, 2.3490, 1.5247, 1.6596, 1.1352, 1.2110, 1.6317, 1.1312], device='cuda:4'), covar=tensor([0.2341, 0.1565, 0.1983, 0.2044, 0.2907, 0.2648, 0.1252, 0.2379], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0218, 0.0173, 0.0206, 0.0207, 0.0185, 0.0164, 0.0189], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-26 21:00:38,852 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.172e+02 1.610e+02 1.994e+02 2.459e+02 7.557e+02, threshold=3.988e+02, percent-clipped=2.0 2023-04-26 21:00:46,719 INFO [finetune.py:976] (4/7) Epoch 8, batch 2150, loss[loss=0.2315, simple_loss=0.3002, pruned_loss=0.08144, over 4828.00 frames. ], tot_loss[loss=0.1971, simple_loss=0.2612, pruned_loss=0.0665, over 956686.83 frames. ], batch size: 47, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:00:47,265 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.3054, 3.2015, 2.4383, 3.8230, 3.3296, 3.3259, 1.4754, 3.2617], device='cuda:4'), covar=tensor([0.1939, 0.1506, 0.3400, 0.2509, 0.3702, 0.2026, 0.6282, 0.2765], device='cuda:4'), in_proj_covar=tensor([0.0245, 0.0218, 0.0250, 0.0310, 0.0301, 0.0250, 0.0273, 0.0272], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 21:00:58,115 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=42262.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:01:14,575 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1984, 1.5634, 2.0584, 2.6067, 1.9466, 1.6098, 1.2130, 1.8937], device='cuda:4'), covar=tensor([0.3664, 0.4035, 0.1870, 0.2678, 0.3282, 0.3102, 0.5004, 0.2587], device='cuda:4'), in_proj_covar=tensor([0.0280, 0.0253, 0.0219, 0.0323, 0.0215, 0.0229, 0.0236, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-26 21:01:20,322 INFO [finetune.py:976] (4/7) Epoch 8, batch 2200, loss[loss=0.2223, simple_loss=0.289, pruned_loss=0.07781, over 4819.00 frames. ], tot_loss[loss=0.201, simple_loss=0.2652, pruned_loss=0.06842, over 956484.16 frames. ], batch size: 38, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:01:26,924 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=42304.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 21:01:30,549 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=42310.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 21:01:43,060 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-26 21:01:44,805 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.211e+02 1.704e+02 2.094e+02 2.427e+02 3.760e+02, threshold=4.188e+02, percent-clipped=0.0 2023-04-26 21:01:47,736 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6224, 2.2286, 1.1891, 1.3313, 2.2801, 1.5124, 1.4875, 1.5799], device='cuda:4'), covar=tensor([0.0659, 0.0309, 0.0323, 0.0665, 0.0259, 0.0727, 0.0684, 0.0644], device='cuda:4'), in_proj_covar=tensor([0.0030, 0.0026, 0.0024, 0.0030, 0.0021, 0.0030, 0.0029, 0.0030], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0045, 0.0038, 0.0050, 0.0038, 0.0049, 0.0048, 0.0050], device='cuda:4') 2023-04-26 21:01:51,892 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.19 vs. limit=5.0 2023-04-26 21:01:52,972 INFO [finetune.py:976] (4/7) Epoch 8, batch 2250, loss[loss=0.168, simple_loss=0.2446, pruned_loss=0.04569, over 4891.00 frames. ], tot_loss[loss=0.2035, simple_loss=0.2679, pruned_loss=0.06949, over 955106.97 frames. ], batch size: 35, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:02:02,509 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=42358.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:02:15,808 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=42379.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:02:21,687 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-26 21:02:26,181 INFO [finetune.py:976] (4/7) Epoch 8, batch 2300, loss[loss=0.182, simple_loss=0.2432, pruned_loss=0.0604, over 4753.00 frames. ], tot_loss[loss=0.2032, simple_loss=0.2682, pruned_loss=0.06915, over 956953.38 frames. ], batch size: 23, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:02:27,973 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.4006, 3.2717, 2.4566, 3.9010, 3.3840, 3.3988, 1.5254, 3.3766], device='cuda:4'), covar=tensor([0.1857, 0.1334, 0.3155, 0.2094, 0.2927, 0.1768, 0.5492, 0.2418], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0217, 0.0249, 0.0308, 0.0299, 0.0250, 0.0272, 0.0271], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 21:02:48,623 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9861, 2.7899, 2.3545, 2.4831, 2.1118, 2.2854, 2.4387, 1.9285], device='cuda:4'), covar=tensor([0.2547, 0.1363, 0.0950, 0.1398, 0.2901, 0.1381, 0.2236, 0.3207], device='cuda:4'), in_proj_covar=tensor([0.0300, 0.0321, 0.0232, 0.0293, 0.0317, 0.0275, 0.0262, 0.0285], device='cuda:4'), out_proj_covar=tensor([1.2166e-04, 1.2974e-04, 9.3914e-05, 1.1716e-04, 1.3033e-04, 1.1103e-04, 1.0751e-04, 1.1469e-04], device='cuda:4') 2023-04-26 21:02:50,917 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.077e+02 1.665e+02 1.908e+02 2.261e+02 6.563e+02, threshold=3.817e+02, percent-clipped=1.0 2023-04-26 21:02:56,804 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=42440.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:02:59,687 INFO [finetune.py:976] (4/7) Epoch 8, batch 2350, loss[loss=0.2082, simple_loss=0.2796, pruned_loss=0.06836, over 4824.00 frames. ], tot_loss[loss=0.1995, simple_loss=0.2644, pruned_loss=0.06728, over 957163.55 frames. ], batch size: 39, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:03:31,266 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5743, 2.4092, 1.9996, 2.2766, 2.4882, 2.0373, 2.9591, 1.7890], device='cuda:4'), covar=tensor([0.3407, 0.1649, 0.3604, 0.2484, 0.1535, 0.2253, 0.1689, 0.4490], device='cuda:4'), in_proj_covar=tensor([0.0350, 0.0354, 0.0436, 0.0366, 0.0392, 0.0386, 0.0386, 0.0424], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 21:03:45,144 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9487, 0.9245, 1.0944, 1.0492, 0.9345, 0.7958, 0.9695, 0.6719], device='cuda:4'), covar=tensor([0.0603, 0.0496, 0.0637, 0.0546, 0.0656, 0.1022, 0.0509, 0.0838], device='cuda:4'), in_proj_covar=tensor([0.0066, 0.0073, 0.0072, 0.0067, 0.0076, 0.0096, 0.0079, 0.0075], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 21:03:48,469 INFO [finetune.py:976] (4/7) Epoch 8, batch 2400, loss[loss=0.1695, simple_loss=0.2369, pruned_loss=0.05107, over 4819.00 frames. ], tot_loss[loss=0.1974, simple_loss=0.2616, pruned_loss=0.06659, over 955214.60 frames. ], batch size: 39, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:04:08,575 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=42506.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:04:23,271 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6327, 1.4359, 1.3533, 1.3221, 1.8449, 1.4717, 1.2327, 1.2605], device='cuda:4'), covar=tensor([0.1459, 0.1101, 0.1807, 0.1263, 0.0719, 0.1348, 0.1773, 0.1754], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0326, 0.0354, 0.0301, 0.0340, 0.0325, 0.0310, 0.0353], device='cuda:4'), out_proj_covar=tensor([6.5182e-05, 6.9289e-05, 7.6561e-05, 6.2502e-05, 7.1549e-05, 7.0031e-05, 6.6860e-05, 7.5936e-05], device='cuda:4') 2023-04-26 21:04:40,672 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.220e+02 1.697e+02 2.030e+02 2.440e+02 3.399e+02, threshold=4.060e+02, percent-clipped=0.0 2023-04-26 21:04:54,582 INFO [finetune.py:976] (4/7) Epoch 8, batch 2450, loss[loss=0.1478, simple_loss=0.2156, pruned_loss=0.03996, over 4758.00 frames. ], tot_loss[loss=0.195, simple_loss=0.2588, pruned_loss=0.06559, over 955722.70 frames. ], batch size: 27, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:05:24,357 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=42562.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:05:27,978 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=42567.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:05:55,846 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-04-26 21:06:00,482 INFO [finetune.py:976] (4/7) Epoch 8, batch 2500, loss[loss=0.1752, simple_loss=0.2254, pruned_loss=0.06252, over 4676.00 frames. ], tot_loss[loss=0.1946, simple_loss=0.2588, pruned_loss=0.06517, over 956494.85 frames. ], batch size: 23, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:06:20,327 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4490, 1.6501, 1.4878, 1.8744, 1.7472, 2.0074, 1.3907, 4.0252], device='cuda:4'), covar=tensor([0.0577, 0.0722, 0.0751, 0.1134, 0.0630, 0.0591, 0.0752, 0.0129], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0039, 0.0040, 0.0044, 0.0041, 0.0040, 0.0039, 0.0059], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-26 21:06:20,334 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=42604.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 21:06:29,256 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=42610.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:06:54,509 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.169e+02 1.868e+02 2.156e+02 2.482e+02 4.292e+02, threshold=4.312e+02, percent-clipped=2.0 2023-04-26 21:07:08,190 INFO [finetune.py:976] (4/7) Epoch 8, batch 2550, loss[loss=0.2016, simple_loss=0.2713, pruned_loss=0.06589, over 4906.00 frames. ], tot_loss[loss=0.1987, simple_loss=0.2637, pruned_loss=0.06689, over 956197.56 frames. ], batch size: 35, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:07:11,907 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=42649.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:07:14,685 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=42652.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 21:07:29,131 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.4114, 4.2495, 3.1014, 5.0157, 4.3531, 4.3713, 2.0204, 4.2970], device='cuda:4'), covar=tensor([0.1589, 0.0963, 0.3391, 0.0907, 0.2333, 0.1557, 0.5372, 0.2156], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0217, 0.0250, 0.0308, 0.0298, 0.0250, 0.0271, 0.0271], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 21:07:41,703 INFO [finetune.py:976] (4/7) Epoch 8, batch 2600, loss[loss=0.1709, simple_loss=0.2459, pruned_loss=0.04799, over 4740.00 frames. ], tot_loss[loss=0.2005, simple_loss=0.2657, pruned_loss=0.06762, over 954076.49 frames. ], batch size: 27, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:07:53,066 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=42710.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:08:07,492 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.967e+01 1.665e+02 1.922e+02 2.451e+02 5.774e+02, threshold=3.843e+02, percent-clipped=4.0 2023-04-26 21:08:09,420 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=42735.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:08:11,556 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.82 vs. limit=5.0 2023-04-26 21:08:13,076 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3886, 1.6713, 1.4772, 2.1220, 1.8381, 2.0559, 1.5505, 4.3829], device='cuda:4'), covar=tensor([0.0587, 0.0792, 0.0806, 0.1145, 0.0694, 0.0544, 0.0768, 0.0137], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0039, 0.0040, 0.0044, 0.0041, 0.0039, 0.0039, 0.0059], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-26 21:08:14,788 INFO [finetune.py:976] (4/7) Epoch 8, batch 2650, loss[loss=0.152, simple_loss=0.215, pruned_loss=0.04451, over 4766.00 frames. ], tot_loss[loss=0.2004, simple_loss=0.2657, pruned_loss=0.06753, over 950855.18 frames. ], batch size: 27, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:08:42,442 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-04-26 21:08:48,362 INFO [finetune.py:976] (4/7) Epoch 8, batch 2700, loss[loss=0.206, simple_loss=0.2543, pruned_loss=0.07887, over 4829.00 frames. ], tot_loss[loss=0.2002, simple_loss=0.2652, pruned_loss=0.06765, over 950933.50 frames. ], batch size: 30, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:09:04,010 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6268, 1.3361, 1.3284, 1.5034, 1.8641, 1.5529, 1.4941, 1.2319], device='cuda:4'), covar=tensor([0.1491, 0.1523, 0.1761, 0.1377, 0.0883, 0.1378, 0.1660, 0.1983], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0326, 0.0355, 0.0303, 0.0343, 0.0327, 0.0312, 0.0355], device='cuda:4'), out_proj_covar=tensor([6.5647e-05, 6.9288e-05, 7.6862e-05, 6.2722e-05, 7.2184e-05, 7.0301e-05, 6.7239e-05, 7.6361e-05], device='cuda:4') 2023-04-26 21:09:14,553 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.600e+01 1.577e+02 1.842e+02 2.218e+02 2.993e+02, threshold=3.685e+02, percent-clipped=0.0 2023-04-26 21:09:21,931 INFO [finetune.py:976] (4/7) Epoch 8, batch 2750, loss[loss=0.1806, simple_loss=0.227, pruned_loss=0.06714, over 4174.00 frames. ], tot_loss[loss=0.1973, simple_loss=0.2614, pruned_loss=0.06659, over 950767.52 frames. ], batch size: 18, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:09:27,314 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1832, 1.4611, 1.2686, 1.7309, 1.4697, 1.7922, 1.3260, 3.2724], device='cuda:4'), covar=tensor([0.0756, 0.1008, 0.1027, 0.1313, 0.0849, 0.0604, 0.0971, 0.0237], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0039, 0.0040, 0.0044, 0.0041, 0.0039, 0.0039, 0.0059], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-26 21:09:29,124 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.9951, 2.7947, 2.2449, 3.3635, 2.9521, 2.9630, 1.1130, 2.8942], device='cuda:4'), covar=tensor([0.2071, 0.1765, 0.3080, 0.2526, 0.2931, 0.2053, 0.6345, 0.2828], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0217, 0.0249, 0.0309, 0.0299, 0.0250, 0.0272, 0.0270], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 21:09:31,556 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=42859.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:09:34,271 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=42862.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:09:57,298 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=42879.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:10:16,362 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.45 vs. limit=5.0 2023-04-26 21:10:17,378 INFO [finetune.py:976] (4/7) Epoch 8, batch 2800, loss[loss=0.1986, simple_loss=0.2609, pruned_loss=0.06813, over 4813.00 frames. ], tot_loss[loss=0.195, simple_loss=0.2587, pruned_loss=0.06563, over 953422.28 frames. ], batch size: 45, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:10:51,627 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=42920.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:11:10,642 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.078e+02 1.665e+02 1.974e+02 2.383e+02 4.846e+02, threshold=3.948e+02, percent-clipped=3.0 2023-04-26 21:11:16,190 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=42940.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:11:17,780 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-26 21:11:18,535 INFO [finetune.py:976] (4/7) Epoch 8, batch 2850, loss[loss=0.183, simple_loss=0.2373, pruned_loss=0.06436, over 4796.00 frames. ], tot_loss[loss=0.1937, simple_loss=0.257, pruned_loss=0.06519, over 951639.92 frames. ], batch size: 25, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:11:56,701 INFO [finetune.py:976] (4/7) Epoch 8, batch 2900, loss[loss=0.1955, simple_loss=0.2702, pruned_loss=0.06044, over 4810.00 frames. ], tot_loss[loss=0.1964, simple_loss=0.26, pruned_loss=0.06639, over 951395.45 frames. ], batch size: 40, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:12:03,841 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1762, 1.9477, 2.2275, 2.6065, 2.5570, 2.0681, 1.6153, 2.2439], device='cuda:4'), covar=tensor([0.0977, 0.1092, 0.0630, 0.0636, 0.0620, 0.1019, 0.0956, 0.0634], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0203, 0.0179, 0.0176, 0.0177, 0.0189, 0.0159, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 21:12:15,326 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=43005.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:12:49,771 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0944, 1.4121, 5.4853, 5.1597, 4.8221, 5.3028, 4.7980, 4.8113], device='cuda:4'), covar=tensor([0.7131, 0.6711, 0.0962, 0.1544, 0.0953, 0.1320, 0.1013, 0.1437], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0311, 0.0413, 0.0418, 0.0354, 0.0409, 0.0319, 0.0374], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-26 21:12:51,483 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.213e+02 1.808e+02 2.091e+02 2.545e+02 3.466e+02, threshold=4.182e+02, percent-clipped=0.0 2023-04-26 21:12:59,358 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=43035.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:12:59,407 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.7255, 1.9349, 1.8579, 1.9717, 1.8209, 2.0969, 2.0337, 1.8501], device='cuda:4'), covar=tensor([0.5121, 0.8134, 0.7386, 0.6357, 0.7662, 0.9869, 0.8474, 0.7941], device='cuda:4'), in_proj_covar=tensor([0.0323, 0.0384, 0.0316, 0.0326, 0.0342, 0.0405, 0.0366, 0.0327], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 21:13:01,762 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-04-26 21:13:10,703 INFO [finetune.py:976] (4/7) Epoch 8, batch 2950, loss[loss=0.2065, simple_loss=0.2688, pruned_loss=0.0721, over 4796.00 frames. ], tot_loss[loss=0.1978, simple_loss=0.2625, pruned_loss=0.06655, over 950094.43 frames. ], batch size: 41, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:13:26,504 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7586, 4.0422, 0.8672, 2.2324, 2.1498, 2.8580, 2.5074, 1.0163], device='cuda:4'), covar=tensor([0.1290, 0.0886, 0.2062, 0.1208, 0.1061, 0.0933, 0.1257, 0.2061], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0255, 0.0144, 0.0125, 0.0137, 0.0157, 0.0122, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 21:13:29,357 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0841, 4.4356, 0.8903, 2.4548, 2.6226, 3.1658, 2.7345, 1.0729], device='cuda:4'), covar=tensor([0.1193, 0.0833, 0.2262, 0.1202, 0.0941, 0.0891, 0.1284, 0.2010], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0255, 0.0144, 0.0125, 0.0137, 0.0157, 0.0122, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 21:13:37,394 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=43083.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:13:44,490 INFO [finetune.py:976] (4/7) Epoch 8, batch 3000, loss[loss=0.2069, simple_loss=0.2683, pruned_loss=0.07277, over 4920.00 frames. ], tot_loss[loss=0.2001, simple_loss=0.2646, pruned_loss=0.06777, over 950110.08 frames. ], batch size: 33, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:13:44,490 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-26 21:13:47,324 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5641, 3.0650, 1.0328, 1.7515, 1.9724, 2.3647, 1.9806, 1.0812], device='cuda:4'), covar=tensor([0.1181, 0.0872, 0.1769, 0.1284, 0.0959, 0.0836, 0.1332, 0.1580], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0255, 0.0144, 0.0125, 0.0137, 0.0157, 0.0122, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 21:13:47,950 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3329, 1.3730, 3.8865, 3.5798, 3.5087, 3.7829, 3.8232, 3.4248], device='cuda:4'), covar=tensor([0.7150, 0.5118, 0.1173, 0.1902, 0.1235, 0.1352, 0.0702, 0.1490], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0309, 0.0411, 0.0415, 0.0351, 0.0407, 0.0318, 0.0372], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-26 21:13:54,962 INFO [finetune.py:1010] (4/7) Epoch 8, validation: loss=0.1551, simple_loss=0.2273, pruned_loss=0.04149, over 2265189.00 frames. 2023-04-26 21:13:54,963 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6336MB 2023-04-26 21:14:18,551 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.136e+02 1.720e+02 2.087e+02 2.508e+02 4.995e+02, threshold=4.174e+02, percent-clipped=1.0 2023-04-26 21:14:27,799 INFO [finetune.py:976] (4/7) Epoch 8, batch 3050, loss[loss=0.2124, simple_loss=0.2507, pruned_loss=0.087, over 4044.00 frames. ], tot_loss[loss=0.2005, simple_loss=0.2653, pruned_loss=0.0679, over 949781.70 frames. ], batch size: 17, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:14:39,686 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=43162.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:15:00,985 INFO [finetune.py:976] (4/7) Epoch 8, batch 3100, loss[loss=0.161, simple_loss=0.2346, pruned_loss=0.04363, over 4902.00 frames. ], tot_loss[loss=0.1985, simple_loss=0.2628, pruned_loss=0.06705, over 948362.90 frames. ], batch size: 46, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:15:11,989 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=43210.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:15:15,103 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=43215.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:15:25,428 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.200e+02 1.610e+02 1.868e+02 2.269e+02 4.698e+02, threshold=3.736e+02, percent-clipped=1.0 2023-04-26 21:15:27,885 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=43235.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:15:34,183 INFO [finetune.py:976] (4/7) Epoch 8, batch 3150, loss[loss=0.1831, simple_loss=0.238, pruned_loss=0.0641, over 4796.00 frames. ], tot_loss[loss=0.1955, simple_loss=0.2593, pruned_loss=0.06583, over 950411.43 frames. ], batch size: 29, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:15:55,725 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6619, 1.2407, 1.7724, 2.0992, 1.8031, 1.6437, 1.7286, 1.7283], device='cuda:4'), covar=tensor([0.6933, 0.8871, 0.8873, 0.8833, 0.8016, 1.1242, 1.1299, 0.9510], device='cuda:4'), in_proj_covar=tensor([0.0406, 0.0423, 0.0505, 0.0527, 0.0436, 0.0456, 0.0468, 0.0465], device='cuda:4'), out_proj_covar=tensor([9.8684e-05, 1.0483e-04, 1.1420e-04, 1.2508e-04, 1.0602e-04, 1.1028e-04, 1.1279e-04, 1.1275e-04], device='cuda:4') 2023-04-26 21:16:18,111 INFO [finetune.py:976] (4/7) Epoch 8, batch 3200, loss[loss=0.1914, simple_loss=0.2502, pruned_loss=0.06625, over 4768.00 frames. ], tot_loss[loss=0.1929, simple_loss=0.2562, pruned_loss=0.06479, over 950349.10 frames. ], batch size: 28, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:16:30,934 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=43305.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:17:03,559 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.128e+02 1.615e+02 2.007e+02 2.471e+02 4.400e+02, threshold=4.014e+02, percent-clipped=2.0 2023-04-26 21:17:20,620 INFO [finetune.py:976] (4/7) Epoch 8, batch 3250, loss[loss=0.1795, simple_loss=0.2357, pruned_loss=0.06164, over 4754.00 frames. ], tot_loss[loss=0.1933, simple_loss=0.2567, pruned_loss=0.065, over 950673.48 frames. ], batch size: 27, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:17:33,418 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=43353.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:17:56,900 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.99 vs. limit=5.0 2023-04-26 21:18:16,670 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0159, 2.0613, 1.7236, 1.9080, 2.1632, 1.7911, 2.6552, 1.5026], device='cuda:4'), covar=tensor([0.4336, 0.2030, 0.5162, 0.3146, 0.1959, 0.2684, 0.1733, 0.5152], device='cuda:4'), in_proj_covar=tensor([0.0348, 0.0351, 0.0431, 0.0363, 0.0387, 0.0384, 0.0384, 0.0420], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 21:18:26,605 INFO [finetune.py:976] (4/7) Epoch 8, batch 3300, loss[loss=0.1865, simple_loss=0.2528, pruned_loss=0.06008, over 4781.00 frames. ], tot_loss[loss=0.1959, simple_loss=0.26, pruned_loss=0.06591, over 950521.90 frames. ], batch size: 29, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:18:59,194 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.201e+02 1.779e+02 2.083e+02 2.593e+02 4.743e+02, threshold=4.166e+02, percent-clipped=4.0 2023-04-26 21:19:06,534 INFO [finetune.py:976] (4/7) Epoch 8, batch 3350, loss[loss=0.1752, simple_loss=0.2502, pruned_loss=0.05003, over 4859.00 frames. ], tot_loss[loss=0.1979, simple_loss=0.2623, pruned_loss=0.06674, over 951025.08 frames. ], batch size: 44, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:19:40,384 INFO [finetune.py:976] (4/7) Epoch 8, batch 3400, loss[loss=0.2384, simple_loss=0.286, pruned_loss=0.09538, over 4681.00 frames. ], tot_loss[loss=0.2001, simple_loss=0.2645, pruned_loss=0.06782, over 949657.47 frames. ], batch size: 23, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:19:54,924 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=43514.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:19:55,552 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=43515.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:20:06,650 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.125e+02 1.715e+02 2.177e+02 2.420e+02 5.380e+02, threshold=4.353e+02, percent-clipped=2.0 2023-04-26 21:20:08,582 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=43535.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:20:13,908 INFO [finetune.py:976] (4/7) Epoch 8, batch 3450, loss[loss=0.1943, simple_loss=0.2521, pruned_loss=0.06828, over 4852.00 frames. ], tot_loss[loss=0.1988, simple_loss=0.2637, pruned_loss=0.06689, over 951371.66 frames. ], batch size: 44, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:20:26,909 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=43563.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:20:33,772 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8912, 2.4020, 1.3749, 1.6775, 2.4065, 1.8449, 1.8373, 1.8406], device='cuda:4'), covar=tensor([0.0506, 0.0334, 0.0305, 0.0552, 0.0233, 0.0521, 0.0491, 0.0556], device='cuda:4'), in_proj_covar=tensor([0.0030, 0.0026, 0.0024, 0.0030, 0.0021, 0.0030, 0.0029, 0.0030], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0050, 0.0038, 0.0048, 0.0048, 0.0050], device='cuda:4') 2023-04-26 21:20:35,643 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=43575.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:20:40,954 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=43583.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:20:47,607 INFO [finetune.py:976] (4/7) Epoch 8, batch 3500, loss[loss=0.2193, simple_loss=0.2894, pruned_loss=0.07463, over 4906.00 frames. ], tot_loss[loss=0.1961, simple_loss=0.2606, pruned_loss=0.06579, over 950513.94 frames. ], batch size: 46, lr: 3.83e-03, grad_scale: 32.0 2023-04-26 21:20:49,546 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3394, 1.2146, 1.5638, 1.5429, 1.2835, 1.1281, 1.3158, 0.9677], device='cuda:4'), covar=tensor([0.0565, 0.0731, 0.0498, 0.0620, 0.0777, 0.1202, 0.0557, 0.0717], device='cuda:4'), in_proj_covar=tensor([0.0065, 0.0073, 0.0072, 0.0067, 0.0076, 0.0095, 0.0078, 0.0075], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 21:21:05,138 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.73 vs. limit=5.0 2023-04-26 21:21:12,598 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.4419, 3.3947, 2.6525, 4.0003, 3.4129, 3.4523, 1.6782, 3.3847], device='cuda:4'), covar=tensor([0.1670, 0.1291, 0.3604, 0.1874, 0.2747, 0.1872, 0.5149, 0.2542], device='cuda:4'), in_proj_covar=tensor([0.0243, 0.0217, 0.0248, 0.0308, 0.0300, 0.0250, 0.0270, 0.0269], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 21:21:13,734 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.083e+02 1.646e+02 1.959e+02 2.415e+02 4.063e+02, threshold=3.918e+02, percent-clipped=0.0 2023-04-26 21:21:21,526 INFO [finetune.py:976] (4/7) Epoch 8, batch 3550, loss[loss=0.1751, simple_loss=0.2385, pruned_loss=0.05583, over 4844.00 frames. ], tot_loss[loss=0.1951, simple_loss=0.2587, pruned_loss=0.06577, over 951854.61 frames. ], batch size: 47, lr: 3.82e-03, grad_scale: 32.0 2023-04-26 21:21:38,683 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4608, 2.4903, 2.1038, 2.3307, 2.7080, 2.2041, 3.4648, 2.0208], device='cuda:4'), covar=tensor([0.4265, 0.2416, 0.4606, 0.3820, 0.1942, 0.2934, 0.1845, 0.4345], device='cuda:4'), in_proj_covar=tensor([0.0344, 0.0349, 0.0428, 0.0362, 0.0385, 0.0382, 0.0381, 0.0417], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 21:22:22,681 INFO [finetune.py:976] (4/7) Epoch 8, batch 3600, loss[loss=0.2416, simple_loss=0.3003, pruned_loss=0.09148, over 4920.00 frames. ], tot_loss[loss=0.1941, simple_loss=0.2575, pruned_loss=0.06541, over 953891.50 frames. ], batch size: 42, lr: 3.82e-03, grad_scale: 32.0 2023-04-26 21:23:15,354 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.186e+02 1.664e+02 2.003e+02 2.703e+02 5.988e+02, threshold=4.006e+02, percent-clipped=3.0 2023-04-26 21:23:34,071 INFO [finetune.py:976] (4/7) Epoch 8, batch 3650, loss[loss=0.2272, simple_loss=0.3027, pruned_loss=0.07588, over 4181.00 frames. ], tot_loss[loss=0.1982, simple_loss=0.2612, pruned_loss=0.06759, over 953549.80 frames. ], batch size: 65, lr: 3.82e-03, grad_scale: 32.0 2023-04-26 21:23:34,181 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4281, 3.2917, 0.9817, 1.7436, 1.7043, 2.2958, 1.8720, 1.0099], device='cuda:4'), covar=tensor([0.1359, 0.0817, 0.1989, 0.1360, 0.1227, 0.1064, 0.1402, 0.2174], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0252, 0.0141, 0.0123, 0.0135, 0.0154, 0.0120, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 21:24:29,572 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=43787.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:24:33,741 INFO [finetune.py:976] (4/7) Epoch 8, batch 3700, loss[loss=0.2025, simple_loss=0.2609, pruned_loss=0.07202, over 4710.00 frames. ], tot_loss[loss=0.2001, simple_loss=0.264, pruned_loss=0.06808, over 950911.24 frames. ], batch size: 23, lr: 3.82e-03, grad_scale: 32.0 2023-04-26 21:25:02,362 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.14 vs. limit=2.0 2023-04-26 21:25:03,743 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.187e+02 1.778e+02 2.044e+02 2.450e+02 4.108e+02, threshold=4.087e+02, percent-clipped=2.0 2023-04-26 21:25:11,089 INFO [finetune.py:976] (4/7) Epoch 8, batch 3750, loss[loss=0.211, simple_loss=0.2759, pruned_loss=0.07307, over 4899.00 frames. ], tot_loss[loss=0.2015, simple_loss=0.2658, pruned_loss=0.06861, over 951852.50 frames. ], batch size: 37, lr: 3.82e-03, grad_scale: 32.0 2023-04-26 21:25:14,030 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=43848.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:25:27,341 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=43870.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:25:36,156 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6608, 1.4116, 1.7234, 2.0202, 1.8282, 1.5927, 1.6416, 1.6930], device='cuda:4'), covar=tensor([0.7000, 0.9169, 0.9117, 1.0306, 0.7600, 1.1719, 1.1649, 0.9816], device='cuda:4'), in_proj_covar=tensor([0.0409, 0.0424, 0.0509, 0.0531, 0.0439, 0.0457, 0.0470, 0.0468], device='cuda:4'), out_proj_covar=tensor([9.9547e-05, 1.0512e-04, 1.1509e-04, 1.2576e-04, 1.0669e-04, 1.1045e-04, 1.1306e-04, 1.1342e-04], device='cuda:4') 2023-04-26 21:25:44,746 INFO [finetune.py:976] (4/7) Epoch 8, batch 3800, loss[loss=0.2604, simple_loss=0.3188, pruned_loss=0.101, over 4810.00 frames. ], tot_loss[loss=0.2027, simple_loss=0.2676, pruned_loss=0.06888, over 951982.64 frames. ], batch size: 38, lr: 3.82e-03, grad_scale: 32.0 2023-04-26 21:25:54,500 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=43909.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:25:56,459 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-26 21:26:09,808 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.001e+02 1.636e+02 2.019e+02 2.329e+02 3.676e+02, threshold=4.038e+02, percent-clipped=0.0 2023-04-26 21:26:11,198 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6773, 1.2405, 1.7975, 2.1381, 1.8345, 1.6390, 1.7006, 1.7349], device='cuda:4'), covar=tensor([0.6940, 0.8949, 0.8393, 0.9075, 0.7909, 1.0988, 1.0624, 0.9767], device='cuda:4'), in_proj_covar=tensor([0.0408, 0.0423, 0.0508, 0.0530, 0.0438, 0.0456, 0.0469, 0.0467], device='cuda:4'), out_proj_covar=tensor([9.9401e-05, 1.0494e-04, 1.1476e-04, 1.2554e-04, 1.0642e-04, 1.1023e-04, 1.1284e-04, 1.1313e-04], device='cuda:4') 2023-04-26 21:26:18,576 INFO [finetune.py:976] (4/7) Epoch 8, batch 3850, loss[loss=0.226, simple_loss=0.2707, pruned_loss=0.09061, over 4814.00 frames. ], tot_loss[loss=0.2, simple_loss=0.2652, pruned_loss=0.06741, over 953119.50 frames. ], batch size: 33, lr: 3.82e-03, grad_scale: 32.0 2023-04-26 21:26:28,331 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=43959.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:26:35,132 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=43970.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:27:02,337 INFO [finetune.py:976] (4/7) Epoch 8, batch 3900, loss[loss=0.2041, simple_loss=0.2607, pruned_loss=0.0737, over 4880.00 frames. ], tot_loss[loss=0.1966, simple_loss=0.2618, pruned_loss=0.06566, over 955443.51 frames. ], batch size: 32, lr: 3.82e-03, grad_scale: 64.0 2023-04-26 21:27:28,193 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.71 vs. limit=2.0 2023-04-26 21:27:35,511 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=44020.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:27:48,404 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.103e+02 1.747e+02 1.960e+02 2.337e+02 4.630e+02, threshold=3.919e+02, percent-clipped=3.0 2023-04-26 21:28:08,675 INFO [finetune.py:976] (4/7) Epoch 8, batch 3950, loss[loss=0.1741, simple_loss=0.2348, pruned_loss=0.05669, over 4755.00 frames. ], tot_loss[loss=0.1941, simple_loss=0.2587, pruned_loss=0.06476, over 956056.11 frames. ], batch size: 23, lr: 3.82e-03, grad_scale: 64.0 2023-04-26 21:28:42,919 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8976, 1.4661, 1.4497, 1.6243, 2.1401, 1.7013, 1.4073, 1.3567], device='cuda:4'), covar=tensor([0.1382, 0.1477, 0.1733, 0.1314, 0.0766, 0.1617, 0.2104, 0.1902], device='cuda:4'), in_proj_covar=tensor([0.0306, 0.0321, 0.0350, 0.0300, 0.0335, 0.0322, 0.0308, 0.0351], device='cuda:4'), out_proj_covar=tensor([6.4882e-05, 6.8114e-05, 7.5765e-05, 6.2152e-05, 7.0180e-05, 6.9179e-05, 6.6306e-05, 7.5464e-05], device='cuda:4') 2023-04-26 21:28:58,843 INFO [finetune.py:976] (4/7) Epoch 8, batch 4000, loss[loss=0.1721, simple_loss=0.2319, pruned_loss=0.05613, over 4738.00 frames. ], tot_loss[loss=0.1953, simple_loss=0.2591, pruned_loss=0.06576, over 956571.20 frames. ], batch size: 23, lr: 3.82e-03, grad_scale: 64.0 2023-04-26 21:29:37,623 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6309, 1.8057, 1.7747, 1.9621, 1.7906, 1.9676, 1.9050, 1.8103], device='cuda:4'), covar=tensor([0.6282, 0.9201, 0.7989, 0.6503, 0.8232, 1.1111, 0.9262, 0.8650], device='cuda:4'), in_proj_covar=tensor([0.0324, 0.0388, 0.0318, 0.0328, 0.0343, 0.0407, 0.0369, 0.0329], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 21:29:49,240 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.897e+01 1.638e+02 2.029e+02 2.445e+02 5.584e+02, threshold=4.057e+02, percent-clipped=1.0 2023-04-26 21:30:02,130 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=44143.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:30:02,663 INFO [finetune.py:976] (4/7) Epoch 8, batch 4050, loss[loss=0.1807, simple_loss=0.2323, pruned_loss=0.06453, over 4329.00 frames. ], tot_loss[loss=0.1976, simple_loss=0.2622, pruned_loss=0.06654, over 956352.44 frames. ], batch size: 19, lr: 3.82e-03, grad_scale: 64.0 2023-04-26 21:30:35,468 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=44170.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:30:45,641 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.82 vs. limit=5.0 2023-04-26 21:31:07,886 INFO [finetune.py:976] (4/7) Epoch 8, batch 4100, loss[loss=0.1778, simple_loss=0.2523, pruned_loss=0.0517, over 4902.00 frames. ], tot_loss[loss=0.1986, simple_loss=0.2636, pruned_loss=0.06678, over 953835.25 frames. ], batch size: 35, lr: 3.82e-03, grad_scale: 64.0 2023-04-26 21:31:26,558 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.69 vs. limit=5.0 2023-04-26 21:31:41,628 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=44218.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:32:00,835 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.207e+02 1.648e+02 2.017e+02 2.380e+02 4.436e+02, threshold=4.034e+02, percent-clipped=1.0 2023-04-26 21:32:10,577 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-04-26 21:32:13,830 INFO [finetune.py:976] (4/7) Epoch 8, batch 4150, loss[loss=0.2418, simple_loss=0.3055, pruned_loss=0.08902, over 4919.00 frames. ], tot_loss[loss=0.2, simple_loss=0.265, pruned_loss=0.06751, over 953399.21 frames. ], batch size: 33, lr: 3.82e-03, grad_scale: 64.0 2023-04-26 21:32:33,289 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.54 vs. limit=2.0 2023-04-26 21:32:34,622 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=44265.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:32:52,671 INFO [finetune.py:976] (4/7) Epoch 8, batch 4200, loss[loss=0.1479, simple_loss=0.2202, pruned_loss=0.03781, over 4756.00 frames. ], tot_loss[loss=0.1993, simple_loss=0.2643, pruned_loss=0.0671, over 953012.33 frames. ], batch size: 28, lr: 3.82e-03, grad_scale: 64.0 2023-04-26 21:32:54,593 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6390, 2.0654, 1.5230, 1.3380, 1.1607, 1.2571, 1.5551, 1.1592], device='cuda:4'), covar=tensor([0.1906, 0.1508, 0.1744, 0.2014, 0.2754, 0.2160, 0.1273, 0.2284], device='cuda:4'), in_proj_covar=tensor([0.0193, 0.0215, 0.0170, 0.0203, 0.0205, 0.0184, 0.0162, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-26 21:33:08,261 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=44315.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:33:18,930 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.191e+02 1.745e+02 1.959e+02 2.417e+02 4.699e+02, threshold=3.917e+02, percent-clipped=3.0 2023-04-26 21:33:26,340 INFO [finetune.py:976] (4/7) Epoch 8, batch 4250, loss[loss=0.1738, simple_loss=0.2513, pruned_loss=0.04816, over 4913.00 frames. ], tot_loss[loss=0.1987, simple_loss=0.2633, pruned_loss=0.06703, over 954023.16 frames. ], batch size: 36, lr: 3.82e-03, grad_scale: 64.0 2023-04-26 21:33:58,432 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4387, 3.6277, 0.6712, 1.9370, 2.0247, 2.5572, 2.0707, 0.8779], device='cuda:4'), covar=tensor([0.1396, 0.0812, 0.2306, 0.1399, 0.1039, 0.1020, 0.1406, 0.2103], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0251, 0.0140, 0.0123, 0.0134, 0.0153, 0.0119, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 21:34:00,167 INFO [finetune.py:976] (4/7) Epoch 8, batch 4300, loss[loss=0.1805, simple_loss=0.2562, pruned_loss=0.05241, over 4914.00 frames. ], tot_loss[loss=0.1953, simple_loss=0.2597, pruned_loss=0.06539, over 954234.19 frames. ], batch size: 37, lr: 3.82e-03, grad_scale: 64.0 2023-04-26 21:34:26,195 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.110e+02 1.667e+02 1.928e+02 2.256e+02 4.744e+02, threshold=3.855e+02, percent-clipped=2.0 2023-04-26 21:34:44,407 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=44443.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:34:44,914 INFO [finetune.py:976] (4/7) Epoch 8, batch 4350, loss[loss=0.1716, simple_loss=0.2363, pruned_loss=0.05339, over 4756.00 frames. ], tot_loss[loss=0.1932, simple_loss=0.2572, pruned_loss=0.06463, over 954409.63 frames. ], batch size: 27, lr: 3.82e-03, grad_scale: 64.0 2023-04-26 21:35:46,506 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=44491.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:35:48,242 INFO [finetune.py:976] (4/7) Epoch 8, batch 4400, loss[loss=0.2033, simple_loss=0.2654, pruned_loss=0.07064, over 4792.00 frames. ], tot_loss[loss=0.1944, simple_loss=0.2583, pruned_loss=0.06525, over 954781.24 frames. ], batch size: 51, lr: 3.82e-03, grad_scale: 64.0 2023-04-26 21:36:11,297 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=44527.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:36:14,647 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.189e+02 1.807e+02 2.142e+02 2.505e+02 5.650e+02, threshold=4.284e+02, percent-clipped=2.0 2023-04-26 21:36:33,283 INFO [finetune.py:976] (4/7) Epoch 8, batch 4450, loss[loss=0.2366, simple_loss=0.2962, pruned_loss=0.08856, over 4753.00 frames. ], tot_loss[loss=0.1971, simple_loss=0.2616, pruned_loss=0.06631, over 954767.58 frames. ], batch size: 54, lr: 3.82e-03, grad_scale: 64.0 2023-04-26 21:36:42,782 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=44550.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:36:48,092 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-04-26 21:36:58,019 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=44565.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:37:30,504 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=44588.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:37:39,770 INFO [finetune.py:976] (4/7) Epoch 8, batch 4500, loss[loss=0.2181, simple_loss=0.2716, pruned_loss=0.08228, over 4816.00 frames. ], tot_loss[loss=0.1992, simple_loss=0.2636, pruned_loss=0.06742, over 954554.78 frames. ], batch size: 25, lr: 3.82e-03, grad_scale: 64.0 2023-04-26 21:37:57,813 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=44611.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:37:58,975 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=44613.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:38:00,250 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=44615.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:38:19,440 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.72 vs. limit=5.0 2023-04-26 21:38:30,731 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.141e+02 1.764e+02 2.099e+02 2.413e+02 4.489e+02, threshold=4.197e+02, percent-clipped=1.0 2023-04-26 21:38:43,667 INFO [finetune.py:976] (4/7) Epoch 8, batch 4550, loss[loss=0.1428, simple_loss=0.217, pruned_loss=0.03426, over 4759.00 frames. ], tot_loss[loss=0.1998, simple_loss=0.2649, pruned_loss=0.06735, over 955216.14 frames. ], batch size: 27, lr: 3.82e-03, grad_scale: 32.0 2023-04-26 21:38:49,388 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-26 21:38:55,369 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=44663.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:39:13,553 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3706, 2.6606, 2.1624, 2.2526, 2.6344, 2.2571, 3.5551, 2.1068], device='cuda:4'), covar=tensor([0.4411, 0.2031, 0.4694, 0.4494, 0.2019, 0.2783, 0.1402, 0.4296], device='cuda:4'), in_proj_covar=tensor([0.0345, 0.0349, 0.0429, 0.0364, 0.0384, 0.0381, 0.0381, 0.0418], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 21:39:17,128 INFO [finetune.py:976] (4/7) Epoch 8, batch 4600, loss[loss=0.1881, simple_loss=0.2523, pruned_loss=0.06195, over 4848.00 frames. ], tot_loss[loss=0.1969, simple_loss=0.2624, pruned_loss=0.06572, over 956703.07 frames. ], batch size: 44, lr: 3.82e-03, grad_scale: 32.0 2023-04-26 21:39:19,143 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8477, 1.6468, 2.0438, 2.2869, 2.0166, 1.7267, 1.8359, 1.8997], device='cuda:4'), covar=tensor([0.7278, 0.9749, 1.0696, 0.9217, 0.8229, 1.3301, 1.2879, 1.1805], device='cuda:4'), in_proj_covar=tensor([0.0409, 0.0423, 0.0505, 0.0527, 0.0438, 0.0457, 0.0469, 0.0465], device='cuda:4'), out_proj_covar=tensor([9.9689e-05, 1.0510e-04, 1.1413e-04, 1.2501e-04, 1.0636e-04, 1.1027e-04, 1.1278e-04, 1.1275e-04], device='cuda:4') 2023-04-26 21:39:43,236 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.165e+02 1.725e+02 1.987e+02 2.350e+02 3.510e+02, threshold=3.973e+02, percent-clipped=0.0 2023-04-26 21:39:55,901 INFO [finetune.py:976] (4/7) Epoch 8, batch 4650, loss[loss=0.2012, simple_loss=0.2661, pruned_loss=0.06816, over 4820.00 frames. ], tot_loss[loss=0.1959, simple_loss=0.2607, pruned_loss=0.06552, over 952531.75 frames. ], batch size: 40, lr: 3.82e-03, grad_scale: 32.0 2023-04-26 21:40:27,103 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.8958, 2.8119, 2.2350, 3.2716, 2.9187, 2.8730, 1.2622, 2.7462], device='cuda:4'), covar=tensor([0.2115, 0.1705, 0.3183, 0.2910, 0.2806, 0.2317, 0.5598, 0.3066], device='cuda:4'), in_proj_covar=tensor([0.0247, 0.0220, 0.0254, 0.0312, 0.0304, 0.0254, 0.0275, 0.0274], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 21:41:01,769 INFO [finetune.py:976] (4/7) Epoch 8, batch 4700, loss[loss=0.165, simple_loss=0.2327, pruned_loss=0.04869, over 4760.00 frames. ], tot_loss[loss=0.1942, simple_loss=0.2585, pruned_loss=0.06492, over 951388.70 frames. ], batch size: 27, lr: 3.82e-03, grad_scale: 32.0 2023-04-26 21:41:26,390 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.155e+02 1.673e+02 1.965e+02 2.397e+02 5.509e+02, threshold=3.929e+02, percent-clipped=2.0 2023-04-26 21:41:35,059 INFO [finetune.py:976] (4/7) Epoch 8, batch 4750, loss[loss=0.185, simple_loss=0.255, pruned_loss=0.05756, over 4939.00 frames. ], tot_loss[loss=0.1923, simple_loss=0.2561, pruned_loss=0.06431, over 952205.64 frames. ], batch size: 33, lr: 3.82e-03, grad_scale: 32.0 2023-04-26 21:41:59,216 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=44883.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:42:08,192 INFO [finetune.py:976] (4/7) Epoch 8, batch 4800, loss[loss=0.185, simple_loss=0.2463, pruned_loss=0.06186, over 4708.00 frames. ], tot_loss[loss=0.1943, simple_loss=0.258, pruned_loss=0.06528, over 951368.07 frames. ], batch size: 23, lr: 3.82e-03, grad_scale: 32.0 2023-04-26 21:42:16,076 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=44906.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:42:32,276 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.212e+02 1.751e+02 2.130e+02 2.730e+02 4.658e+02, threshold=4.261e+02, percent-clipped=2.0 2023-04-26 21:42:40,899 INFO [finetune.py:976] (4/7) Epoch 8, batch 4850, loss[loss=0.2432, simple_loss=0.3075, pruned_loss=0.08944, over 4846.00 frames. ], tot_loss[loss=0.1978, simple_loss=0.2622, pruned_loss=0.06671, over 952220.47 frames. ], batch size: 44, lr: 3.82e-03, grad_scale: 32.0 2023-04-26 21:43:40,816 INFO [finetune.py:976] (4/7) Epoch 8, batch 4900, loss[loss=0.2153, simple_loss=0.2876, pruned_loss=0.07153, over 4908.00 frames. ], tot_loss[loss=0.2006, simple_loss=0.265, pruned_loss=0.06811, over 952799.48 frames. ], batch size: 43, lr: 3.82e-03, grad_scale: 32.0 2023-04-26 21:44:34,742 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.241e+02 1.669e+02 1.942e+02 2.230e+02 4.872e+02, threshold=3.883e+02, percent-clipped=3.0 2023-04-26 21:44:46,772 INFO [finetune.py:976] (4/7) Epoch 8, batch 4950, loss[loss=0.2037, simple_loss=0.2729, pruned_loss=0.06727, over 4883.00 frames. ], tot_loss[loss=0.2, simple_loss=0.2651, pruned_loss=0.06742, over 950273.44 frames. ], batch size: 35, lr: 3.81e-03, grad_scale: 32.0 2023-04-26 21:45:08,561 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=45056.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:45:19,857 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0044, 1.5248, 1.8105, 2.2852, 1.7683, 1.3971, 1.0219, 1.5859], device='cuda:4'), covar=tensor([0.4356, 0.4537, 0.2342, 0.3082, 0.3680, 0.3696, 0.5605, 0.2897], device='cuda:4'), in_proj_covar=tensor([0.0281, 0.0252, 0.0221, 0.0321, 0.0215, 0.0229, 0.0236, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-26 21:45:39,779 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=45081.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:45:47,623 INFO [finetune.py:976] (4/7) Epoch 8, batch 5000, loss[loss=0.1586, simple_loss=0.2208, pruned_loss=0.04819, over 4794.00 frames. ], tot_loss[loss=0.1965, simple_loss=0.2619, pruned_loss=0.06551, over 950508.42 frames. ], batch size: 26, lr: 3.81e-03, grad_scale: 32.0 2023-04-26 21:46:04,183 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=45117.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 21:46:11,457 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3681, 1.5805, 1.5980, 1.7861, 1.6552, 1.8697, 1.7786, 1.6985], device='cuda:4'), covar=tensor([0.4882, 0.6716, 0.6365, 0.5525, 0.6665, 0.9634, 0.6912, 0.6707], device='cuda:4'), in_proj_covar=tensor([0.0323, 0.0385, 0.0317, 0.0327, 0.0343, 0.0404, 0.0366, 0.0327], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 21:46:13,679 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.059e+02 1.645e+02 2.005e+02 2.467e+02 4.350e+02, threshold=4.011e+02, percent-clipped=3.0 2023-04-26 21:46:19,261 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=45142.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:46:20,393 INFO [finetune.py:976] (4/7) Epoch 8, batch 5050, loss[loss=0.208, simple_loss=0.2651, pruned_loss=0.07549, over 4837.00 frames. ], tot_loss[loss=0.1951, simple_loss=0.2596, pruned_loss=0.06531, over 951573.98 frames. ], batch size: 33, lr: 3.81e-03, grad_scale: 32.0 2023-04-26 21:46:36,137 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0594, 2.6956, 1.8427, 1.8792, 1.4680, 1.5266, 2.0248, 1.3975], device='cuda:4'), covar=tensor([0.1789, 0.1574, 0.1762, 0.2068, 0.2670, 0.1986, 0.1177, 0.2211], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0217, 0.0172, 0.0206, 0.0206, 0.0186, 0.0163, 0.0189], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-26 21:46:47,027 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=45183.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:46:53,634 INFO [finetune.py:976] (4/7) Epoch 8, batch 5100, loss[loss=0.1727, simple_loss=0.2261, pruned_loss=0.05967, over 4805.00 frames. ], tot_loss[loss=0.1924, simple_loss=0.2566, pruned_loss=0.06407, over 954601.94 frames. ], batch size: 45, lr: 3.81e-03, grad_scale: 32.0 2023-04-26 21:47:02,073 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=45206.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:47:12,776 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-04-26 21:47:15,723 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4295, 2.0323, 1.9871, 2.4246, 2.1313, 2.1420, 1.6747, 4.7260], device='cuda:4'), covar=tensor([0.0601, 0.0699, 0.0696, 0.1025, 0.0566, 0.0518, 0.0706, 0.0092], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0043, 0.0040, 0.0039, 0.0039, 0.0058], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0012, 0.0013, 0.0016], device='cuda:4') 2023-04-26 21:47:18,726 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=45231.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:47:19,883 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.132e+02 1.617e+02 1.880e+02 2.299e+02 5.532e+02, threshold=3.760e+02, percent-clipped=2.0 2023-04-26 21:47:26,588 INFO [finetune.py:976] (4/7) Epoch 8, batch 5150, loss[loss=0.2062, simple_loss=0.2762, pruned_loss=0.06816, over 4819.00 frames. ], tot_loss[loss=0.1936, simple_loss=0.2572, pruned_loss=0.06495, over 954463.71 frames. ], batch size: 39, lr: 3.81e-03, grad_scale: 32.0 2023-04-26 21:47:32,724 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=45254.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:47:49,207 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-26 21:47:52,233 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3648, 1.2606, 1.7148, 1.6290, 1.2673, 1.0829, 1.4323, 0.8474], device='cuda:4'), covar=tensor([0.0709, 0.0778, 0.0460, 0.0654, 0.0838, 0.1309, 0.0557, 0.0812], device='cuda:4'), in_proj_covar=tensor([0.0066, 0.0073, 0.0072, 0.0067, 0.0077, 0.0096, 0.0078, 0.0074], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 21:48:00,050 INFO [finetune.py:976] (4/7) Epoch 8, batch 5200, loss[loss=0.1833, simple_loss=0.2482, pruned_loss=0.05923, over 4891.00 frames. ], tot_loss[loss=0.1956, simple_loss=0.2601, pruned_loss=0.06559, over 954207.30 frames. ], batch size: 35, lr: 3.81e-03, grad_scale: 32.0 2023-04-26 21:48:44,628 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.724e+02 2.073e+02 2.442e+02 4.626e+02, threshold=4.147e+02, percent-clipped=2.0 2023-04-26 21:48:56,371 INFO [finetune.py:976] (4/7) Epoch 8, batch 5250, loss[loss=0.2195, simple_loss=0.2805, pruned_loss=0.07925, over 4857.00 frames. ], tot_loss[loss=0.1958, simple_loss=0.2613, pruned_loss=0.0652, over 955065.44 frames. ], batch size: 44, lr: 3.81e-03, grad_scale: 32.0 2023-04-26 21:48:58,924 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=45348.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 21:50:01,143 INFO [finetune.py:976] (4/7) Epoch 8, batch 5300, loss[loss=0.1824, simple_loss=0.2512, pruned_loss=0.05685, over 4753.00 frames. ], tot_loss[loss=0.1989, simple_loss=0.2642, pruned_loss=0.06677, over 954808.99 frames. ], batch size: 27, lr: 3.81e-03, grad_scale: 32.0 2023-04-26 21:50:12,228 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.59 vs. limit=5.0 2023-04-26 21:50:21,707 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=45409.0, num_to_drop=1, layers_to_drop={3} 2023-04-26 21:50:23,491 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=45412.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 21:50:55,375 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.314e+02 1.711e+02 1.928e+02 2.396e+02 5.196e+02, threshold=3.857e+02, percent-clipped=2.0 2023-04-26 21:50:57,918 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=45437.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:51:07,772 INFO [finetune.py:976] (4/7) Epoch 8, batch 5350, loss[loss=0.2014, simple_loss=0.2609, pruned_loss=0.07093, over 4811.00 frames. ], tot_loss[loss=0.1994, simple_loss=0.265, pruned_loss=0.06697, over 956001.22 frames. ], batch size: 33, lr: 3.81e-03, grad_scale: 32.0 2023-04-26 21:51:15,247 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.80 vs. limit=2.0 2023-04-26 21:51:45,337 INFO [finetune.py:976] (4/7) Epoch 8, batch 5400, loss[loss=0.2271, simple_loss=0.2824, pruned_loss=0.08592, over 4861.00 frames. ], tot_loss[loss=0.1966, simple_loss=0.2617, pruned_loss=0.06577, over 955894.33 frames. ], batch size: 47, lr: 3.81e-03, grad_scale: 32.0 2023-04-26 21:52:11,224 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.182e+02 1.671e+02 1.868e+02 2.359e+02 4.891e+02, threshold=3.736e+02, percent-clipped=4.0 2023-04-26 21:52:12,403 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0284, 2.4642, 2.1947, 2.2855, 1.8539, 2.0502, 2.0593, 1.7216], device='cuda:4'), covar=tensor([0.2044, 0.1318, 0.0818, 0.1311, 0.3335, 0.1387, 0.2276, 0.2619], device='cuda:4'), in_proj_covar=tensor([0.0300, 0.0321, 0.0231, 0.0293, 0.0319, 0.0274, 0.0260, 0.0283], device='cuda:4'), out_proj_covar=tensor([1.2155e-04, 1.2957e-04, 9.2998e-05, 1.1745e-04, 1.3119e-04, 1.1073e-04, 1.0658e-04, 1.1337e-04], device='cuda:4') 2023-04-26 21:52:18,383 INFO [finetune.py:976] (4/7) Epoch 8, batch 5450, loss[loss=0.1926, simple_loss=0.2567, pruned_loss=0.06425, over 4923.00 frames. ], tot_loss[loss=0.1945, simple_loss=0.2589, pruned_loss=0.06509, over 955338.19 frames. ], batch size: 46, lr: 3.81e-03, grad_scale: 32.0 2023-04-26 21:52:51,573 INFO [finetune.py:976] (4/7) Epoch 8, batch 5500, loss[loss=0.1866, simple_loss=0.2453, pruned_loss=0.06398, over 4867.00 frames. ], tot_loss[loss=0.1918, simple_loss=0.2557, pruned_loss=0.06399, over 955564.05 frames. ], batch size: 31, lr: 3.81e-03, grad_scale: 32.0 2023-04-26 21:53:11,185 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5855, 1.6786, 1.4099, 1.0704, 1.1487, 1.2010, 1.4407, 1.1274], device='cuda:4'), covar=tensor([0.1782, 0.1463, 0.1720, 0.2015, 0.2667, 0.2025, 0.1229, 0.2269], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0216, 0.0171, 0.0204, 0.0205, 0.0184, 0.0161, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-26 21:53:16,426 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.060e+02 1.725e+02 1.934e+02 2.387e+02 4.697e+02, threshold=3.868e+02, percent-clipped=2.0 2023-04-26 21:53:24,572 INFO [finetune.py:976] (4/7) Epoch 8, batch 5550, loss[loss=0.1543, simple_loss=0.2153, pruned_loss=0.04658, over 4689.00 frames. ], tot_loss[loss=0.1932, simple_loss=0.2573, pruned_loss=0.06452, over 956086.69 frames. ], batch size: 23, lr: 3.81e-03, grad_scale: 32.0 2023-04-26 21:53:35,225 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.74 vs. limit=2.0 2023-04-26 21:53:42,472 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2047, 1.5699, 1.7524, 1.8887, 2.3448, 2.0362, 1.7396, 1.5929], device='cuda:4'), covar=tensor([0.1558, 0.1714, 0.1840, 0.1443, 0.0916, 0.1483, 0.2362, 0.2149], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0326, 0.0358, 0.0303, 0.0342, 0.0328, 0.0313, 0.0359], device='cuda:4'), out_proj_covar=tensor([6.6278e-05, 6.9228e-05, 7.7363e-05, 6.2824e-05, 7.1917e-05, 7.0542e-05, 6.7392e-05, 7.7128e-05], device='cuda:4') 2023-04-26 21:54:06,992 INFO [finetune.py:976] (4/7) Epoch 8, batch 5600, loss[loss=0.1816, simple_loss=0.2507, pruned_loss=0.05627, over 4801.00 frames. ], tot_loss[loss=0.1957, simple_loss=0.2606, pruned_loss=0.06542, over 956086.28 frames. ], batch size: 25, lr: 3.81e-03, grad_scale: 32.0 2023-04-26 21:54:12,170 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=45702.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:54:13,288 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=45704.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 21:54:17,973 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=45712.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 21:54:30,164 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.094e+02 1.723e+02 2.023e+02 2.514e+02 5.159e+02, threshold=4.046e+02, percent-clipped=6.0 2023-04-26 21:54:32,580 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=45737.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:54:36,959 INFO [finetune.py:976] (4/7) Epoch 8, batch 5650, loss[loss=0.2386, simple_loss=0.3154, pruned_loss=0.08091, over 4901.00 frames. ], tot_loss[loss=0.1979, simple_loss=0.2634, pruned_loss=0.06619, over 955698.27 frames. ], batch size: 43, lr: 3.81e-03, grad_scale: 32.0 2023-04-26 21:54:46,841 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=45760.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:54:47,505 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4829, 2.0427, 2.5108, 2.9687, 2.3808, 2.0822, 1.7087, 2.2393], device='cuda:4'), covar=tensor([0.3476, 0.3415, 0.1681, 0.2419, 0.2785, 0.2703, 0.4397, 0.2284], device='cuda:4'), in_proj_covar=tensor([0.0282, 0.0253, 0.0220, 0.0320, 0.0214, 0.0229, 0.0236, 0.0189], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-26 21:54:48,654 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=45763.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:54:54,998 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=45774.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:55:07,628 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=45785.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:55:18,507 INFO [finetune.py:976] (4/7) Epoch 8, batch 5700, loss[loss=0.2013, simple_loss=0.2436, pruned_loss=0.07947, over 3976.00 frames. ], tot_loss[loss=0.1967, simple_loss=0.2607, pruned_loss=0.06636, over 940661.07 frames. ], batch size: 17, lr: 3.81e-03, grad_scale: 32.0 2023-04-26 21:55:59,017 INFO [finetune.py:976] (4/7) Epoch 9, batch 0, loss[loss=0.2284, simple_loss=0.3065, pruned_loss=0.07519, over 4901.00 frames. ], tot_loss[loss=0.2284, simple_loss=0.3065, pruned_loss=0.07519, over 4901.00 frames. ], batch size: 46, lr: 3.81e-03, grad_scale: 32.0 2023-04-26 21:55:59,017 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-26 21:56:05,256 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3785, 3.0381, 0.8551, 1.7135, 1.9079, 2.2160, 1.8819, 0.9869], device='cuda:4'), covar=tensor([0.1337, 0.1140, 0.2079, 0.1359, 0.1000, 0.0962, 0.1532, 0.1850], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0253, 0.0142, 0.0124, 0.0135, 0.0156, 0.0120, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 21:56:14,949 INFO [finetune.py:1010] (4/7) Epoch 9, validation: loss=0.1554, simple_loss=0.2289, pruned_loss=0.04093, over 2265189.00 frames. 2023-04-26 21:56:14,949 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6336MB 2023-04-26 21:56:33,616 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.064e+02 1.748e+02 2.064e+02 2.356e+02 2.984e+02, threshold=4.129e+02, percent-clipped=0.0 2023-04-26 21:56:34,966 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=45835.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:56:46,618 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9715, 2.4509, 0.9816, 1.2763, 1.7807, 1.1768, 3.1940, 1.4704], device='cuda:4'), covar=tensor([0.0659, 0.0679, 0.0824, 0.1304, 0.0517, 0.0968, 0.0224, 0.0672], device='cuda:4'), in_proj_covar=tensor([0.0052, 0.0068, 0.0050, 0.0048, 0.0052, 0.0053, 0.0079, 0.0052], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 21:57:05,570 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=45859.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:57:19,424 INFO [finetune.py:976] (4/7) Epoch 9, batch 50, loss[loss=0.2038, simple_loss=0.2635, pruned_loss=0.072, over 4889.00 frames. ], tot_loss[loss=0.1984, simple_loss=0.2669, pruned_loss=0.06495, over 216775.47 frames. ], batch size: 36, lr: 3.81e-03, grad_scale: 32.0 2023-04-26 21:57:38,684 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8178, 1.5323, 1.3952, 1.6120, 2.1707, 1.7025, 1.4763, 1.3334], device='cuda:4'), covar=tensor([0.1606, 0.1564, 0.2064, 0.1382, 0.0822, 0.1588, 0.2118, 0.2204], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0326, 0.0354, 0.0303, 0.0341, 0.0327, 0.0311, 0.0357], device='cuda:4'), out_proj_covar=tensor([6.6122e-05, 6.9124e-05, 7.6568e-05, 6.2735e-05, 7.1757e-05, 7.0199e-05, 6.6910e-05, 7.6727e-05], device='cuda:4') 2023-04-26 21:57:51,858 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=45920.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:57:52,935 INFO [finetune.py:976] (4/7) Epoch 9, batch 100, loss[loss=0.2069, simple_loss=0.2686, pruned_loss=0.0726, over 4741.00 frames. ], tot_loss[loss=0.1917, simple_loss=0.2579, pruned_loss=0.06274, over 381653.72 frames. ], batch size: 54, lr: 3.81e-03, grad_scale: 32.0 2023-04-26 21:58:01,520 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.187e+02 1.655e+02 1.945e+02 2.383e+02 5.251e+02, threshold=3.889e+02, percent-clipped=1.0 2023-04-26 21:58:24,851 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-26 21:58:26,128 INFO [finetune.py:976] (4/7) Epoch 9, batch 150, loss[loss=0.1683, simple_loss=0.2296, pruned_loss=0.05348, over 4908.00 frames. ], tot_loss[loss=0.1895, simple_loss=0.2544, pruned_loss=0.06227, over 508282.97 frames. ], batch size: 36, lr: 3.81e-03, grad_scale: 32.0 2023-04-26 21:58:47,804 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=46004.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 21:58:59,028 INFO [finetune.py:976] (4/7) Epoch 9, batch 200, loss[loss=0.193, simple_loss=0.2626, pruned_loss=0.0617, over 4927.00 frames. ], tot_loss[loss=0.1912, simple_loss=0.2545, pruned_loss=0.06395, over 608618.90 frames. ], batch size: 38, lr: 3.81e-03, grad_scale: 32.0 2023-04-26 21:59:07,994 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.665e+01 1.625e+02 2.034e+02 2.353e+02 6.037e+02, threshold=4.068e+02, percent-clipped=2.0 2023-04-26 21:59:19,778 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=46052.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 21:59:23,462 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=46058.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 21:59:32,393 INFO [finetune.py:976] (4/7) Epoch 9, batch 250, loss[loss=0.1756, simple_loss=0.2429, pruned_loss=0.05411, over 4823.00 frames. ], tot_loss[loss=0.1948, simple_loss=0.2583, pruned_loss=0.0657, over 686779.12 frames. ], batch size: 30, lr: 3.81e-03, grad_scale: 32.0 2023-04-26 21:59:49,105 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.57 vs. limit=5.0 2023-04-26 21:59:51,852 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-26 21:59:53,214 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.14 vs. limit=2.0 2023-04-26 21:59:59,767 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-26 22:00:05,322 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-26 22:00:05,474 INFO [finetune.py:976] (4/7) Epoch 9, batch 300, loss[loss=0.2043, simple_loss=0.2794, pruned_loss=0.0646, over 4812.00 frames. ], tot_loss[loss=0.1971, simple_loss=0.2614, pruned_loss=0.06638, over 746862.49 frames. ], batch size: 45, lr: 3.81e-03, grad_scale: 32.0 2023-04-26 22:00:10,883 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=46130.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:00:12,631 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.245e+02 1.689e+02 1.934e+02 2.285e+02 6.162e+02, threshold=3.867e+02, percent-clipped=1.0 2023-04-26 22:00:24,327 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7658, 1.0965, 1.5280, 1.6468, 1.6018, 1.7320, 1.5461, 1.5589], device='cuda:4'), covar=tensor([0.4949, 0.6958, 0.6223, 0.5854, 0.6721, 0.9268, 0.6834, 0.6329], device='cuda:4'), in_proj_covar=tensor([0.0320, 0.0381, 0.0314, 0.0324, 0.0339, 0.0400, 0.0361, 0.0324], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 22:00:30,406 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4504, 1.7780, 1.3010, 1.1157, 1.1422, 1.1319, 1.3096, 1.0690], device='cuda:4'), covar=tensor([0.1735, 0.1311, 0.1625, 0.1762, 0.2572, 0.2101, 0.1111, 0.2102], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0216, 0.0171, 0.0204, 0.0205, 0.0185, 0.0162, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-26 22:00:42,792 INFO [finetune.py:976] (4/7) Epoch 9, batch 350, loss[loss=0.2681, simple_loss=0.325, pruned_loss=0.1056, over 4847.00 frames. ], tot_loss[loss=0.1992, simple_loss=0.2639, pruned_loss=0.06726, over 792415.11 frames. ], batch size: 49, lr: 3.81e-03, grad_scale: 32.0 2023-04-26 22:01:38,769 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=46215.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:01:48,014 INFO [finetune.py:976] (4/7) Epoch 9, batch 400, loss[loss=0.1874, simple_loss=0.2577, pruned_loss=0.05857, over 4819.00 frames. ], tot_loss[loss=0.1988, simple_loss=0.264, pruned_loss=0.06675, over 829209.96 frames. ], batch size: 38, lr: 3.80e-03, grad_scale: 32.0 2023-04-26 22:02:01,095 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.329e+02 1.830e+02 2.182e+02 2.472e+02 8.213e+02, threshold=4.364e+02, percent-clipped=3.0 2023-04-26 22:02:26,907 INFO [finetune.py:976] (4/7) Epoch 9, batch 450, loss[loss=0.1665, simple_loss=0.2334, pruned_loss=0.04984, over 4708.00 frames. ], tot_loss[loss=0.1977, simple_loss=0.2627, pruned_loss=0.06631, over 856799.91 frames. ], batch size: 23, lr: 3.80e-03, grad_scale: 32.0 2023-04-26 22:02:29,457 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=46276.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:02:54,362 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.23 vs. limit=5.0 2023-04-26 22:03:00,644 INFO [finetune.py:976] (4/7) Epoch 9, batch 500, loss[loss=0.196, simple_loss=0.2529, pruned_loss=0.06955, over 4902.00 frames. ], tot_loss[loss=0.1957, simple_loss=0.2605, pruned_loss=0.06547, over 877327.03 frames. ], batch size: 35, lr: 3.80e-03, grad_scale: 32.0 2023-04-26 22:03:07,763 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.224e+02 1.712e+02 1.942e+02 2.347e+02 4.298e+02, threshold=3.884e+02, percent-clipped=0.0 2023-04-26 22:03:10,769 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=46337.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:03:16,533 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-26 22:03:25,998 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=46358.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:03:26,694 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.63 vs. limit=2.0 2023-04-26 22:03:34,346 INFO [finetune.py:976] (4/7) Epoch 9, batch 550, loss[loss=0.2775, simple_loss=0.3144, pruned_loss=0.1203, over 4829.00 frames. ], tot_loss[loss=0.1933, simple_loss=0.2578, pruned_loss=0.06442, over 894720.57 frames. ], batch size: 38, lr: 3.80e-03, grad_scale: 32.0 2023-04-26 22:03:57,940 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=46406.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:04:01,699 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6741, 1.2680, 1.3772, 1.2957, 1.8454, 1.4851, 1.2243, 1.3181], device='cuda:4'), covar=tensor([0.1663, 0.1478, 0.1749, 0.1377, 0.0781, 0.1602, 0.1960, 0.1868], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0322, 0.0351, 0.0299, 0.0338, 0.0322, 0.0307, 0.0352], device='cuda:4'), out_proj_covar=tensor([6.5362e-05, 6.8273e-05, 7.5811e-05, 6.1986e-05, 7.1194e-05, 6.9239e-05, 6.6182e-05, 7.5651e-05], device='cuda:4') 2023-04-26 22:04:07,673 INFO [finetune.py:976] (4/7) Epoch 9, batch 600, loss[loss=0.1956, simple_loss=0.2695, pruned_loss=0.06087, over 4811.00 frames. ], tot_loss[loss=0.1945, simple_loss=0.2584, pruned_loss=0.06536, over 908756.63 frames. ], batch size: 41, lr: 3.80e-03, grad_scale: 32.0 2023-04-26 22:04:12,574 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=46430.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:04:14,305 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.294e+02 1.692e+02 1.978e+02 2.446e+02 4.653e+02, threshold=3.955e+02, percent-clipped=2.0 2023-04-26 22:04:21,583 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1606, 1.3009, 1.4788, 1.6631, 1.5988, 1.7949, 1.5756, 1.5899], device='cuda:4'), covar=tensor([0.4842, 0.6716, 0.5920, 0.5300, 0.6610, 0.8997, 0.6879, 0.5925], device='cuda:4'), in_proj_covar=tensor([0.0323, 0.0382, 0.0315, 0.0326, 0.0341, 0.0402, 0.0363, 0.0326], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 22:04:40,997 INFO [finetune.py:976] (4/7) Epoch 9, batch 650, loss[loss=0.1831, simple_loss=0.2537, pruned_loss=0.05625, over 4826.00 frames. ], tot_loss[loss=0.1987, simple_loss=0.2631, pruned_loss=0.06713, over 917609.58 frames. ], batch size: 25, lr: 3.80e-03, grad_scale: 32.0 2023-04-26 22:04:44,732 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=46478.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:05:10,685 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=46515.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:05:14,901 INFO [finetune.py:976] (4/7) Epoch 9, batch 700, loss[loss=0.1564, simple_loss=0.2283, pruned_loss=0.04223, over 4806.00 frames. ], tot_loss[loss=0.2004, simple_loss=0.2654, pruned_loss=0.06774, over 927516.88 frames. ], batch size: 51, lr: 3.80e-03, grad_scale: 32.0 2023-04-26 22:05:21,599 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.297e+02 1.714e+02 2.090e+02 2.739e+02 4.841e+02, threshold=4.179e+02, percent-clipped=4.0 2023-04-26 22:05:39,609 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6556, 1.5222, 1.9654, 1.9529, 1.5090, 1.3466, 1.6345, 1.0390], device='cuda:4'), covar=tensor([0.0686, 0.1071, 0.0494, 0.0882, 0.0883, 0.1344, 0.0740, 0.0944], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0074, 0.0072, 0.0067, 0.0076, 0.0097, 0.0079, 0.0074], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 22:05:43,203 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=46563.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:05:53,547 INFO [finetune.py:976] (4/7) Epoch 9, batch 750, loss[loss=0.2016, simple_loss=0.2689, pruned_loss=0.06716, over 4238.00 frames. ], tot_loss[loss=0.2016, simple_loss=0.2666, pruned_loss=0.06834, over 932667.97 frames. ], batch size: 65, lr: 3.80e-03, grad_scale: 32.0 2023-04-26 22:05:54,948 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3226, 1.5821, 1.7193, 1.8969, 1.7078, 1.8434, 1.8674, 1.7981], device='cuda:4'), covar=tensor([0.5042, 0.7239, 0.6295, 0.5436, 0.7213, 1.0330, 0.7246, 0.6708], device='cuda:4'), in_proj_covar=tensor([0.0324, 0.0383, 0.0316, 0.0326, 0.0342, 0.0403, 0.0364, 0.0327], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 22:06:18,437 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=46595.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:06:27,178 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7987, 1.2481, 1.6097, 1.6178, 1.5336, 1.2444, 0.6665, 1.2210], device='cuda:4'), covar=tensor([0.3837, 0.4100, 0.1963, 0.2669, 0.3189, 0.3177, 0.5016, 0.2697], device='cuda:4'), in_proj_covar=tensor([0.0280, 0.0250, 0.0219, 0.0318, 0.0211, 0.0228, 0.0234, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-26 22:06:28,875 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=46602.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:06:36,179 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-26 22:06:50,544 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5273, 1.3821, 1.9120, 1.8270, 1.4189, 1.2560, 1.5830, 1.0315], device='cuda:4'), covar=tensor([0.0674, 0.0906, 0.0458, 0.0848, 0.0801, 0.1220, 0.0726, 0.0881], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0074, 0.0072, 0.0068, 0.0076, 0.0097, 0.0079, 0.0075], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 22:06:58,614 INFO [finetune.py:976] (4/7) Epoch 9, batch 800, loss[loss=0.188, simple_loss=0.2399, pruned_loss=0.06806, over 4925.00 frames. ], tot_loss[loss=0.2001, simple_loss=0.2653, pruned_loss=0.06746, over 936274.51 frames. ], batch size: 33, lr: 3.80e-03, grad_scale: 64.0 2023-04-26 22:06:59,303 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.3358, 3.2549, 2.4124, 3.8601, 3.3577, 3.3992, 1.3581, 3.3168], device='cuda:4'), covar=tensor([0.1819, 0.1397, 0.3483, 0.2290, 0.3252, 0.2030, 0.5760, 0.2438], device='cuda:4'), in_proj_covar=tensor([0.0245, 0.0217, 0.0252, 0.0306, 0.0301, 0.0252, 0.0271, 0.0273], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 22:07:07,533 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5300, 1.2001, 4.4142, 4.0982, 3.8529, 4.2041, 4.0698, 3.8981], device='cuda:4'), covar=tensor([0.7248, 0.6311, 0.0991, 0.1894, 0.1180, 0.1841, 0.1361, 0.1390], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0309, 0.0407, 0.0414, 0.0348, 0.0407, 0.0319, 0.0370], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-26 22:07:09,965 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=46632.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:07:10,489 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.096e+02 1.620e+02 1.991e+02 2.380e+02 4.805e+02, threshold=3.982e+02, percent-clipped=1.0 2023-04-26 22:07:20,375 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5525, 1.1632, 1.6506, 2.0176, 1.7489, 1.5251, 1.5639, 1.5813], device='cuda:4'), covar=tensor([0.5099, 0.6713, 0.6263, 0.6962, 0.5884, 0.8501, 0.7842, 0.7621], device='cuda:4'), in_proj_covar=tensor([0.0409, 0.0421, 0.0503, 0.0524, 0.0437, 0.0456, 0.0467, 0.0466], device='cuda:4'), out_proj_covar=tensor([9.9461e-05, 1.0453e-04, 1.1369e-04, 1.2437e-04, 1.0633e-04, 1.1032e-04, 1.1240e-04, 1.1279e-04], device='cuda:4') 2023-04-26 22:07:26,183 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=46656.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 22:07:31,866 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=46663.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:07:37,211 INFO [finetune.py:976] (4/7) Epoch 9, batch 850, loss[loss=0.1938, simple_loss=0.239, pruned_loss=0.07427, over 3967.00 frames. ], tot_loss[loss=0.1988, simple_loss=0.2635, pruned_loss=0.06709, over 938408.87 frames. ], batch size: 17, lr: 3.80e-03, grad_scale: 64.0 2023-04-26 22:07:53,496 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0460, 2.0047, 1.7824, 1.6778, 2.2724, 1.7226, 2.7054, 1.6457], device='cuda:4'), covar=tensor([0.3785, 0.1899, 0.4746, 0.3045, 0.1652, 0.2507, 0.1219, 0.4019], device='cuda:4'), in_proj_covar=tensor([0.0347, 0.0351, 0.0433, 0.0365, 0.0390, 0.0386, 0.0384, 0.0423], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 22:08:10,358 INFO [finetune.py:976] (4/7) Epoch 9, batch 900, loss[loss=0.1725, simple_loss=0.2302, pruned_loss=0.05743, over 4711.00 frames. ], tot_loss[loss=0.1959, simple_loss=0.26, pruned_loss=0.06584, over 941920.67 frames. ], batch size: 59, lr: 3.80e-03, grad_scale: 64.0 2023-04-26 22:08:14,117 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6299, 1.5442, 1.7206, 2.0196, 2.1193, 1.4873, 1.2878, 1.7708], device='cuda:4'), covar=tensor([0.0936, 0.1217, 0.0766, 0.0613, 0.0512, 0.0973, 0.0868, 0.0618], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0205, 0.0182, 0.0178, 0.0179, 0.0191, 0.0161, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 22:08:17,019 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.197e+02 1.643e+02 2.086e+02 2.450e+02 5.487e+02, threshold=4.172e+02, percent-clipped=3.0 2023-04-26 22:08:43,454 INFO [finetune.py:976] (4/7) Epoch 9, batch 950, loss[loss=0.1324, simple_loss=0.1962, pruned_loss=0.03427, over 4844.00 frames. ], tot_loss[loss=0.194, simple_loss=0.2577, pruned_loss=0.06513, over 942391.53 frames. ], batch size: 25, lr: 3.80e-03, grad_scale: 64.0 2023-04-26 22:08:48,660 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.91 vs. limit=5.0 2023-04-26 22:09:08,170 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=46810.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:09:16,375 INFO [finetune.py:976] (4/7) Epoch 9, batch 1000, loss[loss=0.1955, simple_loss=0.2699, pruned_loss=0.06051, over 4744.00 frames. ], tot_loss[loss=0.1963, simple_loss=0.2611, pruned_loss=0.0658, over 946098.28 frames. ], batch size: 54, lr: 3.80e-03, grad_scale: 64.0 2023-04-26 22:09:22,997 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.171e+02 1.599e+02 1.946e+02 2.521e+02 4.625e+02, threshold=3.893e+02, percent-clipped=2.0 2023-04-26 22:09:33,831 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-26 22:09:49,151 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=46871.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:09:49,628 INFO [finetune.py:976] (4/7) Epoch 9, batch 1050, loss[loss=0.1889, simple_loss=0.2564, pruned_loss=0.0607, over 4897.00 frames. ], tot_loss[loss=0.1991, simple_loss=0.2645, pruned_loss=0.06686, over 947968.33 frames. ], batch size: 35, lr: 3.80e-03, grad_scale: 64.0 2023-04-26 22:10:22,799 INFO [finetune.py:976] (4/7) Epoch 9, batch 1100, loss[loss=0.1832, simple_loss=0.2513, pruned_loss=0.05757, over 4854.00 frames. ], tot_loss[loss=0.1986, simple_loss=0.2644, pruned_loss=0.06641, over 948052.22 frames. ], batch size: 31, lr: 3.80e-03, grad_scale: 32.0 2023-04-26 22:10:29,471 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=46932.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:10:30,585 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.171e+02 1.585e+02 1.927e+02 2.625e+02 4.611e+02, threshold=3.853e+02, percent-clipped=4.0 2023-04-26 22:10:34,517 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.13 vs. limit=2.0 2023-04-26 22:10:40,918 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=46951.0, num_to_drop=1, layers_to_drop={3} 2023-04-26 22:10:45,726 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=46958.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:10:47,592 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2675, 1.2490, 3.8598, 3.5591, 3.4515, 3.6937, 3.7130, 3.4098], device='cuda:4'), covar=tensor([0.7030, 0.5654, 0.1204, 0.2088, 0.1196, 0.1353, 0.1400, 0.1498], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0312, 0.0411, 0.0417, 0.0352, 0.0410, 0.0321, 0.0375], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-26 22:10:56,081 INFO [finetune.py:976] (4/7) Epoch 9, batch 1150, loss[loss=0.1678, simple_loss=0.2432, pruned_loss=0.04626, over 4827.00 frames. ], tot_loss[loss=0.1991, simple_loss=0.2649, pruned_loss=0.06663, over 949831.88 frames. ], batch size: 47, lr: 3.80e-03, grad_scale: 32.0 2023-04-26 22:11:01,459 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=46980.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:11:31,083 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.90 vs. limit=2.0 2023-04-26 22:11:41,251 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.2084, 2.5293, 2.3638, 2.5298, 2.3439, 2.6326, 2.5373, 2.4260], device='cuda:4'), covar=tensor([0.4556, 0.8002, 0.6389, 0.6180, 0.7079, 0.9162, 0.7673, 0.7433], device='cuda:4'), in_proj_covar=tensor([0.0322, 0.0383, 0.0316, 0.0326, 0.0341, 0.0402, 0.0363, 0.0326], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 22:11:43,958 INFO [finetune.py:976] (4/7) Epoch 9, batch 1200, loss[loss=0.1642, simple_loss=0.2344, pruned_loss=0.04698, over 4804.00 frames. ], tot_loss[loss=0.1967, simple_loss=0.2624, pruned_loss=0.0655, over 950113.73 frames. ], batch size: 29, lr: 3.80e-03, grad_scale: 32.0 2023-04-26 22:12:03,720 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.043e+02 1.560e+02 1.899e+02 2.282e+02 5.310e+02, threshold=3.798e+02, percent-clipped=1.0 2023-04-26 22:12:34,498 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.69 vs. limit=5.0 2023-04-26 22:12:55,843 INFO [finetune.py:976] (4/7) Epoch 9, batch 1250, loss[loss=0.196, simple_loss=0.259, pruned_loss=0.06648, over 4892.00 frames. ], tot_loss[loss=0.1929, simple_loss=0.2586, pruned_loss=0.0636, over 952013.23 frames. ], batch size: 35, lr: 3.80e-03, grad_scale: 32.0 2023-04-26 22:12:55,989 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1458, 0.7506, 0.9587, 0.7894, 1.2619, 0.9922, 0.8130, 1.0032], device='cuda:4'), covar=tensor([0.1606, 0.1673, 0.2330, 0.1621, 0.1060, 0.1556, 0.1831, 0.2189], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0325, 0.0355, 0.0303, 0.0340, 0.0325, 0.0309, 0.0355], device='cuda:4'), out_proj_covar=tensor([6.5919e-05, 6.9062e-05, 7.6657e-05, 6.2657e-05, 7.1280e-05, 6.9862e-05, 6.6632e-05, 7.6271e-05], device='cuda:4') 2023-04-26 22:13:21,194 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-26 22:14:00,070 INFO [finetune.py:976] (4/7) Epoch 9, batch 1300, loss[loss=0.2502, simple_loss=0.3018, pruned_loss=0.09933, over 4186.00 frames. ], tot_loss[loss=0.1915, simple_loss=0.2562, pruned_loss=0.06339, over 950647.62 frames. ], batch size: 65, lr: 3.80e-03, grad_scale: 32.0 2023-04-26 22:14:10,081 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8610, 1.7390, 2.1199, 2.3554, 2.0094, 1.8017, 1.9700, 2.0324], device='cuda:4'), covar=tensor([0.6772, 0.9117, 0.9734, 0.8425, 0.8106, 1.2394, 1.2587, 1.0500], device='cuda:4'), in_proj_covar=tensor([0.0407, 0.0418, 0.0502, 0.0522, 0.0436, 0.0453, 0.0466, 0.0463], device='cuda:4'), out_proj_covar=tensor([9.9037e-05, 1.0381e-04, 1.1328e-04, 1.2383e-04, 1.0596e-04, 1.0975e-04, 1.1206e-04, 1.1195e-04], device='cuda:4') 2023-04-26 22:14:15,265 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.174e+02 1.673e+02 1.914e+02 2.315e+02 4.014e+02, threshold=3.828e+02, percent-clipped=2.0 2023-04-26 22:14:58,580 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=47166.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:15:07,953 INFO [finetune.py:976] (4/7) Epoch 9, batch 1350, loss[loss=0.2173, simple_loss=0.2639, pruned_loss=0.08537, over 4761.00 frames. ], tot_loss[loss=0.1934, simple_loss=0.2577, pruned_loss=0.06461, over 952228.64 frames. ], batch size: 27, lr: 3.80e-03, grad_scale: 32.0 2023-04-26 22:15:17,655 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4805, 2.5267, 2.1395, 2.3914, 2.6542, 2.1626, 3.5623, 2.0627], device='cuda:4'), covar=tensor([0.4157, 0.2351, 0.4607, 0.3430, 0.1855, 0.3045, 0.1564, 0.4199], device='cuda:4'), in_proj_covar=tensor([0.0347, 0.0352, 0.0433, 0.0364, 0.0390, 0.0388, 0.0384, 0.0424], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 22:15:29,087 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5598, 1.4706, 0.9732, 1.2353, 1.8549, 1.4058, 1.3327, 1.3512], device='cuda:4'), covar=tensor([0.0532, 0.0394, 0.0358, 0.0597, 0.0279, 0.0558, 0.0511, 0.0596], device='cuda:4'), in_proj_covar=tensor([0.0030, 0.0026, 0.0024, 0.0030, 0.0021, 0.0030, 0.0029, 0.0030], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0050, 0.0038, 0.0048, 0.0048, 0.0050], device='cuda:4') 2023-04-26 22:15:57,491 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4416, 1.6416, 1.4411, 2.0095, 1.8767, 1.9108, 1.4404, 4.2778], device='cuda:4'), covar=tensor([0.0567, 0.0788, 0.0782, 0.1176, 0.0610, 0.0626, 0.0768, 0.0104], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0039, 0.0040, 0.0044, 0.0040, 0.0039, 0.0039, 0.0059], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-26 22:15:58,006 INFO [finetune.py:976] (4/7) Epoch 9, batch 1400, loss[loss=0.1491, simple_loss=0.2284, pruned_loss=0.03488, over 4819.00 frames. ], tot_loss[loss=0.197, simple_loss=0.2622, pruned_loss=0.06588, over 952851.06 frames. ], batch size: 39, lr: 3.80e-03, grad_scale: 32.0 2023-04-26 22:16:06,777 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.248e+02 1.819e+02 2.122e+02 2.454e+02 4.582e+02, threshold=4.244e+02, percent-clipped=5.0 2023-04-26 22:16:11,710 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=47241.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:16:12,958 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7666, 1.8492, 0.9126, 1.4234, 1.9652, 1.6281, 1.4657, 1.6154], device='cuda:4'), covar=tensor([0.0498, 0.0365, 0.0377, 0.0558, 0.0264, 0.0548, 0.0552, 0.0563], device='cuda:4'), in_proj_covar=tensor([0.0030, 0.0026, 0.0024, 0.0030, 0.0021, 0.0030, 0.0029, 0.0030], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0050, 0.0038, 0.0048, 0.0048, 0.0050], device='cuda:4') 2023-04-26 22:16:17,771 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=47251.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:16:21,999 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=47258.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:16:26,251 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9128, 2.3478, 1.0352, 1.2833, 1.8265, 1.1247, 3.3134, 1.5686], device='cuda:4'), covar=tensor([0.0735, 0.0828, 0.0867, 0.1277, 0.0564, 0.1047, 0.0243, 0.0675], device='cuda:4'), in_proj_covar=tensor([0.0052, 0.0068, 0.0050, 0.0048, 0.0052, 0.0053, 0.0079, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0011, 0.0007], device='cuda:4') 2023-04-26 22:16:30,890 INFO [finetune.py:976] (4/7) Epoch 9, batch 1450, loss[loss=0.2067, simple_loss=0.2757, pruned_loss=0.06881, over 4896.00 frames. ], tot_loss[loss=0.1982, simple_loss=0.2639, pruned_loss=0.06622, over 952725.43 frames. ], batch size: 32, lr: 3.80e-03, grad_scale: 32.0 2023-04-26 22:16:43,924 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-04-26 22:16:50,057 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=47299.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:16:51,889 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=47302.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:16:54,274 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=47306.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:17:03,917 INFO [finetune.py:976] (4/7) Epoch 9, batch 1500, loss[loss=0.1874, simple_loss=0.2593, pruned_loss=0.05777, over 4783.00 frames. ], tot_loss[loss=0.2005, simple_loss=0.2664, pruned_loss=0.06734, over 953919.21 frames. ], batch size: 28, lr: 3.80e-03, grad_scale: 32.0 2023-04-26 22:17:08,756 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0456, 2.4146, 1.0194, 1.2922, 1.7254, 1.1923, 3.2811, 1.7085], device='cuda:4'), covar=tensor([0.0701, 0.0640, 0.0811, 0.1313, 0.0579, 0.1035, 0.0287, 0.0648], device='cuda:4'), in_proj_covar=tensor([0.0052, 0.0068, 0.0050, 0.0048, 0.0052, 0.0053, 0.0079, 0.0052], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-26 22:17:12,630 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.146e+02 1.650e+02 1.825e+02 2.290e+02 4.290e+02, threshold=3.651e+02, percent-clipped=1.0 2023-04-26 22:17:21,330 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=5.21 vs. limit=5.0 2023-04-26 22:18:03,822 INFO [finetune.py:976] (4/7) Epoch 9, batch 1550, loss[loss=0.1923, simple_loss=0.252, pruned_loss=0.06627, over 4833.00 frames. ], tot_loss[loss=0.1995, simple_loss=0.2652, pruned_loss=0.06686, over 954928.45 frames. ], batch size: 49, lr: 3.80e-03, grad_scale: 32.0 2023-04-26 22:18:13,648 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=47379.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:19:00,829 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6683, 1.5650, 1.7471, 1.9984, 2.0986, 1.6249, 1.2867, 1.8005], device='cuda:4'), covar=tensor([0.0788, 0.1105, 0.0675, 0.0498, 0.0512, 0.0773, 0.0822, 0.0554], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0204, 0.0182, 0.0177, 0.0178, 0.0190, 0.0161, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 22:19:10,271 INFO [finetune.py:976] (4/7) Epoch 9, batch 1600, loss[loss=0.2573, simple_loss=0.3039, pruned_loss=0.1053, over 4820.00 frames. ], tot_loss[loss=0.1968, simple_loss=0.2623, pruned_loss=0.06565, over 955913.50 frames. ], batch size: 39, lr: 3.80e-03, grad_scale: 32.0 2023-04-26 22:19:23,854 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.166e+02 1.649e+02 1.912e+02 2.310e+02 4.649e+02, threshold=3.824e+02, percent-clipped=3.0 2023-04-26 22:19:34,980 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=47440.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:20:08,747 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=47466.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:20:17,938 INFO [finetune.py:976] (4/7) Epoch 9, batch 1650, loss[loss=0.1993, simple_loss=0.2613, pruned_loss=0.06867, over 4898.00 frames. ], tot_loss[loss=0.193, simple_loss=0.2583, pruned_loss=0.06385, over 956802.00 frames. ], batch size: 32, lr: 3.80e-03, grad_scale: 32.0 2023-04-26 22:21:00,460 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=47514.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:21:05,271 INFO [finetune.py:976] (4/7) Epoch 9, batch 1700, loss[loss=0.1377, simple_loss=0.2092, pruned_loss=0.03308, over 4764.00 frames. ], tot_loss[loss=0.191, simple_loss=0.2559, pruned_loss=0.06305, over 957953.15 frames. ], batch size: 27, lr: 3.80e-03, grad_scale: 32.0 2023-04-26 22:21:12,554 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.160e+02 1.638e+02 2.054e+02 2.588e+02 3.948e+02, threshold=4.108e+02, percent-clipped=3.0 2023-04-26 22:21:33,914 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.47 vs. limit=2.0 2023-04-26 22:21:38,482 INFO [finetune.py:976] (4/7) Epoch 9, batch 1750, loss[loss=0.2042, simple_loss=0.2721, pruned_loss=0.06819, over 4916.00 frames. ], tot_loss[loss=0.1942, simple_loss=0.259, pruned_loss=0.06469, over 957470.13 frames. ], batch size: 36, lr: 3.80e-03, grad_scale: 32.0 2023-04-26 22:21:54,733 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=47597.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:22:11,819 INFO [finetune.py:976] (4/7) Epoch 9, batch 1800, loss[loss=0.2189, simple_loss=0.282, pruned_loss=0.07792, over 4817.00 frames. ], tot_loss[loss=0.1975, simple_loss=0.263, pruned_loss=0.066, over 955386.14 frames. ], batch size: 38, lr: 3.79e-03, grad_scale: 32.0 2023-04-26 22:22:19,164 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.171e+02 1.815e+02 2.187e+02 2.572e+02 4.070e+02, threshold=4.375e+02, percent-clipped=0.0 2023-04-26 22:22:38,163 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5515, 1.7761, 2.3154, 2.9009, 2.3265, 1.7457, 1.5754, 2.1802], device='cuda:4'), covar=tensor([0.3605, 0.3932, 0.1760, 0.2900, 0.3289, 0.2970, 0.4556, 0.2677], device='cuda:4'), in_proj_covar=tensor([0.0282, 0.0251, 0.0220, 0.0319, 0.0213, 0.0228, 0.0235, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-26 22:22:45,311 INFO [finetune.py:976] (4/7) Epoch 9, batch 1850, loss[loss=0.1927, simple_loss=0.2554, pruned_loss=0.06499, over 4904.00 frames. ], tot_loss[loss=0.1977, simple_loss=0.263, pruned_loss=0.06622, over 952366.46 frames. ], batch size: 37, lr: 3.79e-03, grad_scale: 32.0 2023-04-26 22:23:44,011 INFO [finetune.py:976] (4/7) Epoch 9, batch 1900, loss[loss=0.1797, simple_loss=0.2483, pruned_loss=0.05559, over 4736.00 frames. ], tot_loss[loss=0.1984, simple_loss=0.2642, pruned_loss=0.06628, over 953259.54 frames. ], batch size: 54, lr: 3.79e-03, grad_scale: 16.0 2023-04-26 22:24:02,107 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.223e+02 1.717e+02 1.936e+02 2.333e+02 4.969e+02, threshold=3.871e+02, percent-clipped=1.0 2023-04-26 22:24:02,209 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=47735.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:24:04,109 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=47738.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:24:14,270 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2726, 1.1643, 1.5209, 1.4213, 1.2341, 1.0270, 1.2449, 0.9464], device='cuda:4'), covar=tensor([0.0694, 0.0611, 0.0491, 0.0529, 0.0710, 0.1091, 0.0534, 0.0706], device='cuda:4'), in_proj_covar=tensor([0.0066, 0.0073, 0.0072, 0.0067, 0.0076, 0.0096, 0.0078, 0.0074], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 22:24:33,493 INFO [finetune.py:976] (4/7) Epoch 9, batch 1950, loss[loss=0.17, simple_loss=0.2539, pruned_loss=0.04299, over 4852.00 frames. ], tot_loss[loss=0.1981, simple_loss=0.2641, pruned_loss=0.06602, over 955181.81 frames. ], batch size: 34, lr: 3.79e-03, grad_scale: 16.0 2023-04-26 22:24:43,259 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=47788.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:24:50,422 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=47799.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:24:57,539 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=47809.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 22:25:11,663 INFO [finetune.py:976] (4/7) Epoch 9, batch 2000, loss[loss=0.1859, simple_loss=0.2527, pruned_loss=0.05957, over 4829.00 frames. ], tot_loss[loss=0.1968, simple_loss=0.2619, pruned_loss=0.06582, over 957137.23 frames. ], batch size: 33, lr: 3.79e-03, grad_scale: 16.0 2023-04-26 22:25:30,747 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.076e+02 1.686e+02 2.019e+02 2.338e+02 4.477e+02, threshold=4.037e+02, percent-clipped=4.0 2023-04-26 22:25:50,203 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=47849.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:26:17,403 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=47870.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 22:26:24,623 INFO [finetune.py:976] (4/7) Epoch 9, batch 2050, loss[loss=0.2294, simple_loss=0.2697, pruned_loss=0.09451, over 4706.00 frames. ], tot_loss[loss=0.1936, simple_loss=0.2581, pruned_loss=0.06453, over 958564.81 frames. ], batch size: 23, lr: 3.79e-03, grad_scale: 16.0 2023-04-26 22:26:57,922 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=47897.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:27:01,924 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-04-26 22:27:30,785 INFO [finetune.py:976] (4/7) Epoch 9, batch 2100, loss[loss=0.1867, simple_loss=0.2496, pruned_loss=0.06184, over 4873.00 frames. ], tot_loss[loss=0.1944, simple_loss=0.259, pruned_loss=0.06489, over 959409.09 frames. ], batch size: 34, lr: 3.79e-03, grad_scale: 16.0 2023-04-26 22:27:39,268 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.172e+02 1.719e+02 2.114e+02 2.573e+02 5.213e+02, threshold=4.228e+02, percent-clipped=2.0 2023-04-26 22:27:45,505 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=47945.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:27:56,372 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2809, 1.5139, 1.6708, 1.8201, 1.7126, 1.8844, 1.7356, 1.7620], device='cuda:4'), covar=tensor([0.5317, 0.7331, 0.6018, 0.5942, 0.7025, 0.9466, 0.7270, 0.6627], device='cuda:4'), in_proj_covar=tensor([0.0322, 0.0381, 0.0315, 0.0324, 0.0341, 0.0402, 0.0363, 0.0325], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 22:28:20,274 INFO [finetune.py:976] (4/7) Epoch 9, batch 2150, loss[loss=0.2533, simple_loss=0.3135, pruned_loss=0.09655, over 4817.00 frames. ], tot_loss[loss=0.1961, simple_loss=0.2614, pruned_loss=0.06541, over 960074.29 frames. ], batch size: 38, lr: 3.79e-03, grad_scale: 16.0 2023-04-26 22:28:23,963 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2447, 1.6049, 1.4010, 1.7516, 1.6390, 1.8803, 1.3883, 3.5093], device='cuda:4'), covar=tensor([0.0676, 0.0771, 0.0811, 0.1167, 0.0628, 0.0557, 0.0742, 0.0132], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0039, 0.0040, 0.0044, 0.0040, 0.0039, 0.0039, 0.0059], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-26 22:28:48,494 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2218, 1.5008, 1.2846, 1.7507, 1.4773, 1.6358, 1.3230, 3.0262], device='cuda:4'), covar=tensor([0.0665, 0.0863, 0.0879, 0.1244, 0.0708, 0.0545, 0.0829, 0.0223], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0039, 0.0040, 0.0044, 0.0040, 0.0039, 0.0039, 0.0059], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-26 22:28:59,634 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5236, 1.2684, 1.6467, 1.6266, 1.3806, 1.2206, 1.3083, 0.8281], device='cuda:4'), covar=tensor([0.0527, 0.0938, 0.0566, 0.0654, 0.0821, 0.1254, 0.0689, 0.0854], device='cuda:4'), in_proj_covar=tensor([0.0066, 0.0073, 0.0071, 0.0067, 0.0076, 0.0095, 0.0078, 0.0074], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 22:29:07,925 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4790, 0.9831, 0.3562, 1.1850, 1.0985, 1.3717, 1.2668, 1.2502], device='cuda:4'), covar=tensor([0.0539, 0.0446, 0.0473, 0.0620, 0.0327, 0.0573, 0.0544, 0.0622], device='cuda:4'), in_proj_covar=tensor([0.0030, 0.0025, 0.0023, 0.0030, 0.0021, 0.0029, 0.0029, 0.0030], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0050, 0.0037, 0.0048, 0.0048, 0.0049], device='cuda:4') 2023-04-26 22:29:17,268 INFO [finetune.py:976] (4/7) Epoch 9, batch 2200, loss[loss=0.1763, simple_loss=0.2473, pruned_loss=0.05264, over 4815.00 frames. ], tot_loss[loss=0.1972, simple_loss=0.2629, pruned_loss=0.06573, over 958280.40 frames. ], batch size: 39, lr: 3.79e-03, grad_scale: 16.0 2023-04-26 22:29:32,261 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.341e+02 1.685e+02 2.065e+02 2.367e+02 5.445e+02, threshold=4.130e+02, percent-clipped=2.0 2023-04-26 22:29:32,360 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=48035.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:30:13,114 INFO [finetune.py:976] (4/7) Epoch 9, batch 2250, loss[loss=0.2704, simple_loss=0.3176, pruned_loss=0.1116, over 4886.00 frames. ], tot_loss[loss=0.1989, simple_loss=0.2647, pruned_loss=0.06653, over 958786.29 frames. ], batch size: 32, lr: 3.79e-03, grad_scale: 16.0 2023-04-26 22:30:18,388 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=48072.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:30:32,860 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=48083.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:30:45,750 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=48094.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:31:21,271 INFO [finetune.py:976] (4/7) Epoch 9, batch 2300, loss[loss=0.1752, simple_loss=0.2523, pruned_loss=0.04903, over 4870.00 frames. ], tot_loss[loss=0.1975, simple_loss=0.2638, pruned_loss=0.0656, over 957420.37 frames. ], batch size: 34, lr: 3.79e-03, grad_scale: 16.0 2023-04-26 22:31:35,719 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=48133.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:31:36,817 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.203e+02 1.671e+02 1.981e+02 2.243e+02 3.920e+02, threshold=3.963e+02, percent-clipped=0.0 2023-04-26 22:31:50,254 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=48144.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:32:14,433 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=48165.0, num_to_drop=1, layers_to_drop={3} 2023-04-26 22:32:25,262 INFO [finetune.py:976] (4/7) Epoch 9, batch 2350, loss[loss=0.1543, simple_loss=0.2276, pruned_loss=0.04055, over 4751.00 frames. ], tot_loss[loss=0.1945, simple_loss=0.2603, pruned_loss=0.06431, over 956332.00 frames. ], batch size: 23, lr: 3.79e-03, grad_scale: 16.0 2023-04-26 22:32:35,780 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7067, 1.3244, 1.7382, 2.2103, 1.8399, 1.6891, 1.7470, 1.7483], device='cuda:4'), covar=tensor([0.6186, 0.8862, 0.8540, 0.8281, 0.7714, 1.0871, 1.0611, 1.0552], device='cuda:4'), in_proj_covar=tensor([0.0408, 0.0420, 0.0503, 0.0523, 0.0438, 0.0456, 0.0467, 0.0466], device='cuda:4'), out_proj_covar=tensor([9.9324e-05, 1.0414e-04, 1.1354e-04, 1.2412e-04, 1.0649e-04, 1.1021e-04, 1.1231e-04, 1.1269e-04], device='cuda:4') 2023-04-26 22:32:47,418 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7498, 1.3238, 1.8078, 2.2152, 1.8327, 1.7013, 1.7489, 1.8165], device='cuda:4'), covar=tensor([0.6608, 0.8876, 0.9284, 0.8925, 0.8399, 1.0393, 1.1287, 0.9380], device='cuda:4'), in_proj_covar=tensor([0.0407, 0.0419, 0.0502, 0.0522, 0.0438, 0.0455, 0.0466, 0.0465], device='cuda:4'), out_proj_covar=tensor([9.9147e-05, 1.0397e-04, 1.1336e-04, 1.2391e-04, 1.0631e-04, 1.1003e-04, 1.1210e-04, 1.1252e-04], device='cuda:4') 2023-04-26 22:33:00,965 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8085, 2.4692, 1.6894, 1.6103, 1.3538, 1.3371, 1.7075, 1.2464], device='cuda:4'), covar=tensor([0.2107, 0.1404, 0.1857, 0.2111, 0.2798, 0.2374, 0.1343, 0.2278], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0217, 0.0172, 0.0205, 0.0207, 0.0184, 0.0162, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-26 22:33:19,205 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=48212.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 22:33:30,439 INFO [finetune.py:976] (4/7) Epoch 9, batch 2400, loss[loss=0.1906, simple_loss=0.2584, pruned_loss=0.06137, over 4830.00 frames. ], tot_loss[loss=0.193, simple_loss=0.2579, pruned_loss=0.06407, over 958019.78 frames. ], batch size: 33, lr: 3.79e-03, grad_scale: 16.0 2023-04-26 22:33:31,513 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-26 22:33:45,094 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.152e+02 1.712e+02 1.962e+02 2.337e+02 6.803e+02, threshold=3.925e+02, percent-clipped=4.0 2023-04-26 22:34:34,225 INFO [finetune.py:976] (4/7) Epoch 9, batch 2450, loss[loss=0.1924, simple_loss=0.2637, pruned_loss=0.06054, over 4842.00 frames. ], tot_loss[loss=0.1903, simple_loss=0.2548, pruned_loss=0.06288, over 957643.35 frames. ], batch size: 47, lr: 3.79e-03, grad_scale: 16.0 2023-04-26 22:34:34,986 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=48273.0, num_to_drop=1, layers_to_drop={3} 2023-04-26 22:34:45,943 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7576, 1.3263, 1.3960, 1.3951, 1.8892, 1.5170, 1.1843, 1.3369], device='cuda:4'), covar=tensor([0.1904, 0.1751, 0.2353, 0.1564, 0.0968, 0.1917, 0.2894, 0.2691], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0325, 0.0354, 0.0297, 0.0337, 0.0325, 0.0307, 0.0355], device='cuda:4'), out_proj_covar=tensor([6.5279e-05, 6.8971e-05, 7.6539e-05, 6.1292e-05, 7.0704e-05, 6.9693e-05, 6.6005e-05, 7.6128e-05], device='cuda:4') 2023-04-26 22:35:07,202 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-26 22:35:18,099 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5509, 1.0786, 1.6062, 1.9121, 1.6679, 1.5619, 1.6063, 1.6453], device='cuda:4'), covar=tensor([0.6949, 0.9438, 0.9941, 1.0382, 0.7970, 1.1701, 1.1670, 1.1288], device='cuda:4'), in_proj_covar=tensor([0.0410, 0.0421, 0.0505, 0.0524, 0.0439, 0.0458, 0.0470, 0.0468], device='cuda:4'), out_proj_covar=tensor([9.9772e-05, 1.0450e-04, 1.1400e-04, 1.2448e-04, 1.0669e-04, 1.1079e-04, 1.1291e-04, 1.1326e-04], device='cuda:4') 2023-04-26 22:35:25,353 INFO [finetune.py:976] (4/7) Epoch 9, batch 2500, loss[loss=0.2281, simple_loss=0.3002, pruned_loss=0.07797, over 4824.00 frames. ], tot_loss[loss=0.1938, simple_loss=0.2587, pruned_loss=0.06445, over 957093.69 frames. ], batch size: 40, lr: 3.79e-03, grad_scale: 16.0 2023-04-26 22:35:26,648 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=48324.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:35:34,247 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.038e+02 1.734e+02 2.061e+02 2.480e+02 6.002e+02, threshold=4.122e+02, percent-clipped=6.0 2023-04-26 22:36:14,631 INFO [finetune.py:976] (4/7) Epoch 9, batch 2550, loss[loss=0.1704, simple_loss=0.2546, pruned_loss=0.04308, over 4922.00 frames. ], tot_loss[loss=0.1974, simple_loss=0.263, pruned_loss=0.06592, over 956430.15 frames. ], batch size: 38, lr: 3.79e-03, grad_scale: 16.0 2023-04-26 22:36:34,709 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=48385.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:36:53,724 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=48394.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:36:54,394 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7026, 2.3517, 1.8004, 1.7141, 1.2599, 1.3722, 2.0451, 1.2492], device='cuda:4'), covar=tensor([0.2095, 0.1665, 0.1897, 0.2106, 0.2779, 0.2309, 0.1121, 0.2381], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0216, 0.0171, 0.0205, 0.0205, 0.0183, 0.0161, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-26 22:36:55,554 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=48397.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:37:28,005 INFO [finetune.py:976] (4/7) Epoch 9, batch 2600, loss[loss=0.2181, simple_loss=0.2881, pruned_loss=0.07403, over 4916.00 frames. ], tot_loss[loss=0.1977, simple_loss=0.2635, pruned_loss=0.06598, over 954329.95 frames. ], batch size: 36, lr: 3.79e-03, grad_scale: 16.0 2023-04-26 22:37:28,727 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=48423.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:37:31,745 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=48428.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:37:41,881 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.066e+02 1.714e+02 2.123e+02 2.498e+02 4.905e+02, threshold=4.246e+02, percent-clipped=1.0 2023-04-26 22:37:52,939 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=48442.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:37:54,673 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=48444.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:38:06,018 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.4777, 4.3707, 2.9710, 5.1387, 4.4039, 4.4891, 1.9700, 4.4556], device='cuda:4'), covar=tensor([0.1644, 0.0999, 0.3682, 0.1007, 0.2630, 0.1572, 0.5843, 0.2346], device='cuda:4'), in_proj_covar=tensor([0.0243, 0.0218, 0.0253, 0.0307, 0.0301, 0.0251, 0.0272, 0.0271], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 22:38:14,005 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=48458.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:38:15,246 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.8306, 2.0963, 1.9540, 2.1359, 1.9340, 2.1013, 2.0657, 1.9900], device='cuda:4'), covar=tensor([0.5718, 0.8827, 0.7600, 0.6987, 0.7745, 1.0216, 0.8979, 0.8646], device='cuda:4'), in_proj_covar=tensor([0.0323, 0.0383, 0.0316, 0.0326, 0.0342, 0.0403, 0.0364, 0.0326], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 22:38:23,633 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7383, 2.3071, 1.8365, 2.1179, 1.6824, 1.8535, 1.9663, 1.5125], device='cuda:4'), covar=tensor([0.2574, 0.1944, 0.1407, 0.1872, 0.3151, 0.1840, 0.2278, 0.2986], device='cuda:4'), in_proj_covar=tensor([0.0295, 0.0316, 0.0227, 0.0288, 0.0312, 0.0270, 0.0257, 0.0280], device='cuda:4'), out_proj_covar=tensor([1.1947e-04, 1.2715e-04, 9.1432e-05, 1.1542e-04, 1.2812e-04, 1.0876e-04, 1.0535e-04, 1.1236e-04], device='cuda:4') 2023-04-26 22:38:24,205 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=48465.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 22:38:34,661 INFO [finetune.py:976] (4/7) Epoch 9, batch 2650, loss[loss=0.1818, simple_loss=0.2442, pruned_loss=0.05973, over 4302.00 frames. ], tot_loss[loss=0.1988, simple_loss=0.2647, pruned_loss=0.06643, over 953667.05 frames. ], batch size: 65, lr: 3.79e-03, grad_scale: 16.0 2023-04-26 22:38:46,956 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=48484.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:38:57,118 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=48492.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:39:21,191 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=48513.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 22:39:30,133 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6123, 1.1253, 1.6058, 2.0899, 1.7846, 1.5524, 1.5893, 1.6049], device='cuda:4'), covar=tensor([0.5714, 0.7836, 0.7570, 0.8124, 0.6982, 0.9570, 0.9236, 0.9124], device='cuda:4'), in_proj_covar=tensor([0.0409, 0.0420, 0.0503, 0.0523, 0.0438, 0.0457, 0.0469, 0.0467], device='cuda:4'), out_proj_covar=tensor([9.9555e-05, 1.0417e-04, 1.1364e-04, 1.2432e-04, 1.0647e-04, 1.1041e-04, 1.1265e-04, 1.1299e-04], device='cuda:4') 2023-04-26 22:39:32,428 INFO [finetune.py:976] (4/7) Epoch 9, batch 2700, loss[loss=0.1547, simple_loss=0.2286, pruned_loss=0.04037, over 4900.00 frames. ], tot_loss[loss=0.1955, simple_loss=0.2619, pruned_loss=0.0646, over 955546.32 frames. ], batch size: 32, lr: 3.79e-03, grad_scale: 16.0 2023-04-26 22:39:40,872 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.141e+02 1.732e+02 1.990e+02 2.443e+02 3.754e+02, threshold=3.980e+02, percent-clipped=0.0 2023-04-26 22:40:20,119 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=48568.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 22:40:28,387 INFO [finetune.py:976] (4/7) Epoch 9, batch 2750, loss[loss=0.1651, simple_loss=0.2401, pruned_loss=0.04503, over 4816.00 frames. ], tot_loss[loss=0.1942, simple_loss=0.2594, pruned_loss=0.06444, over 955889.09 frames. ], batch size: 40, lr: 3.79e-03, grad_scale: 16.0 2023-04-26 22:41:11,688 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0382, 1.3711, 1.6383, 2.3589, 2.3961, 1.8539, 1.4671, 2.0227], device='cuda:4'), covar=tensor([0.0815, 0.1627, 0.0942, 0.0570, 0.0521, 0.0918, 0.1044, 0.0663], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0206, 0.0184, 0.0178, 0.0180, 0.0191, 0.0161, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 22:41:24,458 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-26 22:41:33,773 INFO [finetune.py:976] (4/7) Epoch 9, batch 2800, loss[loss=0.185, simple_loss=0.2503, pruned_loss=0.05985, over 4911.00 frames. ], tot_loss[loss=0.1904, simple_loss=0.2556, pruned_loss=0.06259, over 954373.65 frames. ], batch size: 32, lr: 3.79e-03, grad_scale: 16.0 2023-04-26 22:41:45,204 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0792, 2.8887, 2.3243, 2.5194, 1.9984, 2.3989, 2.4352, 1.9516], device='cuda:4'), covar=tensor([0.2220, 0.1400, 0.0899, 0.1352, 0.3128, 0.1443, 0.2020, 0.2691], device='cuda:4'), in_proj_covar=tensor([0.0297, 0.0318, 0.0229, 0.0290, 0.0314, 0.0272, 0.0258, 0.0282], device='cuda:4'), out_proj_covar=tensor([1.2018e-04, 1.2812e-04, 9.1977e-05, 1.1603e-04, 1.2911e-04, 1.0968e-04, 1.0583e-04, 1.1296e-04], device='cuda:4') 2023-04-26 22:41:46,303 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.102e+02 1.615e+02 1.895e+02 2.284e+02 5.384e+02, threshold=3.791e+02, percent-clipped=3.0 2023-04-26 22:41:55,126 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.61 vs. limit=2.0 2023-04-26 22:42:38,719 INFO [finetune.py:976] (4/7) Epoch 9, batch 2850, loss[loss=0.1814, simple_loss=0.24, pruned_loss=0.06139, over 4771.00 frames. ], tot_loss[loss=0.1902, simple_loss=0.2553, pruned_loss=0.0626, over 954959.47 frames. ], batch size: 26, lr: 3.79e-03, grad_scale: 16.0 2023-04-26 22:42:49,420 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=48680.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:43:03,805 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6875, 2.2538, 1.8306, 2.1344, 1.4980, 1.7857, 2.0089, 1.4991], device='cuda:4'), covar=tensor([0.2668, 0.1896, 0.1277, 0.1715, 0.3568, 0.1804, 0.2291, 0.3213], device='cuda:4'), in_proj_covar=tensor([0.0296, 0.0317, 0.0228, 0.0289, 0.0313, 0.0271, 0.0257, 0.0280], device='cuda:4'), out_proj_covar=tensor([1.2001e-04, 1.2788e-04, 9.1703e-05, 1.1576e-04, 1.2846e-04, 1.0940e-04, 1.0533e-04, 1.1245e-04], device='cuda:4') 2023-04-26 22:43:44,378 INFO [finetune.py:976] (4/7) Epoch 9, batch 2900, loss[loss=0.1745, simple_loss=0.2494, pruned_loss=0.04983, over 4759.00 frames. ], tot_loss[loss=0.1932, simple_loss=0.2584, pruned_loss=0.06399, over 954301.16 frames. ], batch size: 26, lr: 3.79e-03, grad_scale: 16.0 2023-04-26 22:43:48,146 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=48728.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:43:52,843 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.096e+02 1.707e+02 2.011e+02 2.462e+02 4.094e+02, threshold=4.022e+02, percent-clipped=3.0 2023-04-26 22:43:55,354 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0876, 1.6724, 5.4155, 5.0892, 4.7665, 5.2083, 4.6737, 4.7931], device='cuda:4'), covar=tensor([0.6849, 0.6592, 0.0875, 0.1637, 0.0992, 0.2155, 0.1055, 0.1284], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0311, 0.0411, 0.0414, 0.0353, 0.0408, 0.0318, 0.0373], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-26 22:44:04,272 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=48753.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:44:11,489 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.44 vs. limit=2.0 2023-04-26 22:44:17,828 INFO [finetune.py:976] (4/7) Epoch 9, batch 2950, loss[loss=0.2013, simple_loss=0.2742, pruned_loss=0.06422, over 4830.00 frames. ], tot_loss[loss=0.1947, simple_loss=0.2609, pruned_loss=0.06425, over 955123.51 frames. ], batch size: 33, lr: 3.79e-03, grad_scale: 16.0 2023-04-26 22:44:20,345 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=48776.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:44:22,222 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=48779.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:44:25,541 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.90 vs. limit=2.0 2023-04-26 22:44:52,431 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-26 22:45:04,150 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0322, 1.7620, 2.0270, 2.3489, 2.3160, 1.9816, 1.6863, 2.1059], device='cuda:4'), covar=tensor([0.0732, 0.1055, 0.0539, 0.0443, 0.0484, 0.0715, 0.0785, 0.0497], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0207, 0.0183, 0.0179, 0.0180, 0.0191, 0.0162, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 22:45:10,563 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=48821.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:45:11,036 INFO [finetune.py:976] (4/7) Epoch 9, batch 3000, loss[loss=0.183, simple_loss=0.2509, pruned_loss=0.05756, over 4817.00 frames. ], tot_loss[loss=0.1975, simple_loss=0.2636, pruned_loss=0.06568, over 954762.40 frames. ], batch size: 38, lr: 3.79e-03, grad_scale: 16.0 2023-04-26 22:45:11,036 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-26 22:45:27,359 INFO [finetune.py:1010] (4/7) Epoch 9, validation: loss=0.1543, simple_loss=0.2267, pruned_loss=0.04097, over 2265189.00 frames. 2023-04-26 22:45:27,359 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-26 22:45:46,032 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.141e+02 1.720e+02 1.968e+02 2.331e+02 3.766e+02, threshold=3.936e+02, percent-clipped=0.0 2023-04-26 22:45:47,978 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=48838.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:46:30,398 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=48868.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 22:46:32,696 INFO [finetune.py:976] (4/7) Epoch 9, batch 3050, loss[loss=0.1922, simple_loss=0.2528, pruned_loss=0.06575, over 4791.00 frames. ], tot_loss[loss=0.1967, simple_loss=0.2629, pruned_loss=0.06525, over 953830.21 frames. ], batch size: 54, lr: 3.79e-03, grad_scale: 16.0 2023-04-26 22:46:51,129 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=48882.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:46:53,557 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.7622, 1.8948, 1.6721, 1.3852, 1.8783, 1.4889, 2.4834, 1.3532], device='cuda:4'), covar=tensor([0.3540, 0.1567, 0.4725, 0.3075, 0.1840, 0.2719, 0.1287, 0.4857], device='cuda:4'), in_proj_covar=tensor([0.0349, 0.0353, 0.0435, 0.0364, 0.0394, 0.0387, 0.0384, 0.0425], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 22:47:12,133 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=48899.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:47:27,736 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6116, 1.2985, 1.9157, 1.8789, 1.3946, 1.1589, 1.5502, 1.1070], device='cuda:4'), covar=tensor([0.0596, 0.0866, 0.0478, 0.0673, 0.0800, 0.1464, 0.0719, 0.0887], device='cuda:4'), in_proj_covar=tensor([0.0066, 0.0073, 0.0072, 0.0067, 0.0076, 0.0096, 0.0078, 0.0074], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 22:47:28,838 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=48916.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 22:47:32,443 INFO [finetune.py:976] (4/7) Epoch 9, batch 3100, loss[loss=0.202, simple_loss=0.256, pruned_loss=0.07407, over 4763.00 frames. ], tot_loss[loss=0.1955, simple_loss=0.2612, pruned_loss=0.06489, over 953296.82 frames. ], batch size: 54, lr: 3.79e-03, grad_scale: 16.0 2023-04-26 22:47:35,987 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2024, 1.3176, 1.5739, 1.7224, 1.6735, 1.8185, 1.6450, 1.6751], device='cuda:4'), covar=tensor([0.4320, 0.6014, 0.5074, 0.4693, 0.5747, 0.8341, 0.5646, 0.5378], device='cuda:4'), in_proj_covar=tensor([0.0323, 0.0383, 0.0316, 0.0326, 0.0342, 0.0404, 0.0363, 0.0325], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 22:47:42,242 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.703e+02 2.023e+02 2.523e+02 5.306e+02, threshold=4.046e+02, percent-clipped=4.0 2023-04-26 22:48:05,673 INFO [finetune.py:976] (4/7) Epoch 9, batch 3150, loss[loss=0.175, simple_loss=0.2479, pruned_loss=0.05102, over 4895.00 frames. ], tot_loss[loss=0.1944, simple_loss=0.2594, pruned_loss=0.06468, over 955709.13 frames. ], batch size: 32, lr: 3.78e-03, grad_scale: 16.0 2023-04-26 22:48:11,546 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=48980.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:48:38,603 INFO [finetune.py:976] (4/7) Epoch 9, batch 3200, loss[loss=0.1514, simple_loss=0.2288, pruned_loss=0.03701, over 4786.00 frames. ], tot_loss[loss=0.1908, simple_loss=0.2554, pruned_loss=0.0631, over 953970.96 frames. ], batch size: 26, lr: 3.78e-03, grad_scale: 16.0 2023-04-26 22:48:42,805 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=49028.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:48:47,938 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.162e+02 1.546e+02 1.866e+02 2.430e+02 3.518e+02, threshold=3.733e+02, percent-clipped=0.0 2023-04-26 22:48:58,147 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6627, 1.2804, 1.7674, 2.1286, 1.7819, 1.6843, 1.7323, 1.7189], device='cuda:4'), covar=tensor([0.6192, 0.8393, 0.8042, 0.8401, 0.7360, 0.9448, 0.9757, 0.9574], device='cuda:4'), in_proj_covar=tensor([0.0410, 0.0421, 0.0503, 0.0523, 0.0438, 0.0457, 0.0469, 0.0465], device='cuda:4'), out_proj_covar=tensor([9.9651e-05, 1.0423e-04, 1.1362e-04, 1.2421e-04, 1.0634e-04, 1.1048e-04, 1.1275e-04, 1.1253e-04], device='cuda:4') 2023-04-26 22:48:59,907 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=49053.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:49:12,012 INFO [finetune.py:976] (4/7) Epoch 9, batch 3250, loss[loss=0.192, simple_loss=0.2593, pruned_loss=0.06232, over 4913.00 frames. ], tot_loss[loss=0.1927, simple_loss=0.257, pruned_loss=0.06418, over 953673.16 frames. ], batch size: 36, lr: 3.78e-03, grad_scale: 16.0 2023-04-26 22:49:15,057 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4791, 1.3529, 1.7999, 1.7401, 1.3363, 1.1564, 1.4374, 1.0123], device='cuda:4'), covar=tensor([0.0590, 0.0799, 0.0527, 0.0748, 0.0839, 0.1350, 0.0777, 0.0838], device='cuda:4'), in_proj_covar=tensor([0.0066, 0.0073, 0.0071, 0.0066, 0.0076, 0.0095, 0.0078, 0.0073], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 22:49:16,912 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=49079.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:49:32,077 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=49101.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:49:40,141 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.30 vs. limit=5.0 2023-04-26 22:49:45,282 INFO [finetune.py:976] (4/7) Epoch 9, batch 3300, loss[loss=0.2315, simple_loss=0.2967, pruned_loss=0.08314, over 4863.00 frames. ], tot_loss[loss=0.1971, simple_loss=0.262, pruned_loss=0.06606, over 953865.66 frames. ], batch size: 31, lr: 3.78e-03, grad_scale: 16.0 2023-04-26 22:49:48,901 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=49127.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:49:54,142 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.094e+02 1.776e+02 1.982e+02 2.505e+02 4.368e+02, threshold=3.963e+02, percent-clipped=2.0 2023-04-26 22:50:18,720 INFO [finetune.py:976] (4/7) Epoch 9, batch 3350, loss[loss=0.2118, simple_loss=0.2812, pruned_loss=0.07115, over 4897.00 frames. ], tot_loss[loss=0.1966, simple_loss=0.2625, pruned_loss=0.06535, over 954566.99 frames. ], batch size: 36, lr: 3.78e-03, grad_scale: 16.0 2023-04-26 22:50:21,875 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=49177.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:50:46,869 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=49194.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:51:20,616 INFO [finetune.py:976] (4/7) Epoch 9, batch 3400, loss[loss=0.1915, simple_loss=0.2683, pruned_loss=0.05729, over 4816.00 frames. ], tot_loss[loss=0.1976, simple_loss=0.2637, pruned_loss=0.06577, over 953228.68 frames. ], batch size: 33, lr: 3.78e-03, grad_scale: 16.0 2023-04-26 22:51:38,492 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.071e+02 1.560e+02 1.880e+02 2.304e+02 5.617e+02, threshold=3.759e+02, percent-clipped=2.0 2023-04-26 22:51:39,199 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=49236.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:52:04,330 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9053, 2.3698, 1.9218, 2.2092, 1.6327, 1.8349, 1.9264, 1.5236], device='cuda:4'), covar=tensor([0.1997, 0.1236, 0.1009, 0.1186, 0.3175, 0.1438, 0.1934, 0.2826], device='cuda:4'), in_proj_covar=tensor([0.0295, 0.0316, 0.0227, 0.0287, 0.0312, 0.0270, 0.0256, 0.0280], device='cuda:4'), out_proj_covar=tensor([1.1942e-04, 1.2715e-04, 9.1319e-05, 1.1507e-04, 1.2798e-04, 1.0904e-04, 1.0502e-04, 1.1248e-04], device='cuda:4') 2023-04-26 22:52:13,324 INFO [finetune.py:976] (4/7) Epoch 9, batch 3450, loss[loss=0.1872, simple_loss=0.2541, pruned_loss=0.06016, over 4867.00 frames. ], tot_loss[loss=0.1949, simple_loss=0.2617, pruned_loss=0.06405, over 954729.73 frames. ], batch size: 31, lr: 3.78e-03, grad_scale: 16.0 2023-04-26 22:52:30,761 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=49297.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:52:31,105 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.91 vs. limit=2.0 2023-04-26 22:53:02,578 INFO [finetune.py:976] (4/7) Epoch 9, batch 3500, loss[loss=0.1815, simple_loss=0.252, pruned_loss=0.05554, over 4817.00 frames. ], tot_loss[loss=0.1929, simple_loss=0.2591, pruned_loss=0.0633, over 956134.83 frames. ], batch size: 40, lr: 3.78e-03, grad_scale: 16.0 2023-04-26 22:53:15,826 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.247e+02 1.587e+02 1.893e+02 2.359e+02 3.521e+02, threshold=3.785e+02, percent-clipped=0.0 2023-04-26 22:53:59,436 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=2.00 vs. limit=2.0 2023-04-26 22:54:01,001 INFO [finetune.py:976] (4/7) Epoch 9, batch 3550, loss[loss=0.2063, simple_loss=0.2624, pruned_loss=0.0751, over 4820.00 frames. ], tot_loss[loss=0.1912, simple_loss=0.2568, pruned_loss=0.06283, over 956154.80 frames. ], batch size: 51, lr: 3.78e-03, grad_scale: 16.0 2023-04-26 22:55:07,550 INFO [finetune.py:976] (4/7) Epoch 9, batch 3600, loss[loss=0.2176, simple_loss=0.2795, pruned_loss=0.07783, over 4903.00 frames. ], tot_loss[loss=0.1918, simple_loss=0.2566, pruned_loss=0.06357, over 956930.04 frames. ], batch size: 36, lr: 3.78e-03, grad_scale: 16.0 2023-04-26 22:55:07,656 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=49422.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:55:24,186 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=49432.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 22:55:25,884 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.037e+02 1.649e+02 1.885e+02 2.480e+02 5.265e+02, threshold=3.771e+02, percent-clipped=3.0 2023-04-26 22:56:07,302 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1448, 1.5642, 1.3079, 1.7232, 1.5810, 1.9815, 1.3838, 3.5325], device='cuda:4'), covar=tensor([0.0666, 0.0794, 0.0846, 0.1178, 0.0667, 0.0552, 0.0790, 0.0167], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0043, 0.0040, 0.0039, 0.0039, 0.0058], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0012, 0.0013, 0.0016], device='cuda:4') 2023-04-26 22:56:18,972 INFO [finetune.py:976] (4/7) Epoch 9, batch 3650, loss[loss=0.1852, simple_loss=0.2638, pruned_loss=0.05333, over 4747.00 frames. ], tot_loss[loss=0.1932, simple_loss=0.2583, pruned_loss=0.06407, over 956650.26 frames. ], batch size: 54, lr: 3.78e-03, grad_scale: 16.0 2023-04-26 22:56:22,182 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=49477.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:56:30,064 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=49483.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:56:36,702 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=49493.0, num_to_drop=1, layers_to_drop={3} 2023-04-26 22:56:37,271 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=49494.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:56:56,124 INFO [finetune.py:976] (4/7) Epoch 9, batch 3700, loss[loss=0.1964, simple_loss=0.2616, pruned_loss=0.06556, over 4756.00 frames. ], tot_loss[loss=0.1955, simple_loss=0.2609, pruned_loss=0.06505, over 955675.15 frames. ], batch size: 27, lr: 3.78e-03, grad_scale: 16.0 2023-04-26 22:56:58,029 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=49525.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:57:09,401 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.074e+02 1.745e+02 1.976e+02 2.490e+02 4.358e+02, threshold=3.952e+02, percent-clipped=2.0 2023-04-26 22:57:13,695 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=49542.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:57:45,551 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.50 vs. limit=2.0 2023-04-26 22:57:56,799 INFO [finetune.py:976] (4/7) Epoch 9, batch 3750, loss[loss=0.1891, simple_loss=0.2597, pruned_loss=0.05922, over 4829.00 frames. ], tot_loss[loss=0.1969, simple_loss=0.2629, pruned_loss=0.06547, over 955799.08 frames. ], batch size: 30, lr: 3.78e-03, grad_scale: 16.0 2023-04-26 22:58:23,845 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=49592.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 22:58:44,872 INFO [finetune.py:976] (4/7) Epoch 9, batch 3800, loss[loss=0.1948, simple_loss=0.2634, pruned_loss=0.06311, over 4912.00 frames. ], tot_loss[loss=0.1972, simple_loss=0.2638, pruned_loss=0.06532, over 956381.89 frames. ], batch size: 33, lr: 3.78e-03, grad_scale: 16.0 2023-04-26 22:58:52,832 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.166e+02 1.711e+02 2.053e+02 2.525e+02 6.309e+02, threshold=4.105e+02, percent-clipped=4.0 2023-04-26 22:59:17,905 INFO [finetune.py:976] (4/7) Epoch 9, batch 3850, loss[loss=0.1836, simple_loss=0.2488, pruned_loss=0.05919, over 4810.00 frames. ], tot_loss[loss=0.194, simple_loss=0.2609, pruned_loss=0.06355, over 957347.66 frames. ], batch size: 33, lr: 3.78e-03, grad_scale: 16.0 2023-04-26 22:59:33,176 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8535, 1.0735, 1.4655, 1.6326, 1.5730, 1.7017, 1.5299, 1.5027], device='cuda:4'), covar=tensor([0.4696, 0.6311, 0.5368, 0.4936, 0.6130, 0.8536, 0.5777, 0.5593], device='cuda:4'), in_proj_covar=tensor([0.0322, 0.0381, 0.0314, 0.0325, 0.0341, 0.0401, 0.0361, 0.0322], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 22:59:42,749 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2994, 3.1704, 0.8765, 1.8539, 1.7550, 2.3876, 1.8559, 1.0006], device='cuda:4'), covar=tensor([0.1298, 0.0980, 0.1973, 0.1200, 0.1056, 0.0890, 0.1332, 0.1917], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0248, 0.0140, 0.0121, 0.0135, 0.0152, 0.0117, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 22:59:49,732 INFO [finetune.py:976] (4/7) Epoch 9, batch 3900, loss[loss=0.1774, simple_loss=0.2391, pruned_loss=0.05787, over 4911.00 frames. ], tot_loss[loss=0.1918, simple_loss=0.258, pruned_loss=0.06281, over 959526.10 frames. ], batch size: 35, lr: 3.78e-03, grad_scale: 32.0 2023-04-26 22:59:58,108 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.293e+02 1.654e+02 1.955e+02 2.414e+02 4.265e+02, threshold=3.910e+02, percent-clipped=1.0 2023-04-26 23:00:18,068 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=49766.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:00:21,569 INFO [finetune.py:976] (4/7) Epoch 9, batch 3950, loss[loss=0.1742, simple_loss=0.2317, pruned_loss=0.05842, over 4858.00 frames. ], tot_loss[loss=0.1902, simple_loss=0.2555, pruned_loss=0.0624, over 958518.45 frames. ], batch size: 44, lr: 3.78e-03, grad_scale: 32.0 2023-04-26 23:00:27,205 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=49778.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:00:33,308 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=49788.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 23:00:55,289 INFO [finetune.py:976] (4/7) Epoch 9, batch 4000, loss[loss=0.1736, simple_loss=0.2548, pruned_loss=0.04618, over 4933.00 frames. ], tot_loss[loss=0.1906, simple_loss=0.2553, pruned_loss=0.06291, over 957136.76 frames. ], batch size: 33, lr: 3.78e-03, grad_scale: 32.0 2023-04-26 23:00:59,441 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=49827.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:01:05,129 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.158e+02 1.602e+02 1.886e+02 2.314e+02 3.383e+02, threshold=3.771e+02, percent-clipped=0.0 2023-04-26 23:01:44,300 INFO [finetune.py:976] (4/7) Epoch 9, batch 4050, loss[loss=0.2099, simple_loss=0.2798, pruned_loss=0.06995, over 4845.00 frames. ], tot_loss[loss=0.1942, simple_loss=0.2596, pruned_loss=0.06438, over 956862.38 frames. ], batch size: 47, lr: 3.78e-03, grad_scale: 32.0 2023-04-26 23:02:14,669 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=49892.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:02:17,747 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.6915, 4.6395, 3.0059, 5.3983, 4.7497, 4.5999, 2.8487, 4.5512], device='cuda:4'), covar=tensor([0.1441, 0.0858, 0.3300, 0.0909, 0.2728, 0.1701, 0.4464, 0.2101], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0217, 0.0251, 0.0305, 0.0301, 0.0251, 0.0270, 0.0272], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 23:02:49,741 INFO [finetune.py:976] (4/7) Epoch 9, batch 4100, loss[loss=0.2056, simple_loss=0.2619, pruned_loss=0.07467, over 4817.00 frames. ], tot_loss[loss=0.1963, simple_loss=0.262, pruned_loss=0.06526, over 955655.64 frames. ], batch size: 33, lr: 3.78e-03, grad_scale: 32.0 2023-04-26 23:02:51,695 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.70 vs. limit=2.0 2023-04-26 23:03:10,346 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.345e+02 1.694e+02 2.037e+02 2.558e+02 4.844e+02, threshold=4.074e+02, percent-clipped=3.0 2023-04-26 23:03:19,227 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=49940.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:03:53,583 INFO [finetune.py:976] (4/7) Epoch 9, batch 4150, loss[loss=0.2269, simple_loss=0.2861, pruned_loss=0.08379, over 4819.00 frames. ], tot_loss[loss=0.1974, simple_loss=0.2637, pruned_loss=0.06554, over 955524.24 frames. ], batch size: 39, lr: 3.78e-03, grad_scale: 32.0 2023-04-26 23:04:33,308 INFO [finetune.py:976] (4/7) Epoch 9, batch 4200, loss[loss=0.1926, simple_loss=0.261, pruned_loss=0.06212, over 4756.00 frames. ], tot_loss[loss=0.1965, simple_loss=0.2633, pruned_loss=0.06485, over 955504.92 frames. ], batch size: 28, lr: 3.78e-03, grad_scale: 32.0 2023-04-26 23:04:39,102 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-04-26 23:04:41,638 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.155e+02 1.800e+02 2.150e+02 2.461e+02 7.133e+02, threshold=4.301e+02, percent-clipped=1.0 2023-04-26 23:05:05,670 INFO [finetune.py:976] (4/7) Epoch 9, batch 4250, loss[loss=0.2291, simple_loss=0.2862, pruned_loss=0.08598, over 4884.00 frames. ], tot_loss[loss=0.1941, simple_loss=0.2608, pruned_loss=0.06368, over 957269.92 frames. ], batch size: 32, lr: 3.78e-03, grad_scale: 32.0 2023-04-26 23:05:09,964 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=50078.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:05:16,028 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=50088.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 23:05:37,583 INFO [finetune.py:976] (4/7) Epoch 9, batch 4300, loss[loss=0.1541, simple_loss=0.2249, pruned_loss=0.04165, over 4772.00 frames. ], tot_loss[loss=0.1913, simple_loss=0.2578, pruned_loss=0.06238, over 958214.89 frames. ], batch size: 28, lr: 3.78e-03, grad_scale: 32.0 2023-04-26 23:05:37,652 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=50122.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:05:40,102 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=50126.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:05:41,356 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=50128.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:05:46,446 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.165e+02 1.729e+02 1.986e+02 2.496e+02 5.058e+02, threshold=3.971e+02, percent-clipped=3.0 2023-04-26 23:05:47,120 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=50136.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 23:06:10,816 INFO [finetune.py:976] (4/7) Epoch 9, batch 4350, loss[loss=0.2354, simple_loss=0.2927, pruned_loss=0.08905, over 4855.00 frames. ], tot_loss[loss=0.1888, simple_loss=0.2546, pruned_loss=0.06154, over 957876.33 frames. ], batch size: 44, lr: 3.78e-03, grad_scale: 32.0 2023-04-26 23:06:22,155 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=50189.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:06:27,934 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=50197.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 23:06:44,563 INFO [finetune.py:976] (4/7) Epoch 9, batch 4400, loss[loss=0.216, simple_loss=0.2803, pruned_loss=0.07588, over 4863.00 frames. ], tot_loss[loss=0.1899, simple_loss=0.2558, pruned_loss=0.06204, over 954708.47 frames. ], batch size: 34, lr: 3.78e-03, grad_scale: 32.0 2023-04-26 23:06:52,522 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.016e+02 1.620e+02 1.881e+02 2.281e+02 3.883e+02, threshold=3.762e+02, percent-clipped=0.0 2023-04-26 23:07:03,009 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9390, 1.3140, 1.1470, 1.5709, 1.4050, 1.3429, 1.2109, 2.1559], device='cuda:4'), covar=tensor([0.0568, 0.0716, 0.0732, 0.1086, 0.0579, 0.0493, 0.0681, 0.0233], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0038, 0.0040, 0.0043, 0.0040, 0.0038, 0.0038, 0.0058], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0012, 0.0013, 0.0016], device='cuda:4') 2023-04-26 23:07:13,982 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=50258.0, num_to_drop=1, layers_to_drop={3} 2023-04-26 23:07:27,674 INFO [finetune.py:976] (4/7) Epoch 9, batch 4450, loss[loss=0.2366, simple_loss=0.3071, pruned_loss=0.08311, over 4768.00 frames. ], tot_loss[loss=0.1944, simple_loss=0.2605, pruned_loss=0.0641, over 954923.48 frames. ], batch size: 59, lr: 3.77e-03, grad_scale: 32.0 2023-04-26 23:08:02,141 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1331, 0.7913, 0.9434, 0.7895, 1.2398, 0.9941, 0.8659, 0.9840], device='cuda:4'), covar=tensor([0.1728, 0.1484, 0.2252, 0.1599, 0.1084, 0.1398, 0.1847, 0.2249], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0324, 0.0356, 0.0299, 0.0336, 0.0323, 0.0308, 0.0359], device='cuda:4'), out_proj_covar=tensor([6.5504e-05, 6.8734e-05, 7.6787e-05, 6.1833e-05, 7.0424e-05, 6.9309e-05, 6.6275e-05, 7.6988e-05], device='cuda:4') 2023-04-26 23:08:33,131 INFO [finetune.py:976] (4/7) Epoch 9, batch 4500, loss[loss=0.2057, simple_loss=0.2767, pruned_loss=0.06738, over 4856.00 frames. ], tot_loss[loss=0.1947, simple_loss=0.2615, pruned_loss=0.06393, over 955766.31 frames. ], batch size: 44, lr: 3.77e-03, grad_scale: 32.0 2023-04-26 23:08:46,794 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.096e+02 1.697e+02 2.110e+02 2.509e+02 5.255e+02, threshold=4.219e+02, percent-clipped=3.0 2023-04-26 23:09:12,328 INFO [finetune.py:976] (4/7) Epoch 9, batch 4550, loss[loss=0.2166, simple_loss=0.2766, pruned_loss=0.07828, over 4811.00 frames. ], tot_loss[loss=0.195, simple_loss=0.2617, pruned_loss=0.06412, over 952278.40 frames. ], batch size: 39, lr: 3.77e-03, grad_scale: 32.0 2023-04-26 23:10:20,889 INFO [finetune.py:976] (4/7) Epoch 9, batch 4600, loss[loss=0.181, simple_loss=0.2432, pruned_loss=0.05934, over 4922.00 frames. ], tot_loss[loss=0.1951, simple_loss=0.2615, pruned_loss=0.06437, over 951433.35 frames. ], batch size: 33, lr: 3.77e-03, grad_scale: 32.0 2023-04-26 23:10:20,989 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=50422.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:10:21,635 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3356, 1.6106, 1.4511, 1.8871, 1.7873, 2.0305, 1.4127, 3.8444], device='cuda:4'), covar=tensor([0.0626, 0.0786, 0.0807, 0.1247, 0.0671, 0.0555, 0.0782, 0.0130], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0044, 0.0040, 0.0039, 0.0039, 0.0058], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-26 23:10:39,537 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.116e+02 1.674e+02 1.954e+02 2.272e+02 3.604e+02, threshold=3.908e+02, percent-clipped=0.0 2023-04-26 23:11:10,212 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7960, 2.5074, 1.7968, 1.7526, 1.3607, 1.3887, 1.8999, 1.3375], device='cuda:4'), covar=tensor([0.1612, 0.1295, 0.1393, 0.1753, 0.2346, 0.1866, 0.1003, 0.2017], device='cuda:4'), in_proj_covar=tensor([0.0192, 0.0215, 0.0170, 0.0204, 0.0204, 0.0182, 0.0160, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-26 23:11:14,981 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=50470.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:11:16,116 INFO [finetune.py:976] (4/7) Epoch 9, batch 4650, loss[loss=0.1475, simple_loss=0.2279, pruned_loss=0.03354, over 4779.00 frames. ], tot_loss[loss=0.1933, simple_loss=0.2595, pruned_loss=0.06357, over 954099.41 frames. ], batch size: 28, lr: 3.77e-03, grad_scale: 32.0 2023-04-26 23:11:23,421 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=50484.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:11:31,338 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.7121, 2.0339, 1.8823, 2.1139, 1.8707, 2.0797, 2.0054, 1.8776], device='cuda:4'), covar=tensor([0.5959, 0.8574, 0.7162, 0.6115, 0.7609, 1.0129, 0.8840, 0.8625], device='cuda:4'), in_proj_covar=tensor([0.0322, 0.0381, 0.0314, 0.0324, 0.0337, 0.0399, 0.0360, 0.0322], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 23:11:49,303 INFO [finetune.py:976] (4/7) Epoch 9, batch 4700, loss[loss=0.1969, simple_loss=0.2559, pruned_loss=0.0689, over 4899.00 frames. ], tot_loss[loss=0.191, simple_loss=0.2565, pruned_loss=0.06278, over 955146.80 frames. ], batch size: 35, lr: 3.77e-03, grad_scale: 32.0 2023-04-26 23:11:57,183 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.176e+02 1.640e+02 1.914e+02 2.343e+02 4.102e+02, threshold=3.828e+02, percent-clipped=1.0 2023-04-26 23:12:04,111 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=50546.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:12:08,329 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=50553.0, num_to_drop=1, layers_to_drop={3} 2023-04-26 23:12:22,116 INFO [finetune.py:976] (4/7) Epoch 9, batch 4750, loss[loss=0.1824, simple_loss=0.2613, pruned_loss=0.05175, over 4790.00 frames. ], tot_loss[loss=0.1895, simple_loss=0.2546, pruned_loss=0.06222, over 955860.07 frames. ], batch size: 45, lr: 3.77e-03, grad_scale: 32.0 2023-04-26 23:12:54,838 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=50607.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:13:16,244 INFO [finetune.py:976] (4/7) Epoch 9, batch 4800, loss[loss=0.1944, simple_loss=0.2728, pruned_loss=0.05797, over 4838.00 frames. ], tot_loss[loss=0.1929, simple_loss=0.2581, pruned_loss=0.06385, over 955969.54 frames. ], batch size: 49, lr: 3.77e-03, grad_scale: 32.0 2023-04-26 23:13:30,156 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.165e+02 1.707e+02 2.037e+02 2.400e+02 5.618e+02, threshold=4.074e+02, percent-clipped=3.0 2023-04-26 23:13:33,294 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3259, 2.4815, 1.9418, 2.0521, 2.3148, 1.8210, 3.3551, 1.6886], device='cuda:4'), covar=tensor([0.4606, 0.2234, 0.5123, 0.3971, 0.2495, 0.3200, 0.1506, 0.5009], device='cuda:4'), in_proj_covar=tensor([0.0343, 0.0348, 0.0430, 0.0359, 0.0387, 0.0381, 0.0379, 0.0417], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 23:13:55,575 INFO [finetune.py:976] (4/7) Epoch 9, batch 4850, loss[loss=0.1993, simple_loss=0.2576, pruned_loss=0.07045, over 4733.00 frames. ], tot_loss[loss=0.1941, simple_loss=0.2602, pruned_loss=0.06398, over 955707.48 frames. ], batch size: 23, lr: 3.77e-03, grad_scale: 32.0 2023-04-26 23:14:05,601 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.43 vs. limit=2.0 2023-04-26 23:14:28,054 INFO [finetune.py:976] (4/7) Epoch 9, batch 4900, loss[loss=0.226, simple_loss=0.283, pruned_loss=0.08449, over 4769.00 frames. ], tot_loss[loss=0.195, simple_loss=0.2614, pruned_loss=0.06434, over 955773.33 frames. ], batch size: 54, lr: 3.77e-03, grad_scale: 32.0 2023-04-26 23:14:36,901 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.087e+02 1.673e+02 1.942e+02 2.428e+02 3.700e+02, threshold=3.884e+02, percent-clipped=0.0 2023-04-26 23:14:40,662 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.64 vs. limit=2.0 2023-04-26 23:15:00,118 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4994, 1.4263, 1.8150, 1.8678, 1.4037, 1.1825, 1.5432, 0.9567], device='cuda:4'), covar=tensor([0.0778, 0.0874, 0.0561, 0.0770, 0.0962, 0.1422, 0.0813, 0.1045], device='cuda:4'), in_proj_covar=tensor([0.0066, 0.0073, 0.0071, 0.0067, 0.0076, 0.0096, 0.0078, 0.0074], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 23:15:14,088 INFO [finetune.py:976] (4/7) Epoch 9, batch 4950, loss[loss=0.2327, simple_loss=0.2959, pruned_loss=0.08471, over 4802.00 frames. ], tot_loss[loss=0.1978, simple_loss=0.2641, pruned_loss=0.06577, over 955852.32 frames. ], batch size: 40, lr: 3.77e-03, grad_scale: 32.0 2023-04-26 23:15:33,282 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=50782.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:15:34,472 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=50784.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:16:02,494 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=50808.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 23:16:11,334 INFO [finetune.py:976] (4/7) Epoch 9, batch 5000, loss[loss=0.1866, simple_loss=0.2481, pruned_loss=0.06254, over 4862.00 frames. ], tot_loss[loss=0.1956, simple_loss=0.2615, pruned_loss=0.06488, over 956216.09 frames. ], batch size: 34, lr: 3.77e-03, grad_scale: 32.0 2023-04-26 23:16:19,877 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=50832.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:16:21,616 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.072e+02 1.711e+02 2.099e+02 2.479e+02 5.783e+02, threshold=4.198e+02, percent-clipped=3.0 2023-04-26 23:16:26,683 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=50843.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:16:32,731 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=50853.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 23:16:36,950 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2506, 1.3342, 3.8428, 3.5702, 3.3986, 3.5871, 3.6521, 3.4171], device='cuda:4'), covar=tensor([0.6740, 0.5572, 0.1221, 0.1800, 0.1133, 0.1484, 0.1325, 0.1510], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0306, 0.0402, 0.0409, 0.0347, 0.0403, 0.0312, 0.0369], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-26 23:16:42,423 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=50869.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 23:16:44,512 INFO [finetune.py:976] (4/7) Epoch 9, batch 5050, loss[loss=0.1993, simple_loss=0.2618, pruned_loss=0.0684, over 4826.00 frames. ], tot_loss[loss=0.1931, simple_loss=0.2584, pruned_loss=0.06395, over 955242.73 frames. ], batch size: 39, lr: 3.77e-03, grad_scale: 32.0 2023-04-26 23:16:53,903 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4662, 2.9135, 0.8412, 1.5301, 2.0273, 1.7534, 4.1927, 2.0939], device='cuda:4'), covar=tensor([0.0659, 0.1063, 0.1040, 0.1402, 0.0639, 0.0966, 0.0363, 0.0668], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0049, 0.0047, 0.0051, 0.0052, 0.0078, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0007, 0.0008, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-26 23:17:04,644 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=50901.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 23:17:05,267 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=50902.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:17:13,167 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4257, 2.0833, 2.0017, 2.3634, 2.4404, 2.2191, 1.8913, 4.7637], device='cuda:4'), covar=tensor([0.0618, 0.0713, 0.0735, 0.1095, 0.0552, 0.0522, 0.0671, 0.0109], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0044, 0.0040, 0.0039, 0.0039, 0.0058], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0012, 0.0013, 0.0016], device='cuda:4') 2023-04-26 23:17:17,317 INFO [finetune.py:976] (4/7) Epoch 9, batch 5100, loss[loss=0.2176, simple_loss=0.2715, pruned_loss=0.0819, over 4759.00 frames. ], tot_loss[loss=0.1878, simple_loss=0.2532, pruned_loss=0.06121, over 956039.93 frames. ], batch size: 27, lr: 3.77e-03, grad_scale: 32.0 2023-04-26 23:17:26,163 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.072e+02 1.643e+02 1.891e+02 2.439e+02 4.473e+02, threshold=3.781e+02, percent-clipped=2.0 2023-04-26 23:17:34,473 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=50946.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:17:50,267 INFO [finetune.py:976] (4/7) Epoch 9, batch 5150, loss[loss=0.2196, simple_loss=0.2945, pruned_loss=0.07236, over 4843.00 frames. ], tot_loss[loss=0.1891, simple_loss=0.2541, pruned_loss=0.06202, over 954995.04 frames. ], batch size: 49, lr: 3.77e-03, grad_scale: 32.0 2023-04-26 23:18:06,069 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.43 vs. limit=2.0 2023-04-26 23:18:06,942 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.69 vs. limit=2.0 2023-04-26 23:18:25,494 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=51007.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:18:40,086 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0312, 1.6895, 1.9813, 2.3271, 2.3946, 1.8385, 1.5037, 1.9610], device='cuda:4'), covar=tensor([0.0960, 0.1294, 0.0732, 0.0673, 0.0677, 0.0973, 0.1087, 0.0739], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0207, 0.0183, 0.0177, 0.0179, 0.0190, 0.0161, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 23:18:40,601 INFO [finetune.py:976] (4/7) Epoch 9, batch 5200, loss[loss=0.2333, simple_loss=0.2993, pruned_loss=0.08365, over 4898.00 frames. ], tot_loss[loss=0.1925, simple_loss=0.2586, pruned_loss=0.06323, over 954945.92 frames. ], batch size: 43, lr: 3.77e-03, grad_scale: 32.0 2023-04-26 23:18:49,052 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.109e+02 1.700e+02 2.036e+02 2.415e+02 4.035e+02, threshold=4.072e+02, percent-clipped=2.0 2023-04-26 23:18:58,203 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-26 23:19:08,583 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=51062.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:19:14,486 INFO [finetune.py:976] (4/7) Epoch 9, batch 5250, loss[loss=0.1812, simple_loss=0.2559, pruned_loss=0.05331, over 4849.00 frames. ], tot_loss[loss=0.1946, simple_loss=0.261, pruned_loss=0.06407, over 954974.92 frames. ], batch size: 44, lr: 3.77e-03, grad_scale: 32.0 2023-04-26 23:19:38,148 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8954, 1.3970, 1.4165, 1.6506, 2.1095, 1.7044, 1.4180, 1.3642], device='cuda:4'), covar=tensor([0.1760, 0.1915, 0.1972, 0.1351, 0.0852, 0.1713, 0.2456, 0.2213], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0322, 0.0354, 0.0296, 0.0334, 0.0320, 0.0305, 0.0356], device='cuda:4'), out_proj_covar=tensor([6.4890e-05, 6.8304e-05, 7.6280e-05, 6.0916e-05, 6.9957e-05, 6.8635e-05, 6.5536e-05, 7.6383e-05], device='cuda:4') 2023-04-26 23:19:47,750 INFO [finetune.py:976] (4/7) Epoch 9, batch 5300, loss[loss=0.1555, simple_loss=0.2277, pruned_loss=0.04171, over 4857.00 frames. ], tot_loss[loss=0.195, simple_loss=0.2617, pruned_loss=0.06413, over 954419.40 frames. ], batch size: 31, lr: 3.77e-03, grad_scale: 32.0 2023-04-26 23:19:48,490 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=51123.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:19:50,391 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.90 vs. limit=5.0 2023-04-26 23:19:55,601 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8263, 2.7129, 2.0276, 1.8821, 1.3395, 1.3618, 2.0340, 1.3333], device='cuda:4'), covar=tensor([0.1791, 0.1514, 0.1485, 0.1972, 0.2493, 0.2081, 0.1079, 0.2154], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0217, 0.0171, 0.0205, 0.0205, 0.0184, 0.0161, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-26 23:19:56,068 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.143e+02 1.811e+02 2.045e+02 2.583e+02 4.950e+02, threshold=4.090e+02, percent-clipped=1.0 2023-04-26 23:19:57,961 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=51138.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:20:14,055 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0224, 2.3601, 1.0593, 1.3482, 1.7225, 1.1726, 3.2986, 1.6195], device='cuda:4'), covar=tensor([0.0671, 0.0716, 0.0794, 0.1213, 0.0528, 0.1037, 0.0234, 0.0612], device='cuda:4'), in_proj_covar=tensor([0.0052, 0.0067, 0.0049, 0.0047, 0.0051, 0.0053, 0.0078, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-26 23:20:15,867 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=51164.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 23:20:20,663 INFO [finetune.py:976] (4/7) Epoch 9, batch 5350, loss[loss=0.1874, simple_loss=0.2626, pruned_loss=0.05614, over 4889.00 frames. ], tot_loss[loss=0.1955, simple_loss=0.2621, pruned_loss=0.06441, over 953470.08 frames. ], batch size: 35, lr: 3.77e-03, grad_scale: 32.0 2023-04-26 23:20:20,761 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=51172.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 23:20:31,227 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.91 vs. limit=2.0 2023-04-26 23:20:46,715 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=51202.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:21:10,255 INFO [finetune.py:976] (4/7) Epoch 9, batch 5400, loss[loss=0.1677, simple_loss=0.232, pruned_loss=0.05171, over 4930.00 frames. ], tot_loss[loss=0.1935, simple_loss=0.2596, pruned_loss=0.06373, over 955946.58 frames. ], batch size: 38, lr: 3.77e-03, grad_scale: 32.0 2023-04-26 23:21:21,603 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=51233.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 23:21:22,681 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.244e+02 1.662e+02 1.922e+02 2.270e+02 4.708e+02, threshold=3.844e+02, percent-clipped=3.0 2023-04-26 23:21:44,189 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=51250.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:22:14,586 INFO [finetune.py:976] (4/7) Epoch 9, batch 5450, loss[loss=0.1657, simple_loss=0.2362, pruned_loss=0.04758, over 4819.00 frames. ], tot_loss[loss=0.1902, simple_loss=0.2557, pruned_loss=0.0623, over 956120.76 frames. ], batch size: 38, lr: 3.77e-03, grad_scale: 32.0 2023-04-26 23:22:47,577 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=51302.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:23:19,202 INFO [finetune.py:976] (4/7) Epoch 9, batch 5500, loss[loss=0.1426, simple_loss=0.2156, pruned_loss=0.03481, over 4784.00 frames. ], tot_loss[loss=0.1877, simple_loss=0.2527, pruned_loss=0.06131, over 956074.91 frames. ], batch size: 29, lr: 3.77e-03, grad_scale: 32.0 2023-04-26 23:23:32,910 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.202e+01 1.588e+02 1.902e+02 2.243e+02 3.887e+02, threshold=3.804e+02, percent-clipped=1.0 2023-04-26 23:23:33,076 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3077, 1.5519, 1.7220, 1.8582, 1.6963, 1.7558, 1.8336, 1.7669], device='cuda:4'), covar=tensor([0.5232, 0.7051, 0.5442, 0.5344, 0.6742, 0.9388, 0.6663, 0.6135], device='cuda:4'), in_proj_covar=tensor([0.0325, 0.0385, 0.0317, 0.0327, 0.0339, 0.0403, 0.0361, 0.0325], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 23:24:14,746 INFO [finetune.py:976] (4/7) Epoch 9, batch 5550, loss[loss=0.2209, simple_loss=0.2843, pruned_loss=0.0788, over 4921.00 frames. ], tot_loss[loss=0.1907, simple_loss=0.2561, pruned_loss=0.06265, over 954114.96 frames. ], batch size: 38, lr: 3.77e-03, grad_scale: 32.0 2023-04-26 23:24:20,971 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=51382.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:24:26,985 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=51392.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:24:40,204 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.92 vs. limit=5.0 2023-04-26 23:24:42,935 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=51418.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:24:45,261 INFO [finetune.py:976] (4/7) Epoch 9, batch 5600, loss[loss=0.1904, simple_loss=0.2328, pruned_loss=0.074, over 4223.00 frames. ], tot_loss[loss=0.1948, simple_loss=0.2609, pruned_loss=0.06435, over 954388.13 frames. ], batch size: 18, lr: 3.77e-03, grad_scale: 32.0 2023-04-26 23:24:52,679 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.274e+02 1.747e+02 2.120e+02 2.551e+02 6.497e+02, threshold=4.239e+02, percent-clipped=3.0 2023-04-26 23:24:54,525 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=51438.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:24:57,454 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=51443.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:25:03,253 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=51453.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:25:10,023 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=51464.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 23:25:15,160 INFO [finetune.py:976] (4/7) Epoch 9, batch 5650, loss[loss=0.1724, simple_loss=0.2408, pruned_loss=0.05198, over 4777.00 frames. ], tot_loss[loss=0.1967, simple_loss=0.2633, pruned_loss=0.06506, over 951393.74 frames. ], batch size: 26, lr: 3.77e-03, grad_scale: 32.0 2023-04-26 23:25:23,729 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=51486.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:25:29,052 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=51495.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:25:39,096 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=51512.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 23:25:42,831 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3377, 1.6922, 1.8281, 1.9456, 1.8206, 2.0178, 1.9405, 1.8675], device='cuda:4'), covar=tensor([0.4744, 0.6681, 0.5517, 0.5347, 0.6541, 0.8191, 0.6236, 0.6077], device='cuda:4'), in_proj_covar=tensor([0.0325, 0.0384, 0.0317, 0.0327, 0.0339, 0.0402, 0.0360, 0.0325], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 23:25:45,278 INFO [finetune.py:976] (4/7) Epoch 9, batch 5700, loss[loss=0.146, simple_loss=0.2027, pruned_loss=0.04465, over 4319.00 frames. ], tot_loss[loss=0.1933, simple_loss=0.2581, pruned_loss=0.06421, over 930331.13 frames. ], batch size: 18, lr: 3.77e-03, grad_scale: 32.0 2023-04-26 23:25:48,917 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=51528.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 23:25:53,106 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.028e+02 1.631e+02 1.982e+02 2.330e+02 4.156e+02, threshold=3.963e+02, percent-clipped=0.0 2023-04-26 23:26:16,105 INFO [finetune.py:976] (4/7) Epoch 10, batch 0, loss[loss=0.1681, simple_loss=0.2327, pruned_loss=0.05176, over 4832.00 frames. ], tot_loss[loss=0.1681, simple_loss=0.2327, pruned_loss=0.05176, over 4832.00 frames. ], batch size: 49, lr: 3.76e-03, grad_scale: 32.0 2023-04-26 23:26:16,105 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-26 23:26:31,820 INFO [finetune.py:1010] (4/7) Epoch 10, validation: loss=0.1558, simple_loss=0.2282, pruned_loss=0.04164, over 2265189.00 frames. 2023-04-26 23:26:31,821 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-26 23:26:37,656 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=51556.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:26:41,607 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.49 vs. limit=2.0 2023-04-26 23:26:52,866 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4100, 1.7402, 2.2138, 2.8500, 2.1942, 1.7216, 1.6392, 2.0087], device='cuda:4'), covar=tensor([0.3694, 0.4048, 0.1743, 0.3115, 0.3334, 0.2962, 0.4978, 0.2812], device='cuda:4'), in_proj_covar=tensor([0.0284, 0.0251, 0.0220, 0.0318, 0.0214, 0.0228, 0.0234, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-26 23:26:53,516 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-26 23:27:05,703 INFO [finetune.py:976] (4/7) Epoch 10, batch 50, loss[loss=0.1679, simple_loss=0.2442, pruned_loss=0.04576, over 4806.00 frames. ], tot_loss[loss=0.195, simple_loss=0.2614, pruned_loss=0.06433, over 215135.32 frames. ], batch size: 39, lr: 3.76e-03, grad_scale: 32.0 2023-04-26 23:27:08,591 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=51602.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:27:31,937 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.265e+02 1.709e+02 2.017e+02 2.493e+02 1.011e+03, threshold=4.035e+02, percent-clipped=6.0 2023-04-26 23:27:45,651 INFO [finetune.py:976] (4/7) Epoch 10, batch 100, loss[loss=0.1971, simple_loss=0.262, pruned_loss=0.06612, over 4827.00 frames. ], tot_loss[loss=0.1909, simple_loss=0.2556, pruned_loss=0.06312, over 379119.60 frames. ], batch size: 38, lr: 3.76e-03, grad_scale: 16.0 2023-04-26 23:27:47,277 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=51650.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:28:38,563 INFO [finetune.py:976] (4/7) Epoch 10, batch 150, loss[loss=0.1901, simple_loss=0.2501, pruned_loss=0.06501, over 4799.00 frames. ], tot_loss[loss=0.189, simple_loss=0.253, pruned_loss=0.06248, over 509229.75 frames. ], batch size: 45, lr: 3.76e-03, grad_scale: 16.0 2023-04-26 23:29:04,456 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=51718.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:29:20,964 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.275e+02 1.676e+02 2.023e+02 2.458e+02 4.768e+02, threshold=4.046e+02, percent-clipped=1.0 2023-04-26 23:29:22,217 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=51738.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:29:29,231 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=51747.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:29:29,809 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=51748.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:29:30,343 INFO [finetune.py:976] (4/7) Epoch 10, batch 200, loss[loss=0.2635, simple_loss=0.3022, pruned_loss=0.1124, over 3903.00 frames. ], tot_loss[loss=0.1875, simple_loss=0.2507, pruned_loss=0.06214, over 606865.73 frames. ], batch size: 65, lr: 3.76e-03, grad_scale: 16.0 2023-04-26 23:29:42,682 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=51766.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:30:04,063 INFO [finetune.py:976] (4/7) Epoch 10, batch 250, loss[loss=0.1858, simple_loss=0.2605, pruned_loss=0.05556, over 4755.00 frames. ], tot_loss[loss=0.1897, simple_loss=0.254, pruned_loss=0.06276, over 684945.60 frames. ], batch size: 27, lr: 3.76e-03, grad_scale: 16.0 2023-04-26 23:30:11,187 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=51808.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:30:23,868 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=51828.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 23:30:28,669 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.252e+02 1.711e+02 1.990e+02 2.532e+02 5.377e+02, threshold=3.981e+02, percent-clipped=2.0 2023-04-26 23:30:37,591 INFO [finetune.py:976] (4/7) Epoch 10, batch 300, loss[loss=0.1934, simple_loss=0.2751, pruned_loss=0.05587, over 4807.00 frames. ], tot_loss[loss=0.1913, simple_loss=0.2567, pruned_loss=0.0629, over 745719.50 frames. ], batch size: 51, lr: 3.76e-03, grad_scale: 16.0 2023-04-26 23:30:39,366 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=51851.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:30:44,651 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8264, 1.6068, 2.0531, 2.3578, 1.9294, 1.7826, 1.9044, 1.8944], device='cuda:4'), covar=tensor([0.6647, 0.9630, 1.0035, 0.8449, 0.8053, 1.2015, 1.2242, 1.1142], device='cuda:4'), in_proj_covar=tensor([0.0409, 0.0416, 0.0502, 0.0520, 0.0437, 0.0456, 0.0467, 0.0465], device='cuda:4'), out_proj_covar=tensor([9.9536e-05, 1.0319e-04, 1.1320e-04, 1.2373e-04, 1.0616e-04, 1.1038e-04, 1.1219e-04, 1.1223e-04], device='cuda:4') 2023-04-26 23:30:56,480 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=51876.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 23:31:10,834 INFO [finetune.py:976] (4/7) Epoch 10, batch 350, loss[loss=0.2383, simple_loss=0.2868, pruned_loss=0.09491, over 4806.00 frames. ], tot_loss[loss=0.1942, simple_loss=0.2605, pruned_loss=0.06396, over 792908.89 frames. ], batch size: 25, lr: 3.76e-03, grad_scale: 16.0 2023-04-26 23:31:41,425 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.954e+01 1.688e+02 1.992e+02 2.453e+02 5.822e+02, threshold=3.984e+02, percent-clipped=3.0 2023-04-26 23:32:00,718 INFO [finetune.py:976] (4/7) Epoch 10, batch 400, loss[loss=0.1828, simple_loss=0.2538, pruned_loss=0.05588, over 4837.00 frames. ], tot_loss[loss=0.1941, simple_loss=0.2609, pruned_loss=0.06359, over 827516.84 frames. ], batch size: 49, lr: 3.76e-03, grad_scale: 16.0 2023-04-26 23:32:51,606 INFO [finetune.py:976] (4/7) Epoch 10, batch 450, loss[loss=0.2073, simple_loss=0.2731, pruned_loss=0.07077, over 4895.00 frames. ], tot_loss[loss=0.1923, simple_loss=0.259, pruned_loss=0.06281, over 855222.44 frames. ], batch size: 35, lr: 3.76e-03, grad_scale: 16.0 2023-04-26 23:33:45,355 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.213e+02 1.582e+02 1.952e+02 2.304e+02 4.077e+02, threshold=3.905e+02, percent-clipped=1.0 2023-04-26 23:33:46,696 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=52038.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:33:53,300 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=52040.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:33:58,080 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=52048.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:34:04,123 INFO [finetune.py:976] (4/7) Epoch 10, batch 500, loss[loss=0.1797, simple_loss=0.2461, pruned_loss=0.05671, over 4935.00 frames. ], tot_loss[loss=0.1901, simple_loss=0.2564, pruned_loss=0.06193, over 879361.97 frames. ], batch size: 33, lr: 3.76e-03, grad_scale: 16.0 2023-04-26 23:34:25,673 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.73 vs. limit=5.0 2023-04-26 23:34:51,213 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9631, 1.4610, 1.7565, 1.6388, 1.7091, 1.4296, 0.7962, 1.4062], device='cuda:4'), covar=tensor([0.3822, 0.4136, 0.1939, 0.2829, 0.3111, 0.3120, 0.4777, 0.2654], device='cuda:4'), in_proj_covar=tensor([0.0283, 0.0251, 0.0220, 0.0319, 0.0213, 0.0227, 0.0234, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-26 23:34:51,736 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=52086.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:35:09,318 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=52096.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:35:11,184 INFO [finetune.py:976] (4/7) Epoch 10, batch 550, loss[loss=0.1478, simple_loss=0.2058, pruned_loss=0.04491, over 4439.00 frames. ], tot_loss[loss=0.1883, simple_loss=0.2538, pruned_loss=0.06139, over 897379.22 frames. ], batch size: 19, lr: 3.76e-03, grad_scale: 16.0 2023-04-26 23:35:12,521 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=52101.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:35:13,634 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=52103.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:36:05,086 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.236e+02 1.854e+02 2.167e+02 2.503e+02 5.657e+02, threshold=4.334e+02, percent-clipped=4.0 2023-04-26 23:36:07,113 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6218, 2.1683, 1.7135, 1.5218, 1.2266, 1.2241, 1.7500, 1.2104], device='cuda:4'), covar=tensor([0.1672, 0.1433, 0.1638, 0.1890, 0.2488, 0.2115, 0.1060, 0.2138], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0216, 0.0171, 0.0205, 0.0205, 0.0184, 0.0161, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-26 23:36:18,686 INFO [finetune.py:976] (4/7) Epoch 10, batch 600, loss[loss=0.2018, simple_loss=0.263, pruned_loss=0.07028, over 4825.00 frames. ], tot_loss[loss=0.1904, simple_loss=0.2555, pruned_loss=0.06265, over 910956.53 frames. ], batch size: 33, lr: 3.76e-03, grad_scale: 16.0 2023-04-26 23:36:19,992 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=52151.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:37:10,897 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6864, 1.8823, 1.8317, 1.9850, 1.7528, 1.8979, 1.9498, 1.8394], device='cuda:4'), covar=tensor([0.5270, 0.8467, 0.6665, 0.5803, 0.7364, 0.9877, 0.7859, 0.7478], device='cuda:4'), in_proj_covar=tensor([0.0326, 0.0384, 0.0317, 0.0327, 0.0340, 0.0404, 0.0361, 0.0325], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 23:37:12,642 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6501, 1.6754, 0.9732, 1.2985, 1.8085, 1.4999, 1.4102, 1.4468], device='cuda:4'), covar=tensor([0.0540, 0.0413, 0.0375, 0.0626, 0.0287, 0.0582, 0.0575, 0.0601], device='cuda:4'), in_proj_covar=tensor([0.0030, 0.0025, 0.0023, 0.0030, 0.0021, 0.0029, 0.0029, 0.0030], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0050, 0.0038, 0.0048, 0.0048, 0.0050], device='cuda:4') 2023-04-26 23:37:16,284 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=52193.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:37:25,226 INFO [finetune.py:976] (4/7) Epoch 10, batch 650, loss[loss=0.267, simple_loss=0.3132, pruned_loss=0.1104, over 4103.00 frames. ], tot_loss[loss=0.1926, simple_loss=0.258, pruned_loss=0.06355, over 919760.47 frames. ], batch size: 65, lr: 3.76e-03, grad_scale: 16.0 2023-04-26 23:37:25,288 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=52199.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:38:17,625 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.057e+02 1.623e+02 1.914e+02 2.334e+02 8.066e+02, threshold=3.828e+02, percent-clipped=2.0 2023-04-26 23:38:30,841 INFO [finetune.py:976] (4/7) Epoch 10, batch 700, loss[loss=0.2428, simple_loss=0.3019, pruned_loss=0.09188, over 4855.00 frames. ], tot_loss[loss=0.1941, simple_loss=0.2597, pruned_loss=0.06425, over 927524.20 frames. ], batch size: 31, lr: 3.76e-03, grad_scale: 16.0 2023-04-26 23:38:39,789 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=52254.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:39:44,471 INFO [finetune.py:976] (4/7) Epoch 10, batch 750, loss[loss=0.1726, simple_loss=0.2501, pruned_loss=0.04753, over 4809.00 frames. ], tot_loss[loss=0.1946, simple_loss=0.2605, pruned_loss=0.06433, over 930980.84 frames. ], batch size: 39, lr: 3.76e-03, grad_scale: 16.0 2023-04-26 23:39:54,940 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0899, 2.0088, 1.6995, 1.7815, 2.0982, 1.7292, 2.5208, 1.4621], device='cuda:4'), covar=tensor([0.3860, 0.1882, 0.4841, 0.3178, 0.1904, 0.2645, 0.1750, 0.4920], device='cuda:4'), in_proj_covar=tensor([0.0347, 0.0351, 0.0431, 0.0365, 0.0391, 0.0387, 0.0383, 0.0422], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 23:39:56,801 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=52312.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:40:13,560 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.225e+02 1.731e+02 1.996e+02 2.532e+02 4.524e+02, threshold=3.992e+02, percent-clipped=2.0 2023-04-26 23:40:22,480 INFO [finetune.py:976] (4/7) Epoch 10, batch 800, loss[loss=0.1339, simple_loss=0.2117, pruned_loss=0.028, over 4813.00 frames. ], tot_loss[loss=0.1956, simple_loss=0.2614, pruned_loss=0.06486, over 935991.46 frames. ], batch size: 33, lr: 3.76e-03, grad_scale: 16.0 2023-04-26 23:40:25,695 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6139, 1.4075, 0.4749, 1.2731, 1.4253, 1.4620, 1.4091, 1.3817], device='cuda:4'), covar=tensor([0.0522, 0.0412, 0.0424, 0.0576, 0.0299, 0.0520, 0.0529, 0.0591], device='cuda:4'), in_proj_covar=tensor([0.0030, 0.0025, 0.0023, 0.0030, 0.0021, 0.0029, 0.0029, 0.0030], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0050, 0.0038, 0.0048, 0.0048, 0.0050], device='cuda:4') 2023-04-26 23:40:30,292 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.19 vs. limit=5.0 2023-04-26 23:40:37,979 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=52373.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:40:54,332 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=52396.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:40:56,135 INFO [finetune.py:976] (4/7) Epoch 10, batch 850, loss[loss=0.151, simple_loss=0.2196, pruned_loss=0.04115, over 4819.00 frames. ], tot_loss[loss=0.1923, simple_loss=0.258, pruned_loss=0.06332, over 940197.40 frames. ], batch size: 41, lr: 3.76e-03, grad_scale: 16.0 2023-04-26 23:40:58,773 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=52403.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:41:30,882 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.154e+02 1.587e+02 1.831e+02 2.127e+02 6.660e+02, threshold=3.662e+02, percent-clipped=3.0 2023-04-26 23:41:44,733 INFO [finetune.py:976] (4/7) Epoch 10, batch 900, loss[loss=0.1926, simple_loss=0.2545, pruned_loss=0.06538, over 4819.00 frames. ], tot_loss[loss=0.1894, simple_loss=0.2549, pruned_loss=0.06199, over 943048.12 frames. ], batch size: 39, lr: 3.76e-03, grad_scale: 16.0 2023-04-26 23:41:51,263 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=52451.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:41:52,539 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3762, 3.2715, 0.9341, 1.7209, 1.8753, 2.2665, 1.8516, 1.0012], device='cuda:4'), covar=tensor([0.1295, 0.0945, 0.1912, 0.1266, 0.0981, 0.0975, 0.1492, 0.1878], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0250, 0.0141, 0.0122, 0.0136, 0.0154, 0.0119, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-26 23:42:49,243 INFO [finetune.py:976] (4/7) Epoch 10, batch 950, loss[loss=0.1656, simple_loss=0.2287, pruned_loss=0.05122, over 4764.00 frames. ], tot_loss[loss=0.1881, simple_loss=0.2535, pruned_loss=0.06139, over 946019.48 frames. ], batch size: 26, lr: 3.76e-03, grad_scale: 16.0 2023-04-26 23:43:41,296 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.638e+02 1.935e+02 2.436e+02 4.596e+02, threshold=3.871e+02, percent-clipped=3.0 2023-04-26 23:43:55,595 INFO [finetune.py:976] (4/7) Epoch 10, batch 1000, loss[loss=0.1972, simple_loss=0.2593, pruned_loss=0.06754, over 4784.00 frames. ], tot_loss[loss=0.1911, simple_loss=0.2564, pruned_loss=0.06289, over 948442.02 frames. ], batch size: 28, lr: 3.76e-03, grad_scale: 16.0 2023-04-26 23:44:01,404 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=52549.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:44:05,062 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=52555.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 23:45:07,087 INFO [finetune.py:976] (4/7) Epoch 10, batch 1050, loss[loss=0.1691, simple_loss=0.2403, pruned_loss=0.04891, over 4832.00 frames. ], tot_loss[loss=0.1939, simple_loss=0.26, pruned_loss=0.06387, over 951702.65 frames. ], batch size: 30, lr: 3.76e-03, grad_scale: 16.0 2023-04-26 23:45:11,805 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.87 vs. limit=5.0 2023-04-26 23:45:29,125 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=52616.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 23:45:52,968 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.267e+02 1.739e+02 2.075e+02 2.346e+02 3.891e+02, threshold=4.149e+02, percent-clipped=1.0 2023-04-26 23:46:13,536 INFO [finetune.py:976] (4/7) Epoch 10, batch 1100, loss[loss=0.1741, simple_loss=0.249, pruned_loss=0.04958, over 4795.00 frames. ], tot_loss[loss=0.1941, simple_loss=0.2607, pruned_loss=0.06381, over 951364.48 frames. ], batch size: 51, lr: 3.76e-03, grad_scale: 16.0 2023-04-26 23:46:36,536 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=52668.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:46:51,321 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.75 vs. limit=2.0 2023-04-26 23:46:56,119 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=52696.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:46:57,845 INFO [finetune.py:976] (4/7) Epoch 10, batch 1150, loss[loss=0.1514, simple_loss=0.2302, pruned_loss=0.03632, over 4744.00 frames. ], tot_loss[loss=0.1947, simple_loss=0.2615, pruned_loss=0.06399, over 953181.05 frames. ], batch size: 27, lr: 3.75e-03, grad_scale: 16.0 2023-04-26 23:47:21,727 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.059e+02 1.582e+02 1.906e+02 2.303e+02 4.186e+02, threshold=3.812e+02, percent-clipped=1.0 2023-04-26 23:47:33,672 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=52744.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:47:43,259 INFO [finetune.py:976] (4/7) Epoch 10, batch 1200, loss[loss=0.1805, simple_loss=0.2439, pruned_loss=0.05855, over 4906.00 frames. ], tot_loss[loss=0.192, simple_loss=0.2586, pruned_loss=0.06271, over 953919.70 frames. ], batch size: 36, lr: 3.75e-03, grad_scale: 16.0 2023-04-26 23:47:45,184 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=52752.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:48:48,803 INFO [finetune.py:976] (4/7) Epoch 10, batch 1250, loss[loss=0.1502, simple_loss=0.2136, pruned_loss=0.04343, over 4151.00 frames. ], tot_loss[loss=0.191, simple_loss=0.2569, pruned_loss=0.06256, over 953008.23 frames. ], batch size: 65, lr: 3.75e-03, grad_scale: 16.0 2023-04-26 23:49:09,962 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=52813.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:49:34,630 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.104e+02 1.609e+02 1.926e+02 2.217e+02 3.154e+02, threshold=3.851e+02, percent-clipped=0.0 2023-04-26 23:49:48,761 INFO [finetune.py:976] (4/7) Epoch 10, batch 1300, loss[loss=0.1696, simple_loss=0.2405, pruned_loss=0.0494, over 4849.00 frames. ], tot_loss[loss=0.1878, simple_loss=0.2535, pruned_loss=0.06105, over 953509.07 frames. ], batch size: 47, lr: 3.75e-03, grad_scale: 16.0 2023-04-26 23:49:48,847 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=52849.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:50:11,019 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5250, 1.3670, 1.6348, 1.9083, 1.9601, 1.4813, 1.2324, 1.6338], device='cuda:4'), covar=tensor([0.0884, 0.1223, 0.0739, 0.0621, 0.0638, 0.0930, 0.0892, 0.0699], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0205, 0.0184, 0.0177, 0.0179, 0.0190, 0.0161, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-26 23:50:20,541 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=52897.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:50:21,745 INFO [finetune.py:976] (4/7) Epoch 10, batch 1350, loss[loss=0.1756, simple_loss=0.2499, pruned_loss=0.0506, over 4798.00 frames. ], tot_loss[loss=0.1876, simple_loss=0.253, pruned_loss=0.06106, over 952442.83 frames. ], batch size: 45, lr: 3.75e-03, grad_scale: 16.0 2023-04-26 23:50:25,902 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1027, 1.4633, 1.3420, 1.6702, 1.5211, 1.9546, 1.3082, 3.6271], device='cuda:4'), covar=tensor([0.0673, 0.0811, 0.0829, 0.1185, 0.0669, 0.0550, 0.0788, 0.0140], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0044, 0.0040, 0.0039, 0.0039, 0.0059], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0012, 0.0013, 0.0016], device='cuda:4') 2023-04-26 23:50:31,137 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=52911.0, num_to_drop=1, layers_to_drop={2} 2023-04-26 23:50:46,659 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.089e+02 1.765e+02 2.062e+02 2.587e+02 4.153e+02, threshold=4.125e+02, percent-clipped=1.0 2023-04-26 23:50:55,066 INFO [finetune.py:976] (4/7) Epoch 10, batch 1400, loss[loss=0.2423, simple_loss=0.3077, pruned_loss=0.08842, over 4850.00 frames. ], tot_loss[loss=0.1909, simple_loss=0.2567, pruned_loss=0.06254, over 951512.13 frames. ], batch size: 47, lr: 3.75e-03, grad_scale: 16.0 2023-04-26 23:51:09,516 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=52968.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:51:10,467 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-26 23:51:25,781 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.6951, 3.4801, 2.6825, 4.2261, 3.5107, 3.6263, 1.4768, 3.6231], device='cuda:4'), covar=tensor([0.1665, 0.1354, 0.3736, 0.1155, 0.3012, 0.1700, 0.5496, 0.2213], device='cuda:4'), in_proj_covar=tensor([0.0243, 0.0217, 0.0248, 0.0302, 0.0299, 0.0249, 0.0268, 0.0270], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 23:51:28,729 INFO [finetune.py:976] (4/7) Epoch 10, batch 1450, loss[loss=0.2313, simple_loss=0.2966, pruned_loss=0.083, over 4811.00 frames. ], tot_loss[loss=0.1921, simple_loss=0.2588, pruned_loss=0.06273, over 953217.45 frames. ], batch size: 38, lr: 3.75e-03, grad_scale: 16.0 2023-04-26 23:51:31,456 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4208, 1.7934, 2.2646, 2.8162, 2.1796, 1.7727, 1.6993, 2.0923], device='cuda:4'), covar=tensor([0.3733, 0.3868, 0.1961, 0.3090, 0.3488, 0.3058, 0.4756, 0.2789], device='cuda:4'), in_proj_covar=tensor([0.0284, 0.0252, 0.0220, 0.0319, 0.0215, 0.0229, 0.0234, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-26 23:51:41,184 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=53016.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:52:04,982 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.250e+02 1.692e+02 2.005e+02 2.440e+02 4.945e+02, threshold=4.009e+02, percent-clipped=1.0 2023-04-26 23:52:18,927 INFO [finetune.py:976] (4/7) Epoch 10, batch 1500, loss[loss=0.2112, simple_loss=0.2739, pruned_loss=0.0742, over 4169.00 frames. ], tot_loss[loss=0.1944, simple_loss=0.2612, pruned_loss=0.06378, over 953152.42 frames. ], batch size: 65, lr: 3.75e-03, grad_scale: 16.0 2023-04-26 23:52:28,801 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=53056.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:53:19,945 INFO [finetune.py:976] (4/7) Epoch 10, batch 1550, loss[loss=0.1661, simple_loss=0.2355, pruned_loss=0.04835, over 4727.00 frames. ], tot_loss[loss=0.1942, simple_loss=0.2612, pruned_loss=0.06358, over 955075.85 frames. ], batch size: 23, lr: 3.75e-03, grad_scale: 16.0 2023-04-26 23:53:25,473 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=53108.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:53:32,448 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=53117.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:53:45,210 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.067e+02 1.713e+02 2.006e+02 2.452e+02 5.414e+02, threshold=4.013e+02, percent-clipped=2.0 2023-04-26 23:53:53,128 INFO [finetune.py:976] (4/7) Epoch 10, batch 1600, loss[loss=0.188, simple_loss=0.2591, pruned_loss=0.05843, over 4830.00 frames. ], tot_loss[loss=0.1904, simple_loss=0.2573, pruned_loss=0.06178, over 955088.16 frames. ], batch size: 33, lr: 3.75e-03, grad_scale: 16.0 2023-04-26 23:54:44,154 INFO [finetune.py:976] (4/7) Epoch 10, batch 1650, loss[loss=0.1566, simple_loss=0.2174, pruned_loss=0.04792, over 4908.00 frames. ], tot_loss[loss=0.1888, simple_loss=0.2551, pruned_loss=0.06125, over 954816.77 frames. ], batch size: 46, lr: 3.75e-03, grad_scale: 16.0 2023-04-26 23:54:47,348 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5415, 1.7931, 1.8416, 2.0317, 1.7360, 1.9232, 1.9906, 1.8979], device='cuda:4'), covar=tensor([0.5700, 0.6367, 0.5090, 0.4642, 0.6263, 0.8692, 0.5754, 0.6051], device='cuda:4'), in_proj_covar=tensor([0.0328, 0.0384, 0.0318, 0.0327, 0.0342, 0.0404, 0.0363, 0.0327], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 23:54:49,093 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0385, 1.5549, 4.1144, 3.8978, 3.6416, 3.6927, 3.6640, 3.7076], device='cuda:4'), covar=tensor([0.5537, 0.4994, 0.0936, 0.1340, 0.0974, 0.1906, 0.3378, 0.1211], device='cuda:4'), in_proj_covar=tensor([0.0306, 0.0305, 0.0404, 0.0407, 0.0348, 0.0404, 0.0313, 0.0370], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-26 23:54:52,156 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=53211.0, num_to_drop=1, layers_to_drop={1} 2023-04-26 23:55:09,622 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.078e+02 1.636e+02 1.864e+02 2.338e+02 5.656e+02, threshold=3.728e+02, percent-clipped=1.0 2023-04-26 23:55:12,828 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=53241.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:55:17,567 INFO [finetune.py:976] (4/7) Epoch 10, batch 1700, loss[loss=0.159, simple_loss=0.2284, pruned_loss=0.04481, over 4747.00 frames. ], tot_loss[loss=0.1867, simple_loss=0.2528, pruned_loss=0.06029, over 954276.12 frames. ], batch size: 27, lr: 3.75e-03, grad_scale: 16.0 2023-04-26 23:55:24,104 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=53259.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 23:55:31,319 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1795, 1.5872, 1.3332, 1.7599, 1.6006, 1.9823, 1.4062, 3.3863], device='cuda:4'), covar=tensor([0.0670, 0.0777, 0.0805, 0.1120, 0.0622, 0.0502, 0.0765, 0.0157], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0044, 0.0040, 0.0039, 0.0039, 0.0059], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-26 23:55:31,344 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4705, 1.2931, 1.7580, 1.6694, 1.3703, 1.1250, 1.4902, 1.0417], device='cuda:4'), covar=tensor([0.0617, 0.0790, 0.0476, 0.0871, 0.0801, 0.1145, 0.0824, 0.0729], device='cuda:4'), in_proj_covar=tensor([0.0066, 0.0072, 0.0070, 0.0067, 0.0075, 0.0094, 0.0077, 0.0073], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 23:55:39,514 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9401, 1.9624, 1.0957, 1.6795, 2.0907, 1.8027, 1.7495, 1.7779], device='cuda:4'), covar=tensor([0.0459, 0.0350, 0.0333, 0.0515, 0.0248, 0.0516, 0.0462, 0.0507], device='cuda:4'), in_proj_covar=tensor([0.0030, 0.0025, 0.0023, 0.0030, 0.0021, 0.0029, 0.0029, 0.0030], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0050, 0.0038, 0.0048, 0.0048, 0.0050], device='cuda:4') 2023-04-26 23:55:51,481 INFO [finetune.py:976] (4/7) Epoch 10, batch 1750, loss[loss=0.198, simple_loss=0.2779, pruned_loss=0.05902, over 4903.00 frames. ], tot_loss[loss=0.1895, simple_loss=0.2557, pruned_loss=0.06168, over 955631.18 frames. ], batch size: 43, lr: 3.75e-03, grad_scale: 16.0 2023-04-26 23:55:53,411 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=53302.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:55:54,618 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=53304.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:56:01,777 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=53315.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:56:16,877 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.153e+02 1.803e+02 2.178e+02 2.575e+02 5.782e+02, threshold=4.356e+02, percent-clipped=6.0 2023-04-26 23:56:25,437 INFO [finetune.py:976] (4/7) Epoch 10, batch 1800, loss[loss=0.1429, simple_loss=0.2077, pruned_loss=0.03904, over 4738.00 frames. ], tot_loss[loss=0.1904, simple_loss=0.2572, pruned_loss=0.06178, over 953831.62 frames. ], batch size: 23, lr: 3.75e-03, grad_scale: 16.0 2023-04-26 23:56:35,684 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=53365.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:56:36,906 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=53367.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:56:43,385 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=53376.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:56:59,051 INFO [finetune.py:976] (4/7) Epoch 10, batch 1850, loss[loss=0.2267, simple_loss=0.2895, pruned_loss=0.0819, over 4164.00 frames. ], tot_loss[loss=0.1924, simple_loss=0.2592, pruned_loss=0.06277, over 952646.99 frames. ], batch size: 66, lr: 3.75e-03, grad_scale: 16.0 2023-04-26 23:57:10,982 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=53408.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:57:13,901 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=53412.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:57:35,476 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=53428.0, num_to_drop=1, layers_to_drop={0} 2023-04-26 23:57:45,406 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.032e+02 1.670e+02 2.068e+02 2.536e+02 6.143e+02, threshold=4.136e+02, percent-clipped=4.0 2023-04-26 23:58:05,923 INFO [finetune.py:976] (4/7) Epoch 10, batch 1900, loss[loss=0.2082, simple_loss=0.2804, pruned_loss=0.06802, over 4907.00 frames. ], tot_loss[loss=0.1924, simple_loss=0.26, pruned_loss=0.06244, over 952074.82 frames. ], batch size: 46, lr: 3.75e-03, grad_scale: 16.0 2023-04-26 23:58:15,453 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=53456.0, num_to_drop=0, layers_to_drop=set() 2023-04-26 23:58:48,692 INFO [finetune.py:976] (4/7) Epoch 10, batch 1950, loss[loss=0.2113, simple_loss=0.2662, pruned_loss=0.07821, over 4934.00 frames. ], tot_loss[loss=0.1912, simple_loss=0.2588, pruned_loss=0.06179, over 952728.80 frames. ], batch size: 38, lr: 3.75e-03, grad_scale: 16.0 2023-04-26 23:59:03,225 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.6853, 3.6423, 2.6225, 4.2425, 3.7122, 3.6954, 1.6031, 3.5199], device='cuda:4'), covar=tensor([0.1770, 0.1216, 0.2929, 0.1667, 0.3678, 0.1934, 0.6204, 0.2614], device='cuda:4'), in_proj_covar=tensor([0.0243, 0.0218, 0.0250, 0.0303, 0.0301, 0.0251, 0.0270, 0.0272], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-26 23:59:12,238 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.213e+02 1.722e+02 1.968e+02 2.257e+02 3.746e+02, threshold=3.936e+02, percent-clipped=0.0 2023-04-26 23:59:16,911 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3350, 1.9282, 2.1867, 2.4657, 2.1377, 1.8043, 1.4550, 2.0293], device='cuda:4'), covar=tensor([0.3783, 0.3680, 0.1821, 0.2694, 0.3220, 0.2951, 0.4757, 0.2508], device='cuda:4'), in_proj_covar=tensor([0.0282, 0.0251, 0.0220, 0.0317, 0.0213, 0.0228, 0.0234, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-26 23:59:17,499 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6203, 1.3815, 1.7630, 1.8972, 1.4355, 1.1154, 1.3491, 0.9914], device='cuda:4'), covar=tensor([0.0595, 0.0721, 0.0552, 0.0527, 0.0766, 0.1557, 0.0803, 0.0918], device='cuda:4'), in_proj_covar=tensor([0.0066, 0.0073, 0.0070, 0.0066, 0.0076, 0.0095, 0.0077, 0.0073], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-26 23:59:19,061 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-26 23:59:22,129 INFO [finetune.py:976] (4/7) Epoch 10, batch 2000, loss[loss=0.1587, simple_loss=0.2329, pruned_loss=0.04224, over 4758.00 frames. ], tot_loss[loss=0.1899, simple_loss=0.2568, pruned_loss=0.06148, over 952636.13 frames. ], batch size: 26, lr: 3.75e-03, grad_scale: 16.0 2023-04-26 23:59:38,583 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3045, 1.8020, 1.5821, 2.1147, 1.8416, 1.9469, 1.7061, 4.3488], device='cuda:4'), covar=tensor([0.0595, 0.0783, 0.0774, 0.1166, 0.0651, 0.0575, 0.0720, 0.0128], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0044, 0.0040, 0.0039, 0.0039, 0.0059], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-26 23:59:53,933 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.90 vs. limit=5.0 2023-04-26 23:59:55,967 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.73 vs. limit=5.0 2023-04-27 00:00:04,619 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=53597.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:00:05,760 INFO [finetune.py:976] (4/7) Epoch 10, batch 2050, loss[loss=0.1406, simple_loss=0.206, pruned_loss=0.03761, over 4803.00 frames. ], tot_loss[loss=0.1869, simple_loss=0.2529, pruned_loss=0.06048, over 954302.41 frames. ], batch size: 25, lr: 3.75e-03, grad_scale: 16.0 2023-04-27 00:00:13,471 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9311, 1.2891, 1.5415, 1.5703, 2.0980, 1.6642, 1.3849, 1.4821], device='cuda:4'), covar=tensor([0.1433, 0.1515, 0.1969, 0.1300, 0.0762, 0.1560, 0.2197, 0.1798], device='cuda:4'), in_proj_covar=tensor([0.0306, 0.0321, 0.0355, 0.0298, 0.0335, 0.0322, 0.0308, 0.0359], device='cuda:4'), out_proj_covar=tensor([6.4474e-05, 6.8075e-05, 7.6699e-05, 6.1534e-05, 7.0010e-05, 6.9002e-05, 6.5992e-05, 7.7012e-05], device='cuda:4') 2023-04-27 00:00:34,506 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.104e+02 1.698e+02 1.922e+02 2.404e+02 4.326e+02, threshold=3.844e+02, percent-clipped=2.0 2023-04-27 00:00:44,012 INFO [finetune.py:976] (4/7) Epoch 10, batch 2100, loss[loss=0.1706, simple_loss=0.2516, pruned_loss=0.04475, over 4854.00 frames. ], tot_loss[loss=0.1876, simple_loss=0.2537, pruned_loss=0.06078, over 955966.33 frames. ], batch size: 44, lr: 3.75e-03, grad_scale: 32.0 2023-04-27 00:00:51,820 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=53660.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:00:55,307 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-27 00:00:58,523 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=53671.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:01:16,444 INFO [finetune.py:976] (4/7) Epoch 10, batch 2150, loss[loss=0.1952, simple_loss=0.2534, pruned_loss=0.06855, over 4208.00 frames. ], tot_loss[loss=0.191, simple_loss=0.2574, pruned_loss=0.06226, over 954134.35 frames. ], batch size: 18, lr: 3.75e-03, grad_scale: 32.0 2023-04-27 00:01:22,996 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6368, 2.3092, 1.6822, 1.7714, 1.2779, 1.2707, 1.8077, 1.1941], device='cuda:4'), covar=tensor([0.1629, 0.1303, 0.1502, 0.1633, 0.2514, 0.2011, 0.1005, 0.2111], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0216, 0.0170, 0.0204, 0.0204, 0.0185, 0.0161, 0.0189], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 00:01:26,013 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=53712.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:01:32,727 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=53723.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 00:01:34,669 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2009, 2.1560, 1.7920, 1.9449, 2.4001, 1.7950, 2.7777, 1.6936], device='cuda:4'), covar=tensor([0.4107, 0.1809, 0.4845, 0.3224, 0.1773, 0.2585, 0.1358, 0.4419], device='cuda:4'), in_proj_covar=tensor([0.0342, 0.0345, 0.0429, 0.0361, 0.0386, 0.0382, 0.0377, 0.0419], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 00:01:41,147 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.123e+02 1.745e+02 2.178e+02 2.516e+02 3.729e+02, threshold=4.356e+02, percent-clipped=0.0 2023-04-27 00:01:49,682 INFO [finetune.py:976] (4/7) Epoch 10, batch 2200, loss[loss=0.174, simple_loss=0.2444, pruned_loss=0.05177, over 4839.00 frames. ], tot_loss[loss=0.1922, simple_loss=0.2598, pruned_loss=0.06228, over 956264.43 frames. ], batch size: 49, lr: 3.75e-03, grad_scale: 32.0 2023-04-27 00:01:57,878 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=53760.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:02:22,061 INFO [finetune.py:976] (4/7) Epoch 10, batch 2250, loss[loss=0.2485, simple_loss=0.3151, pruned_loss=0.09099, over 4758.00 frames. ], tot_loss[loss=0.1933, simple_loss=0.2608, pruned_loss=0.06295, over 955156.80 frames. ], batch size: 54, lr: 3.75e-03, grad_scale: 32.0 2023-04-27 00:02:46,629 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.157e+02 1.602e+02 1.955e+02 2.365e+02 4.712e+02, threshold=3.910e+02, percent-clipped=2.0 2023-04-27 00:02:56,520 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.97 vs. limit=5.0 2023-04-27 00:03:06,363 INFO [finetune.py:976] (4/7) Epoch 10, batch 2300, loss[loss=0.1922, simple_loss=0.2647, pruned_loss=0.05986, over 4747.00 frames. ], tot_loss[loss=0.1924, simple_loss=0.2604, pruned_loss=0.06224, over 956429.53 frames. ], batch size: 54, lr: 3.75e-03, grad_scale: 32.0 2023-04-27 00:03:08,169 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0561, 2.5477, 1.0598, 1.3014, 2.0162, 1.1750, 3.1583, 1.5850], device='cuda:4'), covar=tensor([0.0662, 0.0653, 0.0841, 0.1285, 0.0456, 0.1001, 0.0303, 0.0633], device='cuda:4'), in_proj_covar=tensor([0.0052, 0.0067, 0.0049, 0.0047, 0.0051, 0.0052, 0.0078, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-27 00:03:18,384 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2106, 1.8302, 2.1106, 2.4054, 2.0604, 1.7315, 1.4344, 1.8887], device='cuda:4'), covar=tensor([0.3491, 0.3525, 0.1781, 0.2738, 0.3033, 0.2958, 0.4696, 0.2497], device='cuda:4'), in_proj_covar=tensor([0.0282, 0.0250, 0.0220, 0.0315, 0.0212, 0.0227, 0.0234, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 00:04:00,278 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.55 vs. limit=5.0 2023-04-27 00:04:10,841 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=53897.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:04:11,998 INFO [finetune.py:976] (4/7) Epoch 10, batch 2350, loss[loss=0.1594, simple_loss=0.23, pruned_loss=0.04441, over 4824.00 frames. ], tot_loss[loss=0.1909, simple_loss=0.2585, pruned_loss=0.06165, over 954094.00 frames. ], batch size: 33, lr: 3.75e-03, grad_scale: 32.0 2023-04-27 00:04:57,666 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.162e+02 1.640e+02 1.984e+02 2.431e+02 6.591e+02, threshold=3.969e+02, percent-clipped=4.0 2023-04-27 00:05:08,690 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=53945.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:05:16,062 INFO [finetune.py:976] (4/7) Epoch 10, batch 2400, loss[loss=0.1748, simple_loss=0.2434, pruned_loss=0.05313, over 4865.00 frames. ], tot_loss[loss=0.1879, simple_loss=0.2548, pruned_loss=0.06055, over 953050.74 frames. ], batch size: 31, lr: 3.74e-03, grad_scale: 32.0 2023-04-27 00:05:23,337 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=53960.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:05:32,001 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=53971.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:05:49,430 INFO [finetune.py:976] (4/7) Epoch 10, batch 2450, loss[loss=0.1304, simple_loss=0.204, pruned_loss=0.02839, over 4763.00 frames. ], tot_loss[loss=0.1864, simple_loss=0.2525, pruned_loss=0.06011, over 954056.70 frames. ], batch size: 27, lr: 3.74e-03, grad_scale: 32.0 2023-04-27 00:05:56,752 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=54008.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:06:04,941 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=54019.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:06:07,937 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=54023.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 00:06:15,673 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.227e+02 1.726e+02 2.083e+02 2.546e+02 5.806e+02, threshold=4.166e+02, percent-clipped=1.0 2023-04-27 00:06:24,053 INFO [finetune.py:976] (4/7) Epoch 10, batch 2500, loss[loss=0.19, simple_loss=0.2559, pruned_loss=0.06208, over 4777.00 frames. ], tot_loss[loss=0.1889, simple_loss=0.2545, pruned_loss=0.06159, over 953642.70 frames. ], batch size: 29, lr: 3.74e-03, grad_scale: 32.0 2023-04-27 00:06:39,522 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=54071.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:06:39,583 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4198, 0.9443, 0.4335, 1.1862, 1.0803, 1.3225, 1.2166, 1.1924], device='cuda:4'), covar=tensor([0.0506, 0.0391, 0.0403, 0.0540, 0.0286, 0.0473, 0.0490, 0.0565], device='cuda:4'), in_proj_covar=tensor([0.0030, 0.0025, 0.0023, 0.0030, 0.0021, 0.0029, 0.0029, 0.0030], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0045, 0.0038, 0.0050, 0.0038, 0.0048, 0.0048, 0.0050], device='cuda:4') 2023-04-27 00:06:54,300 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3735, 1.3408, 1.7917, 1.6773, 1.3189, 1.0976, 1.4791, 0.9956], device='cuda:4'), covar=tensor([0.0739, 0.0836, 0.0541, 0.0837, 0.1034, 0.1386, 0.0799, 0.0892], device='cuda:4'), in_proj_covar=tensor([0.0065, 0.0072, 0.0070, 0.0066, 0.0075, 0.0095, 0.0077, 0.0073], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 00:06:57,700 INFO [finetune.py:976] (4/7) Epoch 10, batch 2550, loss[loss=0.2054, simple_loss=0.2824, pruned_loss=0.06418, over 4912.00 frames. ], tot_loss[loss=0.1909, simple_loss=0.2573, pruned_loss=0.06224, over 953502.09 frames. ], batch size: 37, lr: 3.74e-03, grad_scale: 32.0 2023-04-27 00:06:57,832 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1601, 2.0128, 2.5127, 2.7559, 1.9553, 1.6183, 2.0873, 1.2119], device='cuda:4'), covar=tensor([0.0716, 0.1033, 0.0558, 0.1033, 0.1070, 0.1553, 0.0974, 0.1086], device='cuda:4'), in_proj_covar=tensor([0.0065, 0.0072, 0.0070, 0.0066, 0.0075, 0.0095, 0.0077, 0.0073], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 00:07:03,763 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=2.09 vs. limit=2.0 2023-04-27 00:07:10,257 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.43 vs. limit=2.0 2023-04-27 00:07:11,142 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.6322, 1.8541, 1.6828, 1.2999, 1.8033, 1.4849, 2.2722, 1.3082], device='cuda:4'), covar=tensor([0.3729, 0.1389, 0.4015, 0.2578, 0.1602, 0.2300, 0.1593, 0.4994], device='cuda:4'), in_proj_covar=tensor([0.0345, 0.0349, 0.0432, 0.0362, 0.0387, 0.0384, 0.0380, 0.0421], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 00:07:22,643 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.109e+02 1.598e+02 1.963e+02 2.419e+02 4.837e+02, threshold=3.926e+02, percent-clipped=1.0 2023-04-27 00:07:30,643 INFO [finetune.py:976] (4/7) Epoch 10, batch 2600, loss[loss=0.1978, simple_loss=0.2647, pruned_loss=0.06546, over 4833.00 frames. ], tot_loss[loss=0.1917, simple_loss=0.2584, pruned_loss=0.06248, over 952086.84 frames. ], batch size: 44, lr: 3.74e-03, grad_scale: 32.0 2023-04-27 00:07:51,276 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1482, 2.8813, 2.0193, 2.1585, 1.5094, 1.4906, 2.2557, 1.4702], device='cuda:4'), covar=tensor([0.1842, 0.1573, 0.1552, 0.1881, 0.2544, 0.2111, 0.1149, 0.2221], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0214, 0.0170, 0.0204, 0.0204, 0.0184, 0.0160, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 00:07:53,233 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.14 vs. limit=5.0 2023-04-27 00:08:04,394 INFO [finetune.py:976] (4/7) Epoch 10, batch 2650, loss[loss=0.2011, simple_loss=0.2715, pruned_loss=0.06535, over 4861.00 frames. ], tot_loss[loss=0.1931, simple_loss=0.26, pruned_loss=0.06309, over 953709.01 frames. ], batch size: 34, lr: 3.74e-03, grad_scale: 32.0 2023-04-27 00:08:39,723 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.224e+02 1.638e+02 1.968e+02 2.302e+02 4.272e+02, threshold=3.936e+02, percent-clipped=1.0 2023-04-27 00:08:40,171 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-27 00:08:52,971 INFO [finetune.py:976] (4/7) Epoch 10, batch 2700, loss[loss=0.1879, simple_loss=0.2577, pruned_loss=0.05903, over 4710.00 frames. ], tot_loss[loss=0.1926, simple_loss=0.2596, pruned_loss=0.06278, over 953859.71 frames. ], batch size: 59, lr: 3.74e-03, grad_scale: 32.0 2023-04-27 00:09:32,419 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0692, 1.0232, 1.2562, 1.1897, 0.9990, 0.8642, 1.0661, 0.6120], device='cuda:4'), covar=tensor([0.0505, 0.0753, 0.0649, 0.0521, 0.0701, 0.1308, 0.0535, 0.0935], device='cuda:4'), in_proj_covar=tensor([0.0065, 0.0072, 0.0070, 0.0066, 0.0075, 0.0095, 0.0077, 0.0073], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 00:10:03,552 INFO [finetune.py:976] (4/7) Epoch 10, batch 2750, loss[loss=0.2014, simple_loss=0.267, pruned_loss=0.0679, over 4768.00 frames. ], tot_loss[loss=0.1903, simple_loss=0.2569, pruned_loss=0.06184, over 954642.35 frames. ], batch size: 59, lr: 3.74e-03, grad_scale: 32.0 2023-04-27 00:10:49,706 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.176e+02 1.579e+02 1.855e+02 2.381e+02 3.900e+02, threshold=3.711e+02, percent-clipped=0.0 2023-04-27 00:10:51,467 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1853, 1.6641, 1.6080, 1.8754, 1.7384, 1.9848, 1.5177, 3.5805], device='cuda:4'), covar=tensor([0.0645, 0.0747, 0.0735, 0.1114, 0.0595, 0.0522, 0.0697, 0.0150], device='cuda:4'), in_proj_covar=tensor([0.0039, 0.0039, 0.0041, 0.0044, 0.0040, 0.0039, 0.0039, 0.0059], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0017], device='cuda:4') 2023-04-27 00:11:09,555 INFO [finetune.py:976] (4/7) Epoch 10, batch 2800, loss[loss=0.2179, simple_loss=0.2853, pruned_loss=0.07531, over 4819.00 frames. ], tot_loss[loss=0.1888, simple_loss=0.2548, pruned_loss=0.06142, over 954715.39 frames. ], batch size: 41, lr: 3.74e-03, grad_scale: 32.0 2023-04-27 00:11:48,133 INFO [finetune.py:976] (4/7) Epoch 10, batch 2850, loss[loss=0.1517, simple_loss=0.2095, pruned_loss=0.04694, over 4726.00 frames. ], tot_loss[loss=0.1868, simple_loss=0.2519, pruned_loss=0.0608, over 953800.35 frames. ], batch size: 23, lr: 3.74e-03, grad_scale: 32.0 2023-04-27 00:11:59,580 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-27 00:12:11,874 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.049e+02 1.761e+02 2.016e+02 2.397e+02 4.207e+02, threshold=4.033e+02, percent-clipped=3.0 2023-04-27 00:12:21,806 INFO [finetune.py:976] (4/7) Epoch 10, batch 2900, loss[loss=0.1824, simple_loss=0.246, pruned_loss=0.05937, over 4833.00 frames. ], tot_loss[loss=0.1904, simple_loss=0.2558, pruned_loss=0.06246, over 951353.32 frames. ], batch size: 30, lr: 3.74e-03, grad_scale: 32.0 2023-04-27 00:12:28,669 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1684, 2.9335, 0.8029, 1.4250, 1.5721, 2.0105, 1.6407, 0.9662], device='cuda:4'), covar=tensor([0.1806, 0.1364, 0.2258, 0.1703, 0.1367, 0.1263, 0.1813, 0.2137], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0249, 0.0140, 0.0123, 0.0135, 0.0154, 0.0119, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 00:12:41,815 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.68 vs. limit=5.0 2023-04-27 00:12:55,780 INFO [finetune.py:976] (4/7) Epoch 10, batch 2950, loss[loss=0.197, simple_loss=0.2749, pruned_loss=0.05956, over 4844.00 frames. ], tot_loss[loss=0.1929, simple_loss=0.2593, pruned_loss=0.06321, over 952327.77 frames. ], batch size: 47, lr: 3.74e-03, grad_scale: 32.0 2023-04-27 00:13:19,226 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.151e+02 1.620e+02 2.110e+02 2.463e+02 5.874e+02, threshold=4.221e+02, percent-clipped=2.0 2023-04-27 00:13:21,678 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2314, 1.4951, 1.5941, 1.7506, 1.6492, 1.7724, 1.7473, 1.6570], device='cuda:4'), covar=tensor([0.4607, 0.5869, 0.5146, 0.4722, 0.5953, 0.8494, 0.5364, 0.5479], device='cuda:4'), in_proj_covar=tensor([0.0322, 0.0377, 0.0312, 0.0322, 0.0335, 0.0397, 0.0355, 0.0320], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 00:13:29,120 INFO [finetune.py:976] (4/7) Epoch 10, batch 3000, loss[loss=0.1788, simple_loss=0.2514, pruned_loss=0.05312, over 4927.00 frames. ], tot_loss[loss=0.1945, simple_loss=0.2612, pruned_loss=0.06391, over 953845.49 frames. ], batch size: 38, lr: 3.74e-03, grad_scale: 32.0 2023-04-27 00:13:29,120 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 00:13:37,483 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7114, 2.1128, 1.7403, 1.9616, 1.6099, 1.7512, 1.7588, 1.4069], device='cuda:4'), covar=tensor([0.1813, 0.1126, 0.0885, 0.1239, 0.3486, 0.1131, 0.1839, 0.2572], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0317, 0.0227, 0.0287, 0.0314, 0.0269, 0.0257, 0.0277], device='cuda:4'), out_proj_covar=tensor([1.1843e-04, 1.2758e-04, 9.1263e-05, 1.1513e-04, 1.2847e-04, 1.0836e-04, 1.0488e-04, 1.1115e-04], device='cuda:4') 2023-04-27 00:13:45,427 INFO [finetune.py:1010] (4/7) Epoch 10, validation: loss=0.1531, simple_loss=0.2257, pruned_loss=0.04026, over 2265189.00 frames. 2023-04-27 00:13:45,428 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-27 00:14:32,350 INFO [finetune.py:976] (4/7) Epoch 10, batch 3050, loss[loss=0.2153, simple_loss=0.2841, pruned_loss=0.07326, over 4847.00 frames. ], tot_loss[loss=0.1946, simple_loss=0.2617, pruned_loss=0.06376, over 954006.65 frames. ], batch size: 44, lr: 3.74e-03, grad_scale: 32.0 2023-04-27 00:14:35,085 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0572, 2.4271, 0.9941, 1.2418, 1.7749, 1.2516, 3.4015, 1.7310], device='cuda:4'), covar=tensor([0.0675, 0.0659, 0.0775, 0.1413, 0.0555, 0.1032, 0.0256, 0.0632], device='cuda:4'), in_proj_covar=tensor([0.0052, 0.0067, 0.0049, 0.0047, 0.0051, 0.0052, 0.0078, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-27 00:14:57,006 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.192e+02 1.553e+02 1.895e+02 2.251e+02 3.622e+02, threshold=3.789e+02, percent-clipped=0.0 2023-04-27 00:15:05,022 INFO [finetune.py:976] (4/7) Epoch 10, batch 3100, loss[loss=0.188, simple_loss=0.2512, pruned_loss=0.06242, over 4742.00 frames. ], tot_loss[loss=0.1938, simple_loss=0.2605, pruned_loss=0.06351, over 955234.74 frames. ], batch size: 27, lr: 3.74e-03, grad_scale: 32.0 2023-04-27 00:15:27,265 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=54664.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 00:15:48,198 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1174, 1.6487, 2.0002, 2.3716, 1.9637, 1.5753, 1.1409, 1.7080], device='cuda:4'), covar=tensor([0.3401, 0.3578, 0.1734, 0.2343, 0.2810, 0.2782, 0.4794, 0.2412], device='cuda:4'), in_proj_covar=tensor([0.0283, 0.0250, 0.0220, 0.0316, 0.0214, 0.0228, 0.0233, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 00:16:11,741 INFO [finetune.py:976] (4/7) Epoch 10, batch 3150, loss[loss=0.1941, simple_loss=0.2596, pruned_loss=0.06432, over 4925.00 frames. ], tot_loss[loss=0.1916, simple_loss=0.2576, pruned_loss=0.06277, over 956841.98 frames. ], batch size: 37, lr: 3.74e-03, grad_scale: 32.0 2023-04-27 00:16:53,623 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=54725.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 00:17:04,898 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.175e+02 1.617e+02 1.915e+02 2.289e+02 4.542e+02, threshold=3.830e+02, percent-clipped=1.0 2023-04-27 00:17:25,059 INFO [finetune.py:976] (4/7) Epoch 10, batch 3200, loss[loss=0.2059, simple_loss=0.2717, pruned_loss=0.07011, over 4828.00 frames. ], tot_loss[loss=0.1886, simple_loss=0.254, pruned_loss=0.0616, over 956862.39 frames. ], batch size: 40, lr: 3.74e-03, grad_scale: 32.0 2023-04-27 00:18:08,047 INFO [finetune.py:976] (4/7) Epoch 10, batch 3250, loss[loss=0.1936, simple_loss=0.2601, pruned_loss=0.06355, over 4862.00 frames. ], tot_loss[loss=0.1893, simple_loss=0.2544, pruned_loss=0.06208, over 954968.57 frames. ], batch size: 34, lr: 3.74e-03, grad_scale: 32.0 2023-04-27 00:18:33,642 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.070e+02 1.703e+02 2.055e+02 2.409e+02 4.051e+02, threshold=4.111e+02, percent-clipped=1.0 2023-04-27 00:18:42,051 INFO [finetune.py:976] (4/7) Epoch 10, batch 3300, loss[loss=0.2088, simple_loss=0.2767, pruned_loss=0.07048, over 4854.00 frames. ], tot_loss[loss=0.1924, simple_loss=0.2582, pruned_loss=0.06334, over 955274.32 frames. ], batch size: 44, lr: 3.74e-03, grad_scale: 32.0 2023-04-27 00:18:45,171 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2572, 3.1203, 0.8017, 1.4589, 1.5701, 2.1721, 1.8314, 0.9310], device='cuda:4'), covar=tensor([0.2062, 0.1754, 0.2637, 0.2109, 0.1593, 0.1469, 0.1963, 0.2339], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0253, 0.0142, 0.0124, 0.0137, 0.0156, 0.0120, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 00:18:50,604 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=54862.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:19:15,377 INFO [finetune.py:976] (4/7) Epoch 10, batch 3350, loss[loss=0.2407, simple_loss=0.2977, pruned_loss=0.09185, over 4843.00 frames. ], tot_loss[loss=0.1945, simple_loss=0.2607, pruned_loss=0.06417, over 953502.09 frames. ], batch size: 49, lr: 3.74e-03, grad_scale: 32.0 2023-04-27 00:19:19,842 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.64 vs. limit=2.0 2023-04-27 00:19:30,580 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=54923.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 00:19:39,810 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.272e+02 1.802e+02 2.053e+02 2.510e+02 4.625e+02, threshold=4.107e+02, percent-clipped=1.0 2023-04-27 00:19:47,703 INFO [finetune.py:976] (4/7) Epoch 10, batch 3400, loss[loss=0.1874, simple_loss=0.2652, pruned_loss=0.05482, over 4856.00 frames. ], tot_loss[loss=0.1954, simple_loss=0.2618, pruned_loss=0.06447, over 955191.44 frames. ], batch size: 44, lr: 3.74e-03, grad_scale: 32.0 2023-04-27 00:19:51,322 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6050, 2.4024, 2.5929, 2.9567, 3.0131, 2.2810, 1.9824, 2.6447], device='cuda:4'), covar=tensor([0.0849, 0.0990, 0.0670, 0.0620, 0.0608, 0.0923, 0.0947, 0.0589], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0206, 0.0185, 0.0178, 0.0181, 0.0191, 0.0162, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 00:20:07,430 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7122, 2.2827, 1.6811, 1.5450, 1.2901, 1.2906, 1.6913, 1.2550], device='cuda:4'), covar=tensor([0.1883, 0.1397, 0.1715, 0.1905, 0.2491, 0.2218, 0.1099, 0.2170], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0215, 0.0170, 0.0204, 0.0204, 0.0184, 0.0160, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 00:20:20,501 INFO [finetune.py:976] (4/7) Epoch 10, batch 3450, loss[loss=0.1854, simple_loss=0.2486, pruned_loss=0.06113, over 4813.00 frames. ], tot_loss[loss=0.1936, simple_loss=0.2608, pruned_loss=0.0632, over 956118.39 frames. ], batch size: 30, lr: 3.74e-03, grad_scale: 32.0 2023-04-27 00:20:34,365 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=55020.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 00:20:36,344 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.68 vs. limit=2.0 2023-04-27 00:20:41,726 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.56 vs. limit=5.0 2023-04-27 00:20:45,489 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.144e+02 1.637e+02 2.017e+02 2.373e+02 5.913e+02, threshold=4.034e+02, percent-clipped=2.0 2023-04-27 00:20:53,439 INFO [finetune.py:976] (4/7) Epoch 10, batch 3500, loss[loss=0.1671, simple_loss=0.2347, pruned_loss=0.04971, over 4778.00 frames. ], tot_loss[loss=0.1921, simple_loss=0.2589, pruned_loss=0.06267, over 955958.18 frames. ], batch size: 26, lr: 3.74e-03, grad_scale: 32.0 2023-04-27 00:21:02,066 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.96 vs. limit=2.0 2023-04-27 00:21:17,487 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3095, 1.1687, 3.8337, 3.5765, 3.3794, 3.6383, 3.7088, 3.4550], device='cuda:4'), covar=tensor([0.6879, 0.5961, 0.1251, 0.1844, 0.1248, 0.1684, 0.1357, 0.1476], device='cuda:4'), in_proj_covar=tensor([0.0315, 0.0315, 0.0413, 0.0417, 0.0356, 0.0413, 0.0320, 0.0376], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 00:21:28,360 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0146, 2.9943, 2.7103, 2.8092, 3.1670, 2.8626, 4.0036, 2.5122], device='cuda:4'), covar=tensor([0.3498, 0.1960, 0.3714, 0.2690, 0.1563, 0.2017, 0.1167, 0.3404], device='cuda:4'), in_proj_covar=tensor([0.0347, 0.0350, 0.0434, 0.0363, 0.0388, 0.0387, 0.0381, 0.0421], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 00:21:46,851 INFO [finetune.py:976] (4/7) Epoch 10, batch 3550, loss[loss=0.1698, simple_loss=0.2381, pruned_loss=0.05078, over 4829.00 frames. ], tot_loss[loss=0.1901, simple_loss=0.2561, pruned_loss=0.06201, over 955989.47 frames. ], batch size: 40, lr: 3.74e-03, grad_scale: 32.0 2023-04-27 00:21:48,191 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5783, 1.3317, 1.6837, 1.8296, 1.4409, 1.1279, 1.3418, 0.8757], device='cuda:4'), covar=tensor([0.0581, 0.0770, 0.0494, 0.0638, 0.0752, 0.1628, 0.0789, 0.0939], device='cuda:4'), in_proj_covar=tensor([0.0066, 0.0073, 0.0071, 0.0067, 0.0076, 0.0096, 0.0077, 0.0073], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 00:22:24,279 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.7273, 1.7966, 1.9256, 2.0668, 1.8060, 2.0089, 2.0270, 2.0077], device='cuda:4'), covar=tensor([0.4541, 0.7296, 0.6538, 0.5540, 0.6784, 0.8801, 0.7274, 0.6215], device='cuda:4'), in_proj_covar=tensor([0.0325, 0.0379, 0.0316, 0.0325, 0.0339, 0.0399, 0.0358, 0.0323], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 00:22:27,074 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.344e+01 1.599e+02 1.898e+02 2.251e+02 4.553e+02, threshold=3.796e+02, percent-clipped=2.0 2023-04-27 00:22:35,882 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=55148.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:22:36,397 INFO [finetune.py:976] (4/7) Epoch 10, batch 3600, loss[loss=0.1479, simple_loss=0.2173, pruned_loss=0.03927, over 4837.00 frames. ], tot_loss[loss=0.1881, simple_loss=0.2539, pruned_loss=0.0612, over 957914.62 frames. ], batch size: 30, lr: 3.74e-03, grad_scale: 32.0 2023-04-27 00:23:26,775 INFO [finetune.py:976] (4/7) Epoch 10, batch 3650, loss[loss=0.2536, simple_loss=0.313, pruned_loss=0.09713, over 4739.00 frames. ], tot_loss[loss=0.1899, simple_loss=0.2557, pruned_loss=0.06208, over 955741.23 frames. ], batch size: 59, lr: 3.74e-03, grad_scale: 32.0 2023-04-27 00:23:33,123 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=55209.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:23:34,360 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=55211.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:23:39,097 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=55218.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 00:23:42,765 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7393, 1.7196, 2.0381, 2.2555, 1.6442, 1.3543, 1.7298, 1.0396], device='cuda:4'), covar=tensor([0.0640, 0.0770, 0.0553, 0.0731, 0.0836, 0.1168, 0.0905, 0.0885], device='cuda:4'), in_proj_covar=tensor([0.0066, 0.0072, 0.0071, 0.0067, 0.0075, 0.0096, 0.0077, 0.0072], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 00:23:50,725 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.180e+02 1.786e+02 2.133e+02 2.599e+02 5.918e+02, threshold=4.267e+02, percent-clipped=5.0 2023-04-27 00:24:00,574 INFO [finetune.py:976] (4/7) Epoch 10, batch 3700, loss[loss=0.2017, simple_loss=0.2768, pruned_loss=0.06331, over 4800.00 frames. ], tot_loss[loss=0.1929, simple_loss=0.2593, pruned_loss=0.06323, over 956506.94 frames. ], batch size: 45, lr: 3.73e-03, grad_scale: 32.0 2023-04-27 00:24:15,040 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=55272.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:24:33,931 INFO [finetune.py:976] (4/7) Epoch 10, batch 3750, loss[loss=0.1908, simple_loss=0.261, pruned_loss=0.06031, over 4889.00 frames. ], tot_loss[loss=0.1944, simple_loss=0.261, pruned_loss=0.06396, over 955280.64 frames. ], batch size: 37, lr: 3.73e-03, grad_scale: 32.0 2023-04-27 00:24:47,379 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=55320.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 00:24:57,145 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.148e+02 1.679e+02 1.930e+02 2.224e+02 3.496e+02, threshold=3.860e+02, percent-clipped=0.0 2023-04-27 00:25:07,102 INFO [finetune.py:976] (4/7) Epoch 10, batch 3800, loss[loss=0.1927, simple_loss=0.2562, pruned_loss=0.06465, over 4765.00 frames. ], tot_loss[loss=0.1939, simple_loss=0.2609, pruned_loss=0.06348, over 954384.62 frames. ], batch size: 28, lr: 3.73e-03, grad_scale: 32.0 2023-04-27 00:25:13,048 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1928, 2.8158, 0.9906, 1.4492, 2.0494, 1.2676, 3.8727, 1.7473], device='cuda:4'), covar=tensor([0.0745, 0.1045, 0.0972, 0.1261, 0.0533, 0.0970, 0.0212, 0.0631], device='cuda:4'), in_proj_covar=tensor([0.0052, 0.0067, 0.0050, 0.0047, 0.0051, 0.0052, 0.0078, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0007, 0.0008, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-27 00:25:17,998 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9332, 1.4816, 1.4574, 1.6766, 2.1836, 1.7220, 1.4225, 1.3831], device='cuda:4'), covar=tensor([0.1376, 0.1721, 0.2082, 0.1290, 0.0735, 0.1844, 0.2234, 0.2113], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0324, 0.0357, 0.0300, 0.0338, 0.0322, 0.0309, 0.0362], device='cuda:4'), out_proj_covar=tensor([6.4614e-05, 6.8718e-05, 7.7054e-05, 6.1956e-05, 7.0801e-05, 6.8960e-05, 6.6254e-05, 7.7744e-05], device='cuda:4') 2023-04-27 00:25:19,735 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=55368.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 00:25:24,603 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7806, 1.3810, 1.6321, 1.5159, 1.6785, 1.3328, 0.7495, 1.4024], device='cuda:4'), covar=tensor([0.3466, 0.3393, 0.1726, 0.2497, 0.2474, 0.2649, 0.4389, 0.1960], device='cuda:4'), in_proj_covar=tensor([0.0283, 0.0249, 0.0219, 0.0316, 0.0213, 0.0227, 0.0233, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 00:25:40,047 INFO [finetune.py:976] (4/7) Epoch 10, batch 3850, loss[loss=0.1398, simple_loss=0.2167, pruned_loss=0.03139, over 4752.00 frames. ], tot_loss[loss=0.1916, simple_loss=0.2585, pruned_loss=0.06232, over 954525.14 frames. ], batch size: 26, lr: 3.73e-03, grad_scale: 32.0 2023-04-27 00:26:04,613 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.147e+02 1.602e+02 1.915e+02 2.260e+02 3.579e+02, threshold=3.830e+02, percent-clipped=0.0 2023-04-27 00:26:12,957 INFO [finetune.py:976] (4/7) Epoch 10, batch 3900, loss[loss=0.2155, simple_loss=0.2696, pruned_loss=0.08069, over 4772.00 frames. ], tot_loss[loss=0.1896, simple_loss=0.2558, pruned_loss=0.06166, over 955428.18 frames. ], batch size: 28, lr: 3.73e-03, grad_scale: 32.0 2023-04-27 00:26:42,299 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6318, 1.5297, 0.6920, 1.2841, 1.6297, 1.4813, 1.3790, 1.4158], device='cuda:4'), covar=tensor([0.0522, 0.0387, 0.0413, 0.0563, 0.0288, 0.0542, 0.0542, 0.0600], device='cuda:4'), in_proj_covar=tensor([0.0030, 0.0026, 0.0023, 0.0030, 0.0021, 0.0029, 0.0029, 0.0030], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0045, 0.0038, 0.0050, 0.0038, 0.0048, 0.0048, 0.0050], device='cuda:4') 2023-04-27 00:26:45,236 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0192, 1.8782, 2.3450, 2.5298, 1.8612, 1.5665, 1.9407, 1.2285], device='cuda:4'), covar=tensor([0.0731, 0.0878, 0.0500, 0.0810, 0.0858, 0.1261, 0.0908, 0.1038], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0073, 0.0071, 0.0067, 0.0076, 0.0096, 0.0078, 0.0073], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 00:27:05,955 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.0161, 4.0097, 2.9043, 4.6179, 4.0940, 4.0141, 1.9886, 3.9816], device='cuda:4'), covar=tensor([0.1811, 0.1060, 0.2894, 0.1452, 0.2603, 0.1888, 0.5577, 0.2203], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0216, 0.0249, 0.0306, 0.0301, 0.0250, 0.0270, 0.0271], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 00:27:07,758 INFO [finetune.py:976] (4/7) Epoch 10, batch 3950, loss[loss=0.1965, simple_loss=0.2565, pruned_loss=0.06822, over 4780.00 frames. ], tot_loss[loss=0.1869, simple_loss=0.2525, pruned_loss=0.06061, over 954345.72 frames. ], batch size: 26, lr: 3.73e-03, grad_scale: 32.0 2023-04-27 00:27:15,664 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=55504.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:27:36,360 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=55518.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 00:27:47,674 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.461e+01 1.659e+02 2.039e+02 2.496e+02 7.335e+02, threshold=4.077e+02, percent-clipped=3.0 2023-04-27 00:27:56,581 INFO [finetune.py:976] (4/7) Epoch 10, batch 4000, loss[loss=0.2652, simple_loss=0.3265, pruned_loss=0.1019, over 4725.00 frames. ], tot_loss[loss=0.1879, simple_loss=0.2531, pruned_loss=0.06138, over 952782.29 frames. ], batch size: 59, lr: 3.73e-03, grad_scale: 32.0 2023-04-27 00:28:09,058 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=55566.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:28:12,130 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=55567.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:28:40,533 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2892, 1.1625, 1.5572, 1.4716, 1.2264, 1.0651, 1.2063, 0.8778], device='cuda:4'), covar=tensor([0.0609, 0.0610, 0.0473, 0.0560, 0.0826, 0.1126, 0.0560, 0.0653], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0073, 0.0072, 0.0068, 0.0076, 0.0096, 0.0078, 0.0073], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 00:28:43,369 INFO [finetune.py:976] (4/7) Epoch 10, batch 4050, loss[loss=0.2559, simple_loss=0.3119, pruned_loss=0.09997, over 4908.00 frames. ], tot_loss[loss=0.1942, simple_loss=0.2594, pruned_loss=0.06456, over 952090.02 frames. ], batch size: 37, lr: 3.73e-03, grad_scale: 32.0 2023-04-27 00:29:09,491 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.009e+02 1.669e+02 1.991e+02 2.505e+02 4.319e+02, threshold=3.981e+02, percent-clipped=1.0 2023-04-27 00:29:13,880 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5015, 1.0443, 4.3107, 4.0401, 3.7594, 4.1235, 4.0101, 3.7662], device='cuda:4'), covar=tensor([0.7278, 0.6698, 0.1009, 0.1592, 0.1054, 0.1873, 0.1663, 0.1416], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0313, 0.0410, 0.0415, 0.0354, 0.0411, 0.0319, 0.0376], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 00:29:16,838 INFO [finetune.py:976] (4/7) Epoch 10, batch 4100, loss[loss=0.1956, simple_loss=0.2572, pruned_loss=0.06705, over 4921.00 frames. ], tot_loss[loss=0.1941, simple_loss=0.2601, pruned_loss=0.06402, over 951731.63 frames. ], batch size: 42, lr: 3.73e-03, grad_scale: 32.0 2023-04-27 00:29:46,010 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4770, 1.7609, 1.7690, 1.9094, 1.6956, 1.8500, 1.8548, 1.8189], device='cuda:4'), covar=tensor([0.4843, 0.7127, 0.6031, 0.5568, 0.6995, 0.8988, 0.7225, 0.6527], device='cuda:4'), in_proj_covar=tensor([0.0324, 0.0378, 0.0314, 0.0323, 0.0337, 0.0398, 0.0357, 0.0322], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 00:29:50,672 INFO [finetune.py:976] (4/7) Epoch 10, batch 4150, loss[loss=0.2268, simple_loss=0.2949, pruned_loss=0.07931, over 4905.00 frames. ], tot_loss[loss=0.1953, simple_loss=0.2616, pruned_loss=0.06452, over 952258.45 frames. ], batch size: 37, lr: 3.73e-03, grad_scale: 32.0 2023-04-27 00:30:16,063 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.104e+02 1.755e+02 2.031e+02 2.318e+02 3.746e+02, threshold=4.063e+02, percent-clipped=0.0 2023-04-27 00:30:23,253 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=55748.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:30:23,769 INFO [finetune.py:976] (4/7) Epoch 10, batch 4200, loss[loss=0.1655, simple_loss=0.2245, pruned_loss=0.05325, over 4817.00 frames. ], tot_loss[loss=0.195, simple_loss=0.2617, pruned_loss=0.06413, over 955079.71 frames. ], batch size: 30, lr: 3.73e-03, grad_scale: 32.0 2023-04-27 00:30:57,483 INFO [finetune.py:976] (4/7) Epoch 10, batch 4250, loss[loss=0.1884, simple_loss=0.2514, pruned_loss=0.06267, over 4868.00 frames. ], tot_loss[loss=0.1941, simple_loss=0.26, pruned_loss=0.06413, over 953130.13 frames. ], batch size: 31, lr: 3.73e-03, grad_scale: 32.0 2023-04-27 00:31:00,767 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=55804.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:31:04,337 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=55809.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:31:23,550 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.518e+01 1.564e+02 1.887e+02 2.474e+02 5.983e+02, threshold=3.773e+02, percent-clipped=1.0 2023-04-27 00:31:29,622 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0077, 1.8478, 2.1385, 2.2816, 2.1747, 1.8468, 1.9920, 1.9435], device='cuda:4'), covar=tensor([0.5311, 0.7373, 0.8701, 0.7298, 0.6396, 1.0637, 1.0277, 0.9391], device='cuda:4'), in_proj_covar=tensor([0.0411, 0.0414, 0.0500, 0.0518, 0.0439, 0.0460, 0.0470, 0.0468], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 00:31:30,772 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=55848.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:31:31,269 INFO [finetune.py:976] (4/7) Epoch 10, batch 4300, loss[loss=0.177, simple_loss=0.2349, pruned_loss=0.05953, over 4904.00 frames. ], tot_loss[loss=0.1903, simple_loss=0.2559, pruned_loss=0.06231, over 951566.03 frames. ], batch size: 37, lr: 3.73e-03, grad_scale: 32.0 2023-04-27 00:31:33,045 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=55852.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:31:48,291 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=55867.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:32:19,025 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1113, 4.3174, 0.8175, 2.1282, 2.3377, 2.8070, 2.4835, 1.0271], device='cuda:4'), covar=tensor([0.1292, 0.0913, 0.2262, 0.1370, 0.1084, 0.1104, 0.1438, 0.2080], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0252, 0.0142, 0.0123, 0.0136, 0.0155, 0.0119, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 00:32:31,696 INFO [finetune.py:976] (4/7) Epoch 10, batch 4350, loss[loss=0.1511, simple_loss=0.2173, pruned_loss=0.04245, over 4798.00 frames. ], tot_loss[loss=0.1889, simple_loss=0.2536, pruned_loss=0.06207, over 952893.33 frames. ], batch size: 29, lr: 3.73e-03, grad_scale: 32.0 2023-04-27 00:32:43,352 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=55909.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:32:46,920 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=55915.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:32:54,387 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.85 vs. limit=2.0 2023-04-27 00:33:02,707 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.042e+02 1.769e+02 2.030e+02 2.376e+02 3.891e+02, threshold=4.060e+02, percent-clipped=2.0 2023-04-27 00:33:10,568 INFO [finetune.py:976] (4/7) Epoch 10, batch 4400, loss[loss=0.173, simple_loss=0.2581, pruned_loss=0.04396, over 4141.00 frames. ], tot_loss[loss=0.1887, simple_loss=0.2539, pruned_loss=0.06179, over 954788.87 frames. ], batch size: 65, lr: 3.73e-03, grad_scale: 32.0 2023-04-27 00:33:42,936 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0238, 2.9508, 1.7908, 2.2944, 1.4655, 1.4870, 2.0828, 1.4752], device='cuda:4'), covar=tensor([0.1860, 0.1520, 0.1889, 0.1776, 0.2795, 0.2376, 0.1247, 0.2227], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0213, 0.0169, 0.0203, 0.0204, 0.0184, 0.0159, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 00:34:18,723 INFO [finetune.py:976] (4/7) Epoch 10, batch 4450, loss[loss=0.2015, simple_loss=0.2497, pruned_loss=0.07665, over 4229.00 frames. ], tot_loss[loss=0.1911, simple_loss=0.257, pruned_loss=0.06265, over 953424.93 frames. ], batch size: 18, lr: 3.73e-03, grad_scale: 32.0 2023-04-27 00:34:57,405 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.139e+02 1.646e+02 1.926e+02 2.335e+02 3.268e+02, threshold=3.851e+02, percent-clipped=0.0 2023-04-27 00:35:03,527 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.0233, 3.9492, 2.7962, 4.6123, 3.9833, 3.9949, 2.0378, 4.0045], device='cuda:4'), covar=tensor([0.1551, 0.1024, 0.3094, 0.1287, 0.2667, 0.1758, 0.4871, 0.1990], device='cuda:4'), in_proj_covar=tensor([0.0240, 0.0213, 0.0245, 0.0301, 0.0296, 0.0247, 0.0264, 0.0266], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 00:35:05,273 INFO [finetune.py:976] (4/7) Epoch 10, batch 4500, loss[loss=0.205, simple_loss=0.2847, pruned_loss=0.06271, over 4815.00 frames. ], tot_loss[loss=0.1916, simple_loss=0.2579, pruned_loss=0.06259, over 954105.58 frames. ], batch size: 39, lr: 3.73e-03, grad_scale: 32.0 2023-04-27 00:35:17,946 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9445, 2.7557, 2.1729, 2.4397, 1.8485, 2.2652, 2.1798, 1.6544], device='cuda:4'), covar=tensor([0.2697, 0.1486, 0.1028, 0.1755, 0.3430, 0.1557, 0.2306, 0.3219], device='cuda:4'), in_proj_covar=tensor([0.0301, 0.0322, 0.0233, 0.0294, 0.0320, 0.0274, 0.0260, 0.0284], device='cuda:4'), out_proj_covar=tensor([1.2173e-04, 1.2958e-04, 9.3466e-05, 1.1777e-04, 1.3104e-04, 1.1044e-04, 1.0588e-04, 1.1409e-04], device='cuda:4') 2023-04-27 00:35:38,736 INFO [finetune.py:976] (4/7) Epoch 10, batch 4550, loss[loss=0.2284, simple_loss=0.2907, pruned_loss=0.08304, over 4813.00 frames. ], tot_loss[loss=0.1936, simple_loss=0.2601, pruned_loss=0.0635, over 953442.00 frames. ], batch size: 30, lr: 3.73e-03, grad_scale: 32.0 2023-04-27 00:35:41,832 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=56104.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:36:03,337 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.622e+01 1.743e+02 2.073e+02 2.397e+02 5.279e+02, threshold=4.145e+02, percent-clipped=1.0 2023-04-27 00:36:09,763 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=56145.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:36:12,138 INFO [finetune.py:976] (4/7) Epoch 10, batch 4600, loss[loss=0.1945, simple_loss=0.2672, pruned_loss=0.06092, over 4795.00 frames. ], tot_loss[loss=0.1933, simple_loss=0.26, pruned_loss=0.06325, over 955656.74 frames. ], batch size: 51, lr: 3.73e-03, grad_scale: 32.0 2023-04-27 00:36:36,872 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.75 vs. limit=2.0 2023-04-27 00:36:45,851 INFO [finetune.py:976] (4/7) Epoch 10, batch 4650, loss[loss=0.1804, simple_loss=0.2429, pruned_loss=0.05898, over 4759.00 frames. ], tot_loss[loss=0.1905, simple_loss=0.2572, pruned_loss=0.06196, over 955469.22 frames. ], batch size: 54, lr: 3.73e-03, grad_scale: 32.0 2023-04-27 00:36:49,600 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=56204.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:36:50,864 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=56206.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:37:04,917 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4455, 1.2481, 4.0148, 3.7081, 3.5477, 3.7227, 3.7659, 3.5650], device='cuda:4'), covar=tensor([0.6604, 0.5884, 0.1019, 0.1740, 0.1076, 0.1567, 0.2024, 0.1434], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0309, 0.0406, 0.0411, 0.0350, 0.0406, 0.0317, 0.0369], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 00:37:16,886 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.791e+01 1.513e+02 1.802e+02 2.222e+02 6.871e+02, threshold=3.603e+02, percent-clipped=2.0 2023-04-27 00:37:24,883 INFO [finetune.py:976] (4/7) Epoch 10, batch 4700, loss[loss=0.1714, simple_loss=0.2316, pruned_loss=0.05558, over 4830.00 frames. ], tot_loss[loss=0.1874, simple_loss=0.2538, pruned_loss=0.06045, over 956617.38 frames. ], batch size: 33, lr: 3.73e-03, grad_scale: 32.0 2023-04-27 00:37:57,893 INFO [finetune.py:976] (4/7) Epoch 10, batch 4750, loss[loss=0.1796, simple_loss=0.2533, pruned_loss=0.0529, over 4857.00 frames. ], tot_loss[loss=0.1861, simple_loss=0.2525, pruned_loss=0.05983, over 955664.40 frames. ], batch size: 44, lr: 3.73e-03, grad_scale: 32.0 2023-04-27 00:38:09,813 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6369, 1.8334, 0.9753, 1.3030, 1.8808, 1.5332, 1.4202, 1.4531], device='cuda:4'), covar=tensor([0.0546, 0.0366, 0.0362, 0.0603, 0.0264, 0.0530, 0.0524, 0.0636], device='cuda:4'), in_proj_covar=tensor([0.0030, 0.0025, 0.0023, 0.0030, 0.0020, 0.0029, 0.0029, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0037, 0.0050, 0.0037, 0.0048, 0.0048, 0.0050], device='cuda:4') 2023-04-27 00:38:14,751 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.8130, 4.6727, 3.2920, 5.5072, 4.8929, 4.7063, 2.5496, 4.6776], device='cuda:4'), covar=tensor([0.1485, 0.0932, 0.3059, 0.0797, 0.4003, 0.1607, 0.4875, 0.1997], device='cuda:4'), in_proj_covar=tensor([0.0240, 0.0214, 0.0245, 0.0302, 0.0296, 0.0247, 0.0264, 0.0266], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 00:38:23,668 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.024e+02 1.579e+02 1.869e+02 2.281e+02 6.054e+02, threshold=3.738e+02, percent-clipped=4.0 2023-04-27 00:38:31,957 INFO [finetune.py:976] (4/7) Epoch 10, batch 4800, loss[loss=0.2305, simple_loss=0.295, pruned_loss=0.08298, over 4807.00 frames. ], tot_loss[loss=0.1886, simple_loss=0.2555, pruned_loss=0.0608, over 954953.71 frames. ], batch size: 41, lr: 3.73e-03, grad_scale: 32.0 2023-04-27 00:39:33,343 INFO [finetune.py:976] (4/7) Epoch 10, batch 4850, loss[loss=0.1764, simple_loss=0.2498, pruned_loss=0.0515, over 4741.00 frames. ], tot_loss[loss=0.1911, simple_loss=0.2586, pruned_loss=0.06177, over 954632.64 frames. ], batch size: 27, lr: 3.73e-03, grad_scale: 32.0 2023-04-27 00:39:43,324 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=56404.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:39:52,805 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-27 00:40:07,080 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5027, 3.0707, 1.1071, 1.6130, 1.7281, 2.3277, 1.8009, 1.0746], device='cuda:4'), covar=tensor([0.1289, 0.1004, 0.1703, 0.1366, 0.1093, 0.0873, 0.1401, 0.1753], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0250, 0.0142, 0.0123, 0.0135, 0.0154, 0.0118, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 00:40:14,076 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.161e+02 1.686e+02 2.043e+02 2.423e+02 5.639e+02, threshold=4.086e+02, percent-clipped=3.0 2023-04-27 00:40:22,380 INFO [finetune.py:976] (4/7) Epoch 10, batch 4900, loss[loss=0.1988, simple_loss=0.2621, pruned_loss=0.0677, over 4693.00 frames. ], tot_loss[loss=0.1934, simple_loss=0.2606, pruned_loss=0.06313, over 953777.17 frames. ], batch size: 23, lr: 3.73e-03, grad_scale: 32.0 2023-04-27 00:40:24,261 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=56452.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:40:56,299 INFO [finetune.py:976] (4/7) Epoch 10, batch 4950, loss[loss=0.202, simple_loss=0.2696, pruned_loss=0.06724, over 4912.00 frames. ], tot_loss[loss=0.1937, simple_loss=0.2607, pruned_loss=0.06338, over 951126.55 frames. ], batch size: 46, lr: 3.72e-03, grad_scale: 32.0 2023-04-27 00:40:57,606 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=56501.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:40:59,496 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=56504.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:41:12,967 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=56523.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 00:41:21,848 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.129e+02 1.631e+02 2.001e+02 2.395e+02 7.277e+02, threshold=4.002e+02, percent-clipped=1.0 2023-04-27 00:41:29,734 INFO [finetune.py:976] (4/7) Epoch 10, batch 5000, loss[loss=0.1991, simple_loss=0.2665, pruned_loss=0.06587, over 4881.00 frames. ], tot_loss[loss=0.1917, simple_loss=0.2585, pruned_loss=0.06245, over 950966.00 frames. ], batch size: 35, lr: 3.72e-03, grad_scale: 32.0 2023-04-27 00:41:32,063 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=56552.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:41:45,223 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1713, 1.5805, 1.9314, 2.2468, 1.9705, 1.5285, 1.1267, 1.7541], device='cuda:4'), covar=tensor([0.3469, 0.3727, 0.1694, 0.2313, 0.2749, 0.2898, 0.4423, 0.2083], device='cuda:4'), in_proj_covar=tensor([0.0284, 0.0249, 0.0220, 0.0316, 0.0214, 0.0228, 0.0232, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 00:41:53,737 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=56584.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 00:41:53,744 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0622, 2.6583, 2.0690, 1.9908, 1.4951, 1.4467, 2.3166, 1.4125], device='cuda:4'), covar=tensor([0.1618, 0.1632, 0.1418, 0.1850, 0.2349, 0.1904, 0.0947, 0.2058], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0214, 0.0169, 0.0203, 0.0203, 0.0184, 0.0159, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 00:41:59,153 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2053, 1.5947, 2.0667, 2.3564, 2.0243, 1.5674, 1.2892, 1.7014], device='cuda:4'), covar=tensor([0.3435, 0.3651, 0.1640, 0.2547, 0.2901, 0.2943, 0.4505, 0.2497], device='cuda:4'), in_proj_covar=tensor([0.0284, 0.0249, 0.0220, 0.0316, 0.0214, 0.0227, 0.0233, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 00:42:03,259 INFO [finetune.py:976] (4/7) Epoch 10, batch 5050, loss[loss=0.1759, simple_loss=0.23, pruned_loss=0.06092, over 4867.00 frames. ], tot_loss[loss=0.1891, simple_loss=0.2551, pruned_loss=0.0615, over 953002.12 frames. ], batch size: 31, lr: 3.72e-03, grad_scale: 32.0 2023-04-27 00:42:56,295 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.103e+02 1.627e+02 1.925e+02 2.295e+02 4.200e+02, threshold=3.850e+02, percent-clipped=1.0 2023-04-27 00:43:09,518 INFO [finetune.py:976] (4/7) Epoch 10, batch 5100, loss[loss=0.1692, simple_loss=0.2355, pruned_loss=0.05149, over 4824.00 frames. ], tot_loss[loss=0.1853, simple_loss=0.2512, pruned_loss=0.05971, over 954101.98 frames. ], batch size: 40, lr: 3.72e-03, grad_scale: 32.0 2023-04-27 00:43:59,157 INFO [finetune.py:976] (4/7) Epoch 10, batch 5150, loss[loss=0.243, simple_loss=0.3044, pruned_loss=0.0908, over 4043.00 frames. ], tot_loss[loss=0.1889, simple_loss=0.2543, pruned_loss=0.06175, over 953841.60 frames. ], batch size: 65, lr: 3.72e-03, grad_scale: 32.0 2023-04-27 00:44:06,982 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2413, 1.3619, 3.7726, 3.4866, 3.3235, 3.5599, 3.5856, 3.3499], device='cuda:4'), covar=tensor([0.7166, 0.5516, 0.1154, 0.1960, 0.1273, 0.1947, 0.2010, 0.1678], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0308, 0.0408, 0.0411, 0.0351, 0.0408, 0.0316, 0.0369], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 00:44:25,420 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.088e+02 1.696e+02 2.008e+02 2.313e+02 3.981e+02, threshold=4.015e+02, percent-clipped=1.0 2023-04-27 00:44:33,209 INFO [finetune.py:976] (4/7) Epoch 10, batch 5200, loss[loss=0.153, simple_loss=0.231, pruned_loss=0.03753, over 4900.00 frames. ], tot_loss[loss=0.1919, simple_loss=0.2581, pruned_loss=0.06285, over 953594.15 frames. ], batch size: 35, lr: 3.72e-03, grad_scale: 32.0 2023-04-27 00:44:51,259 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-27 00:45:18,110 INFO [finetune.py:976] (4/7) Epoch 10, batch 5250, loss[loss=0.195, simple_loss=0.2533, pruned_loss=0.06831, over 3925.00 frames. ], tot_loss[loss=0.1915, simple_loss=0.2587, pruned_loss=0.06214, over 953188.69 frames. ], batch size: 17, lr: 3.72e-03, grad_scale: 32.0 2023-04-27 00:45:20,017 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=56801.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:45:21,228 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=56803.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:45:25,586 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.59 vs. limit=5.0 2023-04-27 00:45:43,652 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.105e+02 1.715e+02 1.995e+02 2.714e+02 4.040e+02, threshold=3.989e+02, percent-clipped=1.0 2023-04-27 00:45:51,393 INFO [finetune.py:976] (4/7) Epoch 10, batch 5300, loss[loss=0.2132, simple_loss=0.2787, pruned_loss=0.07388, over 4756.00 frames. ], tot_loss[loss=0.1941, simple_loss=0.2614, pruned_loss=0.06334, over 954391.44 frames. ], batch size: 59, lr: 3.72e-03, grad_scale: 32.0 2023-04-27 00:45:51,456 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=56849.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:46:00,531 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5553, 3.1624, 2.7485, 2.7200, 2.0284, 2.0227, 3.0551, 2.1377], device='cuda:4'), covar=tensor([0.1588, 0.1679, 0.1362, 0.1519, 0.2187, 0.1839, 0.0799, 0.1773], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0214, 0.0169, 0.0204, 0.0204, 0.0185, 0.0159, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 00:46:01,124 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=56864.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:46:11,638 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=56879.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 00:46:25,221 INFO [finetune.py:976] (4/7) Epoch 10, batch 5350, loss[loss=0.1779, simple_loss=0.2412, pruned_loss=0.05734, over 4893.00 frames. ], tot_loss[loss=0.1931, simple_loss=0.2608, pruned_loss=0.06269, over 956018.21 frames. ], batch size: 32, lr: 3.72e-03, grad_scale: 32.0 2023-04-27 00:46:31,370 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=56908.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:46:50,604 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.133e+02 1.631e+02 1.870e+02 2.324e+02 4.447e+02, threshold=3.741e+02, percent-clipped=2.0 2023-04-27 00:46:58,345 INFO [finetune.py:976] (4/7) Epoch 10, batch 5400, loss[loss=0.1836, simple_loss=0.2487, pruned_loss=0.05927, over 4802.00 frames. ], tot_loss[loss=0.1906, simple_loss=0.2576, pruned_loss=0.06176, over 956877.90 frames. ], batch size: 45, lr: 3.72e-03, grad_scale: 32.0 2023-04-27 00:47:11,575 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=56969.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:47:28,566 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=56993.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:47:29,775 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=56995.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:47:32,096 INFO [finetune.py:976] (4/7) Epoch 10, batch 5450, loss[loss=0.1719, simple_loss=0.245, pruned_loss=0.04942, over 4918.00 frames. ], tot_loss[loss=0.1892, simple_loss=0.2554, pruned_loss=0.06152, over 958029.02 frames. ], batch size: 37, lr: 3.72e-03, grad_scale: 32.0 2023-04-27 00:48:18,181 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.140e+02 1.695e+02 1.974e+02 2.503e+02 6.150e+02, threshold=3.947e+02, percent-clipped=1.0 2023-04-27 00:48:38,384 INFO [finetune.py:976] (4/7) Epoch 10, batch 5500, loss[loss=0.1817, simple_loss=0.2425, pruned_loss=0.06048, over 4763.00 frames. ], tot_loss[loss=0.1871, simple_loss=0.2528, pruned_loss=0.06068, over 955983.85 frames. ], batch size: 23, lr: 3.72e-03, grad_scale: 32.0 2023-04-27 00:48:39,820 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-27 00:48:40,396 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0392, 1.5179, 1.8226, 2.0916, 1.8065, 1.4580, 0.9602, 1.5921], device='cuda:4'), covar=tensor([0.3451, 0.3442, 0.1810, 0.2513, 0.2954, 0.2802, 0.4968, 0.2415], device='cuda:4'), in_proj_covar=tensor([0.0284, 0.0248, 0.0219, 0.0316, 0.0213, 0.0227, 0.0232, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 00:48:47,134 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=57054.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:48:48,382 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=57056.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:49:35,989 INFO [finetune.py:976] (4/7) Epoch 10, batch 5550, loss[loss=0.1517, simple_loss=0.2239, pruned_loss=0.03975, over 4762.00 frames. ], tot_loss[loss=0.1876, simple_loss=0.2534, pruned_loss=0.06087, over 954519.96 frames. ], batch size: 28, lr: 3.72e-03, grad_scale: 32.0 2023-04-27 00:49:47,649 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6319, 3.5312, 1.1813, 1.8479, 1.8741, 2.5734, 2.0154, 1.0751], device='cuda:4'), covar=tensor([0.1365, 0.0958, 0.1850, 0.1348, 0.1141, 0.0984, 0.1462, 0.2012], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0247, 0.0141, 0.0121, 0.0133, 0.0152, 0.0118, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 00:49:59,961 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.204e+02 1.628e+02 1.818e+02 2.115e+02 5.040e+02, threshold=3.636e+02, percent-clipped=1.0 2023-04-27 00:50:06,923 INFO [finetune.py:976] (4/7) Epoch 10, batch 5600, loss[loss=0.2987, simple_loss=0.3412, pruned_loss=0.1281, over 4931.00 frames. ], tot_loss[loss=0.1911, simple_loss=0.2578, pruned_loss=0.06221, over 953646.20 frames. ], batch size: 38, lr: 3.72e-03, grad_scale: 32.0 2023-04-27 00:50:12,763 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=57159.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:50:24,880 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=57179.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 00:50:29,386 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=57186.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:50:35,747 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=57197.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:50:36,872 INFO [finetune.py:976] (4/7) Epoch 10, batch 5650, loss[loss=0.2091, simple_loss=0.2541, pruned_loss=0.08202, over 4116.00 frames. ], tot_loss[loss=0.1936, simple_loss=0.2608, pruned_loss=0.0632, over 951850.02 frames. ], batch size: 18, lr: 3.72e-03, grad_scale: 32.0 2023-04-27 00:50:53,686 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=57227.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 00:50:59,542 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.003e+02 1.673e+02 1.923e+02 2.288e+02 6.250e+02, threshold=3.847e+02, percent-clipped=4.0 2023-04-27 00:51:03,824 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=57244.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:51:05,609 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=57247.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:51:06,699 INFO [finetune.py:976] (4/7) Epoch 10, batch 5700, loss[loss=0.1701, simple_loss=0.2222, pruned_loss=0.059, over 4196.00 frames. ], tot_loss[loss=0.1902, simple_loss=0.2561, pruned_loss=0.06217, over 932858.91 frames. ], batch size: 18, lr: 3.72e-03, grad_scale: 32.0 2023-04-27 00:51:12,268 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=57258.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 00:51:15,769 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=57264.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:51:38,657 INFO [finetune.py:976] (4/7) Epoch 11, batch 0, loss[loss=0.1996, simple_loss=0.2686, pruned_loss=0.0653, over 4824.00 frames. ], tot_loss[loss=0.1996, simple_loss=0.2686, pruned_loss=0.0653, over 4824.00 frames. ], batch size: 33, lr: 3.72e-03, grad_scale: 32.0 2023-04-27 00:51:38,657 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 00:51:55,309 INFO [finetune.py:1010] (4/7) Epoch 11, validation: loss=0.1558, simple_loss=0.2272, pruned_loss=0.04225, over 2265189.00 frames. 2023-04-27 00:51:55,309 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-27 00:52:28,826 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=57305.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:52:42,960 INFO [finetune.py:976] (4/7) Epoch 11, batch 50, loss[loss=0.1587, simple_loss=0.2227, pruned_loss=0.04736, over 4762.00 frames. ], tot_loss[loss=0.1951, simple_loss=0.2608, pruned_loss=0.0647, over 215947.55 frames. ], batch size: 26, lr: 3.72e-03, grad_scale: 32.0 2023-04-27 00:52:49,894 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.045e+02 1.652e+02 2.063e+02 2.429e+02 4.586e+02, threshold=4.127e+02, percent-clipped=3.0 2023-04-27 00:52:57,809 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=57349.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:52:59,006 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=57351.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:53:20,601 INFO [finetune.py:976] (4/7) Epoch 11, batch 100, loss[loss=0.1817, simple_loss=0.241, pruned_loss=0.06124, over 4746.00 frames. ], tot_loss[loss=0.187, simple_loss=0.2527, pruned_loss=0.06064, over 380182.62 frames. ], batch size: 54, lr: 3.72e-03, grad_scale: 32.0 2023-04-27 00:53:40,237 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6308, 1.3855, 0.5892, 1.2415, 1.3944, 1.4931, 1.3639, 1.3114], device='cuda:4'), covar=tensor([0.0514, 0.0392, 0.0417, 0.0559, 0.0307, 0.0511, 0.0488, 0.0583], device='cuda:4'), in_proj_covar=tensor([0.0030, 0.0025, 0.0023, 0.0030, 0.0020, 0.0029, 0.0029, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0037, 0.0050, 0.0037, 0.0048, 0.0048, 0.0050], device='cuda:4') 2023-04-27 00:53:40,253 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=57398.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:54:10,098 INFO [finetune.py:976] (4/7) Epoch 11, batch 150, loss[loss=0.1916, simple_loss=0.2488, pruned_loss=0.06714, over 4861.00 frames. ], tot_loss[loss=0.1862, simple_loss=0.25, pruned_loss=0.06125, over 507997.96 frames. ], batch size: 49, lr: 3.72e-03, grad_scale: 32.0 2023-04-27 00:54:27,174 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.157e+02 1.639e+02 1.968e+02 2.285e+02 5.137e+02, threshold=3.937e+02, percent-clipped=1.0 2023-04-27 00:54:52,722 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=57459.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:54:52,745 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=57459.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:55:00,025 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=57462.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:55:02,965 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4164, 1.4992, 4.1931, 3.8546, 3.6765, 3.9920, 3.9815, 3.6714], device='cuda:4'), covar=tensor([0.7220, 0.5861, 0.1154, 0.2036, 0.1238, 0.1508, 0.1202, 0.1708], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0307, 0.0406, 0.0408, 0.0350, 0.0406, 0.0315, 0.0367], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 00:55:02,995 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6657, 1.4787, 0.7616, 1.2851, 1.6066, 1.5232, 1.3575, 1.3826], device='cuda:4'), covar=tensor([0.0492, 0.0424, 0.0401, 0.0590, 0.0303, 0.0524, 0.0526, 0.0609], device='cuda:4'), in_proj_covar=tensor([0.0030, 0.0025, 0.0023, 0.0030, 0.0020, 0.0029, 0.0029, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0037, 0.0050, 0.0037, 0.0048, 0.0048, 0.0050], device='cuda:4') 2023-04-27 00:55:10,528 INFO [finetune.py:976] (4/7) Epoch 11, batch 200, loss[loss=0.1598, simple_loss=0.2293, pruned_loss=0.04517, over 4762.00 frames. ], tot_loss[loss=0.1839, simple_loss=0.2482, pruned_loss=0.05981, over 606277.95 frames. ], batch size: 28, lr: 3.72e-03, grad_scale: 32.0 2023-04-27 00:55:18,956 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7427, 1.7037, 2.0312, 1.9340, 1.8502, 1.6532, 1.8390, 1.8360], device='cuda:4'), covar=tensor([0.7399, 0.9446, 1.1446, 1.2485, 0.8889, 1.3478, 1.3708, 1.3032], device='cuda:4'), in_proj_covar=tensor([0.0407, 0.0412, 0.0497, 0.0516, 0.0436, 0.0456, 0.0467, 0.0466], device='cuda:4'), out_proj_covar=tensor([9.9191e-05, 1.0213e-04, 1.1216e-04, 1.2270e-04, 1.0566e-04, 1.1033e-04, 1.1196e-04, 1.1214e-04], device='cuda:4') 2023-04-27 00:55:20,210 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.76 vs. limit=5.0 2023-04-27 00:55:22,549 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=57495.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:55:30,376 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=57507.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:55:40,659 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=57523.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:55:42,940 INFO [finetune.py:976] (4/7) Epoch 11, batch 250, loss[loss=0.1497, simple_loss=0.2138, pruned_loss=0.04278, over 4776.00 frames. ], tot_loss[loss=0.1867, simple_loss=0.2518, pruned_loss=0.06077, over 681866.86 frames. ], batch size: 26, lr: 3.71e-03, grad_scale: 32.0 2023-04-27 00:55:50,074 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.267e+02 1.758e+02 2.095e+02 2.670e+02 5.121e+02, threshold=4.190e+02, percent-clipped=8.0 2023-04-27 00:55:52,817 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.7199, 4.6660, 3.1506, 5.3978, 4.7560, 4.6717, 2.2429, 4.6005], device='cuda:4'), covar=tensor([0.1383, 0.0898, 0.3106, 0.0797, 0.3147, 0.1546, 0.5249, 0.2090], device='cuda:4'), in_proj_covar=tensor([0.0241, 0.0214, 0.0248, 0.0301, 0.0296, 0.0248, 0.0267, 0.0269], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 00:55:54,617 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=57542.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:56:01,286 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=57553.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 00:56:03,137 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=57556.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:56:04,896 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4827, 0.6865, 1.4031, 1.8826, 1.5899, 1.4203, 1.3865, 1.4661], device='cuda:4'), covar=tensor([0.4944, 0.7092, 0.6906, 0.6955, 0.6310, 0.8109, 0.8243, 0.8189], device='cuda:4'), in_proj_covar=tensor([0.0407, 0.0412, 0.0497, 0.0516, 0.0437, 0.0457, 0.0467, 0.0466], device='cuda:4'), out_proj_covar=tensor([9.9207e-05, 1.0217e-04, 1.1214e-04, 1.2278e-04, 1.0577e-04, 1.1042e-04, 1.1193e-04, 1.1199e-04], device='cuda:4') 2023-04-27 00:56:07,259 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6078, 1.6182, 4.4371, 4.1599, 3.9224, 4.1508, 4.1388, 3.9353], device='cuda:4'), covar=tensor([0.6638, 0.5612, 0.1136, 0.1713, 0.1102, 0.1872, 0.1063, 0.1421], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0305, 0.0405, 0.0408, 0.0349, 0.0405, 0.0314, 0.0365], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 00:56:08,504 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=57564.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:56:16,735 INFO [finetune.py:976] (4/7) Epoch 11, batch 300, loss[loss=0.1877, simple_loss=0.2603, pruned_loss=0.05752, over 4924.00 frames. ], tot_loss[loss=0.1883, simple_loss=0.2542, pruned_loss=0.0612, over 742723.70 frames. ], batch size: 38, lr: 3.71e-03, grad_scale: 32.0 2023-04-27 00:56:17,686 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.99 vs. limit=5.0 2023-04-27 00:56:27,335 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.6226, 3.5767, 2.6308, 4.2880, 3.5961, 3.6584, 1.7184, 3.5704], device='cuda:4'), covar=tensor([0.1645, 0.1205, 0.3195, 0.1651, 0.2844, 0.1719, 0.5500, 0.2506], device='cuda:4'), in_proj_covar=tensor([0.0242, 0.0215, 0.0249, 0.0303, 0.0297, 0.0249, 0.0268, 0.0270], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 00:56:32,858 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=57600.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:56:34,197 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-27 00:56:40,100 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=57612.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:56:50,091 INFO [finetune.py:976] (4/7) Epoch 11, batch 350, loss[loss=0.195, simple_loss=0.2602, pruned_loss=0.06485, over 4839.00 frames. ], tot_loss[loss=0.1892, simple_loss=0.256, pruned_loss=0.06116, over 790946.51 frames. ], batch size: 44, lr: 3.71e-03, grad_scale: 32.0 2023-04-27 00:56:56,642 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.085e+02 1.613e+02 1.965e+02 2.419e+02 4.062e+02, threshold=3.929e+02, percent-clipped=0.0 2023-04-27 00:57:11,881 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=57649.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:57:13,073 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=57651.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:57:46,784 INFO [finetune.py:976] (4/7) Epoch 11, batch 400, loss[loss=0.2085, simple_loss=0.2737, pruned_loss=0.07168, over 4731.00 frames. ], tot_loss[loss=0.1907, simple_loss=0.2579, pruned_loss=0.06177, over 827041.29 frames. ], batch size: 27, lr: 3.71e-03, grad_scale: 64.0 2023-04-27 00:58:07,047 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8014, 2.6142, 2.1365, 2.3388, 1.9155, 2.2080, 2.1105, 1.6869], device='cuda:4'), covar=tensor([0.2459, 0.1427, 0.1013, 0.1477, 0.3475, 0.1332, 0.2241, 0.2894], device='cuda:4'), in_proj_covar=tensor([0.0294, 0.0315, 0.0228, 0.0285, 0.0314, 0.0267, 0.0253, 0.0276], device='cuda:4'), out_proj_covar=tensor([1.1856e-04, 1.2659e-04, 9.1447e-05, 1.1417e-04, 1.2835e-04, 1.0747e-04, 1.0324e-04, 1.1079e-04], device='cuda:4') 2023-04-27 00:58:10,657 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=57697.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:58:12,862 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=57699.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:58:27,232 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8703, 1.4020, 1.6632, 1.6975, 1.6537, 1.3229, 0.7563, 1.3178], device='cuda:4'), covar=tensor([0.3474, 0.3634, 0.1704, 0.2367, 0.2771, 0.2799, 0.4480, 0.2524], device='cuda:4'), in_proj_covar=tensor([0.0284, 0.0248, 0.0220, 0.0317, 0.0213, 0.0227, 0.0232, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 00:58:30,753 INFO [finetune.py:976] (4/7) Epoch 11, batch 450, loss[loss=0.1964, simple_loss=0.2592, pruned_loss=0.0668, over 4864.00 frames. ], tot_loss[loss=0.1893, simple_loss=0.2564, pruned_loss=0.06115, over 854639.40 frames. ], batch size: 31, lr: 3.71e-03, grad_scale: 64.0 2023-04-27 00:58:33,251 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-27 00:58:34,983 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=57733.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:58:37,319 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.069e+02 1.604e+02 1.948e+02 2.335e+02 4.408e+02, threshold=3.896e+02, percent-clipped=1.0 2023-04-27 00:59:00,416 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=57754.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 00:59:32,059 INFO [finetune.py:976] (4/7) Epoch 11, batch 500, loss[loss=0.1915, simple_loss=0.2476, pruned_loss=0.06773, over 4827.00 frames. ], tot_loss[loss=0.1881, simple_loss=0.2549, pruned_loss=0.06065, over 877412.23 frames. ], batch size: 30, lr: 3.71e-03, grad_scale: 64.0 2023-04-27 00:59:32,368 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-27 00:59:54,243 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=57794.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:00:13,883 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3326, 1.6623, 1.4910, 1.8372, 1.7866, 1.8866, 1.5182, 3.0565], device='cuda:4'), covar=tensor([0.0620, 0.0651, 0.0693, 0.0938, 0.0518, 0.0592, 0.0707, 0.0170], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0043, 0.0040, 0.0038, 0.0038, 0.0058], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0012, 0.0013, 0.0016], device='cuda:4') 2023-04-27 01:00:25,887 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=57818.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:00:31,734 INFO [finetune.py:976] (4/7) Epoch 11, batch 550, loss[loss=0.1974, simple_loss=0.2666, pruned_loss=0.06415, over 4818.00 frames. ], tot_loss[loss=0.1866, simple_loss=0.2528, pruned_loss=0.06022, over 895604.24 frames. ], batch size: 40, lr: 3.71e-03, grad_scale: 64.0 2023-04-27 01:00:38,241 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.096e+02 1.623e+02 1.943e+02 2.373e+02 5.716e+02, threshold=3.887e+02, percent-clipped=4.0 2023-04-27 01:00:41,341 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=57842.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:00:47,294 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=57851.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:00:49,055 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=57853.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 01:00:53,064 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.4806, 3.3875, 2.5013, 4.1111, 3.5214, 3.5002, 1.6815, 3.5134], device='cuda:4'), covar=tensor([0.1697, 0.1424, 0.3627, 0.1902, 0.3983, 0.1978, 0.5258, 0.2562], device='cuda:4'), in_proj_covar=tensor([0.0241, 0.0215, 0.0248, 0.0301, 0.0297, 0.0249, 0.0268, 0.0269], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 01:01:04,584 INFO [finetune.py:976] (4/7) Epoch 11, batch 600, loss[loss=0.1606, simple_loss=0.2432, pruned_loss=0.03901, over 4833.00 frames. ], tot_loss[loss=0.1881, simple_loss=0.2542, pruned_loss=0.06097, over 907514.55 frames. ], batch size: 30, lr: 3.71e-03, grad_scale: 64.0 2023-04-27 01:01:13,413 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=57890.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:01:20,166 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=57900.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:01:20,739 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=57901.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:01:38,047 INFO [finetune.py:976] (4/7) Epoch 11, batch 650, loss[loss=0.1795, simple_loss=0.2546, pruned_loss=0.05217, over 4750.00 frames. ], tot_loss[loss=0.1905, simple_loss=0.2574, pruned_loss=0.06179, over 918255.46 frames. ], batch size: 27, lr: 3.71e-03, grad_scale: 64.0 2023-04-27 01:01:45,068 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.147e+02 1.753e+02 2.161e+02 2.818e+02 6.431e+02, threshold=4.321e+02, percent-clipped=6.0 2023-04-27 01:01:51,798 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=57948.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:01:53,672 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6857, 2.3089, 1.7289, 1.6526, 1.2611, 1.2651, 1.7471, 1.2340], device='cuda:4'), covar=tensor([0.1802, 0.1454, 0.1545, 0.1909, 0.2546, 0.2125, 0.1106, 0.2172], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0215, 0.0170, 0.0204, 0.0204, 0.0185, 0.0159, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 01:02:11,760 INFO [finetune.py:976] (4/7) Epoch 11, batch 700, loss[loss=0.1688, simple_loss=0.2481, pruned_loss=0.04473, over 4817.00 frames. ], tot_loss[loss=0.1913, simple_loss=0.2589, pruned_loss=0.06182, over 926989.43 frames. ], batch size: 39, lr: 3.71e-03, grad_scale: 64.0 2023-04-27 01:02:33,860 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.9993, 3.9544, 2.8652, 4.6910, 4.0572, 3.9991, 1.9365, 3.9625], device='cuda:4'), covar=tensor([0.1726, 0.1116, 0.2978, 0.1437, 0.3241, 0.1908, 0.5447, 0.2382], device='cuda:4'), in_proj_covar=tensor([0.0240, 0.0214, 0.0247, 0.0300, 0.0295, 0.0248, 0.0266, 0.0268], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 01:02:57,543 INFO [finetune.py:976] (4/7) Epoch 11, batch 750, loss[loss=0.1525, simple_loss=0.2092, pruned_loss=0.04786, over 4047.00 frames. ], tot_loss[loss=0.194, simple_loss=0.2612, pruned_loss=0.06337, over 933877.22 frames. ], batch size: 17, lr: 3.71e-03, grad_scale: 32.0 2023-04-27 01:03:04,228 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.211e+02 1.766e+02 1.999e+02 2.458e+02 5.140e+02, threshold=3.998e+02, percent-clipped=2.0 2023-04-27 01:03:06,088 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.50 vs. limit=5.0 2023-04-27 01:03:14,995 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=58054.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:03:31,279 INFO [finetune.py:976] (4/7) Epoch 11, batch 800, loss[loss=0.2069, simple_loss=0.2801, pruned_loss=0.06682, over 4823.00 frames. ], tot_loss[loss=0.1917, simple_loss=0.2592, pruned_loss=0.0621, over 936808.70 frames. ], batch size: 41, lr: 3.71e-03, grad_scale: 32.0 2023-04-27 01:03:34,651 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-27 01:03:38,606 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=58089.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:03:47,436 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=58102.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:03:58,172 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=58118.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:04:04,031 INFO [finetune.py:976] (4/7) Epoch 11, batch 850, loss[loss=0.1862, simple_loss=0.249, pruned_loss=0.0617, over 4832.00 frames. ], tot_loss[loss=0.1915, simple_loss=0.2583, pruned_loss=0.06238, over 939036.76 frames. ], batch size: 47, lr: 3.71e-03, grad_scale: 32.0 2023-04-27 01:04:10,689 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.079e+02 1.590e+02 1.973e+02 2.582e+02 4.894e+02, threshold=3.945e+02, percent-clipped=3.0 2023-04-27 01:04:12,628 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=58141.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:04:14,921 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=58144.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:04:19,565 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=58151.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:04:39,246 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=58166.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:04:52,832 INFO [finetune.py:976] (4/7) Epoch 11, batch 900, loss[loss=0.1701, simple_loss=0.2318, pruned_loss=0.05425, over 4717.00 frames. ], tot_loss[loss=0.1897, simple_loss=0.2561, pruned_loss=0.06166, over 944373.53 frames. ], batch size: 23, lr: 3.71e-03, grad_scale: 32.0 2023-04-27 01:05:02,210 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.76 vs. limit=5.0 2023-04-27 01:05:18,668 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-27 01:05:19,788 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=58199.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:05:22,289 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=58202.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:05:28,584 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=58205.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:05:30,130 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.96 vs. limit=5.0 2023-04-27 01:05:42,974 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4902, 1.4083, 0.6797, 1.2223, 1.3681, 1.3927, 1.2953, 1.3004], device='cuda:4'), covar=tensor([0.0508, 0.0393, 0.0420, 0.0553, 0.0332, 0.0542, 0.0483, 0.0590], device='cuda:4'), in_proj_covar=tensor([0.0030, 0.0025, 0.0023, 0.0030, 0.0020, 0.0029, 0.0029, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0037, 0.0050, 0.0037, 0.0048, 0.0048, 0.0050], device='cuda:4') 2023-04-27 01:05:54,289 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5313, 2.0074, 2.4419, 2.9641, 2.3353, 1.8681, 1.7563, 2.1308], device='cuda:4'), covar=tensor([0.3662, 0.3652, 0.1586, 0.2800, 0.3176, 0.2930, 0.4355, 0.2585], device='cuda:4'), in_proj_covar=tensor([0.0284, 0.0248, 0.0221, 0.0316, 0.0213, 0.0227, 0.0232, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 01:05:54,762 INFO [finetune.py:976] (4/7) Epoch 11, batch 950, loss[loss=0.1435, simple_loss=0.2209, pruned_loss=0.03302, over 4875.00 frames. ], tot_loss[loss=0.1872, simple_loss=0.2535, pruned_loss=0.06041, over 948723.94 frames. ], batch size: 31, lr: 3.71e-03, grad_scale: 32.0 2023-04-27 01:06:10,243 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.115e+02 1.752e+02 2.086e+02 2.428e+02 7.149e+02, threshold=4.172e+02, percent-clipped=3.0 2023-04-27 01:06:47,387 INFO [finetune.py:976] (4/7) Epoch 11, batch 1000, loss[loss=0.2122, simple_loss=0.286, pruned_loss=0.0692, over 4751.00 frames. ], tot_loss[loss=0.189, simple_loss=0.2557, pruned_loss=0.06117, over 950809.19 frames. ], batch size: 59, lr: 3.71e-03, grad_scale: 32.0 2023-04-27 01:07:20,925 INFO [finetune.py:976] (4/7) Epoch 11, batch 1050, loss[loss=0.1782, simple_loss=0.2554, pruned_loss=0.05054, over 4896.00 frames. ], tot_loss[loss=0.1907, simple_loss=0.2582, pruned_loss=0.06157, over 953176.13 frames. ], batch size: 36, lr: 3.71e-03, grad_scale: 32.0 2023-04-27 01:07:28,201 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.162e+02 1.762e+02 2.120e+02 2.435e+02 5.017e+02, threshold=4.240e+02, percent-clipped=2.0 2023-04-27 01:07:37,004 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.43 vs. limit=5.0 2023-04-27 01:07:41,064 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4252, 1.1266, 1.2304, 1.1700, 1.5820, 1.3276, 1.1073, 1.1422], device='cuda:4'), covar=tensor([0.1581, 0.1512, 0.2076, 0.1561, 0.0863, 0.1464, 0.2025, 0.2181], device='cuda:4'), in_proj_covar=tensor([0.0306, 0.0321, 0.0352, 0.0298, 0.0335, 0.0320, 0.0306, 0.0362], device='cuda:4'), out_proj_covar=tensor([6.4240e-05, 6.7929e-05, 7.5606e-05, 6.1568e-05, 7.0034e-05, 6.8307e-05, 6.5496e-05, 7.7729e-05], device='cuda:4') 2023-04-27 01:07:52,945 INFO [finetune.py:976] (4/7) Epoch 11, batch 1100, loss[loss=0.1866, simple_loss=0.2638, pruned_loss=0.05469, over 4890.00 frames. ], tot_loss[loss=0.1914, simple_loss=0.259, pruned_loss=0.06193, over 952138.90 frames. ], batch size: 43, lr: 3.71e-03, grad_scale: 32.0 2023-04-27 01:08:01,690 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=58389.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:08:11,195 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-27 01:08:18,828 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8669, 2.6240, 1.7983, 1.7965, 1.2843, 1.3372, 1.9142, 1.2541], device='cuda:4'), covar=tensor([0.1664, 0.1288, 0.1559, 0.1811, 0.2425, 0.1970, 0.1043, 0.2066], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0214, 0.0169, 0.0204, 0.0203, 0.0184, 0.0158, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 01:08:26,955 INFO [finetune.py:976] (4/7) Epoch 11, batch 1150, loss[loss=0.177, simple_loss=0.2532, pruned_loss=0.05035, over 4807.00 frames. ], tot_loss[loss=0.1922, simple_loss=0.2601, pruned_loss=0.06215, over 954565.11 frames. ], batch size: 40, lr: 3.71e-03, grad_scale: 32.0 2023-04-27 01:08:34,545 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=58437.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:08:34,617 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7257, 1.3014, 1.4027, 1.4664, 1.9348, 1.5410, 1.2116, 1.3020], device='cuda:4'), covar=tensor([0.1687, 0.1467, 0.1871, 0.1223, 0.0802, 0.1556, 0.2341, 0.2213], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0323, 0.0354, 0.0300, 0.0337, 0.0322, 0.0309, 0.0364], device='cuda:4'), out_proj_covar=tensor([6.4787e-05, 6.8291e-05, 7.6149e-05, 6.1830e-05, 7.0516e-05, 6.8865e-05, 6.6060e-05, 7.8282e-05], device='cuda:4') 2023-04-27 01:08:35,082 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.202e+02 1.681e+02 2.024e+02 2.460e+02 8.000e+02, threshold=4.047e+02, percent-clipped=3.0 2023-04-27 01:08:46,851 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4815, 1.6887, 0.5818, 1.2028, 1.4368, 1.3379, 1.2610, 1.2917], device='cuda:4'), covar=tensor([0.0631, 0.0328, 0.0435, 0.0632, 0.0319, 0.0698, 0.0679, 0.0688], device='cuda:4'), in_proj_covar=tensor([0.0030, 0.0025, 0.0023, 0.0030, 0.0020, 0.0029, 0.0029, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0045, 0.0038, 0.0050, 0.0038, 0.0048, 0.0048, 0.0050], device='cuda:4') 2023-04-27 01:08:59,897 INFO [finetune.py:976] (4/7) Epoch 11, batch 1200, loss[loss=0.1823, simple_loss=0.246, pruned_loss=0.05931, over 4812.00 frames. ], tot_loss[loss=0.1905, simple_loss=0.2579, pruned_loss=0.0615, over 953594.18 frames. ], batch size: 38, lr: 3.71e-03, grad_scale: 32.0 2023-04-27 01:09:00,036 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5271, 1.6835, 1.3453, 0.9675, 1.1854, 1.1143, 1.3851, 1.1346], device='cuda:4'), covar=tensor([0.1821, 0.1369, 0.1726, 0.2058, 0.2475, 0.2111, 0.1208, 0.2149], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0214, 0.0169, 0.0203, 0.0203, 0.0184, 0.0158, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 01:09:08,006 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2657, 1.7615, 2.0520, 2.4840, 2.0208, 1.6505, 1.2514, 1.7884], device='cuda:4'), covar=tensor([0.3672, 0.3369, 0.1859, 0.2768, 0.3498, 0.3034, 0.5238, 0.2507], device='cuda:4'), in_proj_covar=tensor([0.0284, 0.0248, 0.0221, 0.0317, 0.0213, 0.0227, 0.0233, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 01:09:13,882 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=58497.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:09:15,684 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=58500.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:09:29,410 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.6825, 1.7702, 1.6847, 1.2893, 1.6887, 1.4624, 2.2198, 1.2435], device='cuda:4'), covar=tensor([0.3079, 0.1280, 0.3667, 0.2311, 0.1365, 0.1839, 0.1130, 0.4240], device='cuda:4'), in_proj_covar=tensor([0.0342, 0.0346, 0.0426, 0.0358, 0.0385, 0.0381, 0.0379, 0.0418], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 01:09:32,953 INFO [finetune.py:976] (4/7) Epoch 11, batch 1250, loss[loss=0.2082, simple_loss=0.2722, pruned_loss=0.07204, over 4121.00 frames. ], tot_loss[loss=0.1894, simple_loss=0.2561, pruned_loss=0.06134, over 954021.77 frames. ], batch size: 65, lr: 3.71e-03, grad_scale: 32.0 2023-04-27 01:09:41,127 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.214e+02 1.641e+02 1.959e+02 2.303e+02 4.487e+02, threshold=3.918e+02, percent-clipped=1.0 2023-04-27 01:10:09,830 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7528, 1.4686, 1.9666, 2.2566, 1.9197, 1.8001, 1.8814, 1.8770], device='cuda:4'), covar=tensor([0.5780, 0.8065, 0.8571, 0.7820, 0.7257, 1.0418, 0.9941, 0.9170], device='cuda:4'), in_proj_covar=tensor([0.0406, 0.0410, 0.0496, 0.0515, 0.0437, 0.0455, 0.0465, 0.0465], device='cuda:4'), out_proj_covar=tensor([9.8817e-05, 1.0169e-04, 1.1178e-04, 1.2235e-04, 1.0562e-04, 1.0997e-04, 1.1137e-04, 1.1162e-04], device='cuda:4') 2023-04-27 01:10:22,918 INFO [finetune.py:976] (4/7) Epoch 11, batch 1300, loss[loss=0.2119, simple_loss=0.2769, pruned_loss=0.07343, over 4823.00 frames. ], tot_loss[loss=0.1867, simple_loss=0.253, pruned_loss=0.06021, over 954523.08 frames. ], batch size: 40, lr: 3.71e-03, grad_scale: 32.0 2023-04-27 01:10:55,523 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3980, 3.4801, 0.8634, 1.8108, 1.8321, 2.5110, 1.9673, 1.0458], device='cuda:4'), covar=tensor([0.1355, 0.0854, 0.2036, 0.1263, 0.1078, 0.0975, 0.1497, 0.2060], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0247, 0.0139, 0.0122, 0.0132, 0.0152, 0.0117, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 01:11:29,576 INFO [finetune.py:976] (4/7) Epoch 11, batch 1350, loss[loss=0.2003, simple_loss=0.2701, pruned_loss=0.06524, over 4823.00 frames. ], tot_loss[loss=0.1864, simple_loss=0.2528, pruned_loss=0.06003, over 956300.48 frames. ], batch size: 39, lr: 3.71e-03, grad_scale: 32.0 2023-04-27 01:11:46,674 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.165e+02 1.586e+02 1.956e+02 2.328e+02 4.668e+02, threshold=3.913e+02, percent-clipped=1.0 2023-04-27 01:12:32,291 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.82 vs. limit=2.0 2023-04-27 01:12:34,429 INFO [finetune.py:976] (4/7) Epoch 11, batch 1400, loss[loss=0.2319, simple_loss=0.2983, pruned_loss=0.08276, over 4905.00 frames. ], tot_loss[loss=0.189, simple_loss=0.256, pruned_loss=0.06098, over 957074.00 frames. ], batch size: 35, lr: 3.71e-03, grad_scale: 32.0 2023-04-27 01:12:42,446 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=58681.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:13:42,639 INFO [finetune.py:976] (4/7) Epoch 11, batch 1450, loss[loss=0.2966, simple_loss=0.3362, pruned_loss=0.1285, over 4907.00 frames. ], tot_loss[loss=0.1922, simple_loss=0.2591, pruned_loss=0.06264, over 957368.32 frames. ], batch size: 37, lr: 3.71e-03, grad_scale: 32.0 2023-04-27 01:13:52,068 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0232, 2.2904, 1.3908, 1.8089, 2.3417, 1.9030, 1.8646, 1.9636], device='cuda:4'), covar=tensor([0.0470, 0.0328, 0.0325, 0.0507, 0.0229, 0.0466, 0.0469, 0.0529], device='cuda:4'), in_proj_covar=tensor([0.0030, 0.0025, 0.0023, 0.0029, 0.0020, 0.0029, 0.0029, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0050, 0.0037, 0.0048, 0.0048, 0.0049], device='cuda:4') 2023-04-27 01:14:01,639 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.919e+01 1.615e+02 1.986e+02 2.296e+02 3.609e+02, threshold=3.972e+02, percent-clipped=0.0 2023-04-27 01:14:05,186 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=58742.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:14:37,143 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=58775.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:14:38,276 INFO [finetune.py:976] (4/7) Epoch 11, batch 1500, loss[loss=0.1892, simple_loss=0.2554, pruned_loss=0.06154, over 4179.00 frames. ], tot_loss[loss=0.1921, simple_loss=0.2593, pruned_loss=0.06243, over 956023.16 frames. ], batch size: 65, lr: 3.70e-03, grad_scale: 32.0 2023-04-27 01:14:52,525 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=58797.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:14:54,839 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3315, 3.2890, 0.9144, 1.7436, 1.8083, 2.3534, 1.8388, 1.0736], device='cuda:4'), covar=tensor([0.1499, 0.0940, 0.2132, 0.1330, 0.1177, 0.1050, 0.1643, 0.2005], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0247, 0.0139, 0.0122, 0.0133, 0.0152, 0.0118, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 01:14:54,854 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=58800.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:14:56,670 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4994, 3.0549, 1.0532, 1.7014, 2.2937, 1.5944, 4.6532, 2.2102], device='cuda:4'), covar=tensor([0.0659, 0.0745, 0.0863, 0.1411, 0.0563, 0.1053, 0.0198, 0.0633], device='cuda:4'), in_proj_covar=tensor([0.0052, 0.0067, 0.0050, 0.0047, 0.0051, 0.0053, 0.0078, 0.0052], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0011, 0.0008], device='cuda:4') 2023-04-27 01:15:08,109 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6259, 1.6273, 0.7862, 1.3037, 1.8521, 1.4992, 1.4023, 1.4384], device='cuda:4'), covar=tensor([0.0518, 0.0392, 0.0386, 0.0597, 0.0270, 0.0567, 0.0549, 0.0615], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0025, 0.0023, 0.0029, 0.0020, 0.0029, 0.0029, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0050, 0.0037, 0.0048, 0.0048, 0.0049], device='cuda:4') 2023-04-27 01:15:11,710 INFO [finetune.py:976] (4/7) Epoch 11, batch 1550, loss[loss=0.1609, simple_loss=0.2321, pruned_loss=0.04481, over 4806.00 frames. ], tot_loss[loss=0.1886, simple_loss=0.2566, pruned_loss=0.06032, over 955889.24 frames. ], batch size: 41, lr: 3.70e-03, grad_scale: 32.0 2023-04-27 01:15:18,220 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=58836.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:15:19,335 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.373e+01 1.631e+02 1.941e+02 2.305e+02 6.407e+02, threshold=3.882e+02, percent-clipped=1.0 2023-04-27 01:15:19,438 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6221, 1.4304, 4.5135, 4.2034, 3.9321, 4.3052, 4.1915, 3.9810], device='cuda:4'), covar=tensor([0.6983, 0.6417, 0.1059, 0.1782, 0.1160, 0.2350, 0.1328, 0.1482], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0306, 0.0402, 0.0406, 0.0348, 0.0404, 0.0312, 0.0367], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 01:15:24,657 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=58845.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:15:26,502 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=58848.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:15:44,973 INFO [finetune.py:976] (4/7) Epoch 11, batch 1600, loss[loss=0.1644, simple_loss=0.2363, pruned_loss=0.04626, over 4809.00 frames. ], tot_loss[loss=0.1871, simple_loss=0.2548, pruned_loss=0.0597, over 955340.91 frames. ], batch size: 29, lr: 3.70e-03, grad_scale: 32.0 2023-04-27 01:16:05,010 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=58906.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:16:18,089 INFO [finetune.py:976] (4/7) Epoch 11, batch 1650, loss[loss=0.1403, simple_loss=0.2035, pruned_loss=0.03856, over 4803.00 frames. ], tot_loss[loss=0.1849, simple_loss=0.2519, pruned_loss=0.05897, over 957002.19 frames. ], batch size: 25, lr: 3.70e-03, grad_scale: 32.0 2023-04-27 01:16:18,899 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.65 vs. limit=2.0 2023-04-27 01:16:25,722 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.044e+02 1.662e+02 1.988e+02 2.360e+02 4.922e+02, threshold=3.975e+02, percent-clipped=3.0 2023-04-27 01:16:44,581 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0466, 2.6886, 2.1251, 2.4295, 1.8421, 2.2124, 2.2264, 1.6511], device='cuda:4'), covar=tensor([0.2114, 0.1235, 0.0885, 0.1506, 0.3162, 0.1389, 0.2023, 0.2732], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0313, 0.0227, 0.0282, 0.0311, 0.0267, 0.0251, 0.0273], device='cuda:4'), out_proj_covar=tensor([1.1812e-04, 1.2568e-04, 9.1132e-05, 1.1304e-04, 1.2722e-04, 1.0753e-04, 1.0248e-04, 1.0953e-04], device='cuda:4') 2023-04-27 01:17:06,253 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8085, 1.4152, 1.9622, 2.2855, 1.9046, 1.7260, 1.8328, 1.8981], device='cuda:4'), covar=tensor([0.5744, 0.8169, 0.8192, 0.7769, 0.7553, 1.0505, 1.0121, 0.8763], device='cuda:4'), in_proj_covar=tensor([0.0408, 0.0412, 0.0498, 0.0517, 0.0439, 0.0457, 0.0468, 0.0467], device='cuda:4'), out_proj_covar=tensor([9.9374e-05, 1.0213e-04, 1.1239e-04, 1.2277e-04, 1.0623e-04, 1.1054e-04, 1.1222e-04, 1.1217e-04], device='cuda:4') 2023-04-27 01:17:07,486 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=58967.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:17:13,434 INFO [finetune.py:976] (4/7) Epoch 11, batch 1700, loss[loss=0.2397, simple_loss=0.2928, pruned_loss=0.09331, over 4821.00 frames. ], tot_loss[loss=0.1836, simple_loss=0.2501, pruned_loss=0.05852, over 957095.82 frames. ], batch size: 40, lr: 3.70e-03, grad_scale: 32.0 2023-04-27 01:17:41,004 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9192, 1.6174, 5.3943, 5.0489, 4.6601, 5.0989, 4.7365, 4.7376], device='cuda:4'), covar=tensor([0.6965, 0.5895, 0.0732, 0.1474, 0.1034, 0.1956, 0.1135, 0.1465], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0305, 0.0402, 0.0406, 0.0349, 0.0405, 0.0312, 0.0366], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 01:17:41,015 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8206, 4.4631, 0.8218, 2.3199, 2.6178, 2.7682, 2.5266, 1.0205], device='cuda:4'), covar=tensor([0.1363, 0.0941, 0.2117, 0.1180, 0.0921, 0.1153, 0.1383, 0.2134], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0249, 0.0141, 0.0122, 0.0134, 0.0154, 0.0119, 0.0123], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 01:17:47,002 INFO [finetune.py:976] (4/7) Epoch 11, batch 1750, loss[loss=0.2526, simple_loss=0.3059, pruned_loss=0.09961, over 4858.00 frames. ], tot_loss[loss=0.1851, simple_loss=0.2516, pruned_loss=0.05926, over 954884.82 frames. ], batch size: 44, lr: 3.70e-03, grad_scale: 32.0 2023-04-27 01:17:53,223 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=59037.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:17:53,741 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.098e+02 1.792e+02 2.172e+02 2.715e+02 5.550e+02, threshold=4.344e+02, percent-clipped=4.0 2023-04-27 01:18:18,316 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5853, 1.4347, 1.6192, 1.9669, 2.0343, 1.5215, 1.1158, 1.7925], device='cuda:4'), covar=tensor([0.0870, 0.1148, 0.0762, 0.0573, 0.0544, 0.0819, 0.0890, 0.0537], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0204, 0.0183, 0.0176, 0.0179, 0.0189, 0.0160, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 01:18:30,829 INFO [finetune.py:976] (4/7) Epoch 11, batch 1800, loss[loss=0.2132, simple_loss=0.2781, pruned_loss=0.07419, over 4186.00 frames. ], tot_loss[loss=0.1887, simple_loss=0.2564, pruned_loss=0.06049, over 954391.09 frames. ], batch size: 65, lr: 3.70e-03, grad_scale: 32.0 2023-04-27 01:19:09,377 INFO [finetune.py:976] (4/7) Epoch 11, batch 1850, loss[loss=0.1593, simple_loss=0.2409, pruned_loss=0.03881, over 4791.00 frames. ], tot_loss[loss=0.1913, simple_loss=0.2591, pruned_loss=0.0617, over 955994.34 frames. ], batch size: 51, lr: 3.70e-03, grad_scale: 32.0 2023-04-27 01:19:11,899 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=59131.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:19:21,290 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.048e+02 1.699e+02 2.019e+02 2.381e+02 7.878e+02, threshold=4.039e+02, percent-clipped=1.0 2023-04-27 01:19:30,567 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6013, 1.2518, 1.8082, 2.0496, 1.7221, 1.5673, 1.6626, 1.6654], device='cuda:4'), covar=tensor([0.5537, 0.7326, 0.7400, 0.7341, 0.6797, 0.9123, 0.9168, 0.8894], device='cuda:4'), in_proj_covar=tensor([0.0409, 0.0411, 0.0499, 0.0517, 0.0438, 0.0458, 0.0468, 0.0467], device='cuda:4'), out_proj_covar=tensor([9.9528e-05, 1.0196e-04, 1.1254e-04, 1.2269e-04, 1.0610e-04, 1.1069e-04, 1.1221e-04, 1.1228e-04], device='cuda:4') 2023-04-27 01:19:30,856 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-27 01:19:52,230 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=59164.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:20:10,991 INFO [finetune.py:976] (4/7) Epoch 11, batch 1900, loss[loss=0.2308, simple_loss=0.2747, pruned_loss=0.09348, over 4138.00 frames. ], tot_loss[loss=0.1929, simple_loss=0.2612, pruned_loss=0.06231, over 957437.96 frames. ], batch size: 65, lr: 3.70e-03, grad_scale: 32.0 2023-04-27 01:21:00,135 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=59225.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:21:01,214 INFO [finetune.py:976] (4/7) Epoch 11, batch 1950, loss[loss=0.1746, simple_loss=0.2422, pruned_loss=0.05352, over 4886.00 frames. ], tot_loss[loss=0.1905, simple_loss=0.2587, pruned_loss=0.0612, over 957216.20 frames. ], batch size: 32, lr: 3.70e-03, grad_scale: 32.0 2023-04-27 01:21:06,098 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3885, 1.5827, 1.7153, 1.8484, 1.6766, 1.8650, 1.8895, 1.7938], device='cuda:4'), covar=tensor([0.4424, 0.6265, 0.5499, 0.5127, 0.6010, 0.8134, 0.5894, 0.6004], device='cuda:4'), in_proj_covar=tensor([0.0327, 0.0380, 0.0315, 0.0326, 0.0339, 0.0400, 0.0359, 0.0323], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 01:21:08,361 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.600e+02 1.889e+02 2.365e+02 4.581e+02, threshold=3.778e+02, percent-clipped=1.0 2023-04-27 01:21:20,314 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-04-27 01:21:24,550 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=59262.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:21:35,025 INFO [finetune.py:976] (4/7) Epoch 11, batch 2000, loss[loss=0.1818, simple_loss=0.2434, pruned_loss=0.06011, over 4827.00 frames. ], tot_loss[loss=0.1882, simple_loss=0.2561, pruned_loss=0.06017, over 957282.00 frames. ], batch size: 33, lr: 3.70e-03, grad_scale: 32.0 2023-04-27 01:22:08,861 INFO [finetune.py:976] (4/7) Epoch 11, batch 2050, loss[loss=0.1818, simple_loss=0.2392, pruned_loss=0.06221, over 4145.00 frames. ], tot_loss[loss=0.1858, simple_loss=0.2528, pruned_loss=0.05943, over 955775.81 frames. ], batch size: 65, lr: 3.70e-03, grad_scale: 32.0 2023-04-27 01:22:08,935 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5931, 1.2553, 4.1266, 3.8569, 3.6368, 3.8366, 3.8184, 3.6508], device='cuda:4'), covar=tensor([0.6761, 0.5955, 0.1005, 0.1703, 0.1073, 0.2042, 0.2102, 0.1516], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0304, 0.0398, 0.0404, 0.0345, 0.0403, 0.0312, 0.0365], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 01:22:15,383 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=59337.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:22:15,895 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.113e+02 1.589e+02 2.008e+02 2.324e+02 3.763e+02, threshold=4.015e+02, percent-clipped=0.0 2023-04-27 01:22:43,270 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7435, 2.3933, 1.6827, 1.5167, 1.2438, 1.2618, 1.7997, 1.1786], device='cuda:4'), covar=tensor([0.1815, 0.1398, 0.1668, 0.2074, 0.2546, 0.2117, 0.1074, 0.2234], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0214, 0.0169, 0.0204, 0.0203, 0.0185, 0.0159, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 01:22:48,352 INFO [finetune.py:976] (4/7) Epoch 11, batch 2100, loss[loss=0.2694, simple_loss=0.3031, pruned_loss=0.1178, over 4856.00 frames. ], tot_loss[loss=0.1868, simple_loss=0.2529, pruned_loss=0.06032, over 955469.75 frames. ], batch size: 49, lr: 3.70e-03, grad_scale: 32.0 2023-04-27 01:22:53,372 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=59385.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:23:32,964 INFO [finetune.py:976] (4/7) Epoch 11, batch 2150, loss[loss=0.172, simple_loss=0.255, pruned_loss=0.04457, over 4908.00 frames. ], tot_loss[loss=0.1893, simple_loss=0.256, pruned_loss=0.06134, over 953708.37 frames. ], batch size: 35, lr: 3.70e-03, grad_scale: 32.0 2023-04-27 01:23:35,533 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=59431.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:23:44,712 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.256e+02 1.698e+02 2.142e+02 2.512e+02 4.265e+02, threshold=4.284e+02, percent-clipped=3.0 2023-04-27 01:24:15,855 INFO [finetune.py:976] (4/7) Epoch 11, batch 2200, loss[loss=0.2246, simple_loss=0.2854, pruned_loss=0.0819, over 4912.00 frames. ], tot_loss[loss=0.1919, simple_loss=0.2588, pruned_loss=0.06249, over 952419.71 frames. ], batch size: 37, lr: 3.70e-03, grad_scale: 32.0 2023-04-27 01:24:18,111 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=59479.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:24:37,674 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=59502.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:24:45,229 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6596, 1.9256, 0.8464, 1.4155, 2.2656, 1.5722, 1.5577, 1.5539], device='cuda:4'), covar=tensor([0.0512, 0.0365, 0.0349, 0.0551, 0.0243, 0.0540, 0.0490, 0.0604], device='cuda:4'), in_proj_covar=tensor([0.0030, 0.0025, 0.0023, 0.0030, 0.0020, 0.0029, 0.0029, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0037, 0.0050, 0.0037, 0.0048, 0.0048, 0.0050], device='cuda:4') 2023-04-27 01:24:54,707 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=59520.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:24:59,359 INFO [finetune.py:976] (4/7) Epoch 11, batch 2250, loss[loss=0.2018, simple_loss=0.2754, pruned_loss=0.06408, over 4830.00 frames. ], tot_loss[loss=0.1927, simple_loss=0.2595, pruned_loss=0.06292, over 954342.95 frames. ], batch size: 30, lr: 3.70e-03, grad_scale: 32.0 2023-04-27 01:25:13,215 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.038e+02 1.641e+02 2.091e+02 2.357e+02 4.911e+02, threshold=4.183e+02, percent-clipped=2.0 2023-04-27 01:25:45,852 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=59562.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:25:46,493 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=59563.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:25:53,409 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3094, 1.6588, 1.6269, 2.1094, 2.3671, 1.9915, 1.9423, 1.6731], device='cuda:4'), covar=tensor([0.1477, 0.1550, 0.1743, 0.1512, 0.1060, 0.1802, 0.2018, 0.1839], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0322, 0.0353, 0.0300, 0.0336, 0.0320, 0.0307, 0.0363], device='cuda:4'), out_proj_covar=tensor([6.4552e-05, 6.7989e-05, 7.5891e-05, 6.1878e-05, 7.0355e-05, 6.8417e-05, 6.5762e-05, 7.7936e-05], device='cuda:4') 2023-04-27 01:25:55,720 INFO [finetune.py:976] (4/7) Epoch 11, batch 2300, loss[loss=0.1932, simple_loss=0.263, pruned_loss=0.06172, over 4889.00 frames. ], tot_loss[loss=0.1911, simple_loss=0.2587, pruned_loss=0.06179, over 953831.61 frames. ], batch size: 43, lr: 3.70e-03, grad_scale: 32.0 2023-04-27 01:26:02,339 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-04-27 01:26:17,898 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=59610.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:26:29,150 INFO [finetune.py:976] (4/7) Epoch 11, batch 2350, loss[loss=0.1606, simple_loss=0.2381, pruned_loss=0.04159, over 4900.00 frames. ], tot_loss[loss=0.1892, simple_loss=0.2567, pruned_loss=0.06086, over 954544.94 frames. ], batch size: 46, lr: 3.70e-03, grad_scale: 32.0 2023-04-27 01:26:37,790 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.079e+02 1.642e+02 1.943e+02 2.312e+02 3.854e+02, threshold=3.885e+02, percent-clipped=0.0 2023-04-27 01:26:58,635 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3839, 1.7409, 2.2222, 2.5328, 2.0442, 1.7261, 1.3082, 1.9384], device='cuda:4'), covar=tensor([0.3579, 0.3612, 0.1823, 0.2930, 0.3143, 0.3006, 0.5104, 0.2497], device='cuda:4'), in_proj_covar=tensor([0.0286, 0.0249, 0.0223, 0.0317, 0.0216, 0.0228, 0.0233, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 01:27:02,578 INFO [finetune.py:976] (4/7) Epoch 11, batch 2400, loss[loss=0.1737, simple_loss=0.2329, pruned_loss=0.05721, over 4665.00 frames. ], tot_loss[loss=0.1865, simple_loss=0.2533, pruned_loss=0.05989, over 954013.37 frames. ], batch size: 23, lr: 3.70e-03, grad_scale: 32.0 2023-04-27 01:27:22,874 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.7359, 1.8067, 2.1957, 3.0625, 3.0159, 2.3755, 1.9234, 2.8623], device='cuda:4'), covar=tensor([0.1022, 0.1702, 0.1015, 0.0707, 0.0618, 0.1121, 0.1106, 0.0662], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0204, 0.0184, 0.0176, 0.0179, 0.0189, 0.0161, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 01:27:36,063 INFO [finetune.py:976] (4/7) Epoch 11, batch 2450, loss[loss=0.1868, simple_loss=0.2506, pruned_loss=0.06156, over 4769.00 frames. ], tot_loss[loss=0.1838, simple_loss=0.2501, pruned_loss=0.0588, over 951994.21 frames. ], batch size: 26, lr: 3.70e-03, grad_scale: 32.0 2023-04-27 01:27:43,829 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.154e+02 1.580e+02 1.900e+02 2.443e+02 4.290e+02, threshold=3.800e+02, percent-clipped=1.0 2023-04-27 01:28:09,960 INFO [finetune.py:976] (4/7) Epoch 11, batch 2500, loss[loss=0.2204, simple_loss=0.2552, pruned_loss=0.09275, over 4116.00 frames. ], tot_loss[loss=0.1884, simple_loss=0.2546, pruned_loss=0.06106, over 952162.73 frames. ], batch size: 18, lr: 3.70e-03, grad_scale: 32.0 2023-04-27 01:28:44,933 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=59820.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:28:54,835 INFO [finetune.py:976] (4/7) Epoch 11, batch 2550, loss[loss=0.1777, simple_loss=0.2491, pruned_loss=0.05321, over 4745.00 frames. ], tot_loss[loss=0.191, simple_loss=0.2585, pruned_loss=0.06176, over 953260.25 frames. ], batch size: 59, lr: 3.70e-03, grad_scale: 32.0 2023-04-27 01:29:12,510 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.167e+02 1.688e+02 2.032e+02 2.494e+02 8.629e+02, threshold=4.065e+02, percent-clipped=5.0 2023-04-27 01:29:23,388 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=59848.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:29:35,317 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=59858.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:29:46,960 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=59868.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:29:57,641 INFO [finetune.py:976] (4/7) Epoch 11, batch 2600, loss[loss=0.2026, simple_loss=0.2648, pruned_loss=0.07024, over 4823.00 frames. ], tot_loss[loss=0.1926, simple_loss=0.2604, pruned_loss=0.0624, over 952638.97 frames. ], batch size: 33, lr: 3.70e-03, grad_scale: 32.0 2023-04-27 01:30:10,824 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=59889.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:30:29,481 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=59909.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:30:51,536 INFO [finetune.py:976] (4/7) Epoch 11, batch 2650, loss[loss=0.2172, simple_loss=0.2923, pruned_loss=0.07102, over 4705.00 frames. ], tot_loss[loss=0.1923, simple_loss=0.2606, pruned_loss=0.06197, over 954590.29 frames. ], batch size: 59, lr: 3.70e-03, grad_scale: 32.0 2023-04-27 01:31:02,810 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5626, 0.9989, 1.6181, 1.9802, 1.6557, 1.5629, 1.5873, 1.6476], device='cuda:4'), covar=tensor([0.5499, 0.7806, 0.7909, 0.7862, 0.6939, 0.9148, 0.9194, 0.8514], device='cuda:4'), in_proj_covar=tensor([0.0408, 0.0411, 0.0498, 0.0517, 0.0439, 0.0458, 0.0467, 0.0468], device='cuda:4'), out_proj_covar=tensor([9.9276e-05, 1.0200e-04, 1.1233e-04, 1.2285e-04, 1.0635e-04, 1.1071e-04, 1.1192e-04, 1.1227e-04], device='cuda:4') 2023-04-27 01:31:04,460 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.138e+02 1.693e+02 1.978e+02 2.480e+02 4.060e+02, threshold=3.956e+02, percent-clipped=0.0 2023-04-27 01:31:21,126 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=59950.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:31:43,485 INFO [finetune.py:976] (4/7) Epoch 11, batch 2700, loss[loss=0.1539, simple_loss=0.2293, pruned_loss=0.03926, over 4911.00 frames. ], tot_loss[loss=0.1899, simple_loss=0.2582, pruned_loss=0.06075, over 955390.89 frames. ], batch size: 36, lr: 3.70e-03, grad_scale: 32.0 2023-04-27 01:31:44,291 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-27 01:32:17,132 INFO [finetune.py:976] (4/7) Epoch 11, batch 2750, loss[loss=0.1928, simple_loss=0.2519, pruned_loss=0.06685, over 4829.00 frames. ], tot_loss[loss=0.1874, simple_loss=0.2552, pruned_loss=0.05984, over 954487.06 frames. ], batch size: 25, lr: 3.69e-03, grad_scale: 64.0 2023-04-27 01:32:24,354 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.196e+02 1.624e+02 1.944e+02 2.490e+02 4.678e+02, threshold=3.889e+02, percent-clipped=1.0 2023-04-27 01:32:25,706 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=60040.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:32:33,993 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=60052.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:32:42,797 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=60065.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:32:50,411 INFO [finetune.py:976] (4/7) Epoch 11, batch 2800, loss[loss=0.1864, simple_loss=0.2466, pruned_loss=0.06314, over 4836.00 frames. ], tot_loss[loss=0.1868, simple_loss=0.2533, pruned_loss=0.06015, over 955768.56 frames. ], batch size: 30, lr: 3.69e-03, grad_scale: 64.0 2023-04-27 01:32:51,092 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=60078.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:33:05,640 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=60101.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 01:33:14,375 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=60113.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 01:33:22,793 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=60126.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:33:23,284 INFO [finetune.py:976] (4/7) Epoch 11, batch 2850, loss[loss=0.166, simple_loss=0.23, pruned_loss=0.05105, over 4869.00 frames. ], tot_loss[loss=0.187, simple_loss=0.2528, pruned_loss=0.06057, over 956077.34 frames. ], batch size: 31, lr: 3.69e-03, grad_scale: 64.0 2023-04-27 01:33:29,990 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.009e+02 1.671e+02 1.998e+02 2.360e+02 5.267e+02, threshold=3.997e+02, percent-clipped=5.0 2023-04-27 01:33:30,724 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=60139.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:33:43,074 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=60158.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:33:55,935 INFO [finetune.py:976] (4/7) Epoch 11, batch 2900, loss[loss=0.2489, simple_loss=0.3149, pruned_loss=0.09147, over 4799.00 frames. ], tot_loss[loss=0.1904, simple_loss=0.2567, pruned_loss=0.06202, over 955513.23 frames. ], batch size: 45, lr: 3.69e-03, grad_scale: 64.0 2023-04-27 01:34:15,784 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=60198.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:34:25,153 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=60204.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:34:26,337 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=60206.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:34:59,279 INFO [finetune.py:976] (4/7) Epoch 11, batch 2950, loss[loss=0.1675, simple_loss=0.2362, pruned_loss=0.04937, over 4890.00 frames. ], tot_loss[loss=0.1922, simple_loss=0.2591, pruned_loss=0.06266, over 956178.78 frames. ], batch size: 32, lr: 3.69e-03, grad_scale: 32.0 2023-04-27 01:35:12,693 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.759e+01 1.634e+02 1.896e+02 2.531e+02 4.222e+02, threshold=3.792e+02, percent-clipped=1.0 2023-04-27 01:35:21,697 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=60245.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:35:41,933 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=60259.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:36:05,415 INFO [finetune.py:976] (4/7) Epoch 11, batch 3000, loss[loss=0.2184, simple_loss=0.2871, pruned_loss=0.07488, over 4920.00 frames. ], tot_loss[loss=0.1936, simple_loss=0.2605, pruned_loss=0.06338, over 953575.17 frames. ], batch size: 33, lr: 3.69e-03, grad_scale: 32.0 2023-04-27 01:36:05,416 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 01:36:12,168 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0778, 1.6339, 1.8586, 2.1937, 1.8959, 1.5364, 1.2217, 1.6774], device='cuda:4'), covar=tensor([0.2767, 0.3233, 0.1620, 0.2083, 0.2644, 0.2581, 0.4499, 0.2341], device='cuda:4'), in_proj_covar=tensor([0.0283, 0.0247, 0.0220, 0.0314, 0.0213, 0.0227, 0.0230, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 01:36:27,793 INFO [finetune.py:1010] (4/7) Epoch 11, validation: loss=0.1531, simple_loss=0.2255, pruned_loss=0.04032, over 2265189.00 frames. 2023-04-27 01:36:27,793 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-27 01:36:58,615 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3758, 1.5638, 1.6951, 1.8553, 1.6653, 1.7924, 1.7107, 1.6890], device='cuda:4'), covar=tensor([0.4645, 0.6553, 0.5383, 0.4898, 0.6392, 0.8829, 0.7256, 0.6062], device='cuda:4'), in_proj_covar=tensor([0.0327, 0.0380, 0.0315, 0.0326, 0.0340, 0.0400, 0.0359, 0.0322], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 01:37:01,198 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.50 vs. limit=2.0 2023-04-27 01:37:31,378 INFO [finetune.py:976] (4/7) Epoch 11, batch 3050, loss[loss=0.1504, simple_loss=0.2178, pruned_loss=0.04153, over 4918.00 frames. ], tot_loss[loss=0.1926, simple_loss=0.26, pruned_loss=0.06263, over 952672.23 frames. ], batch size: 33, lr: 3.69e-03, grad_scale: 32.0 2023-04-27 01:37:45,692 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.005e+02 1.734e+02 2.169e+02 2.489e+02 4.339e+02, threshold=4.338e+02, percent-clipped=2.0 2023-04-27 01:38:26,536 INFO [finetune.py:976] (4/7) Epoch 11, batch 3100, loss[loss=0.1783, simple_loss=0.248, pruned_loss=0.05431, over 4861.00 frames. ], tot_loss[loss=0.1911, simple_loss=0.2579, pruned_loss=0.0621, over 953353.28 frames. ], batch size: 34, lr: 3.69e-03, grad_scale: 32.0 2023-04-27 01:38:34,617 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.99 vs. limit=2.0 2023-04-27 01:38:40,717 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=60396.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 01:38:48,256 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=60408.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 01:38:56,665 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=60421.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:39:00,270 INFO [finetune.py:976] (4/7) Epoch 11, batch 3150, loss[loss=0.164, simple_loss=0.2362, pruned_loss=0.04589, over 4762.00 frames. ], tot_loss[loss=0.1885, simple_loss=0.2551, pruned_loss=0.06095, over 955295.77 frames. ], batch size: 27, lr: 3.69e-03, grad_scale: 32.0 2023-04-27 01:39:06,460 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=60434.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:39:09,791 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.219e+02 1.640e+02 1.939e+02 2.354e+02 4.857e+02, threshold=3.879e+02, percent-clipped=2.0 2023-04-27 01:39:24,476 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3972, 1.5299, 1.7707, 1.9536, 1.8197, 1.9259, 1.8161, 1.7878], device='cuda:4'), covar=tensor([0.4611, 0.7137, 0.5495, 0.5553, 0.6751, 0.9128, 0.7086, 0.6278], device='cuda:4'), in_proj_covar=tensor([0.0324, 0.0378, 0.0314, 0.0325, 0.0338, 0.0397, 0.0356, 0.0320], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 01:39:25,096 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8919, 1.4850, 1.9935, 2.3952, 2.0273, 1.8103, 1.9327, 1.8910], device='cuda:4'), covar=tensor([0.5273, 0.7478, 0.7553, 0.6898, 0.6728, 0.8860, 0.8602, 0.9815], device='cuda:4'), in_proj_covar=tensor([0.0405, 0.0407, 0.0495, 0.0512, 0.0436, 0.0455, 0.0464, 0.0464], device='cuda:4'), out_proj_covar=tensor([9.8479e-05, 1.0105e-04, 1.1157e-04, 1.2155e-04, 1.0568e-04, 1.1009e-04, 1.1113e-04, 1.1129e-04], device='cuda:4') 2023-04-27 01:39:33,961 INFO [finetune.py:976] (4/7) Epoch 11, batch 3200, loss[loss=0.1614, simple_loss=0.2285, pruned_loss=0.04716, over 4913.00 frames. ], tot_loss[loss=0.1857, simple_loss=0.2519, pruned_loss=0.05975, over 956167.17 frames. ], batch size: 43, lr: 3.69e-03, grad_scale: 32.0 2023-04-27 01:39:45,715 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6192, 0.6455, 1.4339, 1.9672, 1.6949, 1.5038, 1.5048, 1.5420], device='cuda:4'), covar=tensor([0.4882, 0.6756, 0.7051, 0.7319, 0.6221, 0.8180, 0.8200, 0.8239], device='cuda:4'), in_proj_covar=tensor([0.0406, 0.0407, 0.0495, 0.0513, 0.0437, 0.0456, 0.0464, 0.0464], device='cuda:4'), out_proj_covar=tensor([9.8611e-05, 1.0113e-04, 1.1167e-04, 1.2177e-04, 1.0581e-04, 1.1014e-04, 1.1121e-04, 1.1143e-04], device='cuda:4') 2023-04-27 01:39:53,438 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=60504.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:40:04,535 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1103, 1.5083, 1.9838, 2.2809, 1.8762, 1.5085, 1.0541, 1.6968], device='cuda:4'), covar=tensor([0.3682, 0.3638, 0.1866, 0.2733, 0.3026, 0.2923, 0.4870, 0.2456], device='cuda:4'), in_proj_covar=tensor([0.0282, 0.0246, 0.0219, 0.0313, 0.0212, 0.0226, 0.0229, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 01:40:07,941 INFO [finetune.py:976] (4/7) Epoch 11, batch 3250, loss[loss=0.2035, simple_loss=0.2744, pruned_loss=0.06632, over 4894.00 frames. ], tot_loss[loss=0.1857, simple_loss=0.2523, pruned_loss=0.05954, over 957108.34 frames. ], batch size: 35, lr: 3.69e-03, grad_scale: 32.0 2023-04-27 01:40:15,722 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.064e+02 1.675e+02 1.982e+02 2.288e+02 4.621e+02, threshold=3.964e+02, percent-clipped=1.0 2023-04-27 01:40:19,344 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.74 vs. limit=2.0 2023-04-27 01:40:20,960 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=60545.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:40:25,628 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=60552.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:40:26,845 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=60554.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:40:51,563 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3398, 1.7510, 2.2162, 2.7279, 2.1859, 1.7936, 1.5965, 1.9340], device='cuda:4'), covar=tensor([0.3636, 0.3802, 0.1734, 0.2983, 0.3129, 0.2934, 0.4531, 0.2810], device='cuda:4'), in_proj_covar=tensor([0.0283, 0.0247, 0.0220, 0.0314, 0.0213, 0.0227, 0.0229, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 01:40:52,025 INFO [finetune.py:976] (4/7) Epoch 11, batch 3300, loss[loss=0.2237, simple_loss=0.2843, pruned_loss=0.08152, over 4744.00 frames. ], tot_loss[loss=0.1906, simple_loss=0.2573, pruned_loss=0.06192, over 955675.03 frames. ], batch size: 54, lr: 3.69e-03, grad_scale: 32.0 2023-04-27 01:41:14,046 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=60593.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:41:58,103 INFO [finetune.py:976] (4/7) Epoch 11, batch 3350, loss[loss=0.2006, simple_loss=0.2638, pruned_loss=0.06871, over 4921.00 frames. ], tot_loss[loss=0.192, simple_loss=0.259, pruned_loss=0.06248, over 956455.24 frames. ], batch size: 33, lr: 3.69e-03, grad_scale: 32.0 2023-04-27 01:42:11,850 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.079e+02 1.646e+02 2.104e+02 2.646e+02 5.419e+02, threshold=4.209e+02, percent-clipped=4.0 2023-04-27 01:42:52,680 INFO [finetune.py:976] (4/7) Epoch 11, batch 3400, loss[loss=0.2229, simple_loss=0.291, pruned_loss=0.07746, over 4893.00 frames. ], tot_loss[loss=0.1939, simple_loss=0.2607, pruned_loss=0.06353, over 954776.30 frames. ], batch size: 35, lr: 3.69e-03, grad_scale: 32.0 2023-04-27 01:42:55,284 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2523, 1.2466, 1.3729, 1.5245, 1.6003, 1.2951, 0.8897, 1.4660], device='cuda:4'), covar=tensor([0.0897, 0.1205, 0.0800, 0.0632, 0.0651, 0.0851, 0.0997, 0.0603], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0204, 0.0184, 0.0177, 0.0181, 0.0190, 0.0161, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 01:43:05,891 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=60696.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:43:14,640 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=60708.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:43:20,708 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=60718.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:43:22,462 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=60721.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:43:25,999 INFO [finetune.py:976] (4/7) Epoch 11, batch 3450, loss[loss=0.167, simple_loss=0.229, pruned_loss=0.0525, over 4282.00 frames. ], tot_loss[loss=0.1926, simple_loss=0.2597, pruned_loss=0.06272, over 955656.33 frames. ], batch size: 66, lr: 3.69e-03, grad_scale: 32.0 2023-04-27 01:43:30,787 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=60734.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:43:32,468 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.44 vs. limit=2.0 2023-04-27 01:43:33,697 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.084e+02 1.685e+02 2.029e+02 2.444e+02 3.873e+02, threshold=4.057e+02, percent-clipped=0.0 2023-04-27 01:43:36,751 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=60744.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:43:45,935 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=60756.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:43:54,318 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=60769.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:43:58,983 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.92 vs. limit=5.0 2023-04-27 01:43:59,111 INFO [finetune.py:976] (4/7) Epoch 11, batch 3500, loss[loss=0.168, simple_loss=0.236, pruned_loss=0.04999, over 4843.00 frames. ], tot_loss[loss=0.1902, simple_loss=0.2571, pruned_loss=0.06167, over 956802.25 frames. ], batch size: 44, lr: 3.69e-03, grad_scale: 32.0 2023-04-27 01:44:00,468 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=60779.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:44:02,197 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=60782.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:44:18,156 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3447, 1.6944, 1.4618, 1.8872, 1.7825, 2.0333, 1.5085, 3.6719], device='cuda:4'), covar=tensor([0.0643, 0.0742, 0.0833, 0.1150, 0.0622, 0.0478, 0.0721, 0.0164], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0044, 0.0040, 0.0038, 0.0039, 0.0058], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0012, 0.0013, 0.0016], device='cuda:4') 2023-04-27 01:44:32,811 INFO [finetune.py:976] (4/7) Epoch 11, batch 3550, loss[loss=0.1864, simple_loss=0.2462, pruned_loss=0.06329, over 4193.00 frames. ], tot_loss[loss=0.1873, simple_loss=0.254, pruned_loss=0.06029, over 955371.21 frames. ], batch size: 18, lr: 3.69e-03, grad_scale: 32.0 2023-04-27 01:44:40,089 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.124e+02 1.561e+02 1.887e+02 2.297e+02 4.767e+02, threshold=3.774e+02, percent-clipped=1.0 2023-04-27 01:44:50,248 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=60854.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:45:06,039 INFO [finetune.py:976] (4/7) Epoch 11, batch 3600, loss[loss=0.2243, simple_loss=0.2812, pruned_loss=0.0837, over 4836.00 frames. ], tot_loss[loss=0.1859, simple_loss=0.2523, pruned_loss=0.05982, over 956874.50 frames. ], batch size: 47, lr: 3.69e-03, grad_scale: 32.0 2023-04-27 01:45:21,748 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=60902.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:45:39,841 INFO [finetune.py:976] (4/7) Epoch 11, batch 3650, loss[loss=0.1725, simple_loss=0.2421, pruned_loss=0.05148, over 4757.00 frames. ], tot_loss[loss=0.1876, simple_loss=0.2543, pruned_loss=0.06041, over 958490.89 frames. ], batch size: 26, lr: 3.69e-03, grad_scale: 32.0 2023-04-27 01:45:47,185 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.103e+02 1.725e+02 2.086e+02 2.393e+02 4.773e+02, threshold=4.172e+02, percent-clipped=3.0 2023-04-27 01:46:05,996 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0235, 2.7658, 2.0107, 1.9974, 1.4069, 1.4550, 2.1743, 1.4147], device='cuda:4'), covar=tensor([0.1791, 0.1488, 0.1626, 0.1970, 0.2525, 0.2009, 0.1110, 0.2180], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0214, 0.0169, 0.0205, 0.0203, 0.0184, 0.0158, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 01:46:25,825 INFO [finetune.py:976] (4/7) Epoch 11, batch 3700, loss[loss=0.1926, simple_loss=0.2556, pruned_loss=0.06481, over 4189.00 frames. ], tot_loss[loss=0.1898, simple_loss=0.2576, pruned_loss=0.06104, over 956439.67 frames. ], batch size: 65, lr: 3.69e-03, grad_scale: 32.0 2023-04-27 01:46:28,430 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.55 vs. limit=5.0 2023-04-27 01:46:52,273 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.58 vs. limit=5.0 2023-04-27 01:47:05,591 INFO [finetune.py:976] (4/7) Epoch 11, batch 3750, loss[loss=0.1448, simple_loss=0.2309, pruned_loss=0.02939, over 4810.00 frames. ], tot_loss[loss=0.1905, simple_loss=0.2585, pruned_loss=0.06123, over 955600.05 frames. ], batch size: 39, lr: 3.69e-03, grad_scale: 32.0 2023-04-27 01:47:17,587 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3217, 1.5061, 4.0229, 3.7229, 3.5453, 3.8666, 3.7960, 3.5841], device='cuda:4'), covar=tensor([0.7407, 0.5265, 0.1154, 0.1833, 0.1194, 0.1410, 0.1787, 0.1529], device='cuda:4'), in_proj_covar=tensor([0.0314, 0.0311, 0.0406, 0.0410, 0.0353, 0.0410, 0.0317, 0.0371], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 01:47:18,728 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.172e+02 1.566e+02 1.879e+02 2.369e+02 3.505e+02, threshold=3.758e+02, percent-clipped=0.0 2023-04-27 01:47:40,517 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=61061.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:48:01,154 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=61074.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:48:05,283 INFO [finetune.py:976] (4/7) Epoch 11, batch 3800, loss[loss=0.1793, simple_loss=0.2448, pruned_loss=0.05689, over 4812.00 frames. ], tot_loss[loss=0.1914, simple_loss=0.2595, pruned_loss=0.06169, over 954368.19 frames. ], batch size: 33, lr: 3.69e-03, grad_scale: 32.0 2023-04-27 01:48:56,481 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=61122.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:48:59,869 INFO [finetune.py:976] (4/7) Epoch 11, batch 3850, loss[loss=0.2126, simple_loss=0.2794, pruned_loss=0.07288, over 4720.00 frames. ], tot_loss[loss=0.1903, simple_loss=0.2582, pruned_loss=0.06123, over 955589.19 frames. ], batch size: 54, lr: 3.69e-03, grad_scale: 32.0 2023-04-27 01:49:08,086 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.671e+02 1.909e+02 2.247e+02 3.528e+02, threshold=3.817e+02, percent-clipped=0.0 2023-04-27 01:49:27,484 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.38 vs. limit=5.0 2023-04-27 01:49:33,125 INFO [finetune.py:976] (4/7) Epoch 11, batch 3900, loss[loss=0.1592, simple_loss=0.2219, pruned_loss=0.04819, over 4756.00 frames. ], tot_loss[loss=0.1884, simple_loss=0.2553, pruned_loss=0.06071, over 955604.49 frames. ], batch size: 28, lr: 3.69e-03, grad_scale: 32.0 2023-04-27 01:49:42,530 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.7542, 3.6412, 2.7842, 4.2893, 3.6242, 3.6950, 1.7436, 3.6387], device='cuda:4'), covar=tensor([0.1506, 0.1163, 0.3410, 0.1368, 0.2526, 0.1688, 0.5185, 0.2208], device='cuda:4'), in_proj_covar=tensor([0.0238, 0.0213, 0.0245, 0.0300, 0.0292, 0.0245, 0.0265, 0.0264], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 01:50:04,199 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3558, 1.4936, 1.7790, 1.8971, 1.7611, 1.9374, 1.9130, 1.8367], device='cuda:4'), covar=tensor([0.4587, 0.5962, 0.4982, 0.4483, 0.6025, 0.7881, 0.5417, 0.5175], device='cuda:4'), in_proj_covar=tensor([0.0325, 0.0376, 0.0314, 0.0325, 0.0338, 0.0399, 0.0357, 0.0321], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 01:50:05,883 INFO [finetune.py:976] (4/7) Epoch 11, batch 3950, loss[loss=0.1988, simple_loss=0.2581, pruned_loss=0.06976, over 4310.00 frames. ], tot_loss[loss=0.1867, simple_loss=0.253, pruned_loss=0.06025, over 955339.73 frames. ], batch size: 65, lr: 3.68e-03, grad_scale: 32.0 2023-04-27 01:50:15,544 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.075e+02 1.581e+02 1.829e+02 2.221e+02 4.088e+02, threshold=3.658e+02, percent-clipped=2.0 2023-04-27 01:50:39,622 INFO [finetune.py:976] (4/7) Epoch 11, batch 4000, loss[loss=0.1847, simple_loss=0.2308, pruned_loss=0.06928, over 3785.00 frames. ], tot_loss[loss=0.1876, simple_loss=0.2536, pruned_loss=0.06082, over 955498.51 frames. ], batch size: 16, lr: 3.68e-03, grad_scale: 32.0 2023-04-27 01:50:48,463 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8145, 2.4185, 2.0062, 2.2045, 1.6342, 1.9237, 1.9524, 1.4064], device='cuda:4'), covar=tensor([0.1773, 0.0962, 0.0753, 0.1099, 0.3090, 0.1164, 0.1737, 0.2570], device='cuda:4'), in_proj_covar=tensor([0.0291, 0.0313, 0.0226, 0.0283, 0.0312, 0.0270, 0.0253, 0.0274], device='cuda:4'), out_proj_covar=tensor([1.1767e-04, 1.2575e-04, 9.0340e-05, 1.1311e-04, 1.2760e-04, 1.0838e-04, 1.0324e-04, 1.0989e-04], device='cuda:4') 2023-04-27 01:51:18,202 INFO [finetune.py:976] (4/7) Epoch 11, batch 4050, loss[loss=0.2245, simple_loss=0.3074, pruned_loss=0.07077, over 4746.00 frames. ], tot_loss[loss=0.1899, simple_loss=0.257, pruned_loss=0.06134, over 955309.35 frames. ], batch size: 54, lr: 3.68e-03, grad_scale: 32.0 2023-04-27 01:51:37,048 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.086e+02 1.723e+02 1.975e+02 2.601e+02 5.201e+02, threshold=3.950e+02, percent-clipped=3.0 2023-04-27 01:51:59,614 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9120, 1.9246, 2.1871, 2.2476, 1.6730, 1.5354, 1.9048, 1.1673], device='cuda:4'), covar=tensor([0.0826, 0.0779, 0.0534, 0.0934, 0.0939, 0.1401, 0.0846, 0.0964], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0071, 0.0071, 0.0067, 0.0075, 0.0097, 0.0076, 0.0072], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 01:52:11,024 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8276, 1.3564, 1.9308, 2.2754, 1.9497, 1.8075, 1.8808, 1.8411], device='cuda:4'), covar=tensor([0.5482, 0.7370, 0.7475, 0.7251, 0.6758, 0.8948, 0.8428, 0.8689], device='cuda:4'), in_proj_covar=tensor([0.0412, 0.0412, 0.0499, 0.0518, 0.0441, 0.0461, 0.0471, 0.0471], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 01:52:21,698 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=61374.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:52:23,440 INFO [finetune.py:976] (4/7) Epoch 11, batch 4100, loss[loss=0.1786, simple_loss=0.2541, pruned_loss=0.05156, over 4796.00 frames. ], tot_loss[loss=0.191, simple_loss=0.2591, pruned_loss=0.06147, over 954156.00 frames. ], batch size: 25, lr: 3.68e-03, grad_scale: 32.0 2023-04-27 01:53:16,502 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=61417.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:53:25,497 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=61422.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 01:53:28,495 INFO [finetune.py:976] (4/7) Epoch 11, batch 4150, loss[loss=0.1853, simple_loss=0.2645, pruned_loss=0.05302, over 4819.00 frames. ], tot_loss[loss=0.1919, simple_loss=0.2598, pruned_loss=0.06199, over 953449.68 frames. ], batch size: 33, lr: 3.68e-03, grad_scale: 32.0 2023-04-27 01:53:48,241 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.138e+02 1.660e+02 1.903e+02 2.358e+02 4.640e+02, threshold=3.807e+02, percent-clipped=2.0 2023-04-27 01:54:37,121 INFO [finetune.py:976] (4/7) Epoch 11, batch 4200, loss[loss=0.1832, simple_loss=0.2427, pruned_loss=0.06182, over 4796.00 frames. ], tot_loss[loss=0.1921, simple_loss=0.2601, pruned_loss=0.06208, over 954928.13 frames. ], batch size: 25, lr: 3.68e-03, grad_scale: 32.0 2023-04-27 01:55:45,102 INFO [finetune.py:976] (4/7) Epoch 11, batch 4250, loss[loss=0.1579, simple_loss=0.2314, pruned_loss=0.04224, over 4758.00 frames. ], tot_loss[loss=0.189, simple_loss=0.2564, pruned_loss=0.06077, over 955084.27 frames. ], batch size: 28, lr: 3.68e-03, grad_scale: 32.0 2023-04-27 01:55:45,848 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8196, 2.5006, 1.9846, 1.8337, 1.3632, 1.3945, 1.9319, 1.3085], device='cuda:4'), covar=tensor([0.1732, 0.1443, 0.1442, 0.1865, 0.2450, 0.2156, 0.1112, 0.2188], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0214, 0.0169, 0.0204, 0.0203, 0.0184, 0.0158, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 01:55:57,237 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.015e+02 1.633e+02 1.870e+02 2.370e+02 4.371e+02, threshold=3.739e+02, percent-clipped=2.0 2023-04-27 01:56:05,323 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-27 01:56:49,827 INFO [finetune.py:976] (4/7) Epoch 11, batch 4300, loss[loss=0.1793, simple_loss=0.2243, pruned_loss=0.0672, over 4202.00 frames. ], tot_loss[loss=0.1866, simple_loss=0.2536, pruned_loss=0.05985, over 954601.59 frames. ], batch size: 18, lr: 3.68e-03, grad_scale: 32.0 2023-04-27 01:57:48,691 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4525, 3.2503, 0.9070, 1.6941, 1.8649, 2.2456, 1.9729, 1.0421], device='cuda:4'), covar=tensor([0.1440, 0.1041, 0.2111, 0.1393, 0.1099, 0.1095, 0.1428, 0.1971], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0250, 0.0141, 0.0123, 0.0135, 0.0154, 0.0119, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 01:57:57,466 INFO [finetune.py:976] (4/7) Epoch 11, batch 4350, loss[loss=0.1482, simple_loss=0.2274, pruned_loss=0.03443, over 4733.00 frames. ], tot_loss[loss=0.1837, simple_loss=0.2506, pruned_loss=0.05835, over 955955.45 frames. ], batch size: 23, lr: 3.68e-03, grad_scale: 32.0 2023-04-27 01:58:10,464 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.036e+02 1.619e+02 1.912e+02 2.174e+02 4.082e+02, threshold=3.823e+02, percent-clipped=2.0 2023-04-27 01:59:01,633 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1704, 1.8349, 2.2148, 2.4557, 2.5304, 1.9955, 1.7715, 2.3833], device='cuda:4'), covar=tensor([0.0833, 0.1100, 0.0635, 0.0596, 0.0565, 0.0965, 0.0829, 0.0547], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0202, 0.0181, 0.0173, 0.0178, 0.0187, 0.0158, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 01:59:02,119 INFO [finetune.py:976] (4/7) Epoch 11, batch 4400, loss[loss=0.1747, simple_loss=0.223, pruned_loss=0.06317, over 4809.00 frames. ], tot_loss[loss=0.1847, simple_loss=0.2521, pruned_loss=0.05867, over 956648.56 frames. ], batch size: 25, lr: 3.68e-03, grad_scale: 32.0 2023-04-27 01:59:11,958 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.8551, 3.8928, 2.7572, 4.5587, 3.9936, 3.8652, 1.7831, 3.8397], device='cuda:4'), covar=tensor([0.1858, 0.1232, 0.3544, 0.1287, 0.4058, 0.1803, 0.5574, 0.2539], device='cuda:4'), in_proj_covar=tensor([0.0238, 0.0213, 0.0246, 0.0300, 0.0292, 0.0246, 0.0265, 0.0265], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 01:59:23,249 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9640, 2.5392, 0.9628, 1.1881, 1.8297, 1.2666, 3.4091, 1.4743], device='cuda:4'), covar=tensor([0.0763, 0.0787, 0.0958, 0.1467, 0.0605, 0.1112, 0.0325, 0.0845], device='cuda:4'), in_proj_covar=tensor([0.0052, 0.0066, 0.0049, 0.0047, 0.0050, 0.0052, 0.0077, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-27 01:59:57,239 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=61717.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:00:05,433 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.85 vs. limit=2.0 2023-04-27 02:00:08,837 INFO [finetune.py:976] (4/7) Epoch 11, batch 4450, loss[loss=0.1497, simple_loss=0.214, pruned_loss=0.04268, over 4723.00 frames. ], tot_loss[loss=0.187, simple_loss=0.2548, pruned_loss=0.0596, over 957437.98 frames. ], batch size: 23, lr: 3.68e-03, grad_scale: 32.0 2023-04-27 02:00:17,361 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=61732.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:00:21,984 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.107e+02 1.778e+02 2.036e+02 2.485e+02 5.642e+02, threshold=4.071e+02, percent-clipped=3.0 2023-04-27 02:00:24,578 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=61743.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:00:29,417 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.51 vs. limit=5.0 2023-04-27 02:00:43,723 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5246, 1.7759, 0.6647, 1.2878, 1.7851, 1.3752, 1.3378, 1.3046], device='cuda:4'), covar=tensor([0.0643, 0.0341, 0.0396, 0.0627, 0.0292, 0.0681, 0.0666, 0.0663], device='cuda:4'), in_proj_covar=tensor([0.0030, 0.0025, 0.0023, 0.0030, 0.0020, 0.0029, 0.0029, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0045, 0.0038, 0.0050, 0.0038, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 02:00:45,397 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=61765.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:01:03,250 INFO [finetune.py:976] (4/7) Epoch 11, batch 4500, loss[loss=0.1923, simple_loss=0.2547, pruned_loss=0.065, over 4768.00 frames. ], tot_loss[loss=0.1876, simple_loss=0.2556, pruned_loss=0.05976, over 956311.79 frames. ], batch size: 26, lr: 3.68e-03, grad_scale: 32.0 2023-04-27 02:01:19,125 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=61793.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:01:20,296 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5352, 1.6136, 0.8885, 1.2749, 1.6312, 1.4544, 1.3534, 1.3687], device='cuda:4'), covar=tensor([0.0498, 0.0346, 0.0387, 0.0529, 0.0307, 0.0505, 0.0487, 0.0564], device='cuda:4'), in_proj_covar=tensor([0.0030, 0.0025, 0.0023, 0.0030, 0.0020, 0.0029, 0.0029, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0045, 0.0038, 0.0050, 0.0037, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 02:01:25,827 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.67 vs. limit=2.0 2023-04-27 02:01:26,323 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=61804.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:01:42,633 INFO [finetune.py:976] (4/7) Epoch 11, batch 4550, loss[loss=0.1601, simple_loss=0.2196, pruned_loss=0.05036, over 4014.00 frames. ], tot_loss[loss=0.1899, simple_loss=0.2578, pruned_loss=0.06098, over 955597.55 frames. ], batch size: 17, lr: 3.68e-03, grad_scale: 32.0 2023-04-27 02:01:48,196 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.7296, 3.7910, 2.6211, 4.4302, 3.8806, 3.7992, 1.7011, 3.6976], device='cuda:4'), covar=tensor([0.1666, 0.1115, 0.3134, 0.1520, 0.2576, 0.1675, 0.5520, 0.2385], device='cuda:4'), in_proj_covar=tensor([0.0240, 0.0215, 0.0248, 0.0303, 0.0295, 0.0247, 0.0267, 0.0267], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 02:01:49,952 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.176e+02 1.694e+02 2.002e+02 2.452e+02 5.471e+02, threshold=4.003e+02, percent-clipped=1.0 2023-04-27 02:02:16,333 INFO [finetune.py:976] (4/7) Epoch 11, batch 4600, loss[loss=0.1889, simple_loss=0.2428, pruned_loss=0.06749, over 4699.00 frames. ], tot_loss[loss=0.19, simple_loss=0.2577, pruned_loss=0.06112, over 955767.68 frames. ], batch size: 23, lr: 3.68e-03, grad_scale: 32.0 2023-04-27 02:02:37,780 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1160, 1.6127, 1.4719, 1.8436, 1.7401, 2.0229, 1.4123, 3.4144], device='cuda:4'), covar=tensor([0.0598, 0.0776, 0.0742, 0.1130, 0.0579, 0.0468, 0.0736, 0.0149], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0044, 0.0040, 0.0038, 0.0038, 0.0057], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0012, 0.0013, 0.0016], device='cuda:4') 2023-04-27 02:02:40,702 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=61914.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:02:45,177 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.64 vs. limit=2.0 2023-04-27 02:02:49,393 INFO [finetune.py:976] (4/7) Epoch 11, batch 4650, loss[loss=0.1877, simple_loss=0.2514, pruned_loss=0.06205, over 4934.00 frames. ], tot_loss[loss=0.1876, simple_loss=0.255, pruned_loss=0.06015, over 955201.68 frames. ], batch size: 38, lr: 3.68e-03, grad_scale: 32.0 2023-04-27 02:02:56,643 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.191e+02 1.561e+02 1.984e+02 2.276e+02 4.966e+02, threshold=3.968e+02, percent-clipped=2.0 2023-04-27 02:03:20,496 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=61975.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:03:22,012 INFO [finetune.py:976] (4/7) Epoch 11, batch 4700, loss[loss=0.1347, simple_loss=0.1971, pruned_loss=0.03619, over 4939.00 frames. ], tot_loss[loss=0.1841, simple_loss=0.2511, pruned_loss=0.05853, over 956085.25 frames. ], batch size: 33, lr: 3.68e-03, grad_scale: 32.0 2023-04-27 02:03:33,473 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1610, 2.7474, 2.1695, 2.1008, 1.7135, 1.6972, 2.2128, 1.6828], device='cuda:4'), covar=tensor([0.1754, 0.1527, 0.1427, 0.1692, 0.2302, 0.1959, 0.1030, 0.1962], device='cuda:4'), in_proj_covar=tensor([0.0193, 0.0212, 0.0167, 0.0202, 0.0200, 0.0182, 0.0157, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 02:04:06,620 INFO [finetune.py:976] (4/7) Epoch 11, batch 4750, loss[loss=0.2007, simple_loss=0.2645, pruned_loss=0.06846, over 4873.00 frames. ], tot_loss[loss=0.1831, simple_loss=0.2495, pruned_loss=0.05834, over 954901.31 frames. ], batch size: 31, lr: 3.68e-03, grad_scale: 32.0 2023-04-27 02:04:20,381 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.139e+02 1.587e+02 1.879e+02 2.301e+02 4.862e+02, threshold=3.757e+02, percent-clipped=3.0 2023-04-27 02:04:56,947 INFO [finetune.py:976] (4/7) Epoch 11, batch 4800, loss[loss=0.1589, simple_loss=0.2278, pruned_loss=0.04497, over 4897.00 frames. ], tot_loss[loss=0.1853, simple_loss=0.2522, pruned_loss=0.05919, over 952731.57 frames. ], batch size: 35, lr: 3.68e-03, grad_scale: 32.0 2023-04-27 02:05:05,179 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=62088.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:05:11,913 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=62099.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:05:27,325 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=62122.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:05:30,804 INFO [finetune.py:976] (4/7) Epoch 11, batch 4850, loss[loss=0.2158, simple_loss=0.2901, pruned_loss=0.0708, over 4926.00 frames. ], tot_loss[loss=0.1869, simple_loss=0.2544, pruned_loss=0.05973, over 954224.72 frames. ], batch size: 42, lr: 3.68e-03, grad_scale: 32.0 2023-04-27 02:05:39,088 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.265e+02 1.826e+02 2.171e+02 2.650e+02 4.437e+02, threshold=4.341e+02, percent-clipped=4.0 2023-04-27 02:05:44,445 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-04-27 02:06:08,686 INFO [finetune.py:976] (4/7) Epoch 11, batch 4900, loss[loss=0.1816, simple_loss=0.2515, pruned_loss=0.0558, over 4818.00 frames. ], tot_loss[loss=0.1882, simple_loss=0.2563, pruned_loss=0.06008, over 953320.56 frames. ], batch size: 30, lr: 3.68e-03, grad_scale: 32.0 2023-04-27 02:06:18,167 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=62183.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:06:52,656 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-27 02:06:52,812 INFO [finetune.py:976] (4/7) Epoch 11, batch 4950, loss[loss=0.2064, simple_loss=0.2697, pruned_loss=0.07156, over 4904.00 frames. ], tot_loss[loss=0.19, simple_loss=0.2584, pruned_loss=0.0608, over 955435.85 frames. ], batch size: 46, lr: 3.68e-03, grad_scale: 64.0 2023-04-27 02:06:58,261 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.63 vs. limit=2.0 2023-04-27 02:07:01,572 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.127e+02 1.624e+02 1.966e+02 2.483e+02 3.537e+02, threshold=3.932e+02, percent-clipped=0.0 2023-04-27 02:07:21,470 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=62270.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:07:23,303 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=62273.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:07:26,149 INFO [finetune.py:976] (4/7) Epoch 11, batch 5000, loss[loss=0.188, simple_loss=0.2609, pruned_loss=0.05758, over 4895.00 frames. ], tot_loss[loss=0.1869, simple_loss=0.2551, pruned_loss=0.05937, over 954269.00 frames. ], batch size: 35, lr: 3.68e-03, grad_scale: 32.0 2023-04-27 02:07:58,883 INFO [finetune.py:976] (4/7) Epoch 11, batch 5050, loss[loss=0.1671, simple_loss=0.23, pruned_loss=0.05211, over 4789.00 frames. ], tot_loss[loss=0.187, simple_loss=0.2543, pruned_loss=0.05981, over 955867.15 frames. ], batch size: 25, lr: 3.68e-03, grad_scale: 32.0 2023-04-27 02:08:04,217 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=62334.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:08:08,228 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.081e+02 1.670e+02 2.040e+02 2.398e+02 4.126e+02, threshold=4.080e+02, percent-clipped=2.0 2023-04-27 02:08:32,080 INFO [finetune.py:976] (4/7) Epoch 11, batch 5100, loss[loss=0.1725, simple_loss=0.2444, pruned_loss=0.05032, over 4832.00 frames. ], tot_loss[loss=0.1859, simple_loss=0.2526, pruned_loss=0.05964, over 956820.41 frames. ], batch size: 30, lr: 3.68e-03, grad_scale: 32.0 2023-04-27 02:08:37,462 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6766, 2.2119, 1.0756, 1.4351, 2.2446, 1.6540, 1.5651, 1.6464], device='cuda:4'), covar=tensor([0.0550, 0.0327, 0.0343, 0.0590, 0.0252, 0.0551, 0.0527, 0.0599], device='cuda:4'), in_proj_covar=tensor([0.0030, 0.0025, 0.0023, 0.0030, 0.0020, 0.0029, 0.0029, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0045, 0.0038, 0.0050, 0.0038, 0.0048, 0.0048, 0.0050], device='cuda:4') 2023-04-27 02:08:40,395 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=62388.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:08:42,188 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=62391.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:08:47,994 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=62399.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:09:06,023 INFO [finetune.py:976] (4/7) Epoch 11, batch 5150, loss[loss=0.1601, simple_loss=0.2289, pruned_loss=0.04565, over 4743.00 frames. ], tot_loss[loss=0.185, simple_loss=0.2517, pruned_loss=0.05915, over 956817.35 frames. ], batch size: 27, lr: 3.67e-03, grad_scale: 32.0 2023-04-27 02:09:12,053 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=62436.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:09:14,851 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.955e+01 1.646e+02 2.010e+02 2.557e+02 5.535e+02, threshold=4.020e+02, percent-clipped=1.0 2023-04-27 02:09:25,277 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=62447.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:09:33,234 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=62452.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:10:01,359 INFO [finetune.py:976] (4/7) Epoch 11, batch 5200, loss[loss=0.1958, simple_loss=0.2709, pruned_loss=0.06031, over 4937.00 frames. ], tot_loss[loss=0.1884, simple_loss=0.2559, pruned_loss=0.06051, over 955510.24 frames. ], batch size: 33, lr: 3.67e-03, grad_scale: 32.0 2023-04-27 02:10:02,043 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=62478.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:10:14,388 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.7551, 1.6208, 1.6010, 1.2351, 1.7276, 1.3915, 2.2537, 1.4029], device='cuda:4'), covar=tensor([0.3509, 0.1744, 0.4698, 0.3011, 0.1724, 0.2429, 0.1358, 0.4910], device='cuda:4'), in_proj_covar=tensor([0.0342, 0.0348, 0.0426, 0.0358, 0.0384, 0.0380, 0.0378, 0.0420], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 02:10:58,656 INFO [finetune.py:976] (4/7) Epoch 11, batch 5250, loss[loss=0.1711, simple_loss=0.2437, pruned_loss=0.0493, over 4170.00 frames. ], tot_loss[loss=0.1908, simple_loss=0.2584, pruned_loss=0.06161, over 955828.53 frames. ], batch size: 65, lr: 3.67e-03, grad_scale: 32.0 2023-04-27 02:11:01,300 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-27 02:11:07,064 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.221e+02 1.615e+02 2.039e+02 2.344e+02 5.619e+02, threshold=4.078e+02, percent-clipped=3.0 2023-04-27 02:11:10,669 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=62545.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:11:28,235 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=62570.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:11:32,451 INFO [finetune.py:976] (4/7) Epoch 11, batch 5300, loss[loss=0.202, simple_loss=0.259, pruned_loss=0.07252, over 4885.00 frames. ], tot_loss[loss=0.1923, simple_loss=0.2599, pruned_loss=0.06236, over 955437.47 frames. ], batch size: 32, lr: 3.67e-03, grad_scale: 32.0 2023-04-27 02:11:37,339 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=62585.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:11:51,651 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4279, 3.2181, 0.8775, 1.7822, 1.8525, 2.2841, 1.8289, 0.9066], device='cuda:4'), covar=tensor([0.1382, 0.0826, 0.2050, 0.1235, 0.1103, 0.0991, 0.1620, 0.2077], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0248, 0.0141, 0.0121, 0.0133, 0.0153, 0.0118, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 02:11:52,272 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=62606.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:12:00,029 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=62618.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:12:05,914 INFO [finetune.py:976] (4/7) Epoch 11, batch 5350, loss[loss=0.1822, simple_loss=0.2495, pruned_loss=0.05749, over 4842.00 frames. ], tot_loss[loss=0.1914, simple_loss=0.2592, pruned_loss=0.06183, over 952913.55 frames. ], batch size: 47, lr: 3.67e-03, grad_scale: 32.0 2023-04-27 02:12:07,196 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=62629.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:12:13,877 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.227e+02 1.690e+02 2.017e+02 2.486e+02 5.260e+02, threshold=4.033e+02, percent-clipped=4.0 2023-04-27 02:12:18,147 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=62646.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:12:32,493 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=62666.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:12:39,515 INFO [finetune.py:976] (4/7) Epoch 11, batch 5400, loss[loss=0.1619, simple_loss=0.2349, pruned_loss=0.0444, over 4905.00 frames. ], tot_loss[loss=0.1885, simple_loss=0.2559, pruned_loss=0.06061, over 953629.63 frames. ], batch size: 46, lr: 3.67e-03, grad_scale: 32.0 2023-04-27 02:13:12,193 INFO [finetune.py:976] (4/7) Epoch 11, batch 5450, loss[loss=0.1932, simple_loss=0.2646, pruned_loss=0.06085, over 4817.00 frames. ], tot_loss[loss=0.1855, simple_loss=0.2525, pruned_loss=0.05926, over 955121.19 frames. ], batch size: 38, lr: 3.67e-03, grad_scale: 32.0 2023-04-27 02:13:12,306 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=62727.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:13:16,516 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.50 vs. limit=5.0 2023-04-27 02:13:20,545 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.017e+02 1.672e+02 1.919e+02 2.198e+02 3.769e+02, threshold=3.837e+02, percent-clipped=0.0 2023-04-27 02:13:24,904 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=62747.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:13:31,018 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.9700, 2.0548, 2.2005, 2.7739, 2.8973, 2.5613, 2.4119, 2.1495], device='cuda:4'), covar=tensor([0.1429, 0.1718, 0.1660, 0.1687, 0.1059, 0.1543, 0.2195, 0.1833], device='cuda:4'), in_proj_covar=tensor([0.0305, 0.0319, 0.0350, 0.0295, 0.0331, 0.0316, 0.0305, 0.0358], device='cuda:4'), out_proj_covar=tensor([6.3853e-05, 6.7118e-05, 7.5331e-05, 6.0566e-05, 6.9051e-05, 6.7479e-05, 6.5152e-05, 7.6837e-05], device='cuda:4') 2023-04-27 02:13:42,648 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=62773.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:13:45,487 INFO [finetune.py:976] (4/7) Epoch 11, batch 5500, loss[loss=0.135, simple_loss=0.2055, pruned_loss=0.03225, over 4930.00 frames. ], tot_loss[loss=0.1834, simple_loss=0.25, pruned_loss=0.05842, over 956466.26 frames. ], batch size: 33, lr: 3.67e-03, grad_scale: 32.0 2023-04-27 02:13:46,159 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=62778.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:14:03,658 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2465, 1.1940, 1.6359, 1.5104, 1.2229, 0.9793, 1.2626, 0.8420], device='cuda:4'), covar=tensor([0.0663, 0.0770, 0.0456, 0.0643, 0.0734, 0.1292, 0.0708, 0.0796], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0071, 0.0070, 0.0067, 0.0075, 0.0097, 0.0076, 0.0071], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 02:14:07,258 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2454, 1.2158, 3.8271, 3.5540, 3.4068, 3.6869, 3.6612, 3.3414], device='cuda:4'), covar=tensor([0.7307, 0.5954, 0.1130, 0.1911, 0.1153, 0.1792, 0.1433, 0.1587], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0303, 0.0398, 0.0404, 0.0346, 0.0404, 0.0310, 0.0366], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 02:14:18,370 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=62826.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:14:18,932 INFO [finetune.py:976] (4/7) Epoch 11, batch 5550, loss[loss=0.1932, simple_loss=0.2731, pruned_loss=0.05667, over 4934.00 frames. ], tot_loss[loss=0.1842, simple_loss=0.2509, pruned_loss=0.05868, over 954822.63 frames. ], batch size: 38, lr: 3.67e-03, grad_scale: 32.0 2023-04-27 02:14:23,708 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=62834.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:14:27,259 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.295e+02 1.658e+02 2.068e+02 2.664e+02 6.166e+02, threshold=4.137e+02, percent-clipped=3.0 2023-04-27 02:14:35,229 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0023, 1.5200, 1.7620, 2.1533, 1.8481, 1.4528, 1.1629, 1.7093], device='cuda:4'), covar=tensor([0.2862, 0.3107, 0.1562, 0.2147, 0.2478, 0.2492, 0.4417, 0.2075], device='cuda:4'), in_proj_covar=tensor([0.0284, 0.0248, 0.0221, 0.0315, 0.0214, 0.0228, 0.0229, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 02:14:45,413 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3260, 3.4694, 0.7845, 1.7609, 1.8439, 2.3625, 1.8367, 0.8869], device='cuda:4'), covar=tensor([0.1469, 0.0939, 0.2214, 0.1275, 0.1096, 0.1027, 0.1674, 0.2251], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0248, 0.0141, 0.0121, 0.0133, 0.0153, 0.0118, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 02:14:56,172 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1633, 2.8983, 0.9603, 1.6278, 1.6033, 2.1482, 1.6679, 0.9983], device='cuda:4'), covar=tensor([0.1425, 0.1031, 0.1917, 0.1283, 0.1145, 0.0911, 0.1577, 0.2003], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0248, 0.0140, 0.0121, 0.0133, 0.0153, 0.0118, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 02:15:05,422 INFO [finetune.py:976] (4/7) Epoch 11, batch 5600, loss[loss=0.186, simple_loss=0.2652, pruned_loss=0.05345, over 4806.00 frames. ], tot_loss[loss=0.1882, simple_loss=0.2559, pruned_loss=0.06028, over 954730.45 frames. ], batch size: 45, lr: 3.67e-03, grad_scale: 32.0 2023-04-27 02:15:15,914 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6607, 1.2557, 1.7495, 2.0907, 1.7707, 1.5758, 1.6369, 1.6648], device='cuda:4'), covar=tensor([0.5571, 0.7610, 0.7890, 0.7384, 0.7008, 0.9197, 0.9630, 1.0078], device='cuda:4'), in_proj_covar=tensor([0.0407, 0.0407, 0.0493, 0.0512, 0.0438, 0.0456, 0.0465, 0.0466], device='cuda:4'), out_proj_covar=tensor([9.8935e-05, 1.0067e-04, 1.1134e-04, 1.2164e-04, 1.0607e-04, 1.1024e-04, 1.1149e-04, 1.1166e-04], device='cuda:4') 2023-04-27 02:15:37,388 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=62901.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:16:10,654 INFO [finetune.py:976] (4/7) Epoch 11, batch 5650, loss[loss=0.1799, simple_loss=0.2596, pruned_loss=0.05009, over 4827.00 frames. ], tot_loss[loss=0.1902, simple_loss=0.2591, pruned_loss=0.0607, over 955439.50 frames. ], batch size: 47, lr: 3.67e-03, grad_scale: 32.0 2023-04-27 02:16:11,880 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=62929.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:16:23,345 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.053e+02 1.567e+02 1.821e+02 2.303e+02 3.535e+02, threshold=3.642e+02, percent-clipped=0.0 2023-04-27 02:16:23,985 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=62941.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:17:06,579 INFO [finetune.py:976] (4/7) Epoch 11, batch 5700, loss[loss=0.138, simple_loss=0.1969, pruned_loss=0.03955, over 3971.00 frames. ], tot_loss[loss=0.1885, simple_loss=0.2554, pruned_loss=0.06078, over 936173.42 frames. ], batch size: 17, lr: 3.67e-03, grad_scale: 32.0 2023-04-27 02:17:06,613 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=62977.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:17:38,299 INFO [finetune.py:976] (4/7) Epoch 12, batch 0, loss[loss=0.1908, simple_loss=0.2476, pruned_loss=0.06705, over 4357.00 frames. ], tot_loss[loss=0.1908, simple_loss=0.2476, pruned_loss=0.06705, over 4357.00 frames. ], batch size: 19, lr: 3.67e-03, grad_scale: 32.0 2023-04-27 02:17:38,300 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 02:17:47,135 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3157, 1.4762, 1.7910, 1.9154, 1.8315, 2.0023, 1.7973, 1.8273], device='cuda:4'), covar=tensor([0.4187, 0.6176, 0.5374, 0.5364, 0.6364, 0.8657, 0.6971, 0.5842], device='cuda:4'), in_proj_covar=tensor([0.0327, 0.0379, 0.0315, 0.0327, 0.0339, 0.0400, 0.0359, 0.0324], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 02:17:53,690 INFO [finetune.py:1010] (4/7) Epoch 12, validation: loss=0.1544, simple_loss=0.2267, pruned_loss=0.04099, over 2265189.00 frames. 2023-04-27 02:17:53,690 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-27 02:18:09,366 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8961, 1.5699, 1.8819, 2.1792, 2.2592, 1.8695, 1.5997, 2.0221], device='cuda:4'), covar=tensor([0.0810, 0.1042, 0.0521, 0.0474, 0.0488, 0.0767, 0.0810, 0.0517], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0204, 0.0183, 0.0175, 0.0179, 0.0189, 0.0159, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 02:18:18,024 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=63022.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:18:32,880 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.8025, 2.0481, 1.9836, 2.2000, 1.8951, 2.1162, 2.1113, 2.0026], device='cuda:4'), covar=tensor([0.5090, 0.7447, 0.5886, 0.5670, 0.6468, 0.8139, 0.7146, 0.6844], device='cuda:4'), in_proj_covar=tensor([0.0327, 0.0378, 0.0315, 0.0327, 0.0338, 0.0399, 0.0358, 0.0324], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 02:18:39,643 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.931e+01 1.655e+02 2.029e+02 2.560e+02 6.942e+02, threshold=4.058e+02, percent-clipped=5.0 2023-04-27 02:18:42,751 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-27 02:18:44,985 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=63047.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:18:53,718 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-27 02:18:54,578 INFO [finetune.py:976] (4/7) Epoch 12, batch 50, loss[loss=0.2016, simple_loss=0.26, pruned_loss=0.07156, over 4859.00 frames. ], tot_loss[loss=0.1889, simple_loss=0.2596, pruned_loss=0.05908, over 217578.18 frames. ], batch size: 34, lr: 3.67e-03, grad_scale: 32.0 2023-04-27 02:19:04,204 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1233, 1.3003, 1.7156, 2.4125, 2.5291, 1.9327, 1.7213, 2.2400], device='cuda:4'), covar=tensor([0.0890, 0.1728, 0.0935, 0.0536, 0.0543, 0.0978, 0.0928, 0.0586], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0204, 0.0183, 0.0175, 0.0180, 0.0189, 0.0159, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 02:19:30,959 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=63095.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:19:37,865 INFO [finetune.py:976] (4/7) Epoch 12, batch 100, loss[loss=0.1438, simple_loss=0.2127, pruned_loss=0.03741, over 4920.00 frames. ], tot_loss[loss=0.1875, simple_loss=0.2543, pruned_loss=0.06035, over 381690.44 frames. ], batch size: 33, lr: 3.67e-03, grad_scale: 32.0 2023-04-27 02:19:54,527 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=63129.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:20:01,184 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.001e+02 1.674e+02 1.936e+02 2.495e+02 3.786e+02, threshold=3.872e+02, percent-clipped=0.0 2023-04-27 02:20:11,745 INFO [finetune.py:976] (4/7) Epoch 12, batch 150, loss[loss=0.1733, simple_loss=0.2314, pruned_loss=0.05755, over 4760.00 frames. ], tot_loss[loss=0.1815, simple_loss=0.248, pruned_loss=0.05754, over 511089.98 frames. ], batch size: 27, lr: 3.67e-03, grad_scale: 32.0 2023-04-27 02:20:21,860 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3040, 3.1877, 2.5671, 2.7383, 2.1966, 2.5005, 2.6857, 1.8919], device='cuda:4'), covar=tensor([0.2280, 0.1245, 0.0835, 0.1267, 0.2808, 0.1404, 0.2005, 0.2716], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0315, 0.0226, 0.0285, 0.0312, 0.0269, 0.0254, 0.0275], device='cuda:4'), out_proj_covar=tensor([1.1818e-04, 1.2643e-04, 9.0329e-05, 1.1409e-04, 1.2755e-04, 1.0817e-04, 1.0324e-04, 1.1014e-04], device='cuda:4') 2023-04-27 02:20:50,072 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.90 vs. limit=2.0 2023-04-27 02:20:58,537 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=63201.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:21:01,806 INFO [finetune.py:976] (4/7) Epoch 12, batch 200, loss[loss=0.1765, simple_loss=0.2414, pruned_loss=0.05574, over 4788.00 frames. ], tot_loss[loss=0.1819, simple_loss=0.2476, pruned_loss=0.05804, over 607646.60 frames. ], batch size: 29, lr: 3.67e-03, grad_scale: 32.0 2023-04-27 02:21:13,190 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=63221.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:21:24,494 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.086e+02 1.629e+02 1.966e+02 2.303e+02 3.666e+02, threshold=3.932e+02, percent-clipped=0.0 2023-04-27 02:21:25,240 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=63241.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:21:30,504 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=63249.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:21:33,534 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1692, 1.5040, 1.3919, 1.7580, 1.7139, 2.1386, 1.3280, 3.6339], device='cuda:4'), covar=tensor([0.0598, 0.0850, 0.0837, 0.1225, 0.0637, 0.0485, 0.0823, 0.0142], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0043, 0.0040, 0.0038, 0.0038, 0.0057], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0012, 0.0013, 0.0016], device='cuda:4') 2023-04-27 02:21:34,071 INFO [finetune.py:976] (4/7) Epoch 12, batch 250, loss[loss=0.219, simple_loss=0.2877, pruned_loss=0.07515, over 4900.00 frames. ], tot_loss[loss=0.1869, simple_loss=0.2537, pruned_loss=0.06009, over 685594.92 frames. ], batch size: 35, lr: 3.67e-03, grad_scale: 32.0 2023-04-27 02:21:38,675 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.9489, 3.7746, 2.9263, 4.5368, 3.9058, 3.9729, 1.7314, 3.9265], device='cuda:4'), covar=tensor([0.1683, 0.1197, 0.3394, 0.1427, 0.3271, 0.1852, 0.5710, 0.2178], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0216, 0.0250, 0.0304, 0.0298, 0.0248, 0.0271, 0.0270], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 02:21:46,375 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3223, 3.4717, 0.7329, 1.9295, 1.7901, 2.2912, 1.9291, 1.0209], device='cuda:4'), covar=tensor([0.1488, 0.0791, 0.2345, 0.1227, 0.1152, 0.1091, 0.1525, 0.2152], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0249, 0.0141, 0.0122, 0.0135, 0.0154, 0.0119, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 02:21:52,374 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=63282.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:21:56,554 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=63289.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:22:12,122 INFO [finetune.py:976] (4/7) Epoch 12, batch 300, loss[loss=0.1912, simple_loss=0.2679, pruned_loss=0.0573, over 4759.00 frames. ], tot_loss[loss=0.1897, simple_loss=0.2576, pruned_loss=0.06083, over 745307.41 frames. ], batch size: 54, lr: 3.67e-03, grad_scale: 32.0 2023-04-27 02:22:36,082 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=63322.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:22:46,901 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.130e+02 1.771e+02 2.081e+02 2.585e+02 5.314e+02, threshold=4.161e+02, percent-clipped=4.0 2023-04-27 02:22:48,254 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5730, 1.9038, 2.4055, 3.0166, 2.3422, 1.8724, 1.7668, 2.3315], device='cuda:4'), covar=tensor([0.3537, 0.3565, 0.1653, 0.2810, 0.3041, 0.2884, 0.4214, 0.2503], device='cuda:4'), in_proj_covar=tensor([0.0283, 0.0247, 0.0221, 0.0313, 0.0213, 0.0227, 0.0229, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 02:23:01,681 INFO [finetune.py:976] (4/7) Epoch 12, batch 350, loss[loss=0.2155, simple_loss=0.2884, pruned_loss=0.07136, over 4791.00 frames. ], tot_loss[loss=0.1893, simple_loss=0.2578, pruned_loss=0.06041, over 793857.65 frames. ], batch size: 45, lr: 3.67e-03, grad_scale: 32.0 2023-04-27 02:23:19,007 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=63370.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:23:57,901 INFO [finetune.py:976] (4/7) Epoch 12, batch 400, loss[loss=0.2057, simple_loss=0.2735, pruned_loss=0.06899, over 4865.00 frames. ], tot_loss[loss=0.1912, simple_loss=0.2595, pruned_loss=0.06139, over 829807.52 frames. ], batch size: 34, lr: 3.66e-03, grad_scale: 32.0 2023-04-27 02:24:00,893 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6468, 1.7114, 1.7319, 2.3659, 2.5537, 2.1298, 2.0449, 1.8441], device='cuda:4'), covar=tensor([0.1354, 0.1745, 0.1941, 0.1504, 0.1094, 0.1861, 0.2112, 0.2222], device='cuda:4'), in_proj_covar=tensor([0.0305, 0.0317, 0.0351, 0.0295, 0.0330, 0.0315, 0.0304, 0.0357], device='cuda:4'), out_proj_covar=tensor([6.3809e-05, 6.6772e-05, 7.5441e-05, 6.0713e-05, 6.8876e-05, 6.7081e-05, 6.4923e-05, 7.6517e-05], device='cuda:4') 2023-04-27 02:24:19,299 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=63420.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:24:21,032 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=63422.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:24:31,386 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=63429.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:24:43,211 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.266e+02 1.590e+02 1.875e+02 2.265e+02 5.610e+02, threshold=3.751e+02, percent-clipped=2.0 2023-04-27 02:24:51,862 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.84 vs. limit=2.0 2023-04-27 02:25:03,611 INFO [finetune.py:976] (4/7) Epoch 12, batch 450, loss[loss=0.1885, simple_loss=0.2606, pruned_loss=0.05814, over 4809.00 frames. ], tot_loss[loss=0.1891, simple_loss=0.2572, pruned_loss=0.06051, over 855993.13 frames. ], batch size: 25, lr: 3.66e-03, grad_scale: 32.0 2023-04-27 02:25:38,225 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=63477.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:25:46,566 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=63481.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:25:47,785 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=63483.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:26:12,255 INFO [finetune.py:976] (4/7) Epoch 12, batch 500, loss[loss=0.1449, simple_loss=0.209, pruned_loss=0.0404, over 4851.00 frames. ], tot_loss[loss=0.1874, simple_loss=0.2553, pruned_loss=0.05981, over 878333.74 frames. ], batch size: 44, lr: 3.66e-03, grad_scale: 32.0 2023-04-27 02:26:41,244 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.737e+01 1.732e+02 1.941e+02 2.281e+02 3.295e+02, threshold=3.881e+02, percent-clipped=0.0 2023-04-27 02:26:50,396 INFO [finetune.py:976] (4/7) Epoch 12, batch 550, loss[loss=0.1894, simple_loss=0.2545, pruned_loss=0.06213, over 4938.00 frames. ], tot_loss[loss=0.1867, simple_loss=0.2534, pruned_loss=0.06003, over 894993.80 frames. ], batch size: 33, lr: 3.66e-03, grad_scale: 32.0 2023-04-27 02:26:59,470 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=63569.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:27:05,806 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=63577.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:27:12,196 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=63586.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:27:23,756 INFO [finetune.py:976] (4/7) Epoch 12, batch 600, loss[loss=0.1791, simple_loss=0.2429, pruned_loss=0.0576, over 4781.00 frames. ], tot_loss[loss=0.1865, simple_loss=0.2534, pruned_loss=0.05981, over 909806.55 frames. ], batch size: 26, lr: 3.66e-03, grad_scale: 32.0 2023-04-27 02:27:29,303 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.6847, 3.5683, 2.7807, 4.2903, 3.6943, 3.6309, 1.7311, 3.7322], device='cuda:4'), covar=tensor([0.1736, 0.1432, 0.3032, 0.1663, 0.3655, 0.1959, 0.5523, 0.2108], device='cuda:4'), in_proj_covar=tensor([0.0243, 0.0216, 0.0248, 0.0302, 0.0297, 0.0246, 0.0269, 0.0269], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 02:27:33,472 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1033, 2.4436, 2.0687, 2.2561, 1.7266, 2.0614, 2.1692, 1.7975], device='cuda:4'), covar=tensor([0.1307, 0.0865, 0.0727, 0.0843, 0.2311, 0.0825, 0.1242, 0.1631], device='cuda:4'), in_proj_covar=tensor([0.0291, 0.0312, 0.0224, 0.0281, 0.0309, 0.0266, 0.0253, 0.0273], device='cuda:4'), out_proj_covar=tensor([1.1746e-04, 1.2520e-04, 8.9817e-05, 1.1254e-04, 1.2631e-04, 1.0671e-04, 1.0286e-04, 1.0902e-04], device='cuda:4') 2023-04-27 02:27:41,599 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=63630.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:27:46,596 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.72 vs. limit=2.0 2023-04-27 02:27:48,534 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.049e+02 1.750e+02 2.019e+02 2.576e+02 5.185e+02, threshold=4.039e+02, percent-clipped=2.0 2023-04-27 02:27:52,948 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=63647.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:27:57,222 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.5218, 3.5281, 2.6235, 4.1190, 3.5805, 3.5578, 1.7739, 3.4993], device='cuda:4'), covar=tensor([0.1698, 0.1302, 0.3249, 0.1906, 0.3554, 0.1737, 0.5029, 0.2482], device='cuda:4'), in_proj_covar=tensor([0.0245, 0.0217, 0.0250, 0.0304, 0.0299, 0.0248, 0.0271, 0.0272], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 02:27:57,765 INFO [finetune.py:976] (4/7) Epoch 12, batch 650, loss[loss=0.2266, simple_loss=0.2944, pruned_loss=0.07941, over 4799.00 frames. ], tot_loss[loss=0.1888, simple_loss=0.2561, pruned_loss=0.06069, over 919891.98 frames. ], batch size: 41, lr: 3.66e-03, grad_scale: 32.0 2023-04-27 02:28:15,869 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.49 vs. limit=5.0 2023-04-27 02:28:42,667 INFO [finetune.py:976] (4/7) Epoch 12, batch 700, loss[loss=0.2132, simple_loss=0.267, pruned_loss=0.07965, over 4831.00 frames. ], tot_loss[loss=0.1893, simple_loss=0.2574, pruned_loss=0.06063, over 926503.97 frames. ], batch size: 49, lr: 3.66e-03, grad_scale: 32.0 2023-04-27 02:29:23,284 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.162e+02 1.645e+02 1.982e+02 2.413e+02 5.979e+02, threshold=3.963e+02, percent-clipped=2.0 2023-04-27 02:29:32,955 INFO [finetune.py:976] (4/7) Epoch 12, batch 750, loss[loss=0.2016, simple_loss=0.2779, pruned_loss=0.06262, over 4798.00 frames. ], tot_loss[loss=0.1909, simple_loss=0.2594, pruned_loss=0.06118, over 933584.88 frames. ], batch size: 51, lr: 3.66e-03, grad_scale: 32.0 2023-04-27 02:29:34,361 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-27 02:29:45,740 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.47 vs. limit=2.0 2023-04-27 02:29:45,878 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=63776.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:29:47,607 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=63778.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:29:52,420 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8711, 2.3947, 2.0045, 2.1489, 1.6538, 1.9840, 2.0186, 1.5304], device='cuda:4'), covar=tensor([0.2059, 0.1366, 0.0941, 0.1397, 0.3097, 0.1188, 0.2139, 0.2496], device='cuda:4'), in_proj_covar=tensor([0.0289, 0.0311, 0.0223, 0.0279, 0.0308, 0.0265, 0.0251, 0.0271], device='cuda:4'), out_proj_covar=tensor([1.1687e-04, 1.2465e-04, 8.9252e-05, 1.1170e-04, 1.2595e-04, 1.0627e-04, 1.0216e-04, 1.0833e-04], device='cuda:4') 2023-04-27 02:30:06,645 INFO [finetune.py:976] (4/7) Epoch 12, batch 800, loss[loss=0.185, simple_loss=0.2547, pruned_loss=0.05761, over 4900.00 frames. ], tot_loss[loss=0.1898, simple_loss=0.2583, pruned_loss=0.06064, over 937306.63 frames. ], batch size: 37, lr: 3.66e-03, grad_scale: 32.0 2023-04-27 02:30:09,826 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=63810.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:30:16,271 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=5.40 vs. limit=5.0 2023-04-27 02:30:40,782 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.251e+02 1.671e+02 1.936e+02 2.427e+02 3.896e+02, threshold=3.873e+02, percent-clipped=0.0 2023-04-27 02:30:47,476 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.47 vs. limit=2.0 2023-04-27 02:30:49,751 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=63845.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:31:01,220 INFO [finetune.py:976] (4/7) Epoch 12, batch 850, loss[loss=0.2261, simple_loss=0.282, pruned_loss=0.0851, over 4806.00 frames. ], tot_loss[loss=0.1884, simple_loss=0.2561, pruned_loss=0.06037, over 939610.57 frames. ], batch size: 45, lr: 3.66e-03, grad_scale: 32.0 2023-04-27 02:31:21,222 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=63871.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:31:30,637 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=63877.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:32:06,884 INFO [finetune.py:976] (4/7) Epoch 12, batch 900, loss[loss=0.1544, simple_loss=0.2342, pruned_loss=0.03728, over 4786.00 frames. ], tot_loss[loss=0.1862, simple_loss=0.2533, pruned_loss=0.05958, over 944263.41 frames. ], batch size: 29, lr: 3.66e-03, grad_scale: 32.0 2023-04-27 02:32:07,623 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=63906.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:32:19,241 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=63916.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:32:24,720 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=63925.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:32:24,738 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=63925.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:32:34,606 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.028e+02 1.661e+02 1.966e+02 2.302e+02 4.372e+02, threshold=3.933e+02, percent-clipped=1.0 2023-04-27 02:32:36,449 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=63942.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:32:45,688 INFO [finetune.py:976] (4/7) Epoch 12, batch 950, loss[loss=0.1722, simple_loss=0.2453, pruned_loss=0.04951, over 4932.00 frames. ], tot_loss[loss=0.1856, simple_loss=0.252, pruned_loss=0.05964, over 947992.41 frames. ], batch size: 33, lr: 3.66e-03, grad_scale: 32.0 2023-04-27 02:32:59,637 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=63977.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:33:20,847 INFO [finetune.py:976] (4/7) Epoch 12, batch 1000, loss[loss=0.1555, simple_loss=0.2205, pruned_loss=0.04521, over 4743.00 frames. ], tot_loss[loss=0.1881, simple_loss=0.2547, pruned_loss=0.06075, over 948280.90 frames. ], batch size: 23, lr: 3.66e-03, grad_scale: 32.0 2023-04-27 02:33:32,469 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=64023.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:33:43,224 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.750e+01 1.665e+02 1.998e+02 2.380e+02 4.049e+02, threshold=3.995e+02, percent-clipped=1.0 2023-04-27 02:33:47,508 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-27 02:33:59,168 INFO [finetune.py:976] (4/7) Epoch 12, batch 1050, loss[loss=0.1759, simple_loss=0.2395, pruned_loss=0.05616, over 4783.00 frames. ], tot_loss[loss=0.1896, simple_loss=0.2567, pruned_loss=0.06124, over 949321.60 frames. ], batch size: 26, lr: 3.66e-03, grad_scale: 32.0 2023-04-27 02:34:29,323 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=64076.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:34:30,597 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=64078.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:34:40,127 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=64084.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:34:50,852 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2014, 2.0347, 2.3224, 2.7046, 2.6890, 2.1626, 1.7802, 2.3147], device='cuda:4'), covar=tensor([0.0978, 0.1082, 0.0671, 0.0581, 0.0645, 0.0948, 0.1035, 0.0632], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0200, 0.0182, 0.0172, 0.0177, 0.0186, 0.0158, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 02:35:05,820 INFO [finetune.py:976] (4/7) Epoch 12, batch 1100, loss[loss=0.1869, simple_loss=0.2583, pruned_loss=0.05773, over 4862.00 frames. ], tot_loss[loss=0.1912, simple_loss=0.2583, pruned_loss=0.06208, over 950003.78 frames. ], batch size: 31, lr: 3.66e-03, grad_scale: 32.0 2023-04-27 02:35:06,450 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4534, 2.1264, 1.7941, 1.9308, 2.2580, 1.8708, 2.5107, 1.5835], device='cuda:4'), covar=tensor([0.3794, 0.1719, 0.4257, 0.3136, 0.1613, 0.2322, 0.1880, 0.4593], device='cuda:4'), in_proj_covar=tensor([0.0343, 0.0349, 0.0431, 0.0361, 0.0385, 0.0382, 0.0377, 0.0423], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 02:35:28,512 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=64124.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:35:29,729 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=64126.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:35:38,342 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.140e+02 1.705e+02 2.099e+02 2.623e+02 4.712e+02, threshold=4.198e+02, percent-clipped=5.0 2023-04-27 02:35:49,424 INFO [finetune.py:976] (4/7) Epoch 12, batch 1150, loss[loss=0.1619, simple_loss=0.2411, pruned_loss=0.04138, over 4927.00 frames. ], tot_loss[loss=0.1909, simple_loss=0.2586, pruned_loss=0.06157, over 950937.19 frames. ], batch size: 42, lr: 3.66e-03, grad_scale: 32.0 2023-04-27 02:35:57,680 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=64166.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:36:19,851 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=64201.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:36:22,239 INFO [finetune.py:976] (4/7) Epoch 12, batch 1200, loss[loss=0.2027, simple_loss=0.2636, pruned_loss=0.07091, over 4900.00 frames. ], tot_loss[loss=0.1889, simple_loss=0.2566, pruned_loss=0.06061, over 952940.37 frames. ], batch size: 46, lr: 3.66e-03, grad_scale: 32.0 2023-04-27 02:36:35,461 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1571, 2.4661, 0.9977, 1.3346, 1.9242, 1.1800, 3.0414, 1.5897], device='cuda:4'), covar=tensor([0.0606, 0.0536, 0.0726, 0.1195, 0.0436, 0.0949, 0.0266, 0.0630], device='cuda:4'), in_proj_covar=tensor([0.0052, 0.0067, 0.0049, 0.0047, 0.0051, 0.0052, 0.0077, 0.0052], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 02:36:36,058 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=64225.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:36:40,387 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=64232.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:36:50,584 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.051e+02 1.610e+02 1.933e+02 2.309e+02 4.139e+02, threshold=3.865e+02, percent-clipped=0.0 2023-04-27 02:36:51,915 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=64242.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:37:10,837 INFO [finetune.py:976] (4/7) Epoch 12, batch 1250, loss[loss=0.1728, simple_loss=0.2395, pruned_loss=0.05301, over 4830.00 frames. ], tot_loss[loss=0.1875, simple_loss=0.2544, pruned_loss=0.06032, over 953489.54 frames. ], batch size: 33, lr: 3.66e-03, grad_scale: 64.0 2023-04-27 02:37:33,223 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=64272.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:37:33,823 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=64273.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:37:36,324 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1476, 1.2952, 1.2226, 1.6316, 1.4668, 1.5173, 1.2709, 2.5412], device='cuda:4'), covar=tensor([0.0648, 0.1033, 0.1021, 0.1305, 0.0823, 0.0576, 0.0942, 0.0303], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0043, 0.0040, 0.0038, 0.0038, 0.0057], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 02:37:55,439 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=64290.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:37:57,318 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=64293.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:38:16,203 INFO [finetune.py:976] (4/7) Epoch 12, batch 1300, loss[loss=0.1751, simple_loss=0.2422, pruned_loss=0.05405, over 4917.00 frames. ], tot_loss[loss=0.185, simple_loss=0.2515, pruned_loss=0.05923, over 955341.62 frames. ], batch size: 37, lr: 3.66e-03, grad_scale: 32.0 2023-04-27 02:38:37,483 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-27 02:38:40,959 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.098e+02 1.564e+02 1.871e+02 2.311e+02 4.814e+02, threshold=3.742e+02, percent-clipped=4.0 2023-04-27 02:38:43,498 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=64345.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:38:49,953 INFO [finetune.py:976] (4/7) Epoch 12, batch 1350, loss[loss=0.1711, simple_loss=0.2403, pruned_loss=0.05098, over 4734.00 frames. ], tot_loss[loss=0.1856, simple_loss=0.2518, pruned_loss=0.0597, over 956611.30 frames. ], batch size: 23, lr: 3.66e-03, grad_scale: 32.0 2023-04-27 02:39:01,816 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9949, 2.6300, 1.9764, 1.9203, 1.4103, 1.3138, 2.2349, 1.3604], device='cuda:4'), covar=tensor([0.1708, 0.1529, 0.1455, 0.1901, 0.2387, 0.2026, 0.0931, 0.2166], device='cuda:4'), in_proj_covar=tensor([0.0193, 0.0213, 0.0168, 0.0203, 0.0200, 0.0183, 0.0157, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 02:39:02,784 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.34 vs. limit=2.0 2023-04-27 02:39:07,143 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=64379.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:39:22,954 INFO [finetune.py:976] (4/7) Epoch 12, batch 1400, loss[loss=0.221, simple_loss=0.2891, pruned_loss=0.07646, over 4820.00 frames. ], tot_loss[loss=0.1881, simple_loss=0.2551, pruned_loss=0.06058, over 956278.98 frames. ], batch size: 40, lr: 3.66e-03, grad_scale: 32.0 2023-04-27 02:39:23,729 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=64406.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:39:48,147 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.081e+02 1.746e+02 2.220e+02 2.632e+02 4.800e+02, threshold=4.440e+02, percent-clipped=5.0 2023-04-27 02:40:07,659 INFO [finetune.py:976] (4/7) Epoch 12, batch 1450, loss[loss=0.2305, simple_loss=0.2867, pruned_loss=0.08714, over 4837.00 frames. ], tot_loss[loss=0.1902, simple_loss=0.2575, pruned_loss=0.06145, over 954859.77 frames. ], batch size: 30, lr: 3.66e-03, grad_scale: 32.0 2023-04-27 02:40:21,226 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=64465.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:40:21,840 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=64466.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:40:55,250 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=64501.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:40:57,611 INFO [finetune.py:976] (4/7) Epoch 12, batch 1500, loss[loss=0.1925, simple_loss=0.2561, pruned_loss=0.06442, over 4923.00 frames. ], tot_loss[loss=0.1922, simple_loss=0.2595, pruned_loss=0.06243, over 954167.91 frames. ], batch size: 33, lr: 3.66e-03, grad_scale: 32.0 2023-04-27 02:41:03,987 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=64514.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:41:06,602 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.12 vs. limit=2.0 2023-04-27 02:41:12,860 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=64526.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:41:22,337 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.729e+02 2.014e+02 2.511e+02 4.004e+02, threshold=4.028e+02, percent-clipped=0.0 2023-04-27 02:41:27,207 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=64549.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:41:30,771 INFO [finetune.py:976] (4/7) Epoch 12, batch 1550, loss[loss=0.1912, simple_loss=0.2487, pruned_loss=0.06685, over 4177.00 frames. ], tot_loss[loss=0.1923, simple_loss=0.26, pruned_loss=0.06226, over 954623.30 frames. ], batch size: 65, lr: 3.66e-03, grad_scale: 32.0 2023-04-27 02:41:42,589 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=64572.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:41:53,768 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=64588.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:41:54,936 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=64589.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:42:10,372 INFO [finetune.py:976] (4/7) Epoch 12, batch 1600, loss[loss=0.1921, simple_loss=0.248, pruned_loss=0.06809, over 4866.00 frames. ], tot_loss[loss=0.1897, simple_loss=0.2571, pruned_loss=0.06117, over 954378.02 frames. ], batch size: 31, lr: 3.65e-03, grad_scale: 32.0 2023-04-27 02:42:32,225 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=64620.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:42:46,405 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.047e+02 1.700e+02 1.959e+02 2.289e+02 5.207e+02, threshold=3.917e+02, percent-clipped=1.0 2023-04-27 02:42:56,558 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=64648.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:42:57,761 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=64650.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:43:06,461 INFO [finetune.py:976] (4/7) Epoch 12, batch 1650, loss[loss=0.1226, simple_loss=0.2004, pruned_loss=0.02238, over 4845.00 frames. ], tot_loss[loss=0.1869, simple_loss=0.254, pruned_loss=0.05994, over 953213.26 frames. ], batch size: 47, lr: 3.65e-03, grad_scale: 32.0 2023-04-27 02:43:10,293 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-27 02:43:18,555 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-27 02:43:36,871 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=64679.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:44:09,944 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=64701.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:44:11,221 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=64703.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:44:12,323 INFO [finetune.py:976] (4/7) Epoch 12, batch 1700, loss[loss=0.2009, simple_loss=0.2662, pruned_loss=0.06784, over 4814.00 frames. ], tot_loss[loss=0.1866, simple_loss=0.2533, pruned_loss=0.05993, over 953519.37 frames. ], batch size: 41, lr: 3.65e-03, grad_scale: 32.0 2023-04-27 02:44:14,917 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=64709.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:44:27,405 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=64727.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:44:37,467 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.062e+02 1.591e+02 1.842e+02 2.298e+02 5.493e+02, threshold=3.684e+02, percent-clipped=1.0 2023-04-27 02:44:40,032 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=64745.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:44:44,739 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=64752.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:44:46,477 INFO [finetune.py:976] (4/7) Epoch 12, batch 1750, loss[loss=0.1704, simple_loss=0.238, pruned_loss=0.05142, over 4899.00 frames. ], tot_loss[loss=0.1877, simple_loss=0.2547, pruned_loss=0.06032, over 955192.63 frames. ], batch size: 35, lr: 3.65e-03, grad_scale: 32.0 2023-04-27 02:44:51,990 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=64764.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:45:04,719 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0513, 2.5518, 2.2660, 2.4422, 1.7925, 2.0502, 2.2038, 1.7437], device='cuda:4'), covar=tensor([0.1940, 0.1422, 0.0823, 0.1005, 0.2935, 0.1103, 0.1844, 0.2344], device='cuda:4'), in_proj_covar=tensor([0.0291, 0.0312, 0.0224, 0.0280, 0.0311, 0.0265, 0.0251, 0.0272], device='cuda:4'), out_proj_covar=tensor([1.1733e-04, 1.2504e-04, 8.9472e-05, 1.1186e-04, 1.2686e-04, 1.0654e-04, 1.0226e-04, 1.0886e-04], device='cuda:4') 2023-04-27 02:45:13,643 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0156, 2.4722, 1.0342, 1.4375, 1.9774, 1.1501, 3.4388, 1.7453], device='cuda:4'), covar=tensor([0.0662, 0.0724, 0.0770, 0.1298, 0.0499, 0.1121, 0.0201, 0.0662], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0049, 0.0047, 0.0051, 0.0053, 0.0076, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 02:45:19,082 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0738, 1.7380, 2.0331, 2.3550, 2.4321, 1.8915, 1.6081, 2.2012], device='cuda:4'), covar=tensor([0.0806, 0.1093, 0.0650, 0.0540, 0.0573, 0.0887, 0.0863, 0.0523], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0201, 0.0182, 0.0173, 0.0177, 0.0187, 0.0158, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 02:45:20,205 INFO [finetune.py:976] (4/7) Epoch 12, batch 1800, loss[loss=0.2136, simple_loss=0.2779, pruned_loss=0.07464, over 4887.00 frames. ], tot_loss[loss=0.1911, simple_loss=0.2584, pruned_loss=0.06192, over 954470.26 frames. ], batch size: 32, lr: 3.65e-03, grad_scale: 32.0 2023-04-27 02:45:20,915 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=64806.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:45:25,149 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=64813.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:45:29,915 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=64821.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:45:36,387 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.8372, 1.8057, 1.8429, 1.5237, 2.0157, 1.6239, 2.6119, 1.5531], device='cuda:4'), covar=tensor([0.4119, 0.1971, 0.4346, 0.3214, 0.1687, 0.2536, 0.1279, 0.4677], device='cuda:4'), in_proj_covar=tensor([0.0342, 0.0349, 0.0433, 0.0360, 0.0388, 0.0383, 0.0378, 0.0424], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 02:45:50,967 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.75 vs. limit=5.0 2023-04-27 02:46:00,005 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.230e+02 1.632e+02 2.043e+02 2.586e+02 5.836e+02, threshold=4.087e+02, percent-clipped=7.0 2023-04-27 02:46:21,070 INFO [finetune.py:976] (4/7) Epoch 12, batch 1850, loss[loss=0.1841, simple_loss=0.2331, pruned_loss=0.06758, over 4387.00 frames. ], tot_loss[loss=0.1916, simple_loss=0.2591, pruned_loss=0.062, over 954877.37 frames. ], batch size: 19, lr: 3.65e-03, grad_scale: 32.0 2023-04-27 02:46:58,521 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=64888.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:47:10,163 INFO [finetune.py:976] (4/7) Epoch 12, batch 1900, loss[loss=0.1644, simple_loss=0.2327, pruned_loss=0.04809, over 4185.00 frames. ], tot_loss[loss=0.19, simple_loss=0.2579, pruned_loss=0.06111, over 952916.26 frames. ], batch size: 66, lr: 3.65e-03, grad_scale: 32.0 2023-04-27 02:47:35,553 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=64936.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:47:37,982 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2715, 1.6507, 1.6864, 1.8185, 1.6485, 1.7836, 1.8024, 1.7288], device='cuda:4'), covar=tensor([0.5270, 0.6621, 0.5741, 0.5123, 0.6588, 0.8814, 0.6909, 0.6370], device='cuda:4'), in_proj_covar=tensor([0.0327, 0.0376, 0.0315, 0.0326, 0.0338, 0.0398, 0.0358, 0.0324], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 02:47:39,045 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.204e+02 1.650e+02 1.962e+02 2.439e+02 4.068e+02, threshold=3.925e+02, percent-clipped=0.0 2023-04-27 02:47:48,349 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=64945.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:48:00,716 INFO [finetune.py:976] (4/7) Epoch 12, batch 1950, loss[loss=0.1618, simple_loss=0.2299, pruned_loss=0.04689, over 4820.00 frames. ], tot_loss[loss=0.1881, simple_loss=0.2559, pruned_loss=0.06008, over 955782.32 frames. ], batch size: 41, lr: 3.65e-03, grad_scale: 32.0 2023-04-27 02:48:31,521 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=64976.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:48:47,448 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4593, 1.5991, 1.6491, 2.3138, 2.4110, 2.0341, 1.9632, 1.7489], device='cuda:4'), covar=tensor([0.1658, 0.1880, 0.2215, 0.1246, 0.1210, 0.2048, 0.2378, 0.2220], device='cuda:4'), in_proj_covar=tensor([0.0302, 0.0314, 0.0349, 0.0292, 0.0329, 0.0314, 0.0303, 0.0357], device='cuda:4'), out_proj_covar=tensor([6.3177e-05, 6.6105e-05, 7.4952e-05, 6.0009e-05, 6.8757e-05, 6.6792e-05, 6.4659e-05, 7.6653e-05], device='cuda:4') 2023-04-27 02:48:49,870 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=65001.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:48:50,194 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.24 vs. limit=5.0 2023-04-27 02:48:51,660 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=65004.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:48:52,189 INFO [finetune.py:976] (4/7) Epoch 12, batch 2000, loss[loss=0.1943, simple_loss=0.2585, pruned_loss=0.06506, over 4877.00 frames. ], tot_loss[loss=0.1858, simple_loss=0.2531, pruned_loss=0.05925, over 954954.78 frames. ], batch size: 31, lr: 3.65e-03, grad_scale: 32.0 2023-04-27 02:48:54,742 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5369, 1.3562, 1.6992, 1.7639, 1.3618, 1.2183, 1.4801, 0.8724], device='cuda:4'), covar=tensor([0.0557, 0.0800, 0.0478, 0.0697, 0.0779, 0.1158, 0.0663, 0.0800], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0072, 0.0071, 0.0068, 0.0076, 0.0098, 0.0076, 0.0072], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 02:48:54,819 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.90 vs. limit=5.0 2023-04-27 02:48:55,971 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7519, 1.3812, 1.4012, 1.6498, 1.9813, 1.6420, 1.4222, 1.3273], device='cuda:4'), covar=tensor([0.1448, 0.1442, 0.1580, 0.1170, 0.0704, 0.1377, 0.1961, 0.1981], device='cuda:4'), in_proj_covar=tensor([0.0302, 0.0315, 0.0349, 0.0292, 0.0329, 0.0314, 0.0303, 0.0357], device='cuda:4'), out_proj_covar=tensor([6.3141e-05, 6.6161e-05, 7.4906e-05, 5.9987e-05, 6.8769e-05, 6.6828e-05, 6.4649e-05, 7.6654e-05], device='cuda:4') 2023-04-27 02:49:12,508 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=65037.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 02:49:15,251 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.045e+02 1.562e+02 1.931e+02 2.366e+02 5.248e+02, threshold=3.862e+02, percent-clipped=2.0 2023-04-27 02:49:21,704 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=65049.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:49:26,343 INFO [finetune.py:976] (4/7) Epoch 12, batch 2050, loss[loss=0.2041, simple_loss=0.2631, pruned_loss=0.0726, over 4900.00 frames. ], tot_loss[loss=0.1824, simple_loss=0.2494, pruned_loss=0.05768, over 957053.18 frames. ], batch size: 35, lr: 3.65e-03, grad_scale: 32.0 2023-04-27 02:49:26,461 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=65055.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:49:28,837 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=65059.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:49:31,311 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.62 vs. limit=5.0 2023-04-27 02:49:34,117 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-27 02:49:34,676 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0471, 2.6363, 2.3133, 2.3429, 1.8682, 2.2076, 2.1996, 1.7649], device='cuda:4'), covar=tensor([0.2143, 0.1390, 0.0848, 0.1413, 0.3480, 0.1445, 0.2172, 0.3175], device='cuda:4'), in_proj_covar=tensor([0.0289, 0.0310, 0.0222, 0.0279, 0.0309, 0.0263, 0.0250, 0.0272], device='cuda:4'), out_proj_covar=tensor([1.1657e-04, 1.2419e-04, 8.8676e-05, 1.1170e-04, 1.2595e-04, 1.0544e-04, 1.0164e-04, 1.0866e-04], device='cuda:4') 2023-04-27 02:49:56,651 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=65101.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:50:00,020 INFO [finetune.py:976] (4/7) Epoch 12, batch 2100, loss[loss=0.1748, simple_loss=0.2466, pruned_loss=0.05145, over 4795.00 frames. ], tot_loss[loss=0.1821, simple_loss=0.2486, pruned_loss=0.05777, over 955483.19 frames. ], batch size: 51, lr: 3.65e-03, grad_scale: 32.0 2023-04-27 02:50:01,929 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=65108.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:50:06,864 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=65116.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:50:10,336 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=65121.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:50:22,438 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.293e+02 1.658e+02 1.987e+02 2.538e+02 5.198e+02, threshold=3.973e+02, percent-clipped=2.0 2023-04-27 02:50:33,404 INFO [finetune.py:976] (4/7) Epoch 12, batch 2150, loss[loss=0.2009, simple_loss=0.2689, pruned_loss=0.06648, over 4857.00 frames. ], tot_loss[loss=0.185, simple_loss=0.2522, pruned_loss=0.05886, over 956309.74 frames. ], batch size: 49, lr: 3.65e-03, grad_scale: 32.0 2023-04-27 02:50:42,398 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=65169.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:51:05,952 INFO [finetune.py:976] (4/7) Epoch 12, batch 2200, loss[loss=0.1997, simple_loss=0.2656, pruned_loss=0.06689, over 4925.00 frames. ], tot_loss[loss=0.1879, simple_loss=0.2554, pruned_loss=0.06015, over 955856.14 frames. ], batch size: 42, lr: 3.65e-03, grad_scale: 32.0 2023-04-27 02:51:37,971 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.82 vs. limit=2.0 2023-04-27 02:51:56,236 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.062e+02 1.743e+02 2.057e+02 2.570e+02 5.251e+02, threshold=4.115e+02, percent-clipped=3.0 2023-04-27 02:51:58,807 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=65245.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:52:11,643 INFO [finetune.py:976] (4/7) Epoch 12, batch 2250, loss[loss=0.1764, simple_loss=0.256, pruned_loss=0.04845, over 4897.00 frames. ], tot_loss[loss=0.1889, simple_loss=0.2571, pruned_loss=0.06031, over 958115.87 frames. ], batch size: 43, lr: 3.65e-03, grad_scale: 32.0 2023-04-27 02:52:53,430 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=65285.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:53:03,654 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=65293.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:53:21,397 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=65304.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:53:21,921 INFO [finetune.py:976] (4/7) Epoch 12, batch 2300, loss[loss=0.1871, simple_loss=0.2575, pruned_loss=0.0584, over 4742.00 frames. ], tot_loss[loss=0.1884, simple_loss=0.2569, pruned_loss=0.05996, over 955294.77 frames. ], batch size: 54, lr: 3.65e-03, grad_scale: 32.0 2023-04-27 02:53:23,795 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3778, 1.7895, 2.2389, 2.8902, 2.2349, 1.7891, 1.8190, 2.1102], device='cuda:4'), covar=tensor([0.3130, 0.3474, 0.1510, 0.2486, 0.2921, 0.2672, 0.4026, 0.2430], device='cuda:4'), in_proj_covar=tensor([0.0288, 0.0251, 0.0224, 0.0318, 0.0215, 0.0231, 0.0232, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 02:53:46,715 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6046, 2.6966, 2.4098, 2.3794, 2.6354, 2.5700, 3.8678, 2.2236], device='cuda:4'), covar=tensor([0.4471, 0.2239, 0.4625, 0.3982, 0.2389, 0.2760, 0.1403, 0.4412], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0347, 0.0429, 0.0359, 0.0385, 0.0382, 0.0375, 0.0422], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 02:53:56,236 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=65332.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 02:54:07,782 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.005e+02 1.597e+02 1.792e+02 2.067e+02 3.757e+02, threshold=3.584e+02, percent-clipped=0.0 2023-04-27 02:54:08,523 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9550, 1.6852, 2.1016, 2.3259, 2.0321, 1.8249, 1.9806, 1.9156], device='cuda:4'), covar=tensor([0.5451, 0.7501, 0.8143, 0.6588, 0.5989, 0.9647, 0.9659, 1.0200], device='cuda:4'), in_proj_covar=tensor([0.0409, 0.0406, 0.0496, 0.0514, 0.0440, 0.0461, 0.0466, 0.0470], device='cuda:4'), out_proj_covar=tensor([9.9441e-05, 1.0064e-04, 1.1192e-04, 1.2186e-04, 1.0658e-04, 1.1124e-04, 1.1166e-04, 1.1246e-04], device='cuda:4') 2023-04-27 02:54:16,353 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=65346.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:54:20,874 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=65352.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:54:22,653 INFO [finetune.py:976] (4/7) Epoch 12, batch 2350, loss[loss=0.1977, simple_loss=0.2573, pruned_loss=0.06902, over 4112.00 frames. ], tot_loss[loss=0.1862, simple_loss=0.2542, pruned_loss=0.05909, over 954411.37 frames. ], batch size: 65, lr: 3.65e-03, grad_scale: 32.0 2023-04-27 02:54:25,649 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=65359.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:54:41,183 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-04-27 02:54:53,204 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=65401.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:54:55,974 INFO [finetune.py:976] (4/7) Epoch 12, batch 2400, loss[loss=0.1529, simple_loss=0.2161, pruned_loss=0.0448, over 4839.00 frames. ], tot_loss[loss=0.1832, simple_loss=0.2508, pruned_loss=0.05781, over 954860.75 frames. ], batch size: 25, lr: 3.65e-03, grad_scale: 32.0 2023-04-27 02:54:57,713 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=65407.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:54:58,362 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=65408.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:55:00,657 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=65411.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:55:01,917 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9937, 1.7525, 2.0711, 2.3694, 2.3744, 1.9111, 1.4864, 2.1118], device='cuda:4'), covar=tensor([0.0854, 0.1111, 0.0665, 0.0546, 0.0569, 0.0907, 0.0913, 0.0557], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0201, 0.0182, 0.0174, 0.0177, 0.0188, 0.0158, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 02:55:03,432 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.88 vs. limit=5.0 2023-04-27 02:55:20,645 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.015e+02 1.563e+02 1.962e+02 2.388e+02 3.739e+02, threshold=3.925e+02, percent-clipped=1.0 2023-04-27 02:55:25,586 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=65449.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:55:29,229 INFO [finetune.py:976] (4/7) Epoch 12, batch 2450, loss[loss=0.1728, simple_loss=0.241, pruned_loss=0.05229, over 4905.00 frames. ], tot_loss[loss=0.1799, simple_loss=0.2472, pruned_loss=0.05634, over 954765.85 frames. ], batch size: 36, lr: 3.65e-03, grad_scale: 32.0 2023-04-27 02:55:30,348 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=65456.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:55:41,385 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.6029, 4.6022, 3.1718, 5.2769, 4.6215, 4.5721, 2.0369, 4.5203], device='cuda:4'), covar=tensor([0.1477, 0.0903, 0.2867, 0.0892, 0.3981, 0.1580, 0.5800, 0.2059], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0217, 0.0250, 0.0305, 0.0299, 0.0249, 0.0272, 0.0271], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 02:55:59,914 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3969, 3.4739, 0.7815, 1.9597, 1.8039, 2.4789, 2.0084, 0.8417], device='cuda:4'), covar=tensor([0.1574, 0.1092, 0.2319, 0.1288, 0.1232, 0.1061, 0.1575, 0.2275], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0249, 0.0141, 0.0123, 0.0135, 0.0154, 0.0119, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 02:56:02,904 INFO [finetune.py:976] (4/7) Epoch 12, batch 2500, loss[loss=0.2338, simple_loss=0.2969, pruned_loss=0.08532, over 4798.00 frames. ], tot_loss[loss=0.183, simple_loss=0.2503, pruned_loss=0.05784, over 952911.56 frames. ], batch size: 51, lr: 3.65e-03, grad_scale: 32.0 2023-04-27 02:56:09,483 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-27 02:56:22,237 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=65532.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 02:56:26,775 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-27 02:56:28,027 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.207e+02 1.847e+02 2.256e+02 2.656e+02 4.911e+02, threshold=4.512e+02, percent-clipped=5.0 2023-04-27 02:56:36,595 INFO [finetune.py:976] (4/7) Epoch 12, batch 2550, loss[loss=0.2012, simple_loss=0.2608, pruned_loss=0.07082, over 4769.00 frames. ], tot_loss[loss=0.186, simple_loss=0.2543, pruned_loss=0.05882, over 952917.99 frames. ], batch size: 54, lr: 3.65e-03, grad_scale: 16.0 2023-04-27 02:56:51,976 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.9900, 4.0686, 3.0796, 4.5728, 4.0157, 3.9417, 2.4727, 4.0327], device='cuda:4'), covar=tensor([0.1827, 0.0966, 0.2869, 0.1670, 0.3884, 0.2107, 0.4758, 0.2266], device='cuda:4'), in_proj_covar=tensor([0.0243, 0.0216, 0.0249, 0.0304, 0.0299, 0.0248, 0.0270, 0.0271], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 02:57:02,305 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=65593.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 02:57:04,487 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.7007, 3.7639, 2.7691, 4.2447, 3.7310, 3.6791, 1.7691, 3.6555], device='cuda:4'), covar=tensor([0.1642, 0.0999, 0.2872, 0.1766, 0.3833, 0.1765, 0.5211, 0.2340], device='cuda:4'), in_proj_covar=tensor([0.0242, 0.0215, 0.0248, 0.0303, 0.0298, 0.0248, 0.0270, 0.0270], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 02:57:10,044 INFO [finetune.py:976] (4/7) Epoch 12, batch 2600, loss[loss=0.2339, simple_loss=0.3026, pruned_loss=0.08254, over 4812.00 frames. ], tot_loss[loss=0.1885, simple_loss=0.2571, pruned_loss=0.05998, over 951408.70 frames. ], batch size: 40, lr: 3.65e-03, grad_scale: 16.0 2023-04-27 02:57:29,335 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=65632.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:57:34,710 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=65641.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:57:35,232 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.096e+02 1.708e+02 2.004e+02 2.304e+02 5.079e+02, threshold=4.008e+02, percent-clipped=1.0 2023-04-27 02:57:54,581 INFO [finetune.py:976] (4/7) Epoch 12, batch 2650, loss[loss=0.2076, simple_loss=0.2878, pruned_loss=0.06366, over 4872.00 frames. ], tot_loss[loss=0.1887, simple_loss=0.2573, pruned_loss=0.06005, over 951053.64 frames. ], batch size: 34, lr: 3.65e-03, grad_scale: 16.0 2023-04-27 02:58:26,147 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=65680.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:58:54,679 INFO [finetune.py:976] (4/7) Epoch 12, batch 2700, loss[loss=0.1901, simple_loss=0.2363, pruned_loss=0.07196, over 4022.00 frames. ], tot_loss[loss=0.1876, simple_loss=0.256, pruned_loss=0.05963, over 948581.32 frames. ], batch size: 17, lr: 3.65e-03, grad_scale: 16.0 2023-04-27 02:59:03,993 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=65711.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 02:59:06,984 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.0484, 3.9419, 2.8488, 4.5750, 4.0245, 3.9635, 1.7931, 3.8936], device='cuda:4'), covar=tensor([0.1642, 0.1143, 0.3197, 0.1661, 0.3587, 0.1761, 0.6153, 0.2662], device='cuda:4'), in_proj_covar=tensor([0.0243, 0.0216, 0.0250, 0.0304, 0.0299, 0.0249, 0.0272, 0.0271], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 02:59:41,282 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.81 vs. limit=2.0 2023-04-27 02:59:42,190 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.037e+02 1.570e+02 1.817e+02 2.252e+02 5.618e+02, threshold=3.635e+02, percent-clipped=2.0 2023-04-27 03:00:01,384 INFO [finetune.py:976] (4/7) Epoch 12, batch 2750, loss[loss=0.1559, simple_loss=0.2181, pruned_loss=0.0468, over 4796.00 frames. ], tot_loss[loss=0.1866, simple_loss=0.2539, pruned_loss=0.05966, over 949022.54 frames. ], batch size: 25, lr: 3.65e-03, grad_scale: 16.0 2023-04-27 03:00:09,045 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=65759.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:01:07,122 INFO [finetune.py:976] (4/7) Epoch 12, batch 2800, loss[loss=0.1735, simple_loss=0.2385, pruned_loss=0.05426, over 4912.00 frames. ], tot_loss[loss=0.1854, simple_loss=0.2519, pruned_loss=0.05946, over 951391.76 frames. ], batch size: 36, lr: 3.64e-03, grad_scale: 16.0 2023-04-27 03:01:51,946 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=65838.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 03:01:54,202 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0256, 2.4172, 0.9595, 1.2996, 1.8811, 1.1981, 3.2508, 1.6232], device='cuda:4'), covar=tensor([0.0638, 0.0604, 0.0758, 0.1251, 0.0486, 0.0967, 0.0225, 0.0621], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0048, 0.0047, 0.0050, 0.0052, 0.0077, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-27 03:01:59,670 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.127e+02 1.664e+02 1.878e+02 2.585e+02 4.047e+02, threshold=3.756e+02, percent-clipped=2.0 2023-04-27 03:02:08,131 INFO [finetune.py:976] (4/7) Epoch 12, batch 2850, loss[loss=0.2284, simple_loss=0.2931, pruned_loss=0.08189, over 4894.00 frames. ], tot_loss[loss=0.1831, simple_loss=0.25, pruned_loss=0.05813, over 954112.16 frames. ], batch size: 35, lr: 3.64e-03, grad_scale: 16.0 2023-04-27 03:02:29,546 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=65888.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 03:02:38,192 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=65899.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 03:02:41,705 INFO [finetune.py:976] (4/7) Epoch 12, batch 2900, loss[loss=0.1502, simple_loss=0.2185, pruned_loss=0.04091, over 4758.00 frames. ], tot_loss[loss=0.1857, simple_loss=0.2531, pruned_loss=0.05912, over 952706.26 frames. ], batch size: 27, lr: 3.64e-03, grad_scale: 16.0 2023-04-27 03:03:05,041 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=65941.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:03:05,559 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.180e+02 1.732e+02 2.023e+02 2.445e+02 4.251e+02, threshold=4.047e+02, percent-clipped=2.0 2023-04-27 03:03:15,314 INFO [finetune.py:976] (4/7) Epoch 12, batch 2950, loss[loss=0.1872, simple_loss=0.2636, pruned_loss=0.05544, over 4799.00 frames. ], tot_loss[loss=0.1883, simple_loss=0.2564, pruned_loss=0.06005, over 950976.02 frames. ], batch size: 51, lr: 3.64e-03, grad_scale: 16.0 2023-04-27 03:03:19,102 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6363, 1.1161, 1.7104, 2.0682, 1.7289, 1.5999, 1.6497, 1.6475], device='cuda:4'), covar=tensor([0.5541, 0.7576, 0.7560, 0.7836, 0.6656, 0.9033, 0.8451, 0.9375], device='cuda:4'), in_proj_covar=tensor([0.0411, 0.0408, 0.0497, 0.0516, 0.0440, 0.0462, 0.0468, 0.0470], device='cuda:4'), out_proj_covar=tensor([9.9975e-05, 1.0124e-04, 1.1220e-04, 1.2246e-04, 1.0647e-04, 1.1148e-04, 1.1218e-04, 1.1245e-04], device='cuda:4') 2023-04-27 03:03:36,945 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=65989.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:03:43,070 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.50 vs. limit=5.0 2023-04-27 03:03:49,753 INFO [finetune.py:976] (4/7) Epoch 12, batch 3000, loss[loss=0.1869, simple_loss=0.2635, pruned_loss=0.0552, over 4811.00 frames. ], tot_loss[loss=0.1895, simple_loss=0.2576, pruned_loss=0.0607, over 951391.50 frames. ], batch size: 40, lr: 3.64e-03, grad_scale: 16.0 2023-04-27 03:03:49,753 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 03:03:51,899 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2051, 1.4141, 1.6923, 1.8267, 1.7562, 1.8917, 1.7091, 1.8017], device='cuda:4'), covar=tensor([0.4675, 0.6941, 0.5886, 0.5556, 0.6878, 0.9341, 0.6938, 0.5764], device='cuda:4'), in_proj_covar=tensor([0.0328, 0.0376, 0.0316, 0.0326, 0.0337, 0.0397, 0.0357, 0.0324], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 03:03:57,768 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2408, 2.5640, 1.0421, 1.4579, 1.9467, 1.4034, 3.0916, 1.7576], device='cuda:4'), covar=tensor([0.0598, 0.0529, 0.0683, 0.1259, 0.0430, 0.0841, 0.0292, 0.0561], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0048, 0.0047, 0.0050, 0.0052, 0.0077, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-27 03:04:00,410 INFO [finetune.py:1010] (4/7) Epoch 12, validation: loss=0.1529, simple_loss=0.2247, pruned_loss=0.04052, over 2265189.00 frames. 2023-04-27 03:04:00,411 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-27 03:04:13,934 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3378, 2.9788, 0.9004, 1.5244, 2.2856, 1.6480, 4.2071, 1.8997], device='cuda:4'), covar=tensor([0.0818, 0.0809, 0.1057, 0.1865, 0.0715, 0.1289, 0.0498, 0.0896], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0048, 0.0047, 0.0050, 0.0052, 0.0077, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-27 03:04:29,438 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.284e+02 1.787e+02 2.243e+02 2.841e+02 1.392e+03, threshold=4.486e+02, percent-clipped=4.0 2023-04-27 03:04:38,274 INFO [finetune.py:976] (4/7) Epoch 12, batch 3050, loss[loss=0.2632, simple_loss=0.3197, pruned_loss=0.1034, over 4849.00 frames. ], tot_loss[loss=0.1905, simple_loss=0.2589, pruned_loss=0.06103, over 951809.52 frames. ], batch size: 44, lr: 3.64e-03, grad_scale: 16.0 2023-04-27 03:05:08,282 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3418, 1.6059, 1.4686, 1.8409, 1.6708, 2.0757, 1.4274, 3.6284], device='cuda:4'), covar=tensor([0.0634, 0.0836, 0.0820, 0.1185, 0.0650, 0.0483, 0.0753, 0.0154], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0044, 0.0040, 0.0039, 0.0039, 0.0058], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 03:05:10,512 INFO [finetune.py:976] (4/7) Epoch 12, batch 3100, loss[loss=0.165, simple_loss=0.2427, pruned_loss=0.04371, over 4820.00 frames. ], tot_loss[loss=0.1894, simple_loss=0.2573, pruned_loss=0.06071, over 949869.01 frames. ], batch size: 41, lr: 3.64e-03, grad_scale: 16.0 2023-04-27 03:05:19,899 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9401, 1.7544, 2.0090, 2.2955, 2.2839, 1.8293, 1.5801, 2.0517], device='cuda:4'), covar=tensor([0.0881, 0.1072, 0.0574, 0.0498, 0.0574, 0.0874, 0.0846, 0.0604], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0205, 0.0186, 0.0177, 0.0181, 0.0192, 0.0160, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 03:05:39,729 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6082, 1.2895, 1.3169, 1.4659, 1.8502, 1.5671, 1.3363, 1.2182], device='cuda:4'), covar=tensor([0.1317, 0.1052, 0.1383, 0.1101, 0.0562, 0.1024, 0.1528, 0.1591], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0321, 0.0353, 0.0297, 0.0334, 0.0318, 0.0307, 0.0360], device='cuda:4'), out_proj_covar=tensor([6.4557e-05, 6.7486e-05, 7.5825e-05, 6.0900e-05, 6.9756e-05, 6.7663e-05, 6.5350e-05, 7.7146e-05], device='cuda:4') 2023-04-27 03:05:40,809 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.130e+02 1.687e+02 1.900e+02 2.210e+02 4.094e+02, threshold=3.799e+02, percent-clipped=0.0 2023-04-27 03:05:41,561 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1571, 2.9407, 2.3657, 2.5864, 2.0617, 2.3422, 2.5421, 1.8283], device='cuda:4'), covar=tensor([0.2322, 0.1027, 0.0813, 0.1268, 0.3137, 0.1153, 0.1855, 0.2550], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0309, 0.0222, 0.0281, 0.0310, 0.0261, 0.0250, 0.0271], device='cuda:4'), out_proj_covar=tensor([1.1668e-04, 1.2366e-04, 8.8855e-05, 1.1227e-04, 1.2669e-04, 1.0473e-04, 1.0155e-04, 1.0839e-04], device='cuda:4') 2023-04-27 03:05:48,029 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=66144.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:05:54,638 INFO [finetune.py:976] (4/7) Epoch 12, batch 3150, loss[loss=0.2033, simple_loss=0.2466, pruned_loss=0.07994, over 4021.00 frames. ], tot_loss[loss=0.187, simple_loss=0.2547, pruned_loss=0.05969, over 952180.41 frames. ], batch size: 17, lr: 3.64e-03, grad_scale: 16.0 2023-04-27 03:05:58,705 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=66160.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 03:06:13,918 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5073, 1.9527, 2.3646, 3.0846, 2.3439, 1.7948, 1.7660, 2.2989], device='cuda:4'), covar=tensor([0.3978, 0.3901, 0.1868, 0.2854, 0.3348, 0.3205, 0.4429, 0.2584], device='cuda:4'), in_proj_covar=tensor([0.0285, 0.0248, 0.0221, 0.0315, 0.0214, 0.0228, 0.0229, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 03:06:27,522 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=66188.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 03:06:36,576 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=66194.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 03:06:38,745 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.90 vs. limit=2.0 2023-04-27 03:06:49,147 INFO [finetune.py:976] (4/7) Epoch 12, batch 3200, loss[loss=0.193, simple_loss=0.2504, pruned_loss=0.06784, over 4783.00 frames. ], tot_loss[loss=0.1841, simple_loss=0.2513, pruned_loss=0.05851, over 954544.36 frames. ], batch size: 26, lr: 3.64e-03, grad_scale: 16.0 2023-04-27 03:06:49,262 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=66205.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 03:07:12,338 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=66221.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 03:07:34,236 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=66236.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 03:07:35,266 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-27 03:07:43,555 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.908e+01 1.563e+02 1.889e+02 2.258e+02 3.990e+02, threshold=3.778e+02, percent-clipped=1.0 2023-04-27 03:07:57,818 INFO [finetune.py:976] (4/7) Epoch 12, batch 3250, loss[loss=0.1964, simple_loss=0.2671, pruned_loss=0.06286, over 4838.00 frames. ], tot_loss[loss=0.1835, simple_loss=0.2508, pruned_loss=0.05808, over 955140.35 frames. ], batch size: 49, lr: 3.64e-03, grad_scale: 16.0 2023-04-27 03:07:59,781 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=66258.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:08:47,970 INFO [finetune.py:976] (4/7) Epoch 12, batch 3300, loss[loss=0.2048, simple_loss=0.2698, pruned_loss=0.06985, over 4930.00 frames. ], tot_loss[loss=0.1868, simple_loss=0.2547, pruned_loss=0.05938, over 955600.03 frames. ], batch size: 33, lr: 3.64e-03, grad_scale: 16.0 2023-04-27 03:08:52,168 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.48 vs. limit=2.0 2023-04-27 03:08:57,552 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=66319.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:09:12,890 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6119, 1.4647, 1.7721, 1.8929, 1.4217, 1.1609, 1.3845, 0.9340], device='cuda:4'), covar=tensor([0.0618, 0.0658, 0.0497, 0.0568, 0.0792, 0.1662, 0.0821, 0.1008], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0072, 0.0071, 0.0067, 0.0076, 0.0098, 0.0077, 0.0072], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 03:09:13,359 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.300e+02 1.708e+02 2.133e+02 2.444e+02 6.112e+02, threshold=4.266e+02, percent-clipped=2.0 2023-04-27 03:09:21,690 INFO [finetune.py:976] (4/7) Epoch 12, batch 3350, loss[loss=0.1723, simple_loss=0.2446, pruned_loss=0.04999, over 4889.00 frames. ], tot_loss[loss=0.1882, simple_loss=0.2566, pruned_loss=0.05992, over 956309.95 frames. ], batch size: 32, lr: 3.64e-03, grad_scale: 16.0 2023-04-27 03:09:30,790 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=5.08 vs. limit=5.0 2023-04-27 03:09:54,778 INFO [finetune.py:976] (4/7) Epoch 12, batch 3400, loss[loss=0.1784, simple_loss=0.2507, pruned_loss=0.05301, over 4824.00 frames. ], tot_loss[loss=0.1898, simple_loss=0.2579, pruned_loss=0.06083, over 956316.51 frames. ], batch size: 30, lr: 3.64e-03, grad_scale: 16.0 2023-04-27 03:10:02,229 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3416, 1.9603, 2.3001, 2.6410, 2.6653, 2.1094, 1.8051, 2.4432], device='cuda:4'), covar=tensor([0.0860, 0.1052, 0.0635, 0.0546, 0.0554, 0.0982, 0.0799, 0.0565], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0201, 0.0184, 0.0174, 0.0179, 0.0188, 0.0157, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 03:10:20,215 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.071e+02 1.592e+02 1.894e+02 2.195e+02 3.276e+02, threshold=3.787e+02, percent-clipped=0.0 2023-04-27 03:10:28,165 INFO [finetune.py:976] (4/7) Epoch 12, batch 3450, loss[loss=0.1719, simple_loss=0.2376, pruned_loss=0.05309, over 4918.00 frames. ], tot_loss[loss=0.1883, simple_loss=0.2568, pruned_loss=0.05987, over 956824.57 frames. ], batch size: 33, lr: 3.64e-03, grad_scale: 16.0 2023-04-27 03:10:31,804 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3099, 1.5359, 1.3710, 1.4740, 1.2916, 1.2466, 1.3913, 1.0937], device='cuda:4'), covar=tensor([0.1602, 0.1333, 0.0899, 0.1154, 0.3289, 0.1083, 0.1466, 0.1972], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0311, 0.0222, 0.0282, 0.0311, 0.0262, 0.0251, 0.0271], device='cuda:4'), out_proj_covar=tensor([1.1692e-04, 1.2439e-04, 8.8816e-05, 1.1267e-04, 1.2697e-04, 1.0493e-04, 1.0203e-04, 1.0838e-04], device='cuda:4') 2023-04-27 03:10:55,270 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=66494.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 03:10:58,894 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=66500.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 03:11:01,854 INFO [finetune.py:976] (4/7) Epoch 12, batch 3500, loss[loss=0.1766, simple_loss=0.2508, pruned_loss=0.05116, over 4868.00 frames. ], tot_loss[loss=0.1865, simple_loss=0.2545, pruned_loss=0.05928, over 957270.52 frames. ], batch size: 31, lr: 3.64e-03, grad_scale: 16.0 2023-04-27 03:11:09,122 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=66516.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 03:11:12,214 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=66521.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 03:11:26,632 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.637e+01 1.554e+02 1.930e+02 2.326e+02 4.798e+02, threshold=3.860e+02, percent-clipped=2.0 2023-04-27 03:11:27,203 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=66542.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 03:11:40,352 INFO [finetune.py:976] (4/7) Epoch 12, batch 3550, loss[loss=0.2543, simple_loss=0.3019, pruned_loss=0.1033, over 4868.00 frames. ], tot_loss[loss=0.1839, simple_loss=0.2512, pruned_loss=0.05832, over 957056.31 frames. ], batch size: 34, lr: 3.64e-03, grad_scale: 16.0 2023-04-27 03:12:02,593 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-27 03:12:12,801 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=66582.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 03:12:45,654 INFO [finetune.py:976] (4/7) Epoch 12, batch 3600, loss[loss=0.1452, simple_loss=0.2229, pruned_loss=0.03374, over 4827.00 frames. ], tot_loss[loss=0.1819, simple_loss=0.2488, pruned_loss=0.05751, over 956088.87 frames. ], batch size: 33, lr: 3.64e-03, grad_scale: 16.0 2023-04-27 03:12:57,587 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=66614.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:13:31,408 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.076e+02 1.620e+02 1.935e+02 2.301e+02 4.046e+02, threshold=3.870e+02, percent-clipped=1.0 2023-04-27 03:13:45,504 INFO [finetune.py:976] (4/7) Epoch 12, batch 3650, loss[loss=0.1858, simple_loss=0.2512, pruned_loss=0.0602, over 4806.00 frames. ], tot_loss[loss=0.1843, simple_loss=0.2515, pruned_loss=0.05852, over 955570.26 frames. ], batch size: 25, lr: 3.64e-03, grad_scale: 16.0 2023-04-27 03:13:49,912 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0912, 1.8735, 2.1382, 2.4675, 2.4582, 1.9019, 1.7369, 2.2854], device='cuda:4'), covar=tensor([0.0915, 0.1091, 0.0665, 0.0593, 0.0589, 0.0941, 0.0802, 0.0520], device='cuda:4'), in_proj_covar=tensor([0.0193, 0.0201, 0.0183, 0.0173, 0.0178, 0.0186, 0.0156, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 03:14:19,349 INFO [finetune.py:976] (4/7) Epoch 12, batch 3700, loss[loss=0.2003, simple_loss=0.2769, pruned_loss=0.06182, over 4852.00 frames. ], tot_loss[loss=0.1851, simple_loss=0.2537, pruned_loss=0.05828, over 955748.64 frames. ], batch size: 44, lr: 3.64e-03, grad_scale: 16.0 2023-04-27 03:14:35,812 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=66731.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 03:14:43,338 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.132e+02 1.645e+02 1.962e+02 2.423e+02 5.947e+02, threshold=3.923e+02, percent-clipped=4.0 2023-04-27 03:14:52,769 INFO [finetune.py:976] (4/7) Epoch 12, batch 3750, loss[loss=0.1862, simple_loss=0.2596, pruned_loss=0.05645, over 4922.00 frames. ], tot_loss[loss=0.1854, simple_loss=0.2544, pruned_loss=0.05815, over 956242.09 frames. ], batch size: 38, lr: 3.64e-03, grad_scale: 16.0 2023-04-27 03:15:16,260 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=66792.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 03:15:18,731 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8141, 1.6802, 2.0389, 2.0588, 1.8113, 1.7247, 1.8254, 1.9271], device='cuda:4'), covar=tensor([0.7596, 0.9595, 1.1379, 1.1331, 0.8625, 1.3600, 1.3610, 1.3153], device='cuda:4'), in_proj_covar=tensor([0.0410, 0.0405, 0.0497, 0.0515, 0.0439, 0.0461, 0.0466, 0.0468], device='cuda:4'), out_proj_covar=tensor([9.9534e-05, 1.0048e-04, 1.1192e-04, 1.2208e-04, 1.0630e-04, 1.1119e-04, 1.1174e-04, 1.1211e-04], device='cuda:4') 2023-04-27 03:15:22,031 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=66800.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 03:15:25,458 INFO [finetune.py:976] (4/7) Epoch 12, batch 3800, loss[loss=0.2708, simple_loss=0.3101, pruned_loss=0.1158, over 4101.00 frames. ], tot_loss[loss=0.1889, simple_loss=0.2575, pruned_loss=0.06016, over 955335.89 frames. ], batch size: 65, lr: 3.64e-03, grad_scale: 16.0 2023-04-27 03:15:31,020 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5079, 1.2526, 4.3494, 4.0110, 3.8297, 4.1509, 4.0941, 3.8324], device='cuda:4'), covar=tensor([0.7030, 0.5991, 0.0924, 0.1728, 0.1153, 0.1546, 0.1040, 0.1270], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0307, 0.0402, 0.0409, 0.0349, 0.0407, 0.0314, 0.0368], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 03:15:32,296 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=66816.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 03:15:38,946 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1166, 2.5375, 2.2222, 2.3485, 1.7422, 2.1634, 2.1355, 1.7866], device='cuda:4'), covar=tensor([0.2250, 0.1330, 0.0798, 0.1396, 0.3619, 0.1189, 0.2180, 0.2500], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0313, 0.0224, 0.0283, 0.0313, 0.0263, 0.0253, 0.0271], device='cuda:4'), out_proj_covar=tensor([1.1785e-04, 1.2503e-04, 8.9459e-05, 1.1324e-04, 1.2772e-04, 1.0539e-04, 1.0288e-04, 1.0832e-04], device='cuda:4') 2023-04-27 03:15:48,323 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.153e+02 1.670e+02 1.892e+02 2.401e+02 5.287e+02, threshold=3.784e+02, percent-clipped=2.0 2023-04-27 03:15:52,969 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=66848.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:15:58,167 INFO [finetune.py:976] (4/7) Epoch 12, batch 3850, loss[loss=0.1996, simple_loss=0.2553, pruned_loss=0.07194, over 4769.00 frames. ], tot_loss[loss=0.1867, simple_loss=0.2553, pruned_loss=0.05899, over 955799.15 frames. ], batch size: 28, lr: 3.64e-03, grad_scale: 16.0 2023-04-27 03:16:04,155 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.90 vs. limit=2.0 2023-04-27 03:16:04,616 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=66864.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 03:16:12,477 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=66877.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 03:16:31,352 INFO [finetune.py:976] (4/7) Epoch 12, batch 3900, loss[loss=0.1771, simple_loss=0.2522, pruned_loss=0.05105, over 4910.00 frames. ], tot_loss[loss=0.1848, simple_loss=0.2526, pruned_loss=0.05843, over 955131.56 frames. ], batch size: 37, lr: 3.64e-03, grad_scale: 16.0 2023-04-27 03:16:38,403 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=66914.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:17:12,162 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.033e+02 1.631e+02 1.862e+02 2.227e+02 3.932e+02, threshold=3.724e+02, percent-clipped=1.0 2023-04-27 03:17:32,061 INFO [finetune.py:976] (4/7) Epoch 12, batch 3950, loss[loss=0.1935, simple_loss=0.2388, pruned_loss=0.07407, over 4063.00 frames. ], tot_loss[loss=0.1818, simple_loss=0.2492, pruned_loss=0.05726, over 954899.82 frames. ], batch size: 17, lr: 3.64e-03, grad_scale: 16.0 2023-04-27 03:17:33,393 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=66957.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:17:42,916 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=66962.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:17:55,584 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.96 vs. limit=5.0 2023-04-27 03:18:28,343 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8487, 2.5719, 1.8060, 1.7299, 1.3070, 1.2930, 2.1055, 1.2789], device='cuda:4'), covar=tensor([0.1828, 0.1481, 0.1651, 0.2063, 0.2713, 0.2196, 0.1029, 0.2262], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0217, 0.0172, 0.0206, 0.0206, 0.0186, 0.0159, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 03:18:39,567 INFO [finetune.py:976] (4/7) Epoch 12, batch 4000, loss[loss=0.191, simple_loss=0.2623, pruned_loss=0.05989, over 4822.00 frames. ], tot_loss[loss=0.181, simple_loss=0.2482, pruned_loss=0.05686, over 955357.80 frames. ], batch size: 38, lr: 3.63e-03, grad_scale: 16.0 2023-04-27 03:18:58,935 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=67018.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 03:19:08,160 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7406, 1.7640, 1.7238, 2.1445, 1.9891, 2.1046, 1.6672, 4.5334], device='cuda:4'), covar=tensor([0.0529, 0.0805, 0.0783, 0.1195, 0.0635, 0.0527, 0.0738, 0.0112], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0043, 0.0040, 0.0038, 0.0038, 0.0057], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0012, 0.0013, 0.0016], device='cuda:4') 2023-04-27 03:19:14,011 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.105e+02 1.700e+02 2.045e+02 2.371e+02 4.368e+02, threshold=4.090e+02, percent-clipped=3.0 2023-04-27 03:19:22,919 INFO [finetune.py:976] (4/7) Epoch 12, batch 4050, loss[loss=0.2399, simple_loss=0.3112, pruned_loss=0.08429, over 4906.00 frames. ], tot_loss[loss=0.1836, simple_loss=0.2515, pruned_loss=0.05783, over 955382.68 frames. ], batch size: 37, lr: 3.63e-03, grad_scale: 16.0 2023-04-27 03:19:32,325 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7024, 1.3492, 1.6028, 1.9823, 2.0851, 1.6328, 1.4024, 1.7839], device='cuda:4'), covar=tensor([0.0785, 0.1190, 0.0761, 0.0503, 0.0507, 0.0783, 0.0704, 0.0512], device='cuda:4'), in_proj_covar=tensor([0.0191, 0.0200, 0.0182, 0.0173, 0.0177, 0.0185, 0.0155, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 03:19:39,561 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0240, 1.4734, 1.8804, 2.1906, 1.8434, 1.4849, 0.9933, 1.5980], device='cuda:4'), covar=tensor([0.3441, 0.3501, 0.1726, 0.2211, 0.2867, 0.2836, 0.4654, 0.2298], device='cuda:4'), in_proj_covar=tensor([0.0287, 0.0248, 0.0223, 0.0317, 0.0214, 0.0229, 0.0231, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 03:19:42,048 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=67083.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:19:44,510 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=67087.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 03:19:56,227 INFO [finetune.py:976] (4/7) Epoch 12, batch 4100, loss[loss=0.23, simple_loss=0.288, pruned_loss=0.086, over 4816.00 frames. ], tot_loss[loss=0.186, simple_loss=0.2542, pruned_loss=0.05886, over 955350.41 frames. ], batch size: 30, lr: 3.63e-03, grad_scale: 16.0 2023-04-27 03:20:02,227 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4243, 3.2157, 0.8812, 1.9605, 1.7294, 2.4030, 1.8347, 0.9617], device='cuda:4'), covar=tensor([0.1329, 0.0917, 0.1839, 0.1134, 0.1124, 0.0906, 0.1522, 0.2201], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0249, 0.0140, 0.0122, 0.0133, 0.0153, 0.0118, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 03:20:21,136 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.154e+02 1.668e+02 1.950e+02 2.451e+02 4.466e+02, threshold=3.899e+02, percent-clipped=3.0 2023-04-27 03:20:22,490 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=67144.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:20:29,535 INFO [finetune.py:976] (4/7) Epoch 12, batch 4150, loss[loss=0.2123, simple_loss=0.2826, pruned_loss=0.07101, over 4930.00 frames. ], tot_loss[loss=0.1897, simple_loss=0.258, pruned_loss=0.06074, over 956896.79 frames. ], batch size: 38, lr: 3.63e-03, grad_scale: 16.0 2023-04-27 03:20:30,820 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.52 vs. limit=2.0 2023-04-27 03:20:45,890 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=67177.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 03:21:03,549 INFO [finetune.py:976] (4/7) Epoch 12, batch 4200, loss[loss=0.1806, simple_loss=0.2406, pruned_loss=0.06027, over 4890.00 frames. ], tot_loss[loss=0.1886, simple_loss=0.2572, pruned_loss=0.06003, over 956395.61 frames. ], batch size: 35, lr: 3.63e-03, grad_scale: 16.0 2023-04-27 03:21:18,107 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=67225.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 03:21:28,836 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.013e+02 1.623e+02 1.873e+02 2.449e+02 3.643e+02, threshold=3.747e+02, percent-clipped=0.0 2023-04-27 03:21:37,114 INFO [finetune.py:976] (4/7) Epoch 12, batch 4250, loss[loss=0.1513, simple_loss=0.224, pruned_loss=0.03928, over 4820.00 frames. ], tot_loss[loss=0.1863, simple_loss=0.2544, pruned_loss=0.05917, over 954368.38 frames. ], batch size: 40, lr: 3.63e-03, grad_scale: 16.0 2023-04-27 03:21:56,129 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.59 vs. limit=2.0 2023-04-27 03:22:10,185 INFO [finetune.py:976] (4/7) Epoch 12, batch 4300, loss[loss=0.1512, simple_loss=0.2223, pruned_loss=0.04003, over 4820.00 frames. ], tot_loss[loss=0.184, simple_loss=0.2513, pruned_loss=0.05832, over 955215.46 frames. ], batch size: 38, lr: 3.63e-03, grad_scale: 16.0 2023-04-27 03:22:15,654 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=67313.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 03:22:23,375 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1273, 1.8881, 2.2450, 2.6047, 2.2111, 2.0301, 2.0975, 2.1136], device='cuda:4'), covar=tensor([0.5422, 0.7719, 0.8448, 0.6866, 0.6868, 0.9493, 0.9505, 1.0234], device='cuda:4'), in_proj_covar=tensor([0.0411, 0.0406, 0.0497, 0.0515, 0.0441, 0.0461, 0.0467, 0.0470], device='cuda:4'), out_proj_covar=tensor([9.9718e-05, 1.0075e-04, 1.1188e-04, 1.2219e-04, 1.0675e-04, 1.1119e-04, 1.1195e-04, 1.1253e-04], device='cuda:4') 2023-04-27 03:22:26,378 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8188, 1.4522, 4.9389, 4.6352, 4.3373, 4.7166, 4.4177, 4.3598], device='cuda:4'), covar=tensor([0.7192, 0.5712, 0.0896, 0.1680, 0.1048, 0.1312, 0.1639, 0.1493], device='cuda:4'), in_proj_covar=tensor([0.0314, 0.0308, 0.0406, 0.0413, 0.0354, 0.0413, 0.0317, 0.0372], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 03:22:31,557 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4686, 1.7141, 1.7540, 1.9305, 1.8078, 1.9610, 1.9050, 1.8162], device='cuda:4'), covar=tensor([0.4453, 0.6804, 0.6174, 0.5811, 0.6655, 0.8616, 0.6563, 0.6204], device='cuda:4'), in_proj_covar=tensor([0.0329, 0.0376, 0.0315, 0.0328, 0.0339, 0.0399, 0.0359, 0.0325], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 03:22:36,119 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.106e+02 1.576e+02 1.859e+02 2.202e+02 4.297e+02, threshold=3.717e+02, percent-clipped=1.0 2023-04-27 03:22:43,955 INFO [finetune.py:976] (4/7) Epoch 12, batch 4350, loss[loss=0.1922, simple_loss=0.2545, pruned_loss=0.06492, over 4869.00 frames. ], tot_loss[loss=0.1816, simple_loss=0.2483, pruned_loss=0.05743, over 955515.67 frames. ], batch size: 34, lr: 3.63e-03, grad_scale: 16.0 2023-04-27 03:23:29,034 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=67387.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 03:23:51,833 INFO [finetune.py:976] (4/7) Epoch 12, batch 4400, loss[loss=0.2039, simple_loss=0.2756, pruned_loss=0.0661, over 4812.00 frames. ], tot_loss[loss=0.184, simple_loss=0.2503, pruned_loss=0.05881, over 955361.42 frames. ], batch size: 45, lr: 3.63e-03, grad_scale: 16.0 2023-04-27 03:24:12,054 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0188, 0.9886, 1.2211, 1.1396, 0.9832, 0.8943, 1.0484, 0.6296], device='cuda:4'), covar=tensor([0.0740, 0.0687, 0.0545, 0.0680, 0.0832, 0.1363, 0.0520, 0.0788], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0071, 0.0070, 0.0067, 0.0075, 0.0096, 0.0075, 0.0070], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 03:24:21,929 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5004, 1.6101, 0.9144, 1.2546, 1.8754, 1.3854, 1.3240, 1.3572], device='cuda:4'), covar=tensor([0.0515, 0.0363, 0.0367, 0.0567, 0.0280, 0.0529, 0.0497, 0.0554], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0025, 0.0023, 0.0029, 0.0020, 0.0029, 0.0029, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0038, 0.0050, 0.0037, 0.0049, 0.0048, 0.0050], device='cuda:4') 2023-04-27 03:24:26,228 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=67435.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 03:24:34,555 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=67439.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:24:36,311 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.200e+02 1.735e+02 1.991e+02 2.507e+02 6.130e+02, threshold=3.981e+02, percent-clipped=5.0 2023-04-27 03:24:56,595 INFO [finetune.py:976] (4/7) Epoch 12, batch 4450, loss[loss=0.1547, simple_loss=0.2271, pruned_loss=0.0411, over 4735.00 frames. ], tot_loss[loss=0.1858, simple_loss=0.2532, pruned_loss=0.05922, over 954159.65 frames. ], batch size: 23, lr: 3.63e-03, grad_scale: 16.0 2023-04-27 03:25:48,054 INFO [finetune.py:976] (4/7) Epoch 12, batch 4500, loss[loss=0.213, simple_loss=0.282, pruned_loss=0.07198, over 4226.00 frames. ], tot_loss[loss=0.1882, simple_loss=0.2556, pruned_loss=0.06041, over 955425.31 frames. ], batch size: 65, lr: 3.63e-03, grad_scale: 16.0 2023-04-27 03:25:53,061 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=67513.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:25:55,736 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.82 vs. limit=2.0 2023-04-27 03:26:12,838 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.294e+02 1.699e+02 1.983e+02 2.527e+02 4.329e+02, threshold=3.965e+02, percent-clipped=1.0 2023-04-27 03:26:22,245 INFO [finetune.py:976] (4/7) Epoch 12, batch 4550, loss[loss=0.1629, simple_loss=0.2396, pruned_loss=0.04311, over 4910.00 frames. ], tot_loss[loss=0.1895, simple_loss=0.2574, pruned_loss=0.06079, over 953827.47 frames. ], batch size: 38, lr: 3.63e-03, grad_scale: 32.0 2023-04-27 03:26:34,574 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=67574.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 03:26:53,833 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.76 vs. limit=2.0 2023-04-27 03:26:56,263 INFO [finetune.py:976] (4/7) Epoch 12, batch 4600, loss[loss=0.1796, simple_loss=0.2335, pruned_loss=0.06287, over 4317.00 frames. ], tot_loss[loss=0.1902, simple_loss=0.258, pruned_loss=0.0612, over 953226.74 frames. ], batch size: 18, lr: 3.63e-03, grad_scale: 32.0 2023-04-27 03:26:57,636 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4379, 2.2989, 2.8200, 3.0301, 2.9124, 2.2820, 2.0235, 2.5087], device='cuda:4'), covar=tensor([0.1021, 0.1050, 0.0534, 0.0606, 0.0593, 0.1032, 0.0917, 0.0673], device='cuda:4'), in_proj_covar=tensor([0.0191, 0.0200, 0.0180, 0.0172, 0.0176, 0.0184, 0.0154, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 03:27:01,283 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=67613.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:27:08,101 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-04-27 03:27:15,736 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6972, 1.1864, 1.3471, 1.4392, 1.8619, 1.5237, 1.2766, 1.3101], device='cuda:4'), covar=tensor([0.1423, 0.1378, 0.1812, 0.1314, 0.0801, 0.1358, 0.1861, 0.1834], device='cuda:4'), in_proj_covar=tensor([0.0306, 0.0317, 0.0352, 0.0295, 0.0333, 0.0315, 0.0304, 0.0361], device='cuda:4'), out_proj_covar=tensor([6.3860e-05, 6.6775e-05, 7.5633e-05, 6.0579e-05, 6.9564e-05, 6.6882e-05, 6.4827e-05, 7.7303e-05], device='cuda:4') 2023-04-27 03:27:20,971 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.110e+02 1.683e+02 1.939e+02 2.322e+02 5.503e+02, threshold=3.878e+02, percent-clipped=1.0 2023-04-27 03:27:29,855 INFO [finetune.py:976] (4/7) Epoch 12, batch 4650, loss[loss=0.1408, simple_loss=0.2054, pruned_loss=0.03812, over 4907.00 frames. ], tot_loss[loss=0.1885, simple_loss=0.2554, pruned_loss=0.06077, over 954184.79 frames. ], batch size: 46, lr: 3.63e-03, grad_scale: 16.0 2023-04-27 03:27:34,156 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=67661.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:27:39,222 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-27 03:27:46,426 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=67680.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:27:46,668 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-04-27 03:27:48,571 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-27 03:27:54,881 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=67692.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:28:04,219 INFO [finetune.py:976] (4/7) Epoch 12, batch 4700, loss[loss=0.1672, simple_loss=0.232, pruned_loss=0.0512, over 4906.00 frames. ], tot_loss[loss=0.186, simple_loss=0.2526, pruned_loss=0.05971, over 955423.34 frames. ], batch size: 32, lr: 3.63e-03, grad_scale: 16.0 2023-04-27 03:28:37,309 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=67739.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:28:38,564 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=67741.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:28:40,070 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.078e+02 1.659e+02 1.914e+02 2.232e+02 4.498e+02, threshold=3.829e+02, percent-clipped=2.0 2023-04-27 03:28:58,097 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=67753.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 03:28:59,228 INFO [finetune.py:976] (4/7) Epoch 12, batch 4750, loss[loss=0.1736, simple_loss=0.2426, pruned_loss=0.05225, over 4749.00 frames. ], tot_loss[loss=0.1841, simple_loss=0.2504, pruned_loss=0.05887, over 954708.55 frames. ], batch size: 54, lr: 3.63e-03, grad_scale: 16.0 2023-04-27 03:29:42,911 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=67787.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:30:05,762 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9815, 2.5676, 1.9198, 1.7402, 1.4257, 1.3853, 2.0322, 1.3894], device='cuda:4'), covar=tensor([0.1580, 0.1143, 0.1491, 0.1763, 0.2219, 0.1959, 0.0989, 0.1961], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0212, 0.0169, 0.0203, 0.0201, 0.0183, 0.0157, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 03:30:06,236 INFO [finetune.py:976] (4/7) Epoch 12, batch 4800, loss[loss=0.1854, simple_loss=0.2502, pruned_loss=0.06035, over 4792.00 frames. ], tot_loss[loss=0.187, simple_loss=0.254, pruned_loss=0.06002, over 954806.41 frames. ], batch size: 29, lr: 3.63e-03, grad_scale: 16.0 2023-04-27 03:30:35,557 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.067e+02 1.720e+02 2.004e+02 2.760e+02 5.659e+02, threshold=4.008e+02, percent-clipped=5.0 2023-04-27 03:30:36,290 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1629, 1.6162, 1.4946, 2.0138, 1.7798, 1.9366, 1.4525, 4.3069], device='cuda:4'), covar=tensor([0.0616, 0.0830, 0.0855, 0.1240, 0.0681, 0.0580, 0.0827, 0.0097], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0044, 0.0040, 0.0038, 0.0038, 0.0057], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0012, 0.0013, 0.0016], device='cuda:4') 2023-04-27 03:30:43,899 INFO [finetune.py:976] (4/7) Epoch 12, batch 4850, loss[loss=0.1595, simple_loss=0.2173, pruned_loss=0.05083, over 4219.00 frames. ], tot_loss[loss=0.1888, simple_loss=0.2567, pruned_loss=0.06045, over 955354.30 frames. ], batch size: 17, lr: 3.63e-03, grad_scale: 16.0 2023-04-27 03:30:48,298 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-04-27 03:30:54,466 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=67869.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 03:31:17,823 INFO [finetune.py:976] (4/7) Epoch 12, batch 4900, loss[loss=0.174, simple_loss=0.2495, pruned_loss=0.04919, over 4814.00 frames. ], tot_loss[loss=0.1894, simple_loss=0.2577, pruned_loss=0.06048, over 953996.63 frames. ], batch size: 38, lr: 3.63e-03, grad_scale: 16.0 2023-04-27 03:31:43,018 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.676e+01 1.571e+02 1.848e+02 2.234e+02 5.872e+02, threshold=3.696e+02, percent-clipped=2.0 2023-04-27 03:31:51,277 INFO [finetune.py:976] (4/7) Epoch 12, batch 4950, loss[loss=0.1671, simple_loss=0.2331, pruned_loss=0.05057, over 4701.00 frames. ], tot_loss[loss=0.1891, simple_loss=0.2579, pruned_loss=0.06012, over 954027.56 frames. ], batch size: 23, lr: 3.63e-03, grad_scale: 16.0 2023-04-27 03:31:59,169 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-27 03:32:09,665 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=67981.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:32:26,419 INFO [finetune.py:976] (4/7) Epoch 12, batch 5000, loss[loss=0.1675, simple_loss=0.242, pruned_loss=0.04647, over 4822.00 frames. ], tot_loss[loss=0.1876, simple_loss=0.2563, pruned_loss=0.05944, over 955892.40 frames. ], batch size: 38, lr: 3.63e-03, grad_scale: 16.0 2023-04-27 03:32:47,900 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=68036.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:32:51,628 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=68042.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:32:52,122 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.570e+01 1.647e+02 1.940e+02 2.431e+02 4.879e+02, threshold=3.879e+02, percent-clipped=2.0 2023-04-27 03:32:55,686 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=68048.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 03:32:59,401 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.43 vs. limit=2.0 2023-04-27 03:32:59,806 INFO [finetune.py:976] (4/7) Epoch 12, batch 5050, loss[loss=0.1854, simple_loss=0.2584, pruned_loss=0.05623, over 4901.00 frames. ], tot_loss[loss=0.1863, simple_loss=0.254, pruned_loss=0.05924, over 955436.95 frames. ], batch size: 32, lr: 3.63e-03, grad_scale: 16.0 2023-04-27 03:33:03,788 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.63 vs. limit=5.0 2023-04-27 03:33:22,289 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=68087.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 03:33:33,614 INFO [finetune.py:976] (4/7) Epoch 12, batch 5100, loss[loss=0.1748, simple_loss=0.2432, pruned_loss=0.05315, over 4907.00 frames. ], tot_loss[loss=0.1826, simple_loss=0.25, pruned_loss=0.05757, over 956345.51 frames. ], batch size: 35, lr: 3.63e-03, grad_scale: 16.0 2023-04-27 03:34:09,661 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.5919, 3.4804, 2.7266, 4.1841, 3.5557, 3.6100, 1.5945, 3.5431], device='cuda:4'), covar=tensor([0.2000, 0.1241, 0.3706, 0.1725, 0.3410, 0.1913, 0.6009, 0.2580], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0216, 0.0249, 0.0304, 0.0299, 0.0248, 0.0272, 0.0268], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 03:34:11,406 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.851e+01 1.609e+02 1.945e+02 2.272e+02 5.546e+02, threshold=3.889e+02, percent-clipped=1.0 2023-04-27 03:34:14,598 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=68148.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 03:34:14,663 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-04-27 03:34:18,645 INFO [finetune.py:976] (4/7) Epoch 12, batch 5150, loss[loss=0.2154, simple_loss=0.2797, pruned_loss=0.07551, over 4908.00 frames. ], tot_loss[loss=0.1834, simple_loss=0.2504, pruned_loss=0.05822, over 956821.23 frames. ], batch size: 37, lr: 3.63e-03, grad_scale: 16.0 2023-04-27 03:34:34,078 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=68169.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:34:48,397 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2659, 2.5453, 0.8828, 1.4822, 1.5701, 1.9577, 1.6448, 0.7743], device='cuda:4'), covar=tensor([0.1350, 0.1410, 0.1719, 0.1369, 0.1094, 0.0979, 0.1503, 0.1662], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0248, 0.0139, 0.0121, 0.0133, 0.0153, 0.0117, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 03:35:01,406 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.54 vs. limit=2.0 2023-04-27 03:35:16,278 INFO [finetune.py:976] (4/7) Epoch 12, batch 5200, loss[loss=0.2107, simple_loss=0.2638, pruned_loss=0.07875, over 4783.00 frames. ], tot_loss[loss=0.1877, simple_loss=0.255, pruned_loss=0.06017, over 958052.59 frames. ], batch size: 28, lr: 3.62e-03, grad_scale: 16.0 2023-04-27 03:35:34,919 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=68217.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:36:09,273 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.144e+02 1.715e+02 2.014e+02 2.436e+02 3.767e+02, threshold=4.027e+02, percent-clipped=0.0 2023-04-27 03:36:22,260 INFO [finetune.py:976] (4/7) Epoch 12, batch 5250, loss[loss=0.2164, simple_loss=0.2835, pruned_loss=0.07459, over 4865.00 frames. ], tot_loss[loss=0.1885, simple_loss=0.2569, pruned_loss=0.06012, over 957606.87 frames. ], batch size: 34, lr: 3.62e-03, grad_scale: 16.0 2023-04-27 03:37:07,992 INFO [finetune.py:976] (4/7) Epoch 12, batch 5300, loss[loss=0.1862, simple_loss=0.2608, pruned_loss=0.05577, over 4810.00 frames. ], tot_loss[loss=0.1889, simple_loss=0.2571, pruned_loss=0.06031, over 956556.96 frames. ], batch size: 38, lr: 3.62e-03, grad_scale: 16.0 2023-04-27 03:37:21,287 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1060, 0.6725, 0.9330, 0.7658, 1.2347, 0.9418, 0.8583, 0.9850], device='cuda:4'), covar=tensor([0.1710, 0.1548, 0.2222, 0.1517, 0.0997, 0.1439, 0.1710, 0.2089], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0322, 0.0355, 0.0299, 0.0336, 0.0318, 0.0311, 0.0365], device='cuda:4'), out_proj_covar=tensor([6.4938e-05, 6.7722e-05, 7.6250e-05, 6.1280e-05, 7.0254e-05, 6.7568e-05, 6.6133e-05, 7.8152e-05], device='cuda:4') 2023-04-27 03:37:29,920 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=68336.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:37:30,465 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=68337.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:37:34,484 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.347e+01 1.659e+02 1.967e+02 2.376e+02 7.370e+02, threshold=3.933e+02, percent-clipped=4.0 2023-04-27 03:37:37,607 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=68348.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:37:41,848 INFO [finetune.py:976] (4/7) Epoch 12, batch 5350, loss[loss=0.197, simple_loss=0.2692, pruned_loss=0.06241, over 4735.00 frames. ], tot_loss[loss=0.1884, simple_loss=0.2572, pruned_loss=0.05984, over 955529.43 frames. ], batch size: 54, lr: 3.62e-03, grad_scale: 16.0 2023-04-27 03:37:49,231 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5660, 1.8613, 1.3515, 1.1259, 1.2092, 1.1543, 1.3124, 1.1060], device='cuda:4'), covar=tensor([0.1943, 0.1305, 0.1790, 0.1976, 0.2630, 0.2362, 0.1169, 0.2193], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0212, 0.0168, 0.0202, 0.0201, 0.0183, 0.0156, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 03:38:00,881 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=68384.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:38:10,025 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=68396.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:38:11,926 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=68399.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:38:15,655 INFO [finetune.py:976] (4/7) Epoch 12, batch 5400, loss[loss=0.1499, simple_loss=0.2235, pruned_loss=0.03817, over 4689.00 frames. ], tot_loss[loss=0.1865, simple_loss=0.2547, pruned_loss=0.05922, over 954801.99 frames. ], batch size: 23, lr: 3.62e-03, grad_scale: 16.0 2023-04-27 03:38:21,804 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5298, 1.2750, 4.0447, 3.5228, 3.5771, 3.7448, 3.7068, 3.3157], device='cuda:4'), covar=tensor([0.9688, 0.8733, 0.1884, 0.3418, 0.2401, 0.3246, 0.2603, 0.3687], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0306, 0.0401, 0.0408, 0.0349, 0.0408, 0.0313, 0.0371], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 03:38:28,268 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0506, 2.4918, 0.9530, 1.3697, 1.8905, 1.1642, 3.4440, 1.6615], device='cuda:4'), covar=tensor([0.0697, 0.0626, 0.0831, 0.1284, 0.0548, 0.1010, 0.0220, 0.0669], device='cuda:4'), in_proj_covar=tensor([0.0052, 0.0067, 0.0049, 0.0047, 0.0051, 0.0052, 0.0077, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 03:38:41,373 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.158e+02 1.609e+02 1.919e+02 2.411e+02 6.706e+02, threshold=3.839e+02, percent-clipped=2.0 2023-04-27 03:38:41,454 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=68443.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 03:38:49,106 INFO [finetune.py:976] (4/7) Epoch 12, batch 5450, loss[loss=0.185, simple_loss=0.2561, pruned_loss=0.05693, over 4829.00 frames. ], tot_loss[loss=0.1849, simple_loss=0.2522, pruned_loss=0.0588, over 953750.95 frames. ], batch size: 33, lr: 3.62e-03, grad_scale: 16.0 2023-04-27 03:38:52,252 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=68460.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:39:33,461 INFO [finetune.py:976] (4/7) Epoch 12, batch 5500, loss[loss=0.1679, simple_loss=0.238, pruned_loss=0.04889, over 4750.00 frames. ], tot_loss[loss=0.1825, simple_loss=0.2494, pruned_loss=0.05775, over 953572.59 frames. ], batch size: 54, lr: 3.62e-03, grad_scale: 16.0 2023-04-27 03:39:58,194 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.589e+02 1.966e+02 2.318e+02 5.103e+02, threshold=3.932e+02, percent-clipped=5.0 2023-04-27 03:40:06,516 INFO [finetune.py:976] (4/7) Epoch 12, batch 5550, loss[loss=0.1865, simple_loss=0.2498, pruned_loss=0.06159, over 4901.00 frames. ], tot_loss[loss=0.1835, simple_loss=0.2506, pruned_loss=0.05826, over 951614.29 frames. ], batch size: 37, lr: 3.62e-03, grad_scale: 16.0 2023-04-27 03:40:12,024 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-27 03:40:15,001 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6056, 1.7742, 1.0694, 1.3469, 2.0196, 1.4936, 1.4613, 1.4757], device='cuda:4'), covar=tensor([0.0497, 0.0338, 0.0306, 0.0544, 0.0250, 0.0499, 0.0468, 0.0548], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0025, 0.0023, 0.0029, 0.0020, 0.0029, 0.0028, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0049, 0.0044, 0.0037, 0.0050, 0.0037, 0.0048, 0.0048, 0.0050], device='cuda:4') 2023-04-27 03:40:21,099 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=68578.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:40:28,872 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6412, 1.1862, 1.2437, 1.3292, 1.8269, 1.4391, 1.2219, 1.2400], device='cuda:4'), covar=tensor([0.1434, 0.1461, 0.2024, 0.1515, 0.0738, 0.1714, 0.1796, 0.1960], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0323, 0.0356, 0.0299, 0.0337, 0.0320, 0.0311, 0.0365], device='cuda:4'), out_proj_covar=tensor([6.4850e-05, 6.7901e-05, 7.6403e-05, 6.1389e-05, 7.0375e-05, 6.7921e-05, 6.6270e-05, 7.8190e-05], device='cuda:4') 2023-04-27 03:40:29,525 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-27 03:40:54,927 INFO [finetune.py:976] (4/7) Epoch 12, batch 5600, loss[loss=0.1868, simple_loss=0.2674, pruned_loss=0.05315, over 4904.00 frames. ], tot_loss[loss=0.1883, simple_loss=0.256, pruned_loss=0.06027, over 951815.10 frames. ], batch size: 43, lr: 3.62e-03, grad_scale: 16.0 2023-04-27 03:41:03,975 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5588, 1.8156, 1.9178, 2.0840, 1.9465, 2.1480, 2.0487, 2.0096], device='cuda:4'), covar=tensor([0.4711, 0.5992, 0.5263, 0.5012, 0.6211, 0.7945, 0.6466, 0.5751], device='cuda:4'), in_proj_covar=tensor([0.0328, 0.0376, 0.0315, 0.0327, 0.0339, 0.0397, 0.0356, 0.0324], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 03:41:36,482 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=68637.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:41:37,659 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=68639.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:41:46,139 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.206e+02 1.587e+02 1.869e+02 2.457e+02 5.644e+02, threshold=3.738e+02, percent-clipped=6.0 2023-04-27 03:41:58,627 INFO [finetune.py:976] (4/7) Epoch 12, batch 5650, loss[loss=0.1898, simple_loss=0.2709, pruned_loss=0.05439, over 4910.00 frames. ], tot_loss[loss=0.1896, simple_loss=0.2577, pruned_loss=0.06076, over 952142.73 frames. ], batch size: 37, lr: 3.62e-03, grad_scale: 16.0 2023-04-27 03:42:01,620 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2741, 2.9863, 0.8255, 1.6882, 1.6205, 2.0687, 1.7105, 0.9301], device='cuda:4'), covar=tensor([0.1483, 0.1032, 0.1985, 0.1353, 0.1191, 0.1088, 0.1552, 0.1921], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0251, 0.0141, 0.0122, 0.0134, 0.0155, 0.0118, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 03:42:33,431 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=68685.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:42:33,472 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=68685.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:42:45,293 INFO [finetune.py:976] (4/7) Epoch 12, batch 5700, loss[loss=0.188, simple_loss=0.2311, pruned_loss=0.07241, over 3583.00 frames. ], tot_loss[loss=0.1882, simple_loss=0.2552, pruned_loss=0.06065, over 934488.51 frames. ], batch size: 15, lr: 3.62e-03, grad_scale: 16.0 2023-04-27 03:43:16,429 INFO [finetune.py:976] (4/7) Epoch 13, batch 0, loss[loss=0.1686, simple_loss=0.2372, pruned_loss=0.04995, over 4849.00 frames. ], tot_loss[loss=0.1686, simple_loss=0.2372, pruned_loss=0.04995, over 4849.00 frames. ], batch size: 31, lr: 3.62e-03, grad_scale: 16.0 2023-04-27 03:43:16,429 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 03:43:32,177 INFO [finetune.py:1010] (4/7) Epoch 13, validation: loss=0.1542, simple_loss=0.2264, pruned_loss=0.04102, over 2265189.00 frames. 2023-04-27 03:43:32,177 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-27 03:43:49,890 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.289e+01 1.578e+02 1.937e+02 2.291e+02 5.419e+02, threshold=3.875e+02, percent-clipped=2.0 2023-04-27 03:43:50,011 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=68743.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 03:43:51,894 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=68746.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:43:56,109 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0973, 0.6679, 0.9486, 0.7574, 1.2200, 1.0037, 0.8777, 1.0135], device='cuda:4'), covar=tensor([0.1597, 0.1509, 0.2304, 0.1740, 0.1038, 0.1558, 0.1714, 0.2169], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0318, 0.0352, 0.0294, 0.0332, 0.0316, 0.0307, 0.0361], device='cuda:4'), out_proj_covar=tensor([6.4078e-05, 6.6941e-05, 7.5476e-05, 6.0355e-05, 6.9367e-05, 6.7107e-05, 6.5417e-05, 7.7128e-05], device='cuda:4') 2023-04-27 03:43:57,289 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=68755.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:44:12,769 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0075, 1.6677, 2.0300, 2.3156, 2.4314, 1.9318, 1.6209, 1.9808], device='cuda:4'), covar=tensor([0.0907, 0.1206, 0.0659, 0.0596, 0.0551, 0.0847, 0.0866, 0.0612], device='cuda:4'), in_proj_covar=tensor([0.0192, 0.0201, 0.0182, 0.0173, 0.0178, 0.0184, 0.0155, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 03:44:15,625 INFO [finetune.py:976] (4/7) Epoch 13, batch 50, loss[loss=0.1659, simple_loss=0.2271, pruned_loss=0.05231, over 4696.00 frames. ], tot_loss[loss=0.1928, simple_loss=0.261, pruned_loss=0.06235, over 217045.64 frames. ], batch size: 23, lr: 3.62e-03, grad_scale: 16.0 2023-04-27 03:44:21,434 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=68791.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 03:44:22,085 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.4014, 1.3106, 1.4327, 0.9929, 1.3677, 1.1615, 1.7336, 1.3462], device='cuda:4'), covar=tensor([0.3758, 0.1846, 0.4885, 0.2731, 0.1684, 0.2289, 0.1759, 0.4456], device='cuda:4'), in_proj_covar=tensor([0.0339, 0.0342, 0.0426, 0.0354, 0.0378, 0.0377, 0.0371, 0.0416], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 03:44:26,399 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1390, 1.9195, 2.2977, 2.6920, 2.2161, 2.0157, 2.1465, 2.1326], device='cuda:4'), covar=tensor([0.5635, 0.7644, 0.8780, 0.6376, 0.6791, 0.9971, 1.0361, 0.9691], device='cuda:4'), in_proj_covar=tensor([0.0412, 0.0407, 0.0499, 0.0514, 0.0442, 0.0462, 0.0470, 0.0473], device='cuda:4'), out_proj_covar=tensor([9.9960e-05, 1.0106e-04, 1.1222e-04, 1.2180e-04, 1.0698e-04, 1.1149e-04, 1.1256e-04, 1.1316e-04], device='cuda:4') 2023-04-27 03:44:46,964 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1927, 1.5879, 2.0346, 2.3598, 1.9538, 1.5774, 1.2157, 1.7655], device='cuda:4'), covar=tensor([0.2994, 0.3279, 0.1585, 0.2195, 0.2754, 0.2560, 0.4313, 0.2050], device='cuda:4'), in_proj_covar=tensor([0.0286, 0.0248, 0.0223, 0.0317, 0.0215, 0.0229, 0.0231, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 03:44:48,042 INFO [finetune.py:976] (4/7) Epoch 13, batch 100, loss[loss=0.1362, simple_loss=0.2074, pruned_loss=0.03249, over 4838.00 frames. ], tot_loss[loss=0.1829, simple_loss=0.2509, pruned_loss=0.05749, over 378849.73 frames. ], batch size: 49, lr: 3.62e-03, grad_scale: 16.0 2023-04-27 03:44:55,551 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.308e+02 1.653e+02 1.937e+02 2.262e+02 3.719e+02, threshold=3.874e+02, percent-clipped=0.0 2023-04-27 03:44:57,708 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-04-27 03:45:21,008 INFO [finetune.py:976] (4/7) Epoch 13, batch 150, loss[loss=0.1594, simple_loss=0.2272, pruned_loss=0.04581, over 4745.00 frames. ], tot_loss[loss=0.1802, simple_loss=0.2472, pruned_loss=0.05661, over 506501.16 frames. ], batch size: 54, lr: 3.62e-03, grad_scale: 16.0 2023-04-27 03:45:43,200 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-04-27 03:45:53,958 INFO [finetune.py:976] (4/7) Epoch 13, batch 200, loss[loss=0.2, simple_loss=0.2689, pruned_loss=0.06553, over 4919.00 frames. ], tot_loss[loss=0.181, simple_loss=0.2469, pruned_loss=0.05758, over 605788.12 frames. ], batch size: 38, lr: 3.62e-03, grad_scale: 16.0 2023-04-27 03:45:55,117 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=68934.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:46:01,420 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.726e+01 1.554e+02 1.965e+02 2.387e+02 1.026e+03, threshold=3.930e+02, percent-clipped=4.0 2023-04-27 03:46:31,805 INFO [finetune.py:976] (4/7) Epoch 13, batch 250, loss[loss=0.1718, simple_loss=0.2475, pruned_loss=0.04804, over 4843.00 frames. ], tot_loss[loss=0.1838, simple_loss=0.2499, pruned_loss=0.05886, over 683987.09 frames. ], batch size: 30, lr: 3.62e-03, grad_scale: 16.0 2023-04-27 03:46:45,305 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3564, 1.7937, 2.2120, 2.9200, 2.2285, 1.7154, 1.8010, 2.1815], device='cuda:4'), covar=tensor([0.3419, 0.3458, 0.1688, 0.2584, 0.3121, 0.2960, 0.4166, 0.2533], device='cuda:4'), in_proj_covar=tensor([0.0286, 0.0248, 0.0224, 0.0317, 0.0214, 0.0230, 0.0232, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 03:47:36,760 INFO [finetune.py:976] (4/7) Epoch 13, batch 300, loss[loss=0.1773, simple_loss=0.2521, pruned_loss=0.05125, over 4828.00 frames. ], tot_loss[loss=0.186, simple_loss=0.2534, pruned_loss=0.05932, over 744022.30 frames. ], batch size: 51, lr: 3.62e-03, grad_scale: 16.0 2023-04-27 03:47:47,288 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=69041.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:47:48,432 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.203e+02 1.674e+02 1.942e+02 2.375e+02 4.255e+02, threshold=3.885e+02, percent-clipped=1.0 2023-04-27 03:47:48,580 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=69043.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:48:05,330 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=69055.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:48:39,070 INFO [finetune.py:976] (4/7) Epoch 13, batch 350, loss[loss=0.2157, simple_loss=0.2899, pruned_loss=0.07077, over 4747.00 frames. ], tot_loss[loss=0.1886, simple_loss=0.2565, pruned_loss=0.06036, over 792594.54 frames. ], batch size: 54, lr: 3.62e-03, grad_scale: 16.0 2023-04-27 03:48:58,746 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-04-27 03:48:59,541 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=69103.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:49:00,190 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=69104.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:49:17,727 INFO [finetune.py:976] (4/7) Epoch 13, batch 400, loss[loss=0.1766, simple_loss=0.2528, pruned_loss=0.05017, over 4774.00 frames. ], tot_loss[loss=0.1887, simple_loss=0.2575, pruned_loss=0.05998, over 829495.68 frames. ], batch size: 26, lr: 3.61e-03, grad_scale: 16.0 2023-04-27 03:49:24,722 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.134e+02 1.707e+02 2.109e+02 2.400e+02 4.777e+02, threshold=4.219e+02, percent-clipped=1.0 2023-04-27 03:49:30,538 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9547, 1.6875, 2.1391, 2.4135, 2.0705, 1.8729, 1.9555, 2.0113], device='cuda:4'), covar=tensor([0.5656, 0.8093, 0.8019, 0.6646, 0.7163, 1.0536, 1.0688, 1.0083], device='cuda:4'), in_proj_covar=tensor([0.0413, 0.0407, 0.0497, 0.0516, 0.0441, 0.0463, 0.0469, 0.0473], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 03:49:38,745 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=69162.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:49:42,425 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=69168.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:49:51,421 INFO [finetune.py:976] (4/7) Epoch 13, batch 450, loss[loss=0.1787, simple_loss=0.2497, pruned_loss=0.05388, over 4741.00 frames. ], tot_loss[loss=0.1872, simple_loss=0.2558, pruned_loss=0.05928, over 857504.04 frames. ], batch size: 27, lr: 3.61e-03, grad_scale: 16.0 2023-04-27 03:49:57,554 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5622, 1.8090, 1.3875, 1.1363, 1.1949, 1.1722, 1.3376, 1.0689], device='cuda:4'), covar=tensor([0.1760, 0.1238, 0.1571, 0.1823, 0.2348, 0.1983, 0.1158, 0.2128], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0214, 0.0170, 0.0204, 0.0204, 0.0185, 0.0157, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 03:49:58,693 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-04-27 03:50:19,113 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=69223.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:50:20,344 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=69225.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:50:22,768 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=69229.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:50:25,122 INFO [finetune.py:976] (4/7) Epoch 13, batch 500, loss[loss=0.1351, simple_loss=0.2019, pruned_loss=0.03418, over 4827.00 frames. ], tot_loss[loss=0.1854, simple_loss=0.2536, pruned_loss=0.05862, over 880612.38 frames. ], batch size: 41, lr: 3.61e-03, grad_scale: 16.0 2023-04-27 03:50:25,829 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=69234.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:50:31,193 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.140e+02 1.700e+02 1.950e+02 2.352e+02 3.819e+02, threshold=3.900e+02, percent-clipped=0.0 2023-04-27 03:50:32,980 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=69245.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 03:50:52,430 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=69273.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:50:57,807 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=69282.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:50:58,358 INFO [finetune.py:976] (4/7) Epoch 13, batch 550, loss[loss=0.1436, simple_loss=0.2273, pruned_loss=0.0299, over 4756.00 frames. ], tot_loss[loss=0.1827, simple_loss=0.25, pruned_loss=0.05767, over 895947.26 frames. ], batch size: 27, lr: 3.61e-03, grad_scale: 16.0 2023-04-27 03:51:00,286 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=69286.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:51:03,051 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-04-27 03:51:12,075 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4775, 1.8025, 1.6716, 1.9903, 1.9489, 2.0737, 1.6888, 3.6743], device='cuda:4'), covar=tensor([0.0543, 0.0700, 0.0697, 0.0987, 0.0539, 0.0634, 0.0663, 0.0155], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0044, 0.0040, 0.0039, 0.0039, 0.0058], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 03:51:14,432 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=69306.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 03:51:24,245 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=69320.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:51:32,010 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.94 vs. limit=2.0 2023-04-27 03:51:32,073 INFO [finetune.py:976] (4/7) Epoch 13, batch 600, loss[loss=0.2608, simple_loss=0.3166, pruned_loss=0.1025, over 3948.00 frames. ], tot_loss[loss=0.1832, simple_loss=0.2502, pruned_loss=0.05813, over 904182.38 frames. ], batch size: 65, lr: 3.61e-03, grad_scale: 16.0 2023-04-27 03:51:32,821 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=69334.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 03:51:37,061 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=69341.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:51:38,140 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.090e+02 1.579e+02 1.985e+02 2.306e+02 4.955e+02, threshold=3.970e+02, percent-clipped=1.0 2023-04-27 03:51:56,869 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=69368.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:52:04,778 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=69381.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:52:05,905 INFO [finetune.py:976] (4/7) Epoch 13, batch 650, loss[loss=0.1536, simple_loss=0.2378, pruned_loss=0.03473, over 4929.00 frames. ], tot_loss[loss=0.1861, simple_loss=0.2536, pruned_loss=0.05925, over 915901.90 frames. ], batch size: 33, lr: 3.61e-03, grad_scale: 16.0 2023-04-27 03:52:09,605 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=69389.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:52:12,664 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=69394.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:52:15,697 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=69399.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:52:25,169 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0311, 2.6679, 2.0806, 2.0201, 1.4294, 1.3531, 2.1729, 1.3266], device='cuda:4'), covar=tensor([0.2023, 0.1724, 0.1677, 0.2067, 0.2687, 0.2208, 0.1153, 0.2353], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0214, 0.0170, 0.0205, 0.0205, 0.0186, 0.0158, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 03:52:36,483 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=69429.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:52:44,420 INFO [finetune.py:976] (4/7) Epoch 13, batch 700, loss[loss=0.1778, simple_loss=0.2636, pruned_loss=0.04596, over 4821.00 frames. ], tot_loss[loss=0.1871, simple_loss=0.2554, pruned_loss=0.0594, over 925061.51 frames. ], batch size: 38, lr: 3.61e-03, grad_scale: 16.0 2023-04-27 03:52:56,107 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.279e+02 1.620e+02 1.977e+02 2.375e+02 4.653e+02, threshold=3.954e+02, percent-clipped=1.0 2023-04-27 03:53:09,902 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=69455.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:53:29,590 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5460, 1.6741, 0.8899, 1.2811, 1.6664, 1.3942, 1.3420, 1.4281], device='cuda:4'), covar=tensor([0.0542, 0.0389, 0.0371, 0.0595, 0.0288, 0.0555, 0.0540, 0.0631], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0025, 0.0023, 0.0029, 0.0020, 0.0029, 0.0028, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0045, 0.0038, 0.0050, 0.0038, 0.0049, 0.0048, 0.0050], device='cuda:4') 2023-04-27 03:53:38,500 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7385, 1.2614, 4.6718, 4.3672, 4.1118, 4.3413, 4.1610, 4.1514], device='cuda:4'), covar=tensor([0.7533, 0.6484, 0.1081, 0.1931, 0.1147, 0.1446, 0.2084, 0.1600], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0303, 0.0399, 0.0408, 0.0348, 0.0407, 0.0311, 0.0366], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 03:53:51,162 INFO [finetune.py:976] (4/7) Epoch 13, batch 750, loss[loss=0.2151, simple_loss=0.2862, pruned_loss=0.07202, over 4898.00 frames. ], tot_loss[loss=0.1888, simple_loss=0.2575, pruned_loss=0.06004, over 933335.04 frames. ], batch size: 37, lr: 3.61e-03, grad_scale: 16.0 2023-04-27 03:54:42,271 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=69518.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:54:45,770 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9091, 2.3233, 0.8902, 1.2274, 1.5907, 1.1718, 3.2147, 1.2960], device='cuda:4'), covar=tensor([0.0912, 0.1142, 0.1096, 0.1741, 0.0787, 0.1417, 0.0411, 0.1045], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0049, 0.0047, 0.0050, 0.0052, 0.0076, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 03:54:46,351 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=69524.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:54:46,844 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.58 vs. limit=2.0 2023-04-27 03:54:57,619 INFO [finetune.py:976] (4/7) Epoch 13, batch 800, loss[loss=0.1954, simple_loss=0.261, pruned_loss=0.06486, over 4765.00 frames. ], tot_loss[loss=0.1876, simple_loss=0.2566, pruned_loss=0.05931, over 937867.94 frames. ], batch size: 28, lr: 3.61e-03, grad_scale: 16.0 2023-04-27 03:55:09,402 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.803e+01 1.651e+02 1.914e+02 2.371e+02 3.866e+02, threshold=3.828e+02, percent-clipped=0.0 2023-04-27 03:55:11,394 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2323, 2.8951, 2.2035, 2.1088, 1.5974, 1.5333, 2.3230, 1.4886], device='cuda:4'), covar=tensor([0.1882, 0.1596, 0.1543, 0.1913, 0.2499, 0.2064, 0.1120, 0.2234], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0213, 0.0169, 0.0203, 0.0203, 0.0185, 0.0156, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 03:55:35,211 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=69581.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:55:36,383 INFO [finetune.py:976] (4/7) Epoch 13, batch 850, loss[loss=0.1541, simple_loss=0.2221, pruned_loss=0.04304, over 4849.00 frames. ], tot_loss[loss=0.1858, simple_loss=0.2545, pruned_loss=0.05858, over 942183.50 frames. ], batch size: 44, lr: 3.61e-03, grad_scale: 16.0 2023-04-27 03:55:40,099 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6252, 1.8121, 1.7815, 2.4654, 2.5457, 2.1313, 2.0584, 1.9720], device='cuda:4'), covar=tensor([0.1275, 0.1717, 0.1998, 0.1519, 0.0993, 0.1684, 0.2063, 0.2025], device='cuda:4'), in_proj_covar=tensor([0.0304, 0.0316, 0.0348, 0.0291, 0.0329, 0.0313, 0.0303, 0.0358], device='cuda:4'), out_proj_covar=tensor([6.3455e-05, 6.6435e-05, 7.4742e-05, 5.9566e-05, 6.8627e-05, 6.6482e-05, 6.4472e-05, 7.6454e-05], device='cuda:4') 2023-04-27 03:55:47,540 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=69601.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 03:55:50,664 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4612, 2.0433, 2.3820, 2.9046, 2.3645, 1.9060, 1.7445, 2.2319], device='cuda:4'), covar=tensor([0.3606, 0.3571, 0.1788, 0.2386, 0.2619, 0.2866, 0.4104, 0.2266], device='cuda:4'), in_proj_covar=tensor([0.0285, 0.0247, 0.0222, 0.0315, 0.0213, 0.0227, 0.0230, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 03:56:03,706 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7064, 1.5107, 1.9828, 1.9619, 1.5549, 1.3463, 1.5767, 1.0816], device='cuda:4'), covar=tensor([0.0607, 0.0804, 0.0459, 0.0624, 0.0719, 0.1305, 0.0670, 0.0750], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0070, 0.0070, 0.0067, 0.0075, 0.0096, 0.0075, 0.0069], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 03:56:07,734 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=69629.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 03:56:10,110 INFO [finetune.py:976] (4/7) Epoch 13, batch 900, loss[loss=0.1974, simple_loss=0.2553, pruned_loss=0.06978, over 4714.00 frames. ], tot_loss[loss=0.184, simple_loss=0.2519, pruned_loss=0.05802, over 944760.83 frames. ], batch size: 23, lr: 3.61e-03, grad_scale: 32.0 2023-04-27 03:56:16,242 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.053e+02 1.649e+02 1.999e+02 2.386e+02 5.563e+02, threshold=3.997e+02, percent-clipped=1.0 2023-04-27 03:56:37,657 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=69674.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 03:56:39,832 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=69676.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:56:44,052 INFO [finetune.py:976] (4/7) Epoch 13, batch 950, loss[loss=0.1727, simple_loss=0.2341, pruned_loss=0.05562, over 4757.00 frames. ], tot_loss[loss=0.1842, simple_loss=0.2514, pruned_loss=0.05851, over 948165.98 frames. ], batch size: 27, lr: 3.61e-03, grad_scale: 32.0 2023-04-27 03:56:53,909 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=69699.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:57:08,815 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9733, 0.9821, 1.1768, 1.0686, 0.9129, 0.7961, 0.9208, 0.6080], device='cuda:4'), covar=tensor([0.0411, 0.0585, 0.0503, 0.0443, 0.0578, 0.1049, 0.0454, 0.0676], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0070, 0.0070, 0.0066, 0.0075, 0.0096, 0.0075, 0.0069], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 03:57:10,861 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-27 03:57:11,100 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=69724.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:57:18,035 INFO [finetune.py:976] (4/7) Epoch 13, batch 1000, loss[loss=0.2639, simple_loss=0.308, pruned_loss=0.1099, over 4830.00 frames. ], tot_loss[loss=0.1852, simple_loss=0.2527, pruned_loss=0.05882, over 951540.59 frames. ], batch size: 39, lr: 3.61e-03, grad_scale: 32.0 2023-04-27 03:57:18,926 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-27 03:57:19,397 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=69735.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 03:57:24,158 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.258e+02 1.699e+02 2.018e+02 2.497e+02 4.277e+02, threshold=4.036e+02, percent-clipped=1.0 2023-04-27 03:57:26,750 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=69747.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:57:28,614 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=69750.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:57:43,369 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.03 vs. limit=5.0 2023-04-27 03:57:50,562 INFO [finetune.py:976] (4/7) Epoch 13, batch 1050, loss[loss=0.1949, simple_loss=0.2698, pruned_loss=0.05994, over 4862.00 frames. ], tot_loss[loss=0.1858, simple_loss=0.2547, pruned_loss=0.05844, over 952479.03 frames. ], batch size: 44, lr: 3.61e-03, grad_scale: 32.0 2023-04-27 03:58:13,073 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=69818.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:58:16,708 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=69824.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:58:23,481 INFO [finetune.py:976] (4/7) Epoch 13, batch 1100, loss[loss=0.1856, simple_loss=0.2582, pruned_loss=0.05645, over 4909.00 frames. ], tot_loss[loss=0.1866, simple_loss=0.2561, pruned_loss=0.0586, over 953142.06 frames. ], batch size: 36, lr: 3.61e-03, grad_scale: 32.0 2023-04-27 03:58:29,604 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0885, 2.5389, 0.9907, 1.4380, 2.0452, 1.2943, 3.3292, 1.5912], device='cuda:4'), covar=tensor([0.0639, 0.0720, 0.0805, 0.1289, 0.0514, 0.0988, 0.0198, 0.0677], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0067, 0.0049, 0.0047, 0.0051, 0.0052, 0.0076, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 03:58:30,712 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.061e+02 1.653e+02 1.942e+02 2.253e+02 5.634e+02, threshold=3.884e+02, percent-clipped=2.0 2023-04-27 03:58:31,515 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.39 vs. limit=5.0 2023-04-27 03:58:50,918 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=69866.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:59:00,300 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=69872.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:59:12,138 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=69881.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 03:59:13,752 INFO [finetune.py:976] (4/7) Epoch 13, batch 1150, loss[loss=0.2251, simple_loss=0.2906, pruned_loss=0.07978, over 4889.00 frames. ], tot_loss[loss=0.1878, simple_loss=0.2573, pruned_loss=0.0591, over 952804.79 frames. ], batch size: 43, lr: 3.61e-03, grad_scale: 32.0 2023-04-27 03:59:41,656 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=69901.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 04:00:16,220 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=69929.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:00:16,268 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=69929.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 04:00:18,639 INFO [finetune.py:976] (4/7) Epoch 13, batch 1200, loss[loss=0.1735, simple_loss=0.2434, pruned_loss=0.05182, over 4882.00 frames. ], tot_loss[loss=0.1876, simple_loss=0.2564, pruned_loss=0.0594, over 953648.71 frames. ], batch size: 32, lr: 3.61e-03, grad_scale: 32.0 2023-04-27 04:00:30,117 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5444, 1.3056, 4.2446, 3.9914, 3.7390, 3.9386, 3.9117, 3.7011], device='cuda:4'), covar=tensor([0.7180, 0.5970, 0.1042, 0.1636, 0.1116, 0.1603, 0.1867, 0.1585], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0302, 0.0398, 0.0406, 0.0345, 0.0404, 0.0310, 0.0364], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 04:00:32,298 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.030e+02 1.675e+02 1.952e+02 2.332e+02 4.660e+02, threshold=3.905e+02, percent-clipped=1.0 2023-04-27 04:00:36,038 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=69949.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 04:00:52,482 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=69976.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:00:53,025 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=69977.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:00:57,154 INFO [finetune.py:976] (4/7) Epoch 13, batch 1250, loss[loss=0.1417, simple_loss=0.2211, pruned_loss=0.03111, over 4765.00 frames. ], tot_loss[loss=0.1849, simple_loss=0.2533, pruned_loss=0.05824, over 956259.23 frames. ], batch size: 26, lr: 3.61e-03, grad_scale: 32.0 2023-04-27 04:01:26,388 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=70024.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:01:26,414 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=70024.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:01:30,023 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=70030.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 04:01:31,778 INFO [finetune.py:976] (4/7) Epoch 13, batch 1300, loss[loss=0.1798, simple_loss=0.2457, pruned_loss=0.05692, over 4928.00 frames. ], tot_loss[loss=0.1832, simple_loss=0.2509, pruned_loss=0.05773, over 957549.52 frames. ], batch size: 38, lr: 3.61e-03, grad_scale: 32.0 2023-04-27 04:01:32,516 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9525, 1.8203, 2.2224, 2.3309, 1.7989, 1.5652, 1.8643, 1.0748], device='cuda:4'), covar=tensor([0.0650, 0.0850, 0.0531, 0.0935, 0.0945, 0.1288, 0.0851, 0.0895], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0070, 0.0071, 0.0066, 0.0075, 0.0096, 0.0075, 0.0069], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 04:01:39,418 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.185e+02 1.660e+02 1.869e+02 2.265e+02 4.379e+02, threshold=3.739e+02, percent-clipped=1.0 2023-04-27 04:01:39,544 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0885, 1.7596, 2.0457, 2.3843, 2.3791, 1.9899, 1.7365, 2.1276], device='cuda:4'), covar=tensor([0.0680, 0.0977, 0.0525, 0.0443, 0.0501, 0.0710, 0.0732, 0.0468], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0204, 0.0184, 0.0175, 0.0179, 0.0187, 0.0157, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 04:01:45,191 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=70050.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:01:50,180 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1905, 1.9328, 2.3714, 2.6971, 2.2505, 2.1098, 2.2591, 2.2415], device='cuda:4'), covar=tensor([0.5836, 0.7625, 0.8317, 0.6705, 0.7144, 1.0582, 0.9730, 0.9834], device='cuda:4'), in_proj_covar=tensor([0.0411, 0.0405, 0.0494, 0.0513, 0.0441, 0.0461, 0.0467, 0.0472], device='cuda:4'), out_proj_covar=tensor([9.9823e-05, 1.0062e-04, 1.1125e-04, 1.2156e-04, 1.0677e-04, 1.1116e-04, 1.1182e-04, 1.1290e-04], device='cuda:4') 2023-04-27 04:01:58,566 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=70072.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:02:05,230 INFO [finetune.py:976] (4/7) Epoch 13, batch 1350, loss[loss=0.1494, simple_loss=0.2207, pruned_loss=0.03908, over 4753.00 frames. ], tot_loss[loss=0.183, simple_loss=0.2507, pruned_loss=0.05761, over 957340.79 frames. ], batch size: 28, lr: 3.61e-03, grad_scale: 32.0 2023-04-27 04:02:16,792 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=70098.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:02:24,493 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=70110.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:02:38,386 INFO [finetune.py:976] (4/7) Epoch 13, batch 1400, loss[loss=0.154, simple_loss=0.2187, pruned_loss=0.04469, over 3912.00 frames. ], tot_loss[loss=0.1854, simple_loss=0.2535, pruned_loss=0.05865, over 954075.07 frames. ], batch size: 17, lr: 3.61e-03, grad_scale: 32.0 2023-04-27 04:02:45,033 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.060e+02 1.629e+02 2.088e+02 2.462e+02 4.428e+02, threshold=4.176e+02, percent-clipped=4.0 2023-04-27 04:03:04,703 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=70171.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:03:11,826 INFO [finetune.py:976] (4/7) Epoch 13, batch 1450, loss[loss=0.1856, simple_loss=0.2536, pruned_loss=0.05874, over 4929.00 frames. ], tot_loss[loss=0.1873, simple_loss=0.2555, pruned_loss=0.05954, over 952062.39 frames. ], batch size: 33, lr: 3.61e-03, grad_scale: 32.0 2023-04-27 04:03:45,323 INFO [finetune.py:976] (4/7) Epoch 13, batch 1500, loss[loss=0.1756, simple_loss=0.2338, pruned_loss=0.0587, over 4346.00 frames. ], tot_loss[loss=0.1871, simple_loss=0.2556, pruned_loss=0.05935, over 953116.58 frames. ], batch size: 19, lr: 3.61e-03, grad_scale: 32.0 2023-04-27 04:03:51,424 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.134e+02 1.692e+02 1.982e+02 2.371e+02 3.829e+02, threshold=3.965e+02, percent-clipped=0.0 2023-04-27 04:04:10,236 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-27 04:04:10,458 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=70260.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:04:18,004 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9760, 2.5194, 2.1225, 2.3781, 1.6721, 2.1421, 2.2220, 1.7503], device='cuda:4'), covar=tensor([0.2084, 0.1276, 0.0731, 0.1169, 0.3247, 0.1021, 0.1977, 0.2620], device='cuda:4'), in_proj_covar=tensor([0.0296, 0.0316, 0.0228, 0.0288, 0.0318, 0.0267, 0.0259, 0.0275], device='cuda:4'), out_proj_covar=tensor([1.1952e-04, 1.2622e-04, 9.1045e-05, 1.1509e-04, 1.2964e-04, 1.0696e-04, 1.0501e-04, 1.1000e-04], device='cuda:4') 2023-04-27 04:04:40,470 INFO [finetune.py:976] (4/7) Epoch 13, batch 1550, loss[loss=0.2098, simple_loss=0.278, pruned_loss=0.07076, over 4864.00 frames. ], tot_loss[loss=0.1863, simple_loss=0.2555, pruned_loss=0.05852, over 954846.24 frames. ], batch size: 34, lr: 3.61e-03, grad_scale: 32.0 2023-04-27 04:05:17,054 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1061, 0.6502, 0.9464, 0.7925, 1.2309, 1.0051, 0.8877, 0.9772], device='cuda:4'), covar=tensor([0.1563, 0.1597, 0.2051, 0.1583, 0.0972, 0.1363, 0.1552, 0.2032], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0317, 0.0350, 0.0291, 0.0330, 0.0314, 0.0304, 0.0359], device='cuda:4'), out_proj_covar=tensor([6.4064e-05, 6.6629e-05, 7.5005e-05, 5.9658e-05, 6.8911e-05, 6.6741e-05, 6.4678e-05, 7.6681e-05], device='cuda:4') 2023-04-27 04:05:17,659 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=70321.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:05:28,465 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=70330.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 04:05:30,176 INFO [finetune.py:976] (4/7) Epoch 13, batch 1600, loss[loss=0.1675, simple_loss=0.2366, pruned_loss=0.04925, over 4768.00 frames. ], tot_loss[loss=0.1849, simple_loss=0.2535, pruned_loss=0.05809, over 956406.75 frames. ], batch size: 28, lr: 3.60e-03, grad_scale: 32.0 2023-04-27 04:05:41,006 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.445e+01 1.654e+02 2.055e+02 2.369e+02 4.520e+02, threshold=4.109e+02, percent-clipped=1.0 2023-04-27 04:06:03,988 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.54 vs. limit=2.0 2023-04-27 04:06:16,251 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=70378.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 04:06:19,128 INFO [finetune.py:976] (4/7) Epoch 13, batch 1650, loss[loss=0.1465, simple_loss=0.2209, pruned_loss=0.03604, over 4722.00 frames. ], tot_loss[loss=0.1838, simple_loss=0.2522, pruned_loss=0.05769, over 957698.13 frames. ], batch size: 23, lr: 3.60e-03, grad_scale: 32.0 2023-04-27 04:06:36,162 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-04-27 04:06:52,936 INFO [finetune.py:976] (4/7) Epoch 13, batch 1700, loss[loss=0.1205, simple_loss=0.1892, pruned_loss=0.02593, over 4795.00 frames. ], tot_loss[loss=0.1809, simple_loss=0.2487, pruned_loss=0.05651, over 958843.08 frames. ], batch size: 25, lr: 3.60e-03, grad_scale: 32.0 2023-04-27 04:06:55,522 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5769, 2.4783, 2.1244, 2.2743, 2.6828, 2.0666, 3.4677, 1.9359], device='cuda:4'), covar=tensor([0.4072, 0.2426, 0.4623, 0.3985, 0.2050, 0.3179, 0.1547, 0.4592], device='cuda:4'), in_proj_covar=tensor([0.0344, 0.0345, 0.0427, 0.0358, 0.0382, 0.0381, 0.0373, 0.0420], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 04:06:59,064 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.006e+02 1.620e+02 1.879e+02 2.417e+02 6.028e+02, threshold=3.758e+02, percent-clipped=1.0 2023-04-27 04:07:16,026 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=70466.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:07:26,840 INFO [finetune.py:976] (4/7) Epoch 13, batch 1750, loss[loss=0.1362, simple_loss=0.2159, pruned_loss=0.02823, over 4792.00 frames. ], tot_loss[loss=0.1818, simple_loss=0.2502, pruned_loss=0.05672, over 959979.53 frames. ], batch size: 29, lr: 3.60e-03, grad_scale: 32.0 2023-04-27 04:08:00,059 INFO [finetune.py:976] (4/7) Epoch 13, batch 1800, loss[loss=0.2343, simple_loss=0.3047, pruned_loss=0.08194, over 4806.00 frames. ], tot_loss[loss=0.1836, simple_loss=0.2526, pruned_loss=0.05726, over 959397.87 frames. ], batch size: 51, lr: 3.60e-03, grad_scale: 32.0 2023-04-27 04:08:06,008 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.076e+02 1.577e+02 1.883e+02 2.332e+02 3.915e+02, threshold=3.766e+02, percent-clipped=2.0 2023-04-27 04:08:12,886 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7290, 1.8448, 0.8826, 1.3945, 2.0365, 1.5910, 1.4541, 1.5256], device='cuda:4'), covar=tensor([0.0500, 0.0364, 0.0363, 0.0569, 0.0259, 0.0536, 0.0539, 0.0578], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0025, 0.0023, 0.0029, 0.0020, 0.0029, 0.0028, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0045, 0.0038, 0.0050, 0.0037, 0.0048, 0.0048, 0.0050], device='cuda:4') 2023-04-27 04:08:33,361 INFO [finetune.py:976] (4/7) Epoch 13, batch 1850, loss[loss=0.1966, simple_loss=0.2652, pruned_loss=0.06407, over 4814.00 frames. ], tot_loss[loss=0.1855, simple_loss=0.2544, pruned_loss=0.05831, over 957184.44 frames. ], batch size: 39, lr: 3.60e-03, grad_scale: 32.0 2023-04-27 04:08:42,656 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5149, 2.0212, 1.0238, 1.2634, 1.9022, 1.3659, 1.3421, 1.3622], device='cuda:4'), covar=tensor([0.0639, 0.0334, 0.0364, 0.0666, 0.0281, 0.0733, 0.0703, 0.0715], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0025, 0.0023, 0.0029, 0.0020, 0.0029, 0.0028, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0045, 0.0038, 0.0050, 0.0037, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 04:08:54,137 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=70616.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:09:06,411 INFO [finetune.py:976] (4/7) Epoch 13, batch 1900, loss[loss=0.1754, simple_loss=0.2514, pruned_loss=0.04971, over 4919.00 frames. ], tot_loss[loss=0.1873, simple_loss=0.2567, pruned_loss=0.05894, over 957608.39 frames. ], batch size: 42, lr: 3.60e-03, grad_scale: 32.0 2023-04-27 04:09:12,475 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.157e+02 1.737e+02 2.053e+02 2.436e+02 7.397e+02, threshold=4.106e+02, percent-clipped=5.0 2023-04-27 04:10:02,034 INFO [finetune.py:976] (4/7) Epoch 13, batch 1950, loss[loss=0.147, simple_loss=0.219, pruned_loss=0.03747, over 4762.00 frames. ], tot_loss[loss=0.1858, simple_loss=0.2552, pruned_loss=0.05816, over 958932.21 frames. ], batch size: 26, lr: 3.60e-03, grad_scale: 32.0 2023-04-27 04:10:08,875 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0728, 1.9485, 2.4672, 2.6435, 1.8621, 1.6899, 2.0342, 1.1093], device='cuda:4'), covar=tensor([0.0722, 0.1453, 0.0590, 0.0791, 0.0999, 0.1452, 0.0861, 0.1065], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0069, 0.0070, 0.0066, 0.0074, 0.0096, 0.0074, 0.0069], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 04:10:16,209 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5413, 1.6875, 1.4008, 0.9538, 1.1762, 1.1587, 1.4062, 1.1088], device='cuda:4'), covar=tensor([0.1708, 0.1460, 0.1516, 0.1996, 0.2428, 0.2026, 0.1112, 0.2052], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0214, 0.0169, 0.0203, 0.0202, 0.0184, 0.0156, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 04:10:39,638 INFO [finetune.py:976] (4/7) Epoch 13, batch 2000, loss[loss=0.1748, simple_loss=0.2423, pruned_loss=0.05361, over 4279.00 frames. ], tot_loss[loss=0.1845, simple_loss=0.2532, pruned_loss=0.05794, over 956312.60 frames. ], batch size: 66, lr: 3.60e-03, grad_scale: 32.0 2023-04-27 04:10:52,177 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.325e+01 1.560e+02 1.780e+02 2.155e+02 4.897e+02, threshold=3.560e+02, percent-clipped=2.0 2023-04-27 04:11:15,891 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=70766.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:11:28,145 INFO [finetune.py:976] (4/7) Epoch 13, batch 2050, loss[loss=0.1638, simple_loss=0.2387, pruned_loss=0.04442, over 4751.00 frames. ], tot_loss[loss=0.1815, simple_loss=0.2496, pruned_loss=0.05666, over 954211.08 frames. ], batch size: 26, lr: 3.60e-03, grad_scale: 32.0 2023-04-27 04:11:28,739 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5043, 3.0673, 1.2300, 2.0081, 1.7854, 2.4501, 1.8968, 1.2842], device='cuda:4'), covar=tensor([0.1171, 0.0802, 0.1586, 0.0959, 0.0987, 0.0805, 0.1328, 0.1765], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0246, 0.0139, 0.0122, 0.0132, 0.0152, 0.0117, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 04:11:48,347 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=70814.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:12:01,815 INFO [finetune.py:976] (4/7) Epoch 13, batch 2100, loss[loss=0.2065, simple_loss=0.2732, pruned_loss=0.06995, over 4932.00 frames. ], tot_loss[loss=0.1819, simple_loss=0.2497, pruned_loss=0.05707, over 952180.93 frames. ], batch size: 38, lr: 3.60e-03, grad_scale: 32.0 2023-04-27 04:12:08,820 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.125e+02 1.590e+02 1.945e+02 2.393e+02 5.833e+02, threshold=3.889e+02, percent-clipped=3.0 2023-04-27 04:12:35,712 INFO [finetune.py:976] (4/7) Epoch 13, batch 2150, loss[loss=0.1725, simple_loss=0.2462, pruned_loss=0.04945, over 4820.00 frames. ], tot_loss[loss=0.1855, simple_loss=0.2538, pruned_loss=0.05859, over 951972.09 frames. ], batch size: 25, lr: 3.60e-03, grad_scale: 32.0 2023-04-27 04:12:56,787 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=70916.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:13:08,519 INFO [finetune.py:976] (4/7) Epoch 13, batch 2200, loss[loss=0.1852, simple_loss=0.2558, pruned_loss=0.05732, over 4815.00 frames. ], tot_loss[loss=0.1882, simple_loss=0.2571, pruned_loss=0.05967, over 952216.31 frames. ], batch size: 51, lr: 3.60e-03, grad_scale: 32.0 2023-04-27 04:13:16,546 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.168e+02 1.648e+02 2.003e+02 2.484e+02 4.937e+02, threshold=4.005e+02, percent-clipped=2.0 2023-04-27 04:13:29,558 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=70964.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:13:41,751 INFO [finetune.py:976] (4/7) Epoch 13, batch 2250, loss[loss=0.2107, simple_loss=0.2764, pruned_loss=0.07248, over 4759.00 frames. ], tot_loss[loss=0.189, simple_loss=0.2578, pruned_loss=0.06009, over 952391.14 frames. ], batch size: 28, lr: 3.60e-03, grad_scale: 32.0 2023-04-27 04:13:49,253 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=70992.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:13:52,709 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-27 04:14:13,322 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=71030.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:14:14,996 INFO [finetune.py:976] (4/7) Epoch 13, batch 2300, loss[loss=0.1861, simple_loss=0.2564, pruned_loss=0.0579, over 4779.00 frames. ], tot_loss[loss=0.1891, simple_loss=0.2584, pruned_loss=0.0599, over 953233.89 frames. ], batch size: 29, lr: 3.60e-03, grad_scale: 32.0 2023-04-27 04:14:23,496 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.148e+02 1.522e+02 1.891e+02 2.315e+02 3.821e+02, threshold=3.782e+02, percent-clipped=0.0 2023-04-27 04:14:30,611 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=71053.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:14:47,219 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9317, 2.5330, 2.0977, 2.3784, 1.8007, 2.0990, 2.0950, 1.8342], device='cuda:4'), covar=tensor([0.2028, 0.1345, 0.0922, 0.1152, 0.3281, 0.1128, 0.2137, 0.2620], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0313, 0.0227, 0.0285, 0.0314, 0.0266, 0.0257, 0.0274], device='cuda:4'), out_proj_covar=tensor([1.1810e-04, 1.2483e-04, 9.0728e-05, 1.1388e-04, 1.2816e-04, 1.0635e-04, 1.0425e-04, 1.0948e-04], device='cuda:4') 2023-04-27 04:14:52,290 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-04-27 04:14:58,300 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1388, 1.5373, 1.9432, 2.2156, 1.8995, 1.5237, 1.0610, 1.6190], device='cuda:4'), covar=tensor([0.3316, 0.3466, 0.1730, 0.2226, 0.2830, 0.2736, 0.4337, 0.2315], device='cuda:4'), in_proj_covar=tensor([0.0284, 0.0246, 0.0221, 0.0314, 0.0212, 0.0227, 0.0229, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 04:15:10,013 INFO [finetune.py:976] (4/7) Epoch 13, batch 2350, loss[loss=0.1297, simple_loss=0.1985, pruned_loss=0.03042, over 4690.00 frames. ], tot_loss[loss=0.1883, simple_loss=0.2571, pruned_loss=0.05977, over 953098.66 frames. ], batch size: 23, lr: 3.60e-03, grad_scale: 32.0 2023-04-27 04:15:20,874 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=71091.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:16:07,372 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.90 vs. limit=2.0 2023-04-27 04:16:08,212 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3076, 1.2324, 1.3601, 1.6365, 1.6040, 1.3186, 1.0593, 1.5232], device='cuda:4'), covar=tensor([0.0817, 0.1231, 0.0865, 0.0592, 0.0643, 0.0742, 0.0784, 0.0550], device='cuda:4'), in_proj_covar=tensor([0.0192, 0.0201, 0.0182, 0.0173, 0.0178, 0.0184, 0.0155, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 04:16:15,378 INFO [finetune.py:976] (4/7) Epoch 13, batch 2400, loss[loss=0.1898, simple_loss=0.2497, pruned_loss=0.06492, over 4802.00 frames. ], tot_loss[loss=0.1855, simple_loss=0.2536, pruned_loss=0.05871, over 954488.42 frames. ], batch size: 51, lr: 3.60e-03, grad_scale: 32.0 2023-04-27 04:16:26,845 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.075e+02 1.510e+02 1.853e+02 2.211e+02 4.308e+02, threshold=3.705e+02, percent-clipped=1.0 2023-04-27 04:16:28,691 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-04-27 04:16:40,342 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9905, 1.6578, 2.1012, 2.4301, 2.0595, 1.8692, 2.0435, 2.0494], device='cuda:4'), covar=tensor([0.5291, 0.7839, 0.8267, 0.6339, 0.6614, 0.9283, 0.9852, 0.9076], device='cuda:4'), in_proj_covar=tensor([0.0414, 0.0409, 0.0497, 0.0516, 0.0444, 0.0464, 0.0471, 0.0475], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 04:16:54,089 INFO [finetune.py:976] (4/7) Epoch 13, batch 2450, loss[loss=0.1576, simple_loss=0.2214, pruned_loss=0.04694, over 4823.00 frames. ], tot_loss[loss=0.1828, simple_loss=0.2505, pruned_loss=0.05758, over 954856.85 frames. ], batch size: 51, lr: 3.60e-03, grad_scale: 32.0 2023-04-27 04:17:07,717 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7338, 4.3225, 0.7535, 2.2517, 2.2345, 2.8319, 2.5276, 1.0605], device='cuda:4'), covar=tensor([0.1615, 0.1649, 0.2701, 0.1455, 0.1256, 0.1314, 0.1533, 0.2148], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0248, 0.0140, 0.0122, 0.0134, 0.0153, 0.0118, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 04:17:09,534 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3355, 2.9554, 0.9397, 1.6272, 2.2484, 1.3689, 3.8666, 1.9715], device='cuda:4'), covar=tensor([0.0657, 0.0824, 0.0892, 0.1220, 0.0486, 0.0983, 0.0180, 0.0576], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0067, 0.0049, 0.0047, 0.0050, 0.0052, 0.0076, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 04:17:23,971 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0529, 1.6828, 2.0619, 2.4129, 2.4280, 1.8587, 1.6599, 1.9977], device='cuda:4'), covar=tensor([0.0876, 0.1182, 0.0655, 0.0552, 0.0602, 0.0848, 0.0828, 0.0640], device='cuda:4'), in_proj_covar=tensor([0.0193, 0.0202, 0.0183, 0.0174, 0.0179, 0.0185, 0.0156, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 04:17:28,066 INFO [finetune.py:976] (4/7) Epoch 13, batch 2500, loss[loss=0.1738, simple_loss=0.2631, pruned_loss=0.0422, over 4835.00 frames. ], tot_loss[loss=0.1847, simple_loss=0.2526, pruned_loss=0.05839, over 955658.78 frames. ], batch size: 47, lr: 3.60e-03, grad_scale: 32.0 2023-04-27 04:17:34,109 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.448e+01 1.657e+02 1.843e+02 2.125e+02 3.906e+02, threshold=3.687e+02, percent-clipped=2.0 2023-04-27 04:18:01,396 INFO [finetune.py:976] (4/7) Epoch 13, batch 2550, loss[loss=0.1979, simple_loss=0.2738, pruned_loss=0.061, over 4159.00 frames. ], tot_loss[loss=0.1871, simple_loss=0.2559, pruned_loss=0.05915, over 954332.42 frames. ], batch size: 65, lr: 3.60e-03, grad_scale: 32.0 2023-04-27 04:18:15,601 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-04-27 04:18:34,377 INFO [finetune.py:976] (4/7) Epoch 13, batch 2600, loss[loss=0.1984, simple_loss=0.2702, pruned_loss=0.06333, over 4903.00 frames. ], tot_loss[loss=0.1899, simple_loss=0.2586, pruned_loss=0.06058, over 955752.99 frames. ], batch size: 43, lr: 3.60e-03, grad_scale: 32.0 2023-04-27 04:18:40,523 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.287e+02 1.728e+02 1.986e+02 2.411e+02 4.998e+02, threshold=3.972e+02, percent-clipped=3.0 2023-04-27 04:18:43,641 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=71348.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:18:55,014 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5296, 1.4888, 1.8170, 1.8522, 1.3306, 1.2076, 1.4401, 0.9578], device='cuda:4'), covar=tensor([0.0616, 0.0745, 0.0461, 0.0661, 0.0952, 0.1367, 0.0703, 0.0755], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0070, 0.0070, 0.0067, 0.0075, 0.0096, 0.0075, 0.0069], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 04:18:56,771 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.62 vs. limit=2.0 2023-04-27 04:19:08,105 INFO [finetune.py:976] (4/7) Epoch 13, batch 2650, loss[loss=0.2252, simple_loss=0.291, pruned_loss=0.07974, over 4803.00 frames. ], tot_loss[loss=0.1893, simple_loss=0.2585, pruned_loss=0.06009, over 955909.26 frames. ], batch size: 41, lr: 3.60e-03, grad_scale: 32.0 2023-04-27 04:19:10,021 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=71386.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:19:40,632 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-27 04:19:41,957 INFO [finetune.py:976] (4/7) Epoch 13, batch 2700, loss[loss=0.1347, simple_loss=0.2028, pruned_loss=0.03332, over 4737.00 frames. ], tot_loss[loss=0.187, simple_loss=0.2561, pruned_loss=0.05898, over 955314.14 frames. ], batch size: 54, lr: 3.60e-03, grad_scale: 32.0 2023-04-27 04:19:43,254 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.0300, 4.0237, 2.8897, 4.6341, 4.1596, 4.0510, 1.8649, 3.9939], device='cuda:4'), covar=tensor([0.1593, 0.0939, 0.2847, 0.1292, 0.2698, 0.1706, 0.5887, 0.2272], device='cuda:4'), in_proj_covar=tensor([0.0240, 0.0212, 0.0247, 0.0299, 0.0294, 0.0244, 0.0267, 0.0267], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 04:19:48,072 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.223e+02 1.575e+02 1.881e+02 2.224e+02 4.491e+02, threshold=3.762e+02, percent-clipped=1.0 2023-04-27 04:20:30,301 INFO [finetune.py:976] (4/7) Epoch 13, batch 2750, loss[loss=0.2024, simple_loss=0.2695, pruned_loss=0.06766, over 4930.00 frames. ], tot_loss[loss=0.1851, simple_loss=0.2534, pruned_loss=0.05844, over 955744.24 frames. ], batch size: 33, lr: 3.59e-03, grad_scale: 32.0 2023-04-27 04:20:51,283 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-27 04:20:53,085 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2727, 3.0380, 2.2718, 2.4038, 1.6460, 1.5710, 2.5882, 1.7231], device='cuda:4'), covar=tensor([0.1804, 0.1565, 0.1377, 0.1698, 0.2400, 0.2014, 0.0947, 0.2032], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0214, 0.0169, 0.0204, 0.0202, 0.0184, 0.0157, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 04:21:05,983 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.94 vs. limit=2.0 2023-04-27 04:21:21,454 INFO [finetune.py:976] (4/7) Epoch 13, batch 2800, loss[loss=0.1994, simple_loss=0.2538, pruned_loss=0.07252, over 4930.00 frames. ], tot_loss[loss=0.1815, simple_loss=0.2494, pruned_loss=0.05686, over 954441.93 frames. ], batch size: 33, lr: 3.59e-03, grad_scale: 32.0 2023-04-27 04:21:33,185 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.066e+02 1.551e+02 1.952e+02 2.336e+02 5.892e+02, threshold=3.903e+02, percent-clipped=4.0 2023-04-27 04:21:40,257 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1399, 2.7433, 2.1363, 2.0622, 1.5850, 1.4888, 2.2856, 1.5657], device='cuda:4'), covar=tensor([0.1568, 0.1477, 0.1317, 0.1777, 0.2222, 0.1776, 0.0921, 0.1875], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0212, 0.0168, 0.0202, 0.0201, 0.0183, 0.0156, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 04:22:06,642 INFO [finetune.py:976] (4/7) Epoch 13, batch 2850, loss[loss=0.2022, simple_loss=0.2656, pruned_loss=0.06941, over 4828.00 frames. ], tot_loss[loss=0.1805, simple_loss=0.2483, pruned_loss=0.05637, over 954566.35 frames. ], batch size: 30, lr: 3.59e-03, grad_scale: 32.0 2023-04-27 04:22:28,015 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=71616.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:22:38,579 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.12 vs. limit=2.0 2023-04-27 04:22:40,695 INFO [finetune.py:976] (4/7) Epoch 13, batch 2900, loss[loss=0.1571, simple_loss=0.2278, pruned_loss=0.04321, over 4761.00 frames. ], tot_loss[loss=0.1841, simple_loss=0.2523, pruned_loss=0.05789, over 952351.23 frames. ], batch size: 27, lr: 3.59e-03, grad_scale: 64.0 2023-04-27 04:22:46,792 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.322e+02 1.703e+02 2.022e+02 2.297e+02 3.791e+02, threshold=4.044e+02, percent-clipped=0.0 2023-04-27 04:22:49,910 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=71648.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:23:03,682 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8548, 2.5647, 1.8817, 1.7716, 1.3125, 1.3347, 1.9934, 1.2532], device='cuda:4'), covar=tensor([0.1918, 0.1484, 0.1571, 0.1998, 0.2595, 0.2171, 0.1070, 0.2297], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0212, 0.0168, 0.0202, 0.0200, 0.0182, 0.0156, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 04:23:09,457 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=71677.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:23:13,484 INFO [finetune.py:976] (4/7) Epoch 13, batch 2950, loss[loss=0.2031, simple_loss=0.2607, pruned_loss=0.0728, over 4215.00 frames. ], tot_loss[loss=0.1866, simple_loss=0.2553, pruned_loss=0.05899, over 951097.74 frames. ], batch size: 18, lr: 3.59e-03, grad_scale: 64.0 2023-04-27 04:23:15,389 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=71686.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:23:20,268 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8775, 2.3921, 1.9536, 2.3330, 1.6777, 1.9458, 1.9103, 1.6783], device='cuda:4'), covar=tensor([0.2216, 0.1482, 0.0966, 0.1215, 0.3514, 0.1337, 0.2200, 0.2777], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0313, 0.0226, 0.0284, 0.0314, 0.0266, 0.0256, 0.0273], device='cuda:4'), out_proj_covar=tensor([1.1795e-04, 1.2488e-04, 9.0418e-05, 1.1346e-04, 1.2803e-04, 1.0653e-04, 1.0364e-04, 1.0903e-04], device='cuda:4') 2023-04-27 04:23:21,435 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=71696.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:23:45,590 INFO [finetune.py:976] (4/7) Epoch 13, batch 3000, loss[loss=0.2421, simple_loss=0.3082, pruned_loss=0.08805, over 4892.00 frames. ], tot_loss[loss=0.1884, simple_loss=0.2572, pruned_loss=0.05982, over 953356.32 frames. ], batch size: 43, lr: 3.59e-03, grad_scale: 64.0 2023-04-27 04:23:45,590 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 04:23:49,200 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4449, 1.7005, 1.6442, 1.8991, 1.8052, 1.9053, 1.5261, 3.0477], device='cuda:4'), covar=tensor([0.0515, 0.0669, 0.0670, 0.0993, 0.0493, 0.0397, 0.0635, 0.0222], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0043, 0.0040, 0.0038, 0.0038, 0.0057], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0012, 0.0013, 0.0016], device='cuda:4') 2023-04-27 04:23:56,065 INFO [finetune.py:1010] (4/7) Epoch 13, validation: loss=0.1517, simple_loss=0.224, pruned_loss=0.03973, over 2265189.00 frames. 2023-04-27 04:23:56,066 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-27 04:23:56,741 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=71734.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:24:00,247 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2350, 1.7298, 2.1544, 2.7246, 2.1528, 1.6658, 1.5157, 2.1001], device='cuda:4'), covar=tensor([0.3694, 0.3741, 0.1752, 0.2543, 0.2923, 0.2962, 0.4363, 0.2211], device='cuda:4'), in_proj_covar=tensor([0.0285, 0.0247, 0.0221, 0.0315, 0.0214, 0.0229, 0.0230, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 04:24:03,112 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.215e+02 1.669e+02 1.992e+02 2.305e+02 4.949e+02, threshold=3.985e+02, percent-clipped=1.0 2023-04-27 04:24:12,121 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.66 vs. limit=5.0 2023-04-27 04:24:27,545 INFO [finetune.py:976] (4/7) Epoch 13, batch 3050, loss[loss=0.2206, simple_loss=0.2831, pruned_loss=0.07907, over 4774.00 frames. ], tot_loss[loss=0.1883, simple_loss=0.2574, pruned_loss=0.05963, over 953805.44 frames. ], batch size: 51, lr: 3.59e-03, grad_scale: 64.0 2023-04-27 04:24:48,509 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-27 04:24:49,645 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3122, 1.6484, 1.5275, 1.8583, 1.8241, 2.0845, 1.5033, 3.7266], device='cuda:4'), covar=tensor([0.0602, 0.0767, 0.0814, 0.1149, 0.0579, 0.0488, 0.0718, 0.0114], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0043, 0.0040, 0.0038, 0.0038, 0.0057], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0012, 0.0013, 0.0016], device='cuda:4') 2023-04-27 04:24:57,039 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.69 vs. limit=2.0 2023-04-27 04:25:00,523 INFO [finetune.py:976] (4/7) Epoch 13, batch 3100, loss[loss=0.2088, simple_loss=0.2506, pruned_loss=0.08347, over 4808.00 frames. ], tot_loss[loss=0.1854, simple_loss=0.2543, pruned_loss=0.05827, over 953655.81 frames. ], batch size: 51, lr: 3.59e-03, grad_scale: 64.0 2023-04-27 04:25:08,970 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.880e+01 1.511e+02 1.834e+02 2.152e+02 3.267e+02, threshold=3.669e+02, percent-clipped=0.0 2023-04-27 04:25:14,567 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.81 vs. limit=2.0 2023-04-27 04:25:42,159 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5924, 1.5750, 0.7044, 1.2508, 1.4768, 1.3930, 1.3411, 1.3430], device='cuda:4'), covar=tensor([0.0607, 0.0365, 0.0376, 0.0633, 0.0291, 0.0639, 0.0648, 0.0683], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0025, 0.0023, 0.0029, 0.0020, 0.0029, 0.0028, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0044, 0.0037, 0.0050, 0.0037, 0.0048, 0.0049, 0.0050], device='cuda:4') 2023-04-27 04:25:46,573 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-27 04:25:54,816 INFO [finetune.py:976] (4/7) Epoch 13, batch 3150, loss[loss=0.1754, simple_loss=0.2482, pruned_loss=0.05128, over 4912.00 frames. ], tot_loss[loss=0.183, simple_loss=0.2512, pruned_loss=0.05738, over 954023.28 frames. ], batch size: 36, lr: 3.59e-03, grad_scale: 64.0 2023-04-27 04:26:48,051 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=71919.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:27:02,052 INFO [finetune.py:976] (4/7) Epoch 13, batch 3200, loss[loss=0.2151, simple_loss=0.2807, pruned_loss=0.07476, over 4909.00 frames. ], tot_loss[loss=0.1806, simple_loss=0.2482, pruned_loss=0.05654, over 955798.28 frames. ], batch size: 36, lr: 3.59e-03, grad_scale: 64.0 2023-04-27 04:27:08,187 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.251e+02 1.591e+02 1.837e+02 2.276e+02 4.272e+02, threshold=3.675e+02, percent-clipped=2.0 2023-04-27 04:27:24,998 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-27 04:27:26,369 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.4896, 3.4772, 2.6677, 4.0784, 3.4939, 3.5334, 1.6352, 3.4439], device='cuda:4'), covar=tensor([0.1938, 0.1281, 0.3649, 0.1849, 0.3553, 0.1875, 0.5871, 0.2743], device='cuda:4'), in_proj_covar=tensor([0.0242, 0.0213, 0.0247, 0.0300, 0.0296, 0.0246, 0.0269, 0.0268], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 04:27:28,807 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=71972.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:27:33,747 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=71980.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:27:35,467 INFO [finetune.py:976] (4/7) Epoch 13, batch 3250, loss[loss=0.1378, simple_loss=0.209, pruned_loss=0.0333, over 4830.00 frames. ], tot_loss[loss=0.1812, simple_loss=0.2488, pruned_loss=0.05677, over 953418.79 frames. ], batch size: 30, lr: 3.59e-03, grad_scale: 64.0 2023-04-27 04:27:57,811 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-04-27 04:28:01,949 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.13 vs. limit=5.0 2023-04-27 04:28:10,359 INFO [finetune.py:976] (4/7) Epoch 13, batch 3300, loss[loss=0.1965, simple_loss=0.272, pruned_loss=0.06048, over 4901.00 frames. ], tot_loss[loss=0.1842, simple_loss=0.2526, pruned_loss=0.05794, over 951444.53 frames. ], batch size: 35, lr: 3.59e-03, grad_scale: 32.0 2023-04-27 04:28:16,989 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.593e+02 1.877e+02 2.325e+02 3.643e+02, threshold=3.754e+02, percent-clipped=0.0 2023-04-27 04:28:28,476 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1290, 2.7547, 1.0860, 1.5179, 2.3850, 1.2617, 3.8004, 1.9104], device='cuda:4'), covar=tensor([0.0694, 0.0615, 0.0799, 0.1347, 0.0438, 0.1044, 0.0227, 0.0643], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0049, 0.0047, 0.0050, 0.0052, 0.0076, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 04:28:43,991 INFO [finetune.py:976] (4/7) Epoch 13, batch 3350, loss[loss=0.1943, simple_loss=0.2601, pruned_loss=0.06423, over 4920.00 frames. ], tot_loss[loss=0.186, simple_loss=0.255, pruned_loss=0.0585, over 952627.07 frames. ], batch size: 38, lr: 3.59e-03, grad_scale: 32.0 2023-04-27 04:29:15,419 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=72129.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:29:17,727 INFO [finetune.py:976] (4/7) Epoch 13, batch 3400, loss[loss=0.1712, simple_loss=0.2401, pruned_loss=0.05117, over 4724.00 frames. ], tot_loss[loss=0.1864, simple_loss=0.2556, pruned_loss=0.05854, over 951913.54 frames. ], batch size: 54, lr: 3.59e-03, grad_scale: 32.0 2023-04-27 04:29:24,412 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.141e+02 1.571e+02 1.878e+02 2.335e+02 4.983e+02, threshold=3.756e+02, percent-clipped=1.0 2023-04-27 04:29:51,374 INFO [finetune.py:976] (4/7) Epoch 13, batch 3450, loss[loss=0.1632, simple_loss=0.241, pruned_loss=0.04268, over 4813.00 frames. ], tot_loss[loss=0.1858, simple_loss=0.2556, pruned_loss=0.05805, over 954487.19 frames. ], batch size: 39, lr: 3.59e-03, grad_scale: 32.0 2023-04-27 04:29:55,729 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=72190.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:30:24,560 INFO [finetune.py:976] (4/7) Epoch 13, batch 3500, loss[loss=0.2124, simple_loss=0.2697, pruned_loss=0.07752, over 4816.00 frames. ], tot_loss[loss=0.1851, simple_loss=0.2536, pruned_loss=0.05824, over 955610.25 frames. ], batch size: 25, lr: 3.59e-03, grad_scale: 32.0 2023-04-27 04:30:31,094 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.136e+02 1.651e+02 2.080e+02 2.507e+02 4.410e+02, threshold=4.160e+02, percent-clipped=6.0 2023-04-27 04:30:50,581 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4188, 1.2589, 1.6882, 1.6001, 1.3140, 1.2251, 1.2815, 0.7929], device='cuda:4'), covar=tensor([0.0549, 0.0782, 0.0466, 0.0576, 0.0787, 0.1297, 0.0589, 0.0681], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0071, 0.0070, 0.0068, 0.0076, 0.0097, 0.0076, 0.0069], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 04:31:08,036 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=72272.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:31:09,847 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=72275.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:31:20,405 INFO [finetune.py:976] (4/7) Epoch 13, batch 3550, loss[loss=0.1807, simple_loss=0.2492, pruned_loss=0.05607, over 4859.00 frames. ], tot_loss[loss=0.1835, simple_loss=0.2509, pruned_loss=0.05804, over 957429.76 frames. ], batch size: 44, lr: 3.59e-03, grad_scale: 32.0 2023-04-27 04:31:54,479 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-27 04:32:06,131 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=72320.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:32:16,756 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=72326.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:32:26,340 INFO [finetune.py:976] (4/7) Epoch 13, batch 3600, loss[loss=0.1724, simple_loss=0.2479, pruned_loss=0.04849, over 4747.00 frames. ], tot_loss[loss=0.1828, simple_loss=0.2501, pruned_loss=0.05772, over 956298.43 frames. ], batch size: 23, lr: 3.59e-03, grad_scale: 32.0 2023-04-27 04:32:33,139 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.090e+02 1.673e+02 2.029e+02 2.570e+02 4.003e+02, threshold=4.058e+02, percent-clipped=0.0 2023-04-27 04:32:34,533 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4448, 1.9748, 2.4006, 2.7752, 2.3442, 1.9329, 1.7101, 2.1745], device='cuda:4'), covar=tensor([0.3231, 0.3112, 0.1574, 0.2196, 0.2621, 0.2523, 0.3928, 0.2009], device='cuda:4'), in_proj_covar=tensor([0.0284, 0.0246, 0.0221, 0.0315, 0.0213, 0.0228, 0.0229, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 04:32:59,901 INFO [finetune.py:976] (4/7) Epoch 13, batch 3650, loss[loss=0.1707, simple_loss=0.2482, pruned_loss=0.04659, over 4911.00 frames. ], tot_loss[loss=0.184, simple_loss=0.2516, pruned_loss=0.05819, over 956011.23 frames. ], batch size: 37, lr: 3.59e-03, grad_scale: 32.0 2023-04-27 04:33:02,540 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=72387.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:33:33,726 INFO [finetune.py:976] (4/7) Epoch 13, batch 3700, loss[loss=0.1855, simple_loss=0.2563, pruned_loss=0.05738, over 4932.00 frames. ], tot_loss[loss=0.1859, simple_loss=0.2545, pruned_loss=0.05865, over 952819.86 frames. ], batch size: 33, lr: 3.59e-03, grad_scale: 32.0 2023-04-27 04:33:40,456 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.072e+02 1.574e+02 1.888e+02 2.267e+02 6.702e+02, threshold=3.776e+02, percent-clipped=3.0 2023-04-27 04:33:44,199 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5586, 1.7637, 0.7642, 1.2982, 1.7042, 1.4419, 1.3458, 1.4239], device='cuda:4'), covar=tensor([0.0521, 0.0347, 0.0394, 0.0559, 0.0300, 0.0522, 0.0494, 0.0564], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0025, 0.0023, 0.0029, 0.0020, 0.0029, 0.0028, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0045, 0.0038, 0.0050, 0.0037, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 04:34:07,020 INFO [finetune.py:976] (4/7) Epoch 13, batch 3750, loss[loss=0.1986, simple_loss=0.2781, pruned_loss=0.05954, over 4768.00 frames. ], tot_loss[loss=0.1877, simple_loss=0.2563, pruned_loss=0.05949, over 953586.37 frames. ], batch size: 28, lr: 3.59e-03, grad_scale: 32.0 2023-04-27 04:34:08,311 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=72485.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:34:27,689 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6082, 2.4082, 1.8920, 2.1618, 2.3173, 1.8589, 2.8594, 1.7795], device='cuda:4'), covar=tensor([0.3565, 0.1768, 0.3482, 0.2808, 0.1819, 0.2490, 0.1894, 0.4107], device='cuda:4'), in_proj_covar=tensor([0.0344, 0.0345, 0.0428, 0.0355, 0.0380, 0.0382, 0.0373, 0.0418], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 04:34:39,242 INFO [finetune.py:976] (4/7) Epoch 13, batch 3800, loss[loss=0.1776, simple_loss=0.2501, pruned_loss=0.05256, over 4917.00 frames. ], tot_loss[loss=0.1872, simple_loss=0.2564, pruned_loss=0.05903, over 952976.81 frames. ], batch size: 41, lr: 3.59e-03, grad_scale: 32.0 2023-04-27 04:34:46,446 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.067e+02 1.671e+02 1.907e+02 2.254e+02 4.045e+02, threshold=3.813e+02, percent-clipped=1.0 2023-04-27 04:35:05,801 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=72575.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:35:11,561 INFO [finetune.py:976] (4/7) Epoch 13, batch 3850, loss[loss=0.1717, simple_loss=0.239, pruned_loss=0.05217, over 4828.00 frames. ], tot_loss[loss=0.1849, simple_loss=0.254, pruned_loss=0.05789, over 954980.33 frames. ], batch size: 30, lr: 3.59e-03, grad_scale: 32.0 2023-04-27 04:35:37,010 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=72623.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:35:44,026 INFO [finetune.py:976] (4/7) Epoch 13, batch 3900, loss[loss=0.1208, simple_loss=0.1972, pruned_loss=0.02221, over 4816.00 frames. ], tot_loss[loss=0.1824, simple_loss=0.2509, pruned_loss=0.05692, over 955037.29 frames. ], batch size: 51, lr: 3.59e-03, grad_scale: 32.0 2023-04-27 04:35:51,958 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.075e+02 1.669e+02 1.861e+02 2.404e+02 3.488e+02, threshold=3.722e+02, percent-clipped=0.0 2023-04-27 04:36:17,228 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4211, 1.5753, 1.7658, 1.9255, 1.7610, 1.9395, 1.9096, 1.8043], device='cuda:4'), covar=tensor([0.4205, 0.5484, 0.4405, 0.4955, 0.5752, 0.7220, 0.5369, 0.5177], device='cuda:4'), in_proj_covar=tensor([0.0332, 0.0379, 0.0320, 0.0332, 0.0345, 0.0399, 0.0359, 0.0328], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 04:36:27,957 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-27 04:36:32,484 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=72682.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:36:33,065 INFO [finetune.py:976] (4/7) Epoch 13, batch 3950, loss[loss=0.1483, simple_loss=0.2229, pruned_loss=0.03682, over 4788.00 frames. ], tot_loss[loss=0.1802, simple_loss=0.2482, pruned_loss=0.05614, over 956048.53 frames. ], batch size: 29, lr: 3.58e-03, grad_scale: 32.0 2023-04-27 04:37:38,719 INFO [finetune.py:976] (4/7) Epoch 13, batch 4000, loss[loss=0.2115, simple_loss=0.2805, pruned_loss=0.07121, over 4902.00 frames. ], tot_loss[loss=0.1795, simple_loss=0.2477, pruned_loss=0.05571, over 954903.93 frames. ], batch size: 43, lr: 3.58e-03, grad_scale: 32.0 2023-04-27 04:37:56,698 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6641, 1.5473, 1.7357, 2.0922, 2.0892, 1.7575, 1.3282, 1.8257], device='cuda:4'), covar=tensor([0.0842, 0.1229, 0.0783, 0.0563, 0.0613, 0.0775, 0.0779, 0.0577], device='cuda:4'), in_proj_covar=tensor([0.0191, 0.0201, 0.0181, 0.0172, 0.0177, 0.0182, 0.0155, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 04:37:57,173 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 7.525e+01 1.494e+02 1.807e+02 2.123e+02 3.249e+02, threshold=3.613e+02, percent-clipped=0.0 2023-04-27 04:38:27,851 INFO [finetune.py:976] (4/7) Epoch 13, batch 4050, loss[loss=0.136, simple_loss=0.2082, pruned_loss=0.03187, over 4800.00 frames. ], tot_loss[loss=0.1836, simple_loss=0.252, pruned_loss=0.05761, over 955622.23 frames. ], batch size: 25, lr: 3.58e-03, grad_scale: 32.0 2023-04-27 04:38:29,181 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=72785.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:38:34,285 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4862, 1.3358, 0.5568, 1.2179, 1.3979, 1.3697, 1.2579, 1.2692], device='cuda:4'), covar=tensor([0.0521, 0.0401, 0.0427, 0.0588, 0.0322, 0.0533, 0.0512, 0.0583], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0025, 0.0023, 0.0030, 0.0020, 0.0029, 0.0029, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0045, 0.0038, 0.0051, 0.0038, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 04:38:45,468 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-04-27 04:39:01,351 INFO [finetune.py:976] (4/7) Epoch 13, batch 4100, loss[loss=0.2111, simple_loss=0.2798, pruned_loss=0.0712, over 4894.00 frames. ], tot_loss[loss=0.1853, simple_loss=0.2541, pruned_loss=0.05828, over 954029.80 frames. ], batch size: 35, lr: 3.58e-03, grad_scale: 32.0 2023-04-27 04:39:01,412 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=72833.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:39:09,034 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.148e+02 1.644e+02 1.849e+02 2.325e+02 6.848e+02, threshold=3.698e+02, percent-clipped=3.0 2023-04-27 04:39:34,768 INFO [finetune.py:976] (4/7) Epoch 13, batch 4150, loss[loss=0.1882, simple_loss=0.259, pruned_loss=0.05873, over 4904.00 frames. ], tot_loss[loss=0.1862, simple_loss=0.2554, pruned_loss=0.05853, over 953103.01 frames. ], batch size: 36, lr: 3.58e-03, grad_scale: 32.0 2023-04-27 04:39:42,026 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=72894.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:40:08,511 INFO [finetune.py:976] (4/7) Epoch 13, batch 4200, loss[loss=0.1812, simple_loss=0.2596, pruned_loss=0.05139, over 4832.00 frames. ], tot_loss[loss=0.1854, simple_loss=0.2551, pruned_loss=0.05781, over 954927.36 frames. ], batch size: 39, lr: 3.58e-03, grad_scale: 32.0 2023-04-27 04:40:15,133 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.034e+02 1.607e+02 1.987e+02 2.411e+02 4.371e+02, threshold=3.975e+02, percent-clipped=4.0 2023-04-27 04:40:24,276 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=72955.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:40:25,073 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.93 vs. limit=2.0 2023-04-27 04:40:30,253 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=72964.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 04:40:41,145 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=72982.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:40:41,680 INFO [finetune.py:976] (4/7) Epoch 13, batch 4250, loss[loss=0.1842, simple_loss=0.2519, pruned_loss=0.05829, over 4895.00 frames. ], tot_loss[loss=0.1839, simple_loss=0.2526, pruned_loss=0.05758, over 955576.95 frames. ], batch size: 32, lr: 3.58e-03, grad_scale: 32.0 2023-04-27 04:40:58,904 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8498, 2.5792, 1.8338, 1.9157, 1.3442, 1.2909, 1.8021, 1.2532], device='cuda:4'), covar=tensor([0.1689, 0.1303, 0.1539, 0.1775, 0.2371, 0.1961, 0.1062, 0.2109], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0213, 0.0168, 0.0202, 0.0201, 0.0183, 0.0156, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 04:41:10,613 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6533, 1.6961, 0.6982, 1.3292, 1.8187, 1.5097, 1.3457, 1.4606], device='cuda:4'), covar=tensor([0.0503, 0.0377, 0.0368, 0.0568, 0.0248, 0.0548, 0.0511, 0.0579], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0025, 0.0023, 0.0029, 0.0020, 0.0029, 0.0028, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0044, 0.0038, 0.0050, 0.0037, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 04:41:10,639 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=73025.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 04:41:13,602 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=73030.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:41:15,356 INFO [finetune.py:976] (4/7) Epoch 13, batch 4300, loss[loss=0.2281, simple_loss=0.2941, pruned_loss=0.08108, over 4943.00 frames. ], tot_loss[loss=0.1825, simple_loss=0.2505, pruned_loss=0.05724, over 954781.48 frames. ], batch size: 38, lr: 3.58e-03, grad_scale: 32.0 2023-04-27 04:41:15,478 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5134, 1.6699, 1.4055, 1.1196, 1.2082, 1.1045, 1.3459, 1.1504], device='cuda:4'), covar=tensor([0.1796, 0.1317, 0.1531, 0.1859, 0.2348, 0.2059, 0.1113, 0.2060], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0214, 0.0169, 0.0203, 0.0201, 0.0184, 0.0157, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 04:41:16,670 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1161, 1.3480, 1.3034, 1.6567, 1.5295, 1.6218, 1.3153, 2.4090], device='cuda:4'), covar=tensor([0.0614, 0.0819, 0.0779, 0.1143, 0.0618, 0.0498, 0.0719, 0.0224], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0039, 0.0043, 0.0039, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0012, 0.0013, 0.0016], device='cuda:4') 2023-04-27 04:41:17,343 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7354, 1.3428, 1.8299, 2.2484, 1.8548, 1.7343, 1.7912, 1.7522], device='cuda:4'), covar=tensor([0.5186, 0.6996, 0.7151, 0.6500, 0.6629, 0.8364, 0.8391, 0.8879], device='cuda:4'), in_proj_covar=tensor([0.0412, 0.0407, 0.0496, 0.0512, 0.0442, 0.0463, 0.0469, 0.0473], device='cuda:4'), out_proj_covar=tensor([9.9864e-05, 1.0081e-04, 1.1158e-04, 1.2157e-04, 1.0672e-04, 1.1154e-04, 1.1204e-04, 1.1310e-04], device='cuda:4') 2023-04-27 04:41:22,019 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.082e+02 1.560e+02 1.991e+02 2.397e+02 3.754e+02, threshold=3.982e+02, percent-clipped=0.0 2023-04-27 04:41:28,421 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.90 vs. limit=2.0 2023-04-27 04:41:49,370 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7994, 1.3090, 1.8267, 2.3029, 1.9107, 1.7539, 1.8200, 1.7560], device='cuda:4'), covar=tensor([0.5470, 0.7553, 0.7565, 0.6863, 0.6738, 0.9359, 0.8747, 0.9291], device='cuda:4'), in_proj_covar=tensor([0.0413, 0.0408, 0.0497, 0.0513, 0.0443, 0.0464, 0.0470, 0.0475], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 04:41:59,412 INFO [finetune.py:976] (4/7) Epoch 13, batch 4350, loss[loss=0.185, simple_loss=0.248, pruned_loss=0.06105, over 4864.00 frames. ], tot_loss[loss=0.1813, simple_loss=0.2486, pruned_loss=0.05701, over 956243.31 frames. ], batch size: 31, lr: 3.58e-03, grad_scale: 32.0 2023-04-27 04:42:38,505 INFO [finetune.py:976] (4/7) Epoch 13, batch 4400, loss[loss=0.1693, simple_loss=0.244, pruned_loss=0.04726, over 4808.00 frames. ], tot_loss[loss=0.1824, simple_loss=0.2495, pruned_loss=0.05762, over 954569.20 frames. ], batch size: 45, lr: 3.58e-03, grad_scale: 32.0 2023-04-27 04:42:47,738 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.074e+02 1.646e+02 2.036e+02 2.477e+02 5.410e+02, threshold=4.072e+02, percent-clipped=5.0 2023-04-27 04:43:10,046 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0701, 1.7108, 2.1802, 2.5656, 2.2469, 1.9815, 2.0373, 1.9976], device='cuda:4'), covar=tensor([0.5057, 0.6721, 0.7546, 0.6395, 0.6379, 0.8515, 0.8663, 0.8432], device='cuda:4'), in_proj_covar=tensor([0.0412, 0.0407, 0.0495, 0.0512, 0.0442, 0.0463, 0.0468, 0.0474], device='cuda:4'), out_proj_covar=tensor([9.9771e-05, 1.0080e-04, 1.1135e-04, 1.2158e-04, 1.0673e-04, 1.1145e-04, 1.1195e-04, 1.1334e-04], device='cuda:4') 2023-04-27 04:43:38,172 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6822, 1.9127, 1.1285, 1.4409, 2.2147, 1.5179, 1.4900, 1.5495], device='cuda:4'), covar=tensor([0.0490, 0.0343, 0.0298, 0.0522, 0.0222, 0.0503, 0.0462, 0.0539], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0025, 0.0023, 0.0029, 0.0020, 0.0029, 0.0028, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0044, 0.0038, 0.0050, 0.0037, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 04:43:39,920 INFO [finetune.py:976] (4/7) Epoch 13, batch 4450, loss[loss=0.2029, simple_loss=0.258, pruned_loss=0.07396, over 4711.00 frames. ], tot_loss[loss=0.1854, simple_loss=0.2533, pruned_loss=0.05871, over 954320.23 frames. ], batch size: 23, lr: 3.58e-03, grad_scale: 32.0 2023-04-27 04:44:30,376 INFO [finetune.py:976] (4/7) Epoch 13, batch 4500, loss[loss=0.1867, simple_loss=0.2698, pruned_loss=0.05178, over 4847.00 frames. ], tot_loss[loss=0.1856, simple_loss=0.254, pruned_loss=0.05858, over 953824.46 frames. ], batch size: 44, lr: 3.58e-03, grad_scale: 32.0 2023-04-27 04:44:37,121 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.883e+01 1.575e+02 1.982e+02 2.231e+02 3.715e+02, threshold=3.964e+02, percent-clipped=0.0 2023-04-27 04:44:40,828 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=73250.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:44:44,883 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.49 vs. limit=2.0 2023-04-27 04:44:48,655 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7245, 2.0970, 0.8576, 1.1558, 1.5248, 1.0507, 2.2839, 1.2476], device='cuda:4'), covar=tensor([0.0662, 0.0926, 0.0649, 0.0984, 0.0424, 0.0891, 0.0348, 0.0606], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0049, 0.0047, 0.0051, 0.0052, 0.0076, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 04:44:57,087 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2660, 2.3609, 1.6064, 1.9764, 2.3849, 2.0452, 1.9885, 2.0633], device='cuda:4'), covar=tensor([0.0422, 0.0328, 0.0293, 0.0499, 0.0228, 0.0469, 0.0491, 0.0492], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0025, 0.0023, 0.0029, 0.0020, 0.0029, 0.0028, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0044, 0.0037, 0.0050, 0.0037, 0.0048, 0.0048, 0.0050], device='cuda:4') 2023-04-27 04:45:04,242 INFO [finetune.py:976] (4/7) Epoch 13, batch 4550, loss[loss=0.1833, simple_loss=0.256, pruned_loss=0.05527, over 4826.00 frames. ], tot_loss[loss=0.187, simple_loss=0.2558, pruned_loss=0.05912, over 953751.08 frames. ], batch size: 39, lr: 3.58e-03, grad_scale: 32.0 2023-04-27 04:45:10,364 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=73293.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:45:24,827 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0533, 1.0324, 1.2051, 1.1502, 0.9380, 0.8814, 1.0401, 0.7980], device='cuda:4'), covar=tensor([0.0571, 0.0605, 0.0490, 0.0517, 0.0709, 0.1247, 0.0446, 0.0613], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0070, 0.0070, 0.0068, 0.0075, 0.0096, 0.0075, 0.0069], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 04:45:28,146 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=73320.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 04:45:37,743 INFO [finetune.py:976] (4/7) Epoch 13, batch 4600, loss[loss=0.1786, simple_loss=0.2478, pruned_loss=0.05468, over 4908.00 frames. ], tot_loss[loss=0.1867, simple_loss=0.2555, pruned_loss=0.05899, over 954496.46 frames. ], batch size: 37, lr: 3.58e-03, grad_scale: 32.0 2023-04-27 04:45:42,485 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.30 vs. limit=5.0 2023-04-27 04:45:44,461 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.205e+02 1.640e+02 1.996e+02 2.306e+02 3.998e+02, threshold=3.993e+02, percent-clipped=1.0 2023-04-27 04:45:50,705 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=73354.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:46:10,622 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3985, 2.9726, 0.9005, 1.6980, 2.2724, 1.4056, 4.0926, 1.9829], device='cuda:4'), covar=tensor([0.0622, 0.0850, 0.0893, 0.1220, 0.0554, 0.0981, 0.0287, 0.0614], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0048, 0.0046, 0.0050, 0.0051, 0.0075, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0007, 0.0008, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-27 04:46:11,138 INFO [finetune.py:976] (4/7) Epoch 13, batch 4650, loss[loss=0.1839, simple_loss=0.2519, pruned_loss=0.05795, over 4864.00 frames. ], tot_loss[loss=0.1851, simple_loss=0.253, pruned_loss=0.05855, over 955797.69 frames. ], batch size: 34, lr: 3.58e-03, grad_scale: 32.0 2023-04-27 04:46:15,363 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8568, 1.8545, 1.1384, 1.5388, 1.7907, 1.6610, 1.5808, 1.6450], device='cuda:4'), covar=tensor([0.0494, 0.0363, 0.0351, 0.0537, 0.0284, 0.0506, 0.0493, 0.0572], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0025, 0.0023, 0.0029, 0.0020, 0.0029, 0.0028, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0044, 0.0037, 0.0051, 0.0037, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 04:46:44,503 INFO [finetune.py:976] (4/7) Epoch 13, batch 4700, loss[loss=0.1498, simple_loss=0.2228, pruned_loss=0.03843, over 4819.00 frames. ], tot_loss[loss=0.1821, simple_loss=0.2497, pruned_loss=0.05731, over 955510.77 frames. ], batch size: 41, lr: 3.58e-03, grad_scale: 32.0 2023-04-27 04:46:51,528 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.168e+02 1.646e+02 2.000e+02 2.343e+02 4.660e+02, threshold=4.000e+02, percent-clipped=2.0 2023-04-27 04:47:29,931 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9254, 1.6557, 2.1700, 2.4388, 2.0510, 1.8324, 2.0394, 2.0048], device='cuda:4'), covar=tensor([0.5444, 0.7926, 0.8025, 0.6726, 0.6792, 1.0106, 1.0322, 1.0055], device='cuda:4'), in_proj_covar=tensor([0.0410, 0.0407, 0.0493, 0.0511, 0.0442, 0.0463, 0.0469, 0.0473], device='cuda:4'), out_proj_covar=tensor([9.9406e-05, 1.0078e-04, 1.1117e-04, 1.2138e-04, 1.0661e-04, 1.1142e-04, 1.1206e-04, 1.1298e-04], device='cuda:4') 2023-04-27 04:47:33,204 INFO [finetune.py:976] (4/7) Epoch 13, batch 4750, loss[loss=0.1333, simple_loss=0.2102, pruned_loss=0.02821, over 4857.00 frames. ], tot_loss[loss=0.1803, simple_loss=0.2477, pruned_loss=0.05645, over 955567.44 frames. ], batch size: 49, lr: 3.58e-03, grad_scale: 32.0 2023-04-27 04:47:42,217 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0085, 2.4619, 1.0932, 1.3424, 1.9824, 1.0907, 3.2081, 1.7057], device='cuda:4'), covar=tensor([0.0695, 0.0752, 0.0796, 0.1267, 0.0499, 0.1081, 0.0253, 0.0668], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0049, 0.0047, 0.0050, 0.0052, 0.0076, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0007, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 04:48:13,156 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4029, 1.2770, 1.6048, 1.5749, 1.2998, 1.1746, 1.2870, 0.8358], device='cuda:4'), covar=tensor([0.0599, 0.0796, 0.0530, 0.0714, 0.0808, 0.1263, 0.0660, 0.0709], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0070, 0.0070, 0.0067, 0.0075, 0.0096, 0.0075, 0.0069], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 04:48:28,608 INFO [finetune.py:976] (4/7) Epoch 13, batch 4800, loss[loss=0.1343, simple_loss=0.2061, pruned_loss=0.03125, over 4777.00 frames. ], tot_loss[loss=0.1819, simple_loss=0.2493, pruned_loss=0.05723, over 953007.53 frames. ], batch size: 28, lr: 3.58e-03, grad_scale: 32.0 2023-04-27 04:48:36,791 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.040e+02 1.600e+02 1.891e+02 2.207e+02 3.425e+02, threshold=3.783e+02, percent-clipped=0.0 2023-04-27 04:48:41,061 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=73550.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:48:49,669 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.9305, 2.2547, 2.1540, 2.2665, 2.0426, 2.1657, 2.2539, 2.1643], device='cuda:4'), covar=tensor([0.4564, 0.6570, 0.5482, 0.5432, 0.6513, 0.8317, 0.6652, 0.6067], device='cuda:4'), in_proj_covar=tensor([0.0326, 0.0371, 0.0313, 0.0326, 0.0337, 0.0393, 0.0353, 0.0320], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 04:49:02,205 INFO [finetune.py:976] (4/7) Epoch 13, batch 4850, loss[loss=0.1839, simple_loss=0.2593, pruned_loss=0.0542, over 4810.00 frames. ], tot_loss[loss=0.1837, simple_loss=0.2522, pruned_loss=0.05757, over 953209.41 frames. ], batch size: 51, lr: 3.58e-03, grad_scale: 32.0 2023-04-27 04:49:17,648 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=73590.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:49:28,459 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=73598.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:49:53,457 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=73620.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 04:50:13,624 INFO [finetune.py:976] (4/7) Epoch 13, batch 4900, loss[loss=0.2034, simple_loss=0.268, pruned_loss=0.06941, over 4895.00 frames. ], tot_loss[loss=0.1864, simple_loss=0.255, pruned_loss=0.05892, over 950534.79 frames. ], batch size: 35, lr: 3.58e-03, grad_scale: 32.0 2023-04-27 04:50:28,449 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.161e+02 1.796e+02 2.177e+02 2.620e+02 4.604e+02, threshold=4.354e+02, percent-clipped=4.0 2023-04-27 04:50:38,436 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=73649.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:50:39,740 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=73651.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:50:55,037 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=73668.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 04:51:04,150 INFO [finetune.py:976] (4/7) Epoch 13, batch 4950, loss[loss=0.2367, simple_loss=0.3089, pruned_loss=0.08226, over 4702.00 frames. ], tot_loss[loss=0.1869, simple_loss=0.2558, pruned_loss=0.05904, over 953143.46 frames. ], batch size: 59, lr: 3.58e-03, grad_scale: 32.0 2023-04-27 04:51:19,615 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6482, 1.5666, 0.7215, 1.3159, 1.8613, 1.4917, 1.3782, 1.4100], device='cuda:4'), covar=tensor([0.0521, 0.0391, 0.0380, 0.0589, 0.0284, 0.0538, 0.0535, 0.0609], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0025, 0.0023, 0.0029, 0.0020, 0.0029, 0.0028, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0045, 0.0038, 0.0051, 0.0037, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 04:51:28,283 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.72 vs. limit=2.0 2023-04-27 04:51:36,599 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=73732.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:51:37,096 INFO [finetune.py:976] (4/7) Epoch 13, batch 5000, loss[loss=0.1622, simple_loss=0.2353, pruned_loss=0.04458, over 4896.00 frames. ], tot_loss[loss=0.1849, simple_loss=0.2535, pruned_loss=0.05814, over 951914.47 frames. ], batch size: 36, lr: 3.58e-03, grad_scale: 32.0 2023-04-27 04:51:45,206 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.282e+02 1.739e+02 2.081e+02 2.360e+02 4.914e+02, threshold=4.162e+02, percent-clipped=1.0 2023-04-27 04:52:01,411 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1223, 2.6799, 0.9817, 1.4902, 1.9322, 1.3083, 3.4553, 1.9186], device='cuda:4'), covar=tensor([0.0596, 0.0745, 0.0915, 0.1149, 0.0524, 0.0948, 0.0190, 0.0585], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0065, 0.0048, 0.0046, 0.0050, 0.0051, 0.0075, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0007, 0.0008, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-27 04:52:02,983 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.75 vs. limit=5.0 2023-04-27 04:52:11,152 INFO [finetune.py:976] (4/7) Epoch 13, batch 5050, loss[loss=0.1828, simple_loss=0.2427, pruned_loss=0.06142, over 4756.00 frames. ], tot_loss[loss=0.1822, simple_loss=0.2499, pruned_loss=0.05727, over 951779.43 frames. ], batch size: 23, lr: 3.58e-03, grad_scale: 32.0 2023-04-27 04:52:17,789 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=73793.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:52:56,562 INFO [finetune.py:976] (4/7) Epoch 13, batch 5100, loss[loss=0.1384, simple_loss=0.2096, pruned_loss=0.03362, over 4760.00 frames. ], tot_loss[loss=0.1792, simple_loss=0.2469, pruned_loss=0.05577, over 954046.21 frames. ], batch size: 23, lr: 3.57e-03, grad_scale: 32.0 2023-04-27 04:53:03,638 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.084e+02 1.578e+02 1.847e+02 2.242e+02 3.779e+02, threshold=3.694e+02, percent-clipped=0.0 2023-04-27 04:53:44,068 INFO [finetune.py:976] (4/7) Epoch 13, batch 5150, loss[loss=0.2106, simple_loss=0.2668, pruned_loss=0.07721, over 4866.00 frames. ], tot_loss[loss=0.1792, simple_loss=0.2468, pruned_loss=0.05581, over 954110.29 frames. ], batch size: 31, lr: 3.57e-03, grad_scale: 32.0 2023-04-27 04:53:54,611 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.59 vs. limit=2.0 2023-04-27 04:54:18,052 INFO [finetune.py:976] (4/7) Epoch 13, batch 5200, loss[loss=0.2085, simple_loss=0.2834, pruned_loss=0.06674, over 4854.00 frames. ], tot_loss[loss=0.1829, simple_loss=0.2503, pruned_loss=0.05776, over 951809.74 frames. ], batch size: 44, lr: 3.57e-03, grad_scale: 32.0 2023-04-27 04:54:24,748 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.235e+02 1.597e+02 1.940e+02 2.333e+02 4.411e+02, threshold=3.880e+02, percent-clipped=2.0 2023-04-27 04:54:26,022 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=73946.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:54:28,351 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=73949.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:55:07,561 INFO [finetune.py:976] (4/7) Epoch 13, batch 5250, loss[loss=0.1091, simple_loss=0.1652, pruned_loss=0.02651, over 4134.00 frames. ], tot_loss[loss=0.1852, simple_loss=0.2531, pruned_loss=0.05864, over 949564.40 frames. ], batch size: 17, lr: 3.57e-03, grad_scale: 32.0 2023-04-27 04:55:16,181 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=73997.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:55:16,890 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4471, 1.7904, 1.7836, 1.8909, 1.7143, 1.8245, 1.8909, 1.7827], device='cuda:4'), covar=tensor([0.3936, 0.6037, 0.5210, 0.4823, 0.5912, 0.8105, 0.6364, 0.5897], device='cuda:4'), in_proj_covar=tensor([0.0329, 0.0375, 0.0316, 0.0328, 0.0340, 0.0397, 0.0355, 0.0323], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 04:55:42,313 INFO [finetune.py:976] (4/7) Epoch 13, batch 5300, loss[loss=0.2222, simple_loss=0.2784, pruned_loss=0.083, over 4737.00 frames. ], tot_loss[loss=0.1869, simple_loss=0.255, pruned_loss=0.05943, over 952062.97 frames. ], batch size: 54, lr: 3.57e-03, grad_scale: 64.0 2023-04-27 04:55:54,437 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.148e+02 1.549e+02 1.803e+02 2.231e+02 3.853e+02, threshold=3.607e+02, percent-clipped=0.0 2023-04-27 04:56:19,410 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=74064.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:56:36,836 INFO [finetune.py:976] (4/7) Epoch 13, batch 5350, loss[loss=0.1911, simple_loss=0.2558, pruned_loss=0.06321, over 4921.00 frames. ], tot_loss[loss=0.1864, simple_loss=0.2554, pruned_loss=0.05871, over 952670.80 frames. ], batch size: 33, lr: 3.57e-03, grad_scale: 64.0 2023-04-27 04:56:39,880 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=74088.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:56:46,556 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2042, 1.6279, 1.5762, 1.9773, 1.8670, 2.0325, 1.6165, 3.9486], device='cuda:4'), covar=tensor([0.0564, 0.0827, 0.0776, 0.1108, 0.0621, 0.0531, 0.0726, 0.0147], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0044, 0.0040, 0.0038, 0.0038, 0.0057], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 04:56:47,993 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.04 vs. limit=5.0 2023-04-27 04:57:05,261 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=74125.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:57:10,417 INFO [finetune.py:976] (4/7) Epoch 13, batch 5400, loss[loss=0.2082, simple_loss=0.2712, pruned_loss=0.07259, over 4907.00 frames. ], tot_loss[loss=0.1842, simple_loss=0.253, pruned_loss=0.05771, over 951616.04 frames. ], batch size: 37, lr: 3.57e-03, grad_scale: 64.0 2023-04-27 04:57:17,121 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.047e+02 1.609e+02 1.962e+02 2.381e+02 6.341e+02, threshold=3.923e+02, percent-clipped=2.0 2023-04-27 04:57:43,731 INFO [finetune.py:976] (4/7) Epoch 13, batch 5450, loss[loss=0.155, simple_loss=0.2183, pruned_loss=0.04582, over 4903.00 frames. ], tot_loss[loss=0.1813, simple_loss=0.2495, pruned_loss=0.05655, over 951474.74 frames. ], batch size: 35, lr: 3.57e-03, grad_scale: 64.0 2023-04-27 04:58:35,027 INFO [finetune.py:976] (4/7) Epoch 13, batch 5500, loss[loss=0.2113, simple_loss=0.2668, pruned_loss=0.07792, over 4099.00 frames. ], tot_loss[loss=0.1797, simple_loss=0.2476, pruned_loss=0.05587, over 952289.26 frames. ], batch size: 65, lr: 3.57e-03, grad_scale: 64.0 2023-04-27 04:58:39,427 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=74240.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:58:42,297 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.038e+02 1.578e+02 1.994e+02 2.336e+02 4.147e+02, threshold=3.989e+02, percent-clipped=1.0 2023-04-27 04:58:43,610 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=74246.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:59:08,625 INFO [finetune.py:976] (4/7) Epoch 13, batch 5550, loss[loss=0.2021, simple_loss=0.2751, pruned_loss=0.06453, over 4832.00 frames. ], tot_loss[loss=0.182, simple_loss=0.25, pruned_loss=0.05699, over 954139.21 frames. ], batch size: 47, lr: 3.57e-03, grad_scale: 64.0 2023-04-27 04:59:15,399 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=74294.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:59:20,227 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=74301.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 04:59:39,644 INFO [finetune.py:976] (4/7) Epoch 13, batch 5600, loss[loss=0.1995, simple_loss=0.2612, pruned_loss=0.0689, over 4185.00 frames. ], tot_loss[loss=0.184, simple_loss=0.2532, pruned_loss=0.05743, over 952224.60 frames. ], batch size: 65, lr: 3.57e-03, grad_scale: 64.0 2023-04-27 04:59:46,069 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.158e+02 1.613e+02 1.797e+02 2.128e+02 3.644e+02, threshold=3.594e+02, percent-clipped=1.0 2023-04-27 05:00:16,160 INFO [finetune.py:976] (4/7) Epoch 13, batch 5650, loss[loss=0.2456, simple_loss=0.2989, pruned_loss=0.09609, over 4884.00 frames. ], tot_loss[loss=0.1847, simple_loss=0.2548, pruned_loss=0.05727, over 953951.02 frames. ], batch size: 32, lr: 3.57e-03, grad_scale: 64.0 2023-04-27 05:00:25,009 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=74388.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:00:44,667 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=74420.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:00:47,738 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-04-27 05:00:51,208 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.2565, 3.2449, 2.5491, 3.8414, 3.3992, 3.3148, 1.5854, 3.2240], device='cuda:4'), covar=tensor([0.2262, 0.1430, 0.3156, 0.2218, 0.3537, 0.2114, 0.6413, 0.2961], device='cuda:4'), in_proj_covar=tensor([0.0243, 0.0214, 0.0248, 0.0302, 0.0296, 0.0245, 0.0270, 0.0272], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 05:00:52,322 INFO [finetune.py:976] (4/7) Epoch 13, batch 5700, loss[loss=0.1562, simple_loss=0.2126, pruned_loss=0.04996, over 3957.00 frames. ], tot_loss[loss=0.1822, simple_loss=0.2512, pruned_loss=0.05661, over 937346.31 frames. ], batch size: 17, lr: 3.57e-03, grad_scale: 64.0 2023-04-27 05:00:54,134 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=74436.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:00:58,756 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.054e+02 1.564e+02 1.898e+02 2.359e+02 3.680e+02, threshold=3.797e+02, percent-clipped=1.0 2023-04-27 05:01:23,646 INFO [finetune.py:976] (4/7) Epoch 14, batch 0, loss[loss=0.1752, simple_loss=0.2547, pruned_loss=0.04784, over 4855.00 frames. ], tot_loss[loss=0.1752, simple_loss=0.2547, pruned_loss=0.04784, over 4855.00 frames. ], batch size: 31, lr: 3.57e-03, grad_scale: 64.0 2023-04-27 05:01:23,646 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 05:01:45,229 INFO [finetune.py:1010] (4/7) Epoch 14, validation: loss=0.1535, simple_loss=0.226, pruned_loss=0.04054, over 2265189.00 frames. 2023-04-27 05:01:45,230 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-27 05:01:57,204 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4625, 1.9021, 1.8856, 2.2889, 2.1408, 2.3461, 1.7577, 4.7163], device='cuda:4'), covar=tensor([0.0588, 0.0769, 0.0825, 0.1182, 0.0636, 0.0465, 0.0741, 0.0170], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0043, 0.0040, 0.0038, 0.0038, 0.0057], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0012, 0.0013, 0.0016], device='cuda:4') 2023-04-27 05:02:18,140 INFO [finetune.py:976] (4/7) Epoch 14, batch 50, loss[loss=0.1485, simple_loss=0.2227, pruned_loss=0.03717, over 4841.00 frames. ], tot_loss[loss=0.1904, simple_loss=0.2572, pruned_loss=0.06177, over 216655.26 frames. ], batch size: 49, lr: 3.57e-03, grad_scale: 32.0 2023-04-27 05:02:41,160 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.646e+01 1.673e+02 1.901e+02 2.437e+02 4.267e+02, threshold=3.802e+02, percent-clipped=2.0 2023-04-27 05:02:51,798 INFO [finetune.py:976] (4/7) Epoch 14, batch 100, loss[loss=0.1473, simple_loss=0.2144, pruned_loss=0.04007, over 4179.00 frames. ], tot_loss[loss=0.1822, simple_loss=0.2494, pruned_loss=0.05755, over 380579.74 frames. ], batch size: 65, lr: 3.57e-03, grad_scale: 32.0 2023-04-27 05:02:55,995 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.82 vs. limit=2.0 2023-04-27 05:02:56,519 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2177, 2.5602, 0.8428, 1.4428, 1.5826, 1.9326, 1.6548, 0.8441], device='cuda:4'), covar=tensor([0.1472, 0.1200, 0.1835, 0.1351, 0.1149, 0.0990, 0.1573, 0.1721], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0244, 0.0137, 0.0120, 0.0132, 0.0151, 0.0117, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 05:03:29,434 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=74595.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:03:30,510 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=74596.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:03:43,825 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-04-27 05:03:51,942 INFO [finetune.py:976] (4/7) Epoch 14, batch 150, loss[loss=0.2054, simple_loss=0.2603, pruned_loss=0.07529, over 4800.00 frames. ], tot_loss[loss=0.1802, simple_loss=0.2453, pruned_loss=0.05751, over 505680.86 frames. ], batch size: 51, lr: 3.57e-03, grad_scale: 32.0 2023-04-27 05:04:20,051 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.093e+02 1.548e+02 1.852e+02 2.253e+02 4.630e+02, threshold=3.704e+02, percent-clipped=3.0 2023-04-27 05:04:27,975 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=74656.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:04:31,410 INFO [finetune.py:976] (4/7) Epoch 14, batch 200, loss[loss=0.1514, simple_loss=0.2289, pruned_loss=0.03697, over 4867.00 frames. ], tot_loss[loss=0.18, simple_loss=0.2462, pruned_loss=0.05687, over 605835.92 frames. ], batch size: 34, lr: 3.57e-03, grad_scale: 32.0 2023-04-27 05:05:05,316 INFO [finetune.py:976] (4/7) Epoch 14, batch 250, loss[loss=0.1404, simple_loss=0.2066, pruned_loss=0.03707, over 4708.00 frames. ], tot_loss[loss=0.1843, simple_loss=0.2512, pruned_loss=0.05871, over 683493.73 frames. ], batch size: 23, lr: 3.57e-03, grad_scale: 32.0 2023-04-27 05:05:11,829 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=74720.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:05:39,269 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.695e+01 1.601e+02 1.840e+02 2.318e+02 4.456e+02, threshold=3.680e+02, percent-clipped=1.0 2023-04-27 05:05:40,583 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3366, 2.8988, 0.8413, 1.5575, 2.0561, 1.4939, 4.0410, 1.9339], device='cuda:4'), covar=tensor([0.0682, 0.0912, 0.0929, 0.1337, 0.0613, 0.1029, 0.0301, 0.0699], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0049, 0.0047, 0.0050, 0.0052, 0.0075, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 05:05:49,927 INFO [finetune.py:976] (4/7) Epoch 14, batch 300, loss[loss=0.2131, simple_loss=0.2844, pruned_loss=0.07091, over 4910.00 frames. ], tot_loss[loss=0.1855, simple_loss=0.2535, pruned_loss=0.05873, over 743817.17 frames. ], batch size: 36, lr: 3.56e-03, grad_scale: 32.0 2023-04-27 05:05:54,766 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=74768.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:06:10,155 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0740, 1.9586, 1.7955, 1.7466, 2.1857, 1.7371, 2.6927, 1.5289], device='cuda:4'), covar=tensor([0.3878, 0.2007, 0.4680, 0.3470, 0.1766, 0.2725, 0.1559, 0.4533], device='cuda:4'), in_proj_covar=tensor([0.0341, 0.0343, 0.0425, 0.0354, 0.0378, 0.0380, 0.0371, 0.0417], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 05:06:16,383 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.3694, 1.2950, 1.4207, 0.9611, 1.2496, 1.1226, 1.7486, 1.2217], device='cuda:4'), covar=tensor([0.3649, 0.1814, 0.5347, 0.2776, 0.1664, 0.2395, 0.1632, 0.5257], device='cuda:4'), in_proj_covar=tensor([0.0341, 0.0344, 0.0426, 0.0355, 0.0379, 0.0381, 0.0371, 0.0418], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 05:06:23,251 INFO [finetune.py:976] (4/7) Epoch 14, batch 350, loss[loss=0.1888, simple_loss=0.2544, pruned_loss=0.06163, over 4868.00 frames. ], tot_loss[loss=0.1884, simple_loss=0.2564, pruned_loss=0.06019, over 790514.30 frames. ], batch size: 31, lr: 3.56e-03, grad_scale: 32.0 2023-04-27 05:07:09,534 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.209e+02 1.682e+02 1.913e+02 2.290e+02 4.465e+02, threshold=3.826e+02, percent-clipped=1.0 2023-04-27 05:07:25,449 INFO [finetune.py:976] (4/7) Epoch 14, batch 400, loss[loss=0.1956, simple_loss=0.2691, pruned_loss=0.06104, over 4163.00 frames. ], tot_loss[loss=0.1884, simple_loss=0.2567, pruned_loss=0.06002, over 825362.72 frames. ], batch size: 65, lr: 3.56e-03, grad_scale: 32.0 2023-04-27 05:07:49,800 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=74896.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:07:57,106 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0763, 2.4982, 1.0348, 1.3741, 2.0083, 1.3014, 3.3766, 1.7685], device='cuda:4'), covar=tensor([0.0639, 0.0657, 0.0770, 0.1276, 0.0502, 0.0974, 0.0251, 0.0600], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0067, 0.0049, 0.0047, 0.0051, 0.0052, 0.0076, 0.0052], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 05:07:57,736 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=74909.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:07:59,305 INFO [finetune.py:976] (4/7) Epoch 14, batch 450, loss[loss=0.1984, simple_loss=0.2678, pruned_loss=0.06453, over 4803.00 frames. ], tot_loss[loss=0.1859, simple_loss=0.2544, pruned_loss=0.05869, over 854359.01 frames. ], batch size: 41, lr: 3.56e-03, grad_scale: 32.0 2023-04-27 05:08:37,890 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=74944.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:08:38,409 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.611e+01 1.647e+02 1.819e+02 2.107e+02 4.773e+02, threshold=3.638e+02, percent-clipped=2.0 2023-04-27 05:08:39,247 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-27 05:08:42,176 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=74951.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:08:42,231 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.3665, 1.2946, 1.4264, 0.9850, 1.3814, 1.1983, 1.6947, 1.3141], device='cuda:4'), covar=tensor([0.4180, 0.1952, 0.5702, 0.2922, 0.1658, 0.2422, 0.1737, 0.5157], device='cuda:4'), in_proj_covar=tensor([0.0344, 0.0345, 0.0428, 0.0357, 0.0381, 0.0383, 0.0372, 0.0419], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 05:08:53,659 INFO [finetune.py:976] (4/7) Epoch 14, batch 500, loss[loss=0.1416, simple_loss=0.2124, pruned_loss=0.03541, over 4834.00 frames. ], tot_loss[loss=0.1838, simple_loss=0.2517, pruned_loss=0.05795, over 878547.45 frames. ], batch size: 47, lr: 3.56e-03, grad_scale: 32.0 2023-04-27 05:08:56,094 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=74964.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:08:59,784 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=74970.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:09:04,452 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6799, 2.0219, 1.6502, 1.3978, 1.2372, 1.2372, 1.7213, 1.1469], device='cuda:4'), covar=tensor([0.1778, 0.1395, 0.1695, 0.2080, 0.2571, 0.2202, 0.1139, 0.2381], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0212, 0.0168, 0.0203, 0.0201, 0.0183, 0.0156, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 05:09:27,830 INFO [finetune.py:976] (4/7) Epoch 14, batch 550, loss[loss=0.1767, simple_loss=0.249, pruned_loss=0.05218, over 4807.00 frames. ], tot_loss[loss=0.1821, simple_loss=0.2496, pruned_loss=0.05731, over 896669.79 frames. ], batch size: 51, lr: 3.56e-03, grad_scale: 32.0 2023-04-27 05:09:30,405 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-27 05:09:37,444 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=75025.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:09:51,447 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.023e+02 1.675e+02 1.910e+02 2.434e+02 4.302e+02, threshold=3.821e+02, percent-clipped=4.0 2023-04-27 05:10:01,744 INFO [finetune.py:976] (4/7) Epoch 14, batch 600, loss[loss=0.1596, simple_loss=0.2265, pruned_loss=0.04634, over 4909.00 frames. ], tot_loss[loss=0.183, simple_loss=0.2506, pruned_loss=0.05772, over 909582.44 frames. ], batch size: 32, lr: 3.56e-03, grad_scale: 32.0 2023-04-27 05:10:35,754 INFO [finetune.py:976] (4/7) Epoch 14, batch 650, loss[loss=0.1486, simple_loss=0.2121, pruned_loss=0.04259, over 4714.00 frames. ], tot_loss[loss=0.1846, simple_loss=0.2531, pruned_loss=0.05809, over 920315.37 frames. ], batch size: 23, lr: 3.56e-03, grad_scale: 32.0 2023-04-27 05:10:54,111 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2907, 2.0228, 2.4703, 2.7031, 2.8569, 2.1670, 1.8109, 2.2118], device='cuda:4'), covar=tensor([0.0898, 0.1092, 0.0566, 0.0655, 0.0572, 0.0845, 0.0817, 0.0656], device='cuda:4'), in_proj_covar=tensor([0.0191, 0.0203, 0.0183, 0.0173, 0.0179, 0.0184, 0.0156, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 05:10:59,292 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.138e+02 1.654e+02 2.108e+02 2.537e+02 4.925e+02, threshold=4.216e+02, percent-clipped=6.0 2023-04-27 05:11:09,460 INFO [finetune.py:976] (4/7) Epoch 14, batch 700, loss[loss=0.1667, simple_loss=0.2457, pruned_loss=0.04386, over 4877.00 frames. ], tot_loss[loss=0.1852, simple_loss=0.2547, pruned_loss=0.05782, over 927647.88 frames. ], batch size: 32, lr: 3.56e-03, grad_scale: 32.0 2023-04-27 05:11:10,789 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1710, 2.4490, 0.7956, 1.4529, 1.5560, 1.7967, 1.6542, 0.8047], device='cuda:4'), covar=tensor([0.1346, 0.1388, 0.1679, 0.1230, 0.1085, 0.0930, 0.1383, 0.1516], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0244, 0.0137, 0.0120, 0.0132, 0.0151, 0.0117, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 05:11:46,822 INFO [finetune.py:976] (4/7) Epoch 14, batch 750, loss[loss=0.1997, simple_loss=0.2691, pruned_loss=0.06517, over 4799.00 frames. ], tot_loss[loss=0.1854, simple_loss=0.2551, pruned_loss=0.0578, over 934965.64 frames. ], batch size: 40, lr: 3.56e-03, grad_scale: 32.0 2023-04-27 05:12:09,250 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=75228.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:12:13,670 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5289, 0.6436, 1.4677, 1.9206, 1.6479, 1.4081, 1.4352, 1.4778], device='cuda:4'), covar=tensor([0.4499, 0.7056, 0.5721, 0.6131, 0.5481, 0.7308, 0.7312, 0.8239], device='cuda:4'), in_proj_covar=tensor([0.0415, 0.0410, 0.0497, 0.0514, 0.0445, 0.0467, 0.0472, 0.0477], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 05:12:31,822 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.094e+02 1.600e+02 1.912e+02 2.289e+02 3.679e+02, threshold=3.824e+02, percent-clipped=0.0 2023-04-27 05:12:35,601 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=75251.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:12:41,570 INFO [finetune.py:976] (4/7) Epoch 14, batch 800, loss[loss=0.1704, simple_loss=0.246, pruned_loss=0.04739, over 4823.00 frames. ], tot_loss[loss=0.1845, simple_loss=0.2544, pruned_loss=0.05732, over 940635.64 frames. ], batch size: 39, lr: 3.56e-03, grad_scale: 32.0 2023-04-27 05:12:44,560 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=75265.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:12:47,617 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=75270.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:12:57,716 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9643, 1.9891, 1.8691, 1.7868, 2.1828, 1.8832, 2.7381, 1.5235], device='cuda:4'), covar=tensor([0.3719, 0.1808, 0.4899, 0.2868, 0.1736, 0.2423, 0.1271, 0.4980], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0345, 0.0424, 0.0353, 0.0379, 0.0378, 0.0369, 0.0416], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 05:13:00,650 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=75289.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:13:07,657 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=75299.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:13:14,865 INFO [finetune.py:976] (4/7) Epoch 14, batch 850, loss[loss=0.2028, simple_loss=0.2626, pruned_loss=0.07151, over 4899.00 frames. ], tot_loss[loss=0.1841, simple_loss=0.253, pruned_loss=0.05759, over 943964.51 frames. ], batch size: 36, lr: 3.56e-03, grad_scale: 32.0 2023-04-27 05:13:26,908 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=75320.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:13:40,183 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=75331.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 05:13:56,949 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.246e+02 1.618e+02 2.006e+02 2.388e+02 8.621e+02, threshold=4.012e+02, percent-clipped=6.0 2023-04-27 05:13:58,543 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.13 vs. limit=2.0 2023-04-27 05:14:12,834 INFO [finetune.py:976] (4/7) Epoch 14, batch 900, loss[loss=0.1684, simple_loss=0.2401, pruned_loss=0.04838, over 4822.00 frames. ], tot_loss[loss=0.1805, simple_loss=0.2492, pruned_loss=0.05591, over 947589.39 frames. ], batch size: 30, lr: 3.56e-03, grad_scale: 32.0 2023-04-27 05:14:57,517 INFO [finetune.py:976] (4/7) Epoch 14, batch 950, loss[loss=0.1761, simple_loss=0.2367, pruned_loss=0.05781, over 4053.00 frames. ], tot_loss[loss=0.1786, simple_loss=0.2472, pruned_loss=0.05497, over 949410.75 frames. ], batch size: 17, lr: 3.56e-03, grad_scale: 32.0 2023-04-27 05:15:02,565 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5886, 1.3118, 1.2560, 1.3935, 1.8763, 1.4714, 1.2455, 1.2173], device='cuda:4'), covar=tensor([0.1459, 0.1450, 0.1850, 0.1371, 0.0758, 0.1605, 0.1880, 0.2028], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0317, 0.0354, 0.0293, 0.0332, 0.0315, 0.0307, 0.0363], device='cuda:4'), out_proj_covar=tensor([6.4060e-05, 6.6676e-05, 7.5968e-05, 5.9990e-05, 6.9202e-05, 6.6675e-05, 6.4967e-05, 7.7640e-05], device='cuda:4') 2023-04-27 05:15:04,282 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.60 vs. limit=2.0 2023-04-27 05:15:20,051 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.068e+02 1.489e+02 1.870e+02 2.238e+02 3.708e+02, threshold=3.739e+02, percent-clipped=0.0 2023-04-27 05:15:30,329 INFO [finetune.py:976] (4/7) Epoch 14, batch 1000, loss[loss=0.1644, simple_loss=0.2442, pruned_loss=0.04231, over 4922.00 frames. ], tot_loss[loss=0.1811, simple_loss=0.2499, pruned_loss=0.05613, over 950860.37 frames. ], batch size: 37, lr: 3.56e-03, grad_scale: 32.0 2023-04-27 05:16:03,286 INFO [finetune.py:976] (4/7) Epoch 14, batch 1050, loss[loss=0.1898, simple_loss=0.2706, pruned_loss=0.05449, over 4933.00 frames. ], tot_loss[loss=0.1828, simple_loss=0.2526, pruned_loss=0.05652, over 954979.31 frames. ], batch size: 38, lr: 3.56e-03, grad_scale: 32.0 2023-04-27 05:16:25,348 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.198e+02 1.730e+02 1.979e+02 2.287e+02 8.214e+02, threshold=3.958e+02, percent-clipped=1.0 2023-04-27 05:16:36,970 INFO [finetune.py:976] (4/7) Epoch 14, batch 1100, loss[loss=0.1832, simple_loss=0.2459, pruned_loss=0.06028, over 4863.00 frames. ], tot_loss[loss=0.1833, simple_loss=0.2533, pruned_loss=0.05664, over 953763.46 frames. ], batch size: 34, lr: 3.56e-03, grad_scale: 32.0 2023-04-27 05:16:39,542 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=75565.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:16:57,394 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=75584.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:17:31,384 INFO [finetune.py:976] (4/7) Epoch 14, batch 1150, loss[loss=0.1636, simple_loss=0.2184, pruned_loss=0.05437, over 4315.00 frames. ], tot_loss[loss=0.1845, simple_loss=0.2543, pruned_loss=0.0574, over 951850.75 frames. ], batch size: 18, lr: 3.56e-03, grad_scale: 32.0 2023-04-27 05:17:32,672 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=75613.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:17:36,931 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=75620.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:17:40,633 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=75626.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 05:17:53,271 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.071e+02 1.637e+02 1.916e+02 2.260e+02 4.790e+02, threshold=3.833e+02, percent-clipped=2.0 2023-04-27 05:18:05,378 INFO [finetune.py:976] (4/7) Epoch 14, batch 1200, loss[loss=0.164, simple_loss=0.2365, pruned_loss=0.04574, over 4876.00 frames. ], tot_loss[loss=0.1821, simple_loss=0.2519, pruned_loss=0.05621, over 952647.41 frames. ], batch size: 34, lr: 3.56e-03, grad_scale: 32.0 2023-04-27 05:18:09,670 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=75668.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:18:24,671 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3505, 3.3347, 0.8251, 1.9068, 1.9814, 2.3742, 1.9797, 0.9596], device='cuda:4'), covar=tensor([0.1507, 0.0882, 0.1948, 0.1221, 0.1059, 0.1073, 0.1461, 0.2140], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0246, 0.0138, 0.0121, 0.0133, 0.0153, 0.0118, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 05:18:43,842 INFO [finetune.py:976] (4/7) Epoch 14, batch 1250, loss[loss=0.1635, simple_loss=0.2337, pruned_loss=0.04667, over 4894.00 frames. ], tot_loss[loss=0.1802, simple_loss=0.2491, pruned_loss=0.0556, over 952948.40 frames. ], batch size: 32, lr: 3.56e-03, grad_scale: 32.0 2023-04-27 05:19:24,353 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.269e+02 1.670e+02 2.009e+02 2.395e+02 6.270e+02, threshold=4.018e+02, percent-clipped=3.0 2023-04-27 05:19:45,791 INFO [finetune.py:976] (4/7) Epoch 14, batch 1300, loss[loss=0.1787, simple_loss=0.2398, pruned_loss=0.05879, over 4896.00 frames. ], tot_loss[loss=0.1782, simple_loss=0.2464, pruned_loss=0.05501, over 953495.90 frames. ], batch size: 32, lr: 3.56e-03, grad_scale: 32.0 2023-04-27 05:20:24,984 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1785, 1.8104, 1.9706, 2.3757, 2.4006, 2.0162, 1.4881, 2.1207], device='cuda:4'), covar=tensor([0.0741, 0.1135, 0.0712, 0.0571, 0.0564, 0.0789, 0.0888, 0.0571], device='cuda:4'), in_proj_covar=tensor([0.0191, 0.0205, 0.0184, 0.0174, 0.0180, 0.0184, 0.0157, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 05:20:50,515 INFO [finetune.py:976] (4/7) Epoch 14, batch 1350, loss[loss=0.1961, simple_loss=0.2632, pruned_loss=0.06448, over 4906.00 frames. ], tot_loss[loss=0.1802, simple_loss=0.2481, pruned_loss=0.05609, over 954818.40 frames. ], batch size: 37, lr: 3.56e-03, grad_scale: 32.0 2023-04-27 05:21:39,791 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.162e+02 1.563e+02 1.910e+02 2.423e+02 5.618e+02, threshold=3.821e+02, percent-clipped=3.0 2023-04-27 05:21:55,186 INFO [finetune.py:976] (4/7) Epoch 14, batch 1400, loss[loss=0.2372, simple_loss=0.3047, pruned_loss=0.08486, over 4731.00 frames. ], tot_loss[loss=0.1821, simple_loss=0.2509, pruned_loss=0.05662, over 955000.95 frames. ], batch size: 59, lr: 3.56e-03, grad_scale: 32.0 2023-04-27 05:22:34,310 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=75884.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:22:51,836 INFO [finetune.py:976] (4/7) Epoch 14, batch 1450, loss[loss=0.1674, simple_loss=0.2477, pruned_loss=0.04352, over 4931.00 frames. ], tot_loss[loss=0.184, simple_loss=0.2533, pruned_loss=0.05735, over 953705.89 frames. ], batch size: 42, lr: 3.55e-03, grad_scale: 32.0 2023-04-27 05:23:03,091 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=75926.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 05:23:06,210 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8894, 2.5306, 1.8509, 1.7557, 1.3183, 1.3489, 1.9242, 1.2675], device='cuda:4'), covar=tensor([0.1792, 0.1415, 0.1526, 0.1880, 0.2516, 0.2016, 0.1047, 0.2205], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0213, 0.0168, 0.0204, 0.0201, 0.0184, 0.0157, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 05:23:06,747 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=75932.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:23:14,582 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.134e+02 1.664e+02 2.009e+02 2.321e+02 4.322e+02, threshold=4.018e+02, percent-clipped=2.0 2023-04-27 05:23:25,302 INFO [finetune.py:976] (4/7) Epoch 14, batch 1500, loss[loss=0.1727, simple_loss=0.2447, pruned_loss=0.05031, over 4922.00 frames. ], tot_loss[loss=0.1862, simple_loss=0.2552, pruned_loss=0.05856, over 952791.08 frames. ], batch size: 42, lr: 3.55e-03, grad_scale: 32.0 2023-04-27 05:23:34,210 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=75974.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:23:57,933 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.70 vs. limit=5.0 2023-04-27 05:24:10,687 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-27 05:24:11,483 INFO [finetune.py:976] (4/7) Epoch 14, batch 1550, loss[loss=0.2032, simple_loss=0.2612, pruned_loss=0.07263, over 4808.00 frames. ], tot_loss[loss=0.1855, simple_loss=0.2546, pruned_loss=0.05819, over 950881.23 frames. ], batch size: 40, lr: 3.55e-03, grad_scale: 32.0 2023-04-27 05:24:18,774 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4824, 1.7619, 1.7080, 2.3053, 2.6227, 2.1109, 1.9762, 1.8129], device='cuda:4'), covar=tensor([0.1841, 0.1590, 0.1994, 0.1457, 0.0853, 0.1793, 0.2246, 0.2164], device='cuda:4'), in_proj_covar=tensor([0.0305, 0.0313, 0.0349, 0.0289, 0.0327, 0.0311, 0.0302, 0.0360], device='cuda:4'), out_proj_covar=tensor([6.3397e-05, 6.5678e-05, 7.4897e-05, 5.9140e-05, 6.8136e-05, 6.5882e-05, 6.3988e-05, 7.6902e-05], device='cuda:4') 2023-04-27 05:24:40,106 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.063e+02 1.630e+02 1.920e+02 2.278e+02 6.745e+02, threshold=3.839e+02, percent-clipped=3.0 2023-04-27 05:25:01,381 INFO [finetune.py:976] (4/7) Epoch 14, batch 1600, loss[loss=0.1678, simple_loss=0.2428, pruned_loss=0.04642, over 4769.00 frames. ], tot_loss[loss=0.1837, simple_loss=0.2529, pruned_loss=0.0573, over 951797.74 frames. ], batch size: 26, lr: 3.55e-03, grad_scale: 32.0 2023-04-27 05:25:34,941 INFO [finetune.py:976] (4/7) Epoch 14, batch 1650, loss[loss=0.1968, simple_loss=0.2597, pruned_loss=0.06695, over 4908.00 frames. ], tot_loss[loss=0.182, simple_loss=0.251, pruned_loss=0.05644, over 953382.20 frames. ], batch size: 32, lr: 3.55e-03, grad_scale: 32.0 2023-04-27 05:25:58,460 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.045e+02 1.559e+02 1.837e+02 2.256e+02 6.477e+02, threshold=3.674e+02, percent-clipped=1.0 2023-04-27 05:26:08,246 INFO [finetune.py:976] (4/7) Epoch 14, batch 1700, loss[loss=0.1564, simple_loss=0.2173, pruned_loss=0.0478, over 4826.00 frames. ], tot_loss[loss=0.1799, simple_loss=0.2485, pruned_loss=0.05567, over 953261.74 frames. ], batch size: 30, lr: 3.55e-03, grad_scale: 32.0 2023-04-27 05:26:42,127 INFO [finetune.py:976] (4/7) Epoch 14, batch 1750, loss[loss=0.1707, simple_loss=0.2343, pruned_loss=0.05353, over 4718.00 frames. ], tot_loss[loss=0.1824, simple_loss=0.2507, pruned_loss=0.05705, over 953186.45 frames. ], batch size: 23, lr: 3.55e-03, grad_scale: 32.0 2023-04-27 05:26:50,617 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5515, 3.0217, 2.3165, 2.2983, 1.8019, 1.7162, 2.5220, 1.7127], device='cuda:4'), covar=tensor([0.1597, 0.1595, 0.1371, 0.1791, 0.2244, 0.1892, 0.0951, 0.1993], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0212, 0.0168, 0.0204, 0.0201, 0.0183, 0.0156, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 05:26:51,822 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6232, 2.2797, 2.8570, 3.0076, 2.9446, 2.5190, 2.1407, 2.4865], device='cuda:4'), covar=tensor([0.0850, 0.1063, 0.0565, 0.0682, 0.0642, 0.0844, 0.0807, 0.0653], device='cuda:4'), in_proj_covar=tensor([0.0190, 0.0203, 0.0183, 0.0173, 0.0179, 0.0183, 0.0155, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 05:27:06,455 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.142e+02 1.610e+02 1.943e+02 2.415e+02 4.122e+02, threshold=3.886e+02, percent-clipped=3.0 2023-04-27 05:27:16,247 INFO [finetune.py:976] (4/7) Epoch 14, batch 1800, loss[loss=0.1845, simple_loss=0.2476, pruned_loss=0.06076, over 4825.00 frames. ], tot_loss[loss=0.1838, simple_loss=0.2532, pruned_loss=0.05722, over 952674.27 frames. ], batch size: 30, lr: 3.55e-03, grad_scale: 32.0 2023-04-27 05:27:27,570 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=76269.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 05:28:02,612 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=76294.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:28:03,784 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1294, 1.2985, 5.2262, 4.8453, 4.5482, 4.9799, 4.5806, 4.6071], device='cuda:4'), covar=tensor([0.6172, 0.6271, 0.0797, 0.1639, 0.1043, 0.1200, 0.1187, 0.1418], device='cuda:4'), in_proj_covar=tensor([0.0306, 0.0306, 0.0400, 0.0401, 0.0345, 0.0402, 0.0312, 0.0363], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 05:28:12,928 INFO [finetune.py:976] (4/7) Epoch 14, batch 1850, loss[loss=0.1834, simple_loss=0.2527, pruned_loss=0.05708, over 4880.00 frames. ], tot_loss[loss=0.1852, simple_loss=0.2545, pruned_loss=0.058, over 954430.24 frames. ], batch size: 32, lr: 3.55e-03, grad_scale: 32.0 2023-04-27 05:28:20,903 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=76323.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 05:28:21,577 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7338, 1.3709, 1.8539, 2.1970, 1.8549, 1.6915, 1.8030, 1.7725], device='cuda:4'), covar=tensor([0.4884, 0.7239, 0.7052, 0.6552, 0.6268, 0.8745, 0.9167, 0.9044], device='cuda:4'), in_proj_covar=tensor([0.0417, 0.0411, 0.0498, 0.0517, 0.0446, 0.0469, 0.0475, 0.0478], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 05:28:25,638 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9727, 0.9817, 3.1572, 2.7313, 2.8491, 2.9028, 2.9863, 2.6541], device='cuda:4'), covar=tensor([0.9183, 0.8261, 0.2213, 0.3796, 0.2832, 0.3932, 0.3239, 0.3754], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0307, 0.0402, 0.0403, 0.0347, 0.0404, 0.0314, 0.0365], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 05:28:25,691 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=76330.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 05:28:36,248 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.095e+02 1.653e+02 2.056e+02 2.488e+02 6.955e+02, threshold=4.112e+02, percent-clipped=5.0 2023-04-27 05:28:43,366 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=76355.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:28:46,906 INFO [finetune.py:976] (4/7) Epoch 14, batch 1900, loss[loss=0.188, simple_loss=0.2433, pruned_loss=0.06634, over 4765.00 frames. ], tot_loss[loss=0.1856, simple_loss=0.2551, pruned_loss=0.05802, over 956012.16 frames. ], batch size: 28, lr: 3.55e-03, grad_scale: 32.0 2023-04-27 05:28:56,636 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.4019, 1.3509, 1.4037, 0.9698, 1.3510, 1.2180, 1.6873, 1.3117], device='cuda:4'), covar=tensor([0.4109, 0.1829, 0.5358, 0.2979, 0.1825, 0.2310, 0.1698, 0.4988], device='cuda:4'), in_proj_covar=tensor([0.0342, 0.0348, 0.0428, 0.0359, 0.0382, 0.0384, 0.0374, 0.0421], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 05:28:57,804 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3319, 1.2538, 3.7519, 3.5120, 3.2931, 3.5515, 3.4849, 3.3017], device='cuda:4'), covar=tensor([0.6416, 0.5561, 0.1079, 0.1624, 0.1165, 0.1859, 0.2409, 0.1474], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0305, 0.0399, 0.0401, 0.0345, 0.0401, 0.0312, 0.0363], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 05:29:01,986 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=76384.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 05:29:20,620 INFO [finetune.py:976] (4/7) Epoch 14, batch 1950, loss[loss=0.1941, simple_loss=0.2655, pruned_loss=0.06134, over 4895.00 frames. ], tot_loss[loss=0.1839, simple_loss=0.2536, pruned_loss=0.05713, over 957165.02 frames. ], batch size: 43, lr: 3.55e-03, grad_scale: 32.0 2023-04-27 05:29:42,667 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.125e+01 1.567e+02 1.836e+02 2.067e+02 4.304e+02, threshold=3.671e+02, percent-clipped=1.0 2023-04-27 05:29:55,871 INFO [finetune.py:976] (4/7) Epoch 14, batch 2000, loss[loss=0.2045, simple_loss=0.2563, pruned_loss=0.07636, over 4789.00 frames. ], tot_loss[loss=0.1817, simple_loss=0.2509, pruned_loss=0.05629, over 956312.12 frames. ], batch size: 29, lr: 3.55e-03, grad_scale: 32.0 2023-04-27 05:29:56,008 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6020, 1.5190, 1.3040, 1.5020, 1.9091, 1.5994, 1.3374, 1.2055], device='cuda:4'), covar=tensor([0.2144, 0.1207, 0.2047, 0.1267, 0.0700, 0.1714, 0.2141, 0.1963], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0315, 0.0353, 0.0292, 0.0329, 0.0313, 0.0304, 0.0362], device='cuda:4'), out_proj_covar=tensor([6.3774e-05, 6.6152e-05, 7.5699e-05, 5.9683e-05, 6.8583e-05, 6.6378e-05, 6.4403e-05, 7.7295e-05], device='cuda:4') 2023-04-27 05:30:57,068 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4056, 1.3901, 1.7180, 1.6807, 1.3091, 1.1743, 1.4206, 0.9232], device='cuda:4'), covar=tensor([0.0626, 0.0711, 0.0511, 0.0644, 0.0840, 0.1219, 0.0767, 0.0721], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0070, 0.0069, 0.0068, 0.0075, 0.0096, 0.0075, 0.0070], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 05:30:57,657 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.4068, 1.4133, 1.4382, 0.9807, 1.3354, 1.1588, 1.7289, 1.2300], device='cuda:4'), covar=tensor([0.3495, 0.1744, 0.5016, 0.2679, 0.1631, 0.2201, 0.1623, 0.5293], device='cuda:4'), in_proj_covar=tensor([0.0345, 0.0351, 0.0430, 0.0360, 0.0384, 0.0386, 0.0375, 0.0424], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 05:30:58,157 INFO [finetune.py:976] (4/7) Epoch 14, batch 2050, loss[loss=0.206, simple_loss=0.2756, pruned_loss=0.06819, over 4909.00 frames. ], tot_loss[loss=0.1793, simple_loss=0.2478, pruned_loss=0.05535, over 956792.91 frames. ], batch size: 43, lr: 3.55e-03, grad_scale: 64.0 2023-04-27 05:31:19,388 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5518, 0.9023, 1.5921, 1.9463, 1.6239, 1.5007, 1.5495, 1.5515], device='cuda:4'), covar=tensor([0.4523, 0.6547, 0.6362, 0.6451, 0.5743, 0.7549, 0.7398, 0.7631], device='cuda:4'), in_proj_covar=tensor([0.0416, 0.0410, 0.0496, 0.0515, 0.0445, 0.0468, 0.0474, 0.0476], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 05:31:19,795 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.549e+02 1.874e+02 2.355e+02 5.627e+02, threshold=3.748e+02, percent-clipped=2.0 2023-04-27 05:31:20,541 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0003, 2.7380, 1.9927, 2.0951, 1.4660, 1.3855, 2.0918, 1.3865], device='cuda:4'), covar=tensor([0.1496, 0.1337, 0.1384, 0.1606, 0.2129, 0.1757, 0.0904, 0.1950], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0212, 0.0167, 0.0203, 0.0200, 0.0182, 0.0156, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 05:31:32,055 INFO [finetune.py:976] (4/7) Epoch 14, batch 2100, loss[loss=0.126, simple_loss=0.2042, pruned_loss=0.0239, over 4744.00 frames. ], tot_loss[loss=0.1788, simple_loss=0.247, pruned_loss=0.0553, over 957005.41 frames. ], batch size: 26, lr: 3.55e-03, grad_scale: 64.0 2023-04-27 05:32:06,104 INFO [finetune.py:976] (4/7) Epoch 14, batch 2150, loss[loss=0.181, simple_loss=0.2647, pruned_loss=0.04861, over 4902.00 frames. ], tot_loss[loss=0.1826, simple_loss=0.2514, pruned_loss=0.05691, over 956223.46 frames. ], batch size: 37, lr: 3.55e-03, grad_scale: 64.0 2023-04-27 05:32:14,748 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=76625.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 05:32:27,725 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.061e+02 1.703e+02 1.997e+02 2.488e+02 3.640e+02, threshold=3.993e+02, percent-clipped=1.0 2023-04-27 05:32:28,466 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0269, 2.6252, 2.2361, 2.5728, 1.8388, 2.3416, 2.3948, 1.8149], device='cuda:4'), covar=tensor([0.2147, 0.1242, 0.0794, 0.1170, 0.3021, 0.1070, 0.1925, 0.2423], device='cuda:4'), in_proj_covar=tensor([0.0295, 0.0314, 0.0228, 0.0287, 0.0315, 0.0268, 0.0258, 0.0274], device='cuda:4'), out_proj_covar=tensor([1.1900e-04, 1.2514e-04, 9.1033e-05, 1.1469e-04, 1.2845e-04, 1.0735e-04, 1.0463e-04, 1.0927e-04], device='cuda:4') 2023-04-27 05:32:30,844 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=76650.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:32:44,024 INFO [finetune.py:976] (4/7) Epoch 14, batch 2200, loss[loss=0.2106, simple_loss=0.2664, pruned_loss=0.07741, over 4794.00 frames. ], tot_loss[loss=0.1836, simple_loss=0.2529, pruned_loss=0.0571, over 954002.73 frames. ], batch size: 26, lr: 3.55e-03, grad_scale: 64.0 2023-04-27 05:32:45,739 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5507, 1.5382, 1.8771, 1.8569, 1.4013, 1.1905, 1.6351, 1.0666], device='cuda:4'), covar=tensor([0.0585, 0.0714, 0.0445, 0.0742, 0.0928, 0.1161, 0.0701, 0.0716], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0070, 0.0070, 0.0068, 0.0075, 0.0096, 0.0076, 0.0070], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 05:33:06,763 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=76679.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 05:33:08,602 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9097, 2.3757, 1.9935, 2.3662, 1.6879, 2.0049, 1.9074, 1.5217], device='cuda:4'), covar=tensor([0.1862, 0.1276, 0.0892, 0.1058, 0.3026, 0.1106, 0.1956, 0.2676], device='cuda:4'), in_proj_covar=tensor([0.0294, 0.0313, 0.0227, 0.0287, 0.0314, 0.0268, 0.0257, 0.0273], device='cuda:4'), out_proj_covar=tensor([1.1851e-04, 1.2481e-04, 9.0739e-05, 1.1438e-04, 1.2814e-04, 1.0706e-04, 1.0436e-04, 1.0903e-04], device='cuda:4') 2023-04-27 05:33:41,684 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5916, 1.1199, 1.2018, 1.3095, 1.7088, 1.3681, 1.1871, 1.1194], device='cuda:4'), covar=tensor([0.1537, 0.1567, 0.1913, 0.1326, 0.0874, 0.1682, 0.2118, 0.2117], device='cuda:4'), in_proj_covar=tensor([0.0306, 0.0314, 0.0351, 0.0290, 0.0328, 0.0313, 0.0303, 0.0360], device='cuda:4'), out_proj_covar=tensor([6.3612e-05, 6.6030e-05, 7.5353e-05, 5.9226e-05, 6.8420e-05, 6.6439e-05, 6.4167e-05, 7.6829e-05], device='cuda:4') 2023-04-27 05:33:46,580 INFO [finetune.py:976] (4/7) Epoch 14, batch 2250, loss[loss=0.2049, simple_loss=0.2763, pruned_loss=0.0668, over 4741.00 frames. ], tot_loss[loss=0.1853, simple_loss=0.2546, pruned_loss=0.05797, over 952550.79 frames. ], batch size: 54, lr: 3.55e-03, grad_scale: 64.0 2023-04-27 05:33:56,843 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1568, 1.6098, 2.0376, 2.4863, 1.9432, 1.5785, 1.1981, 1.8139], device='cuda:4'), covar=tensor([0.3293, 0.3365, 0.1741, 0.2220, 0.2799, 0.2709, 0.4296, 0.2253], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0248, 0.0224, 0.0317, 0.0217, 0.0230, 0.0231, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 05:34:30,282 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.017e+02 1.594e+02 1.898e+02 2.176e+02 4.652e+02, threshold=3.795e+02, percent-clipped=1.0 2023-04-27 05:34:47,192 INFO [finetune.py:976] (4/7) Epoch 14, batch 2300, loss[loss=0.1828, simple_loss=0.2609, pruned_loss=0.0524, over 4771.00 frames. ], tot_loss[loss=0.1848, simple_loss=0.254, pruned_loss=0.05785, over 952389.45 frames. ], batch size: 26, lr: 3.55e-03, grad_scale: 64.0 2023-04-27 05:34:58,180 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-04-27 05:34:59,765 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.6525, 3.5230, 2.7904, 4.1998, 3.6275, 3.6331, 1.6047, 3.5129], device='cuda:4'), covar=tensor([0.1583, 0.1298, 0.3210, 0.1526, 0.2887, 0.1649, 0.5527, 0.2344], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0212, 0.0247, 0.0301, 0.0295, 0.0245, 0.0270, 0.0270], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 05:35:26,138 INFO [finetune.py:976] (4/7) Epoch 14, batch 2350, loss[loss=0.1731, simple_loss=0.24, pruned_loss=0.05315, over 4817.00 frames. ], tot_loss[loss=0.1829, simple_loss=0.2516, pruned_loss=0.05709, over 952637.15 frames. ], batch size: 41, lr: 3.55e-03, grad_scale: 64.0 2023-04-27 05:36:05,984 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=76839.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:36:09,505 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.125e+02 1.656e+02 1.888e+02 2.270e+02 4.545e+02, threshold=3.776e+02, percent-clipped=2.0 2023-04-27 05:36:20,261 INFO [finetune.py:976] (4/7) Epoch 14, batch 2400, loss[loss=0.1944, simple_loss=0.2595, pruned_loss=0.06467, over 4753.00 frames. ], tot_loss[loss=0.1805, simple_loss=0.2488, pruned_loss=0.05612, over 952169.22 frames. ], batch size: 27, lr: 3.55e-03, grad_scale: 64.0 2023-04-27 05:36:26,448 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-04-27 05:36:28,516 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3641, 1.9288, 2.2976, 2.8971, 2.2957, 1.7957, 1.7721, 2.1994], device='cuda:4'), covar=tensor([0.3222, 0.3237, 0.1582, 0.2403, 0.2809, 0.2698, 0.3913, 0.2400], device='cuda:4'), in_proj_covar=tensor([0.0287, 0.0246, 0.0222, 0.0314, 0.0215, 0.0229, 0.0229, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 05:36:46,490 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=76900.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:36:54,077 INFO [finetune.py:976] (4/7) Epoch 14, batch 2450, loss[loss=0.1859, simple_loss=0.2251, pruned_loss=0.07332, over 3701.00 frames. ], tot_loss[loss=0.1786, simple_loss=0.2464, pruned_loss=0.05541, over 952398.89 frames. ], batch size: 16, lr: 3.55e-03, grad_scale: 64.0 2023-04-27 05:37:04,097 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=76925.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 05:37:04,868 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=5.03 vs. limit=5.0 2023-04-27 05:37:17,022 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.085e+02 1.514e+02 1.972e+02 2.483e+02 3.800e+02, threshold=3.944e+02, percent-clipped=1.0 2023-04-27 05:37:20,125 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=76950.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:37:27,669 INFO [finetune.py:976] (4/7) Epoch 14, batch 2500, loss[loss=0.2381, simple_loss=0.3078, pruned_loss=0.08417, over 4838.00 frames. ], tot_loss[loss=0.18, simple_loss=0.2476, pruned_loss=0.05623, over 949210.76 frames. ], batch size: 47, lr: 3.55e-03, grad_scale: 64.0 2023-04-27 05:37:36,502 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=76973.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 05:37:40,712 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=76979.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 05:37:52,652 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=76998.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:38:06,758 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3196, 1.5917, 1.6517, 1.7937, 1.6396, 1.7669, 1.7739, 1.7314], device='cuda:4'), covar=tensor([0.4296, 0.5809, 0.4843, 0.4810, 0.6087, 0.8130, 0.5854, 0.5384], device='cuda:4'), in_proj_covar=tensor([0.0330, 0.0375, 0.0315, 0.0328, 0.0342, 0.0398, 0.0355, 0.0324], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 05:38:07,211 INFO [finetune.py:976] (4/7) Epoch 14, batch 2550, loss[loss=0.1804, simple_loss=0.2577, pruned_loss=0.05156, over 4815.00 frames. ], tot_loss[loss=0.1819, simple_loss=0.2509, pruned_loss=0.05647, over 951020.79 frames. ], batch size: 38, lr: 3.55e-03, grad_scale: 64.0 2023-04-27 05:38:18,429 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=77027.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 05:38:30,330 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.907e+01 1.587e+02 1.892e+02 2.346e+02 6.910e+02, threshold=3.784e+02, percent-clipped=5.0 2023-04-27 05:38:37,121 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0329, 1.6001, 1.9131, 2.1593, 1.8391, 1.5133, 1.1222, 1.6255], device='cuda:4'), covar=tensor([0.3118, 0.3150, 0.1573, 0.2167, 0.2445, 0.2575, 0.4089, 0.2058], device='cuda:4'), in_proj_covar=tensor([0.0286, 0.0246, 0.0221, 0.0312, 0.0213, 0.0228, 0.0227, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 05:38:40,005 INFO [finetune.py:976] (4/7) Epoch 14, batch 2600, loss[loss=0.2198, simple_loss=0.295, pruned_loss=0.07231, over 4828.00 frames. ], tot_loss[loss=0.1826, simple_loss=0.252, pruned_loss=0.0566, over 951125.05 frames. ], batch size: 30, lr: 3.55e-03, grad_scale: 64.0 2023-04-27 05:38:41,342 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4566, 1.3879, 1.7101, 1.6971, 1.3077, 1.2267, 1.3939, 1.0657], device='cuda:4'), covar=tensor([0.0647, 0.0665, 0.0469, 0.0733, 0.0886, 0.1189, 0.0600, 0.0681], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0071, 0.0070, 0.0068, 0.0075, 0.0097, 0.0076, 0.0070], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 05:39:20,545 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=77095.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:39:41,101 INFO [finetune.py:976] (4/7) Epoch 14, batch 2650, loss[loss=0.1763, simple_loss=0.258, pruned_loss=0.04736, over 4817.00 frames. ], tot_loss[loss=0.1836, simple_loss=0.2536, pruned_loss=0.05679, over 952973.20 frames. ], batch size: 38, lr: 3.54e-03, grad_scale: 64.0 2023-04-27 05:40:19,436 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.209e+02 1.638e+02 1.948e+02 2.279e+02 4.007e+02, threshold=3.896e+02, percent-clipped=1.0 2023-04-27 05:40:22,663 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7702, 1.4678, 1.9137, 2.2931, 1.9116, 1.7457, 1.9001, 1.7897], device='cuda:4'), covar=tensor([0.4377, 0.6261, 0.6434, 0.5463, 0.5570, 0.7885, 0.7049, 0.8243], device='cuda:4'), in_proj_covar=tensor([0.0414, 0.0407, 0.0496, 0.0512, 0.0444, 0.0465, 0.0473, 0.0475], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 05:40:26,293 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=77156.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 05:40:29,225 INFO [finetune.py:976] (4/7) Epoch 14, batch 2700, loss[loss=0.1409, simple_loss=0.2163, pruned_loss=0.03272, over 4919.00 frames. ], tot_loss[loss=0.1823, simple_loss=0.2524, pruned_loss=0.05613, over 952394.49 frames. ], batch size: 38, lr: 3.54e-03, grad_scale: 64.0 2023-04-27 05:40:53,253 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=77195.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:41:03,078 INFO [finetune.py:976] (4/7) Epoch 14, batch 2750, loss[loss=0.16, simple_loss=0.2294, pruned_loss=0.04529, over 4827.00 frames. ], tot_loss[loss=0.1809, simple_loss=0.2501, pruned_loss=0.05585, over 953193.21 frames. ], batch size: 41, lr: 3.54e-03, grad_scale: 64.0 2023-04-27 05:41:03,203 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=77211.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:41:37,791 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.161e+02 1.631e+02 1.907e+02 2.391e+02 4.996e+02, threshold=3.813e+02, percent-clipped=2.0 2023-04-27 05:41:40,226 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8880, 2.0629, 1.3487, 1.6558, 2.2237, 1.7907, 1.7346, 1.7604], device='cuda:4'), covar=tensor([0.0424, 0.0302, 0.0292, 0.0458, 0.0267, 0.0430, 0.0414, 0.0458], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0025, 0.0023, 0.0029, 0.0020, 0.0029, 0.0028, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0045, 0.0038, 0.0051, 0.0038, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 05:41:48,207 INFO [finetune.py:976] (4/7) Epoch 14, batch 2800, loss[loss=0.1444, simple_loss=0.2206, pruned_loss=0.03405, over 4774.00 frames. ], tot_loss[loss=0.1774, simple_loss=0.2464, pruned_loss=0.05421, over 954225.96 frames. ], batch size: 28, lr: 3.54e-03, grad_scale: 64.0 2023-04-27 05:41:55,170 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=77272.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:41:55,471 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.34 vs. limit=5.0 2023-04-27 05:41:55,731 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=77273.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:41:58,630 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=77277.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:42:22,268 INFO [finetune.py:976] (4/7) Epoch 14, batch 2850, loss[loss=0.141, simple_loss=0.2058, pruned_loss=0.03809, over 4744.00 frames. ], tot_loss[loss=0.1768, simple_loss=0.2451, pruned_loss=0.05422, over 953827.04 frames. ], batch size: 27, lr: 3.54e-03, grad_scale: 64.0 2023-04-27 05:42:27,874 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5803, 3.9084, 0.7544, 2.0576, 2.0113, 2.6554, 2.2679, 0.8135], device='cuda:4'), covar=tensor([0.1492, 0.0928, 0.2150, 0.1267, 0.1151, 0.1108, 0.1397, 0.2422], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0247, 0.0140, 0.0122, 0.0133, 0.0154, 0.0119, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 05:42:37,432 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=77334.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:42:39,859 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=77338.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:42:44,408 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.158e+02 1.666e+02 1.986e+02 2.404e+02 4.400e+02, threshold=3.972e+02, percent-clipped=2.0 2023-04-27 05:42:52,729 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3044, 1.5722, 1.3789, 1.5150, 1.2718, 1.3105, 1.3234, 1.1547], device='cuda:4'), covar=tensor([0.1883, 0.1191, 0.1029, 0.1193, 0.3809, 0.1291, 0.1790, 0.2069], device='cuda:4'), in_proj_covar=tensor([0.0288, 0.0308, 0.0224, 0.0281, 0.0309, 0.0263, 0.0252, 0.0268], device='cuda:4'), out_proj_covar=tensor([1.1612e-04, 1.2254e-04, 8.9271e-05, 1.1217e-04, 1.2617e-04, 1.0507e-04, 1.0196e-04, 1.0698e-04], device='cuda:4') 2023-04-27 05:42:55,658 INFO [finetune.py:976] (4/7) Epoch 14, batch 2900, loss[loss=0.2244, simple_loss=0.2948, pruned_loss=0.077, over 4860.00 frames. ], tot_loss[loss=0.1824, simple_loss=0.2509, pruned_loss=0.05694, over 955183.72 frames. ], batch size: 44, lr: 3.54e-03, grad_scale: 64.0 2023-04-27 05:43:23,686 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-27 05:43:28,972 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-04-27 05:43:29,377 INFO [finetune.py:976] (4/7) Epoch 14, batch 2950, loss[loss=0.2332, simple_loss=0.2991, pruned_loss=0.08367, over 4809.00 frames. ], tot_loss[loss=0.1838, simple_loss=0.2529, pruned_loss=0.0574, over 953919.53 frames. ], batch size: 45, lr: 3.54e-03, grad_scale: 64.0 2023-04-27 05:43:50,925 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.161e+02 1.615e+02 1.939e+02 2.289e+02 5.440e+02, threshold=3.878e+02, percent-clipped=2.0 2023-04-27 05:43:56,099 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=77451.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 05:44:03,068 INFO [finetune.py:976] (4/7) Epoch 14, batch 3000, loss[loss=0.1994, simple_loss=0.2548, pruned_loss=0.07198, over 4727.00 frames. ], tot_loss[loss=0.185, simple_loss=0.2541, pruned_loss=0.05794, over 953858.91 frames. ], batch size: 23, lr: 3.54e-03, grad_scale: 64.0 2023-04-27 05:44:03,068 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 05:44:11,843 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.6664, 1.8455, 1.8120, 1.4181, 1.9121, 1.5334, 2.3636, 1.5840], device='cuda:4'), covar=tensor([0.4033, 0.1701, 0.4994, 0.2754, 0.1653, 0.2429, 0.1453, 0.5259], device='cuda:4'), in_proj_covar=tensor([0.0343, 0.0347, 0.0429, 0.0356, 0.0383, 0.0385, 0.0373, 0.0422], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 05:44:19,534 INFO [finetune.py:1010] (4/7) Epoch 14, validation: loss=0.1527, simple_loss=0.224, pruned_loss=0.04073, over 2265189.00 frames. 2023-04-27 05:44:19,535 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-27 05:45:02,946 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=77495.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:45:24,239 INFO [finetune.py:976] (4/7) Epoch 14, batch 3050, loss[loss=0.1561, simple_loss=0.2394, pruned_loss=0.03641, over 4849.00 frames. ], tot_loss[loss=0.185, simple_loss=0.2542, pruned_loss=0.05786, over 950960.03 frames. ], batch size: 44, lr: 3.54e-03, grad_scale: 64.0 2023-04-27 05:45:45,785 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=77543.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:45:46,942 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.289e+02 1.813e+02 2.136e+02 2.467e+02 4.229e+02, threshold=4.273e+02, percent-clipped=1.0 2023-04-27 05:45:57,727 INFO [finetune.py:976] (4/7) Epoch 14, batch 3100, loss[loss=0.1594, simple_loss=0.2347, pruned_loss=0.04204, over 4858.00 frames. ], tot_loss[loss=0.1833, simple_loss=0.2526, pruned_loss=0.05697, over 953115.71 frames. ], batch size: 47, lr: 3.54e-03, grad_scale: 64.0 2023-04-27 05:46:02,417 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.8950, 1.2009, 3.2862, 3.0317, 2.9693, 3.2041, 3.1949, 2.8955], device='cuda:4'), covar=tensor([0.7211, 0.5331, 0.1466, 0.2255, 0.1488, 0.1793, 0.1791, 0.1725], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0306, 0.0403, 0.0406, 0.0346, 0.0405, 0.0313, 0.0364], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 05:46:02,420 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=77567.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:46:36,288 INFO [finetune.py:976] (4/7) Epoch 14, batch 3150, loss[loss=0.1614, simple_loss=0.2243, pruned_loss=0.04926, over 4824.00 frames. ], tot_loss[loss=0.1818, simple_loss=0.2503, pruned_loss=0.05661, over 951212.78 frames. ], batch size: 25, lr: 3.54e-03, grad_scale: 32.0 2023-04-27 05:46:45,228 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=77616.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:47:05,390 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=77629.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:47:07,853 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=77633.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:47:20,987 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.865e+01 1.595e+02 1.944e+02 2.388e+02 4.637e+02, threshold=3.889e+02, percent-clipped=1.0 2023-04-27 05:47:38,383 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9700, 1.7453, 2.0622, 2.2864, 2.4504, 1.9119, 1.5278, 2.0311], device='cuda:4'), covar=tensor([0.0840, 0.1100, 0.0571, 0.0550, 0.0560, 0.0767, 0.0781, 0.0573], device='cuda:4'), in_proj_covar=tensor([0.0191, 0.0202, 0.0182, 0.0172, 0.0178, 0.0182, 0.0154, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 05:47:40,237 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-27 05:47:41,184 INFO [finetune.py:976] (4/7) Epoch 14, batch 3200, loss[loss=0.1613, simple_loss=0.2252, pruned_loss=0.04869, over 4834.00 frames. ], tot_loss[loss=0.1786, simple_loss=0.2466, pruned_loss=0.05528, over 953179.70 frames. ], batch size: 30, lr: 3.54e-03, grad_scale: 32.0 2023-04-27 05:48:10,697 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=77677.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:48:10,707 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=77677.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:48:37,435 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0071, 2.4311, 1.9210, 1.9131, 1.4561, 1.4648, 2.0150, 1.3568], device='cuda:4'), covar=tensor([0.1515, 0.1370, 0.1418, 0.1665, 0.2176, 0.1895, 0.0945, 0.2003], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0213, 0.0168, 0.0205, 0.0200, 0.0184, 0.0156, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 05:48:44,396 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-04-27 05:48:48,226 INFO [finetune.py:976] (4/7) Epoch 14, batch 3250, loss[loss=0.1897, simple_loss=0.2766, pruned_loss=0.0514, over 4821.00 frames. ], tot_loss[loss=0.1793, simple_loss=0.2475, pruned_loss=0.05558, over 954529.58 frames. ], batch size: 39, lr: 3.54e-03, grad_scale: 32.0 2023-04-27 05:48:58,735 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-04-27 05:49:28,776 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=77738.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:49:33,562 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.092e+02 1.660e+02 2.088e+02 2.486e+02 4.990e+02, threshold=4.175e+02, percent-clipped=6.0 2023-04-27 05:49:36,703 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=77751.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 05:49:42,679 INFO [finetune.py:976] (4/7) Epoch 14, batch 3300, loss[loss=0.2165, simple_loss=0.285, pruned_loss=0.074, over 4903.00 frames. ], tot_loss[loss=0.1826, simple_loss=0.2511, pruned_loss=0.05701, over 954799.55 frames. ], batch size: 35, lr: 3.54e-03, grad_scale: 32.0 2023-04-27 05:49:49,690 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.74 vs. limit=2.0 2023-04-27 05:50:08,404 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=77799.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:50:08,482 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8350, 2.4555, 1.8280, 1.7166, 1.3581, 1.3458, 2.0403, 1.3021], device='cuda:4'), covar=tensor([0.1675, 0.1325, 0.1500, 0.1798, 0.2270, 0.2018, 0.0957, 0.2151], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0213, 0.0168, 0.0204, 0.0201, 0.0184, 0.0157, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 05:50:11,066 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9190, 0.8912, 1.0686, 1.0162, 0.8812, 0.7630, 0.8347, 0.5626], device='cuda:4'), covar=tensor([0.0553, 0.0431, 0.0542, 0.0481, 0.0650, 0.0944, 0.0489, 0.0649], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0070, 0.0070, 0.0069, 0.0076, 0.0097, 0.0076, 0.0070], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 05:50:15,839 INFO [finetune.py:976] (4/7) Epoch 14, batch 3350, loss[loss=0.2126, simple_loss=0.2746, pruned_loss=0.07527, over 4903.00 frames. ], tot_loss[loss=0.1837, simple_loss=0.2523, pruned_loss=0.05758, over 954826.80 frames. ], batch size: 35, lr: 3.54e-03, grad_scale: 32.0 2023-04-27 05:50:21,860 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.7150, 4.7354, 3.2379, 5.4481, 4.8504, 4.5914, 2.4122, 4.6479], device='cuda:4'), covar=tensor([0.1551, 0.0904, 0.2822, 0.0940, 0.3347, 0.1768, 0.5331, 0.2005], device='cuda:4'), in_proj_covar=tensor([0.0246, 0.0215, 0.0251, 0.0305, 0.0300, 0.0248, 0.0273, 0.0275], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 05:50:39,906 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.052e+02 1.645e+02 1.956e+02 2.306e+02 5.075e+02, threshold=3.912e+02, percent-clipped=1.0 2023-04-27 05:50:45,583 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3356, 1.7722, 2.0908, 2.8150, 2.2204, 1.6719, 1.7715, 2.0199], device='cuda:4'), covar=tensor([0.3422, 0.3580, 0.1781, 0.2498, 0.3031, 0.2828, 0.3926, 0.2382], device='cuda:4'), in_proj_covar=tensor([0.0287, 0.0246, 0.0222, 0.0316, 0.0214, 0.0229, 0.0228, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 05:50:49,108 INFO [finetune.py:976] (4/7) Epoch 14, batch 3400, loss[loss=0.218, simple_loss=0.2887, pruned_loss=0.07367, over 4712.00 frames. ], tot_loss[loss=0.1858, simple_loss=0.2541, pruned_loss=0.05871, over 953345.38 frames. ], batch size: 59, lr: 3.54e-03, grad_scale: 32.0 2023-04-27 05:50:52,852 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=77867.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:51:22,437 INFO [finetune.py:976] (4/7) Epoch 14, batch 3450, loss[loss=0.1984, simple_loss=0.2713, pruned_loss=0.06271, over 4815.00 frames. ], tot_loss[loss=0.186, simple_loss=0.255, pruned_loss=0.05851, over 953285.21 frames. ], batch size: 38, lr: 3.54e-03, grad_scale: 32.0 2023-04-27 05:51:24,876 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=77915.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:51:33,834 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=77929.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:51:36,802 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=77933.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:51:45,497 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.078e+02 1.602e+02 1.891e+02 2.396e+02 5.235e+02, threshold=3.783e+02, percent-clipped=1.0 2023-04-27 05:51:54,677 INFO [finetune.py:976] (4/7) Epoch 14, batch 3500, loss[loss=0.1605, simple_loss=0.2279, pruned_loss=0.04652, over 4850.00 frames. ], tot_loss[loss=0.1846, simple_loss=0.2531, pruned_loss=0.05799, over 953028.02 frames. ], batch size: 49, lr: 3.54e-03, grad_scale: 32.0 2023-04-27 05:52:01,406 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=77972.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:52:04,957 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=77977.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:52:07,434 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=77981.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:52:13,766 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.45 vs. limit=2.0 2023-04-27 05:52:28,917 INFO [finetune.py:976] (4/7) Epoch 14, batch 3550, loss[loss=0.157, simple_loss=0.2319, pruned_loss=0.04108, over 4821.00 frames. ], tot_loss[loss=0.1831, simple_loss=0.2508, pruned_loss=0.05774, over 953785.72 frames. ], batch size: 41, lr: 3.54e-03, grad_scale: 32.0 2023-04-27 05:52:37,018 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-27 05:52:42,971 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=78033.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:52:56,899 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.150e+02 1.685e+02 1.967e+02 2.304e+02 5.361e+02, threshold=3.934e+02, percent-clipped=3.0 2023-04-27 05:53:17,519 INFO [finetune.py:976] (4/7) Epoch 14, batch 3600, loss[loss=0.1381, simple_loss=0.2192, pruned_loss=0.02848, over 4752.00 frames. ], tot_loss[loss=0.181, simple_loss=0.2482, pruned_loss=0.05689, over 954570.27 frames. ], batch size: 28, lr: 3.54e-03, grad_scale: 32.0 2023-04-27 05:54:11,969 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7048, 1.9976, 1.1188, 1.3872, 2.2675, 1.5367, 1.4892, 1.5503], device='cuda:4'), covar=tensor([0.0521, 0.0359, 0.0316, 0.0568, 0.0242, 0.0523, 0.0495, 0.0579], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0025, 0.0023, 0.0029, 0.0020, 0.0029, 0.0028, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0045, 0.0038, 0.0051, 0.0038, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 05:54:18,586 INFO [finetune.py:976] (4/7) Epoch 14, batch 3650, loss[loss=0.1844, simple_loss=0.2584, pruned_loss=0.05524, over 4905.00 frames. ], tot_loss[loss=0.183, simple_loss=0.2505, pruned_loss=0.05776, over 956131.17 frames. ], batch size: 37, lr: 3.54e-03, grad_scale: 32.0 2023-04-27 05:54:56,789 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.146e+02 1.566e+02 1.842e+02 2.272e+02 4.887e+02, threshold=3.683e+02, percent-clipped=1.0 2023-04-27 05:55:18,208 INFO [finetune.py:976] (4/7) Epoch 14, batch 3700, loss[loss=0.1785, simple_loss=0.2539, pruned_loss=0.05154, over 4784.00 frames. ], tot_loss[loss=0.1841, simple_loss=0.2525, pruned_loss=0.05782, over 955293.69 frames. ], batch size: 29, lr: 3.54e-03, grad_scale: 32.0 2023-04-27 05:55:19,278 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.43 vs. limit=2.0 2023-04-27 05:55:19,888 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.77 vs. limit=5.0 2023-04-27 05:55:56,844 INFO [finetune.py:976] (4/7) Epoch 14, batch 3750, loss[loss=0.201, simple_loss=0.2765, pruned_loss=0.06273, over 4742.00 frames. ], tot_loss[loss=0.1846, simple_loss=0.2534, pruned_loss=0.05791, over 955219.84 frames. ], batch size: 54, lr: 3.54e-03, grad_scale: 32.0 2023-04-27 05:56:09,772 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=78232.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:56:15,841 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1854, 1.4081, 1.3631, 1.6360, 1.5369, 1.7422, 1.3114, 3.0353], device='cuda:4'), covar=tensor([0.0630, 0.0805, 0.0808, 0.1207, 0.0642, 0.0488, 0.0721, 0.0169], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0012, 0.0013, 0.0016], device='cuda:4') 2023-04-27 05:56:18,617 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.305e+02 1.679e+02 1.977e+02 2.535e+02 5.677e+02, threshold=3.954e+02, percent-clipped=2.0 2023-04-27 05:56:30,168 INFO [finetune.py:976] (4/7) Epoch 14, batch 3800, loss[loss=0.2046, simple_loss=0.2731, pruned_loss=0.06802, over 4841.00 frames. ], tot_loss[loss=0.1863, simple_loss=0.2551, pruned_loss=0.05872, over 955659.46 frames. ], batch size: 44, lr: 3.53e-03, grad_scale: 32.0 2023-04-27 05:56:37,434 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=78272.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:56:50,292 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=78293.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 05:57:03,403 INFO [finetune.py:976] (4/7) Epoch 14, batch 3850, loss[loss=0.1667, simple_loss=0.2311, pruned_loss=0.05112, over 4888.00 frames. ], tot_loss[loss=0.1843, simple_loss=0.2536, pruned_loss=0.05751, over 956382.20 frames. ], batch size: 35, lr: 3.53e-03, grad_scale: 32.0 2023-04-27 05:57:04,570 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6466, 4.1012, 0.8337, 2.0966, 2.0242, 2.6014, 2.3040, 0.9152], device='cuda:4'), covar=tensor([0.1441, 0.0931, 0.2070, 0.1275, 0.1129, 0.1180, 0.1479, 0.2330], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0245, 0.0138, 0.0121, 0.0132, 0.0153, 0.0118, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 05:57:09,827 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=78320.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:57:16,653 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0486, 1.6582, 1.4561, 1.9670, 2.1809, 1.9172, 1.7898, 1.4321], device='cuda:4'), covar=tensor([0.1407, 0.1364, 0.1360, 0.1272, 0.0812, 0.1400, 0.1805, 0.1527], device='cuda:4'), in_proj_covar=tensor([0.0305, 0.0314, 0.0350, 0.0290, 0.0327, 0.0313, 0.0302, 0.0358], device='cuda:4'), out_proj_covar=tensor([6.3270e-05, 6.5978e-05, 7.5112e-05, 5.9231e-05, 6.7923e-05, 6.6257e-05, 6.3884e-05, 7.6573e-05], device='cuda:4') 2023-04-27 05:57:17,838 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=78333.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:57:25,785 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.080e+02 1.524e+02 1.884e+02 2.249e+02 3.539e+02, threshold=3.767e+02, percent-clipped=0.0 2023-04-27 05:57:36,835 INFO [finetune.py:976] (4/7) Epoch 14, batch 3900, loss[loss=0.1691, simple_loss=0.2411, pruned_loss=0.04857, over 4824.00 frames. ], tot_loss[loss=0.1828, simple_loss=0.2513, pruned_loss=0.05714, over 956521.40 frames. ], batch size: 25, lr: 3.53e-03, grad_scale: 32.0 2023-04-27 05:57:38,585 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.7151, 3.6796, 2.7646, 4.2557, 3.7005, 3.6921, 1.4902, 3.7073], device='cuda:4'), covar=tensor([0.2164, 0.1166, 0.3547, 0.1911, 0.3350, 0.2129, 0.6499, 0.2304], device='cuda:4'), in_proj_covar=tensor([0.0247, 0.0216, 0.0253, 0.0306, 0.0300, 0.0249, 0.0274, 0.0274], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 05:57:50,413 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=78381.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:57:50,471 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.8013, 2.4395, 2.9051, 3.2043, 3.1991, 2.6904, 2.1616, 2.7510], device='cuda:4'), covar=tensor([0.0817, 0.0956, 0.0542, 0.0498, 0.0506, 0.0709, 0.0705, 0.0563], device='cuda:4'), in_proj_covar=tensor([0.0191, 0.0202, 0.0182, 0.0172, 0.0177, 0.0182, 0.0153, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 05:58:09,807 INFO [finetune.py:976] (4/7) Epoch 14, batch 3950, loss[loss=0.1689, simple_loss=0.2271, pruned_loss=0.05533, over 4868.00 frames. ], tot_loss[loss=0.1798, simple_loss=0.2477, pruned_loss=0.05597, over 954231.64 frames. ], batch size: 34, lr: 3.53e-03, grad_scale: 32.0 2023-04-27 05:58:35,562 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.80 vs. limit=2.0 2023-04-27 05:58:50,267 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.090e+02 1.573e+02 1.912e+02 2.257e+02 5.641e+02, threshold=3.824e+02, percent-clipped=2.0 2023-04-27 05:59:02,406 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.06 vs. limit=5.0 2023-04-27 05:59:04,073 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5900, 2.1818, 1.4532, 1.3087, 1.1953, 1.1679, 1.4493, 1.0674], device='cuda:4'), covar=tensor([0.1797, 0.1220, 0.1780, 0.2021, 0.2624, 0.2506, 0.1176, 0.2284], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0213, 0.0168, 0.0205, 0.0201, 0.0184, 0.0157, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 05:59:04,641 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=78460.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 05:59:05,121 INFO [finetune.py:976] (4/7) Epoch 14, batch 4000, loss[loss=0.2459, simple_loss=0.2902, pruned_loss=0.1007, over 4038.00 frames. ], tot_loss[loss=0.1792, simple_loss=0.2467, pruned_loss=0.05585, over 952575.84 frames. ], batch size: 65, lr: 3.53e-03, grad_scale: 32.0 2023-04-27 05:59:05,719 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9662, 4.2572, 0.8232, 2.1650, 2.3497, 2.6137, 2.4966, 0.8882], device='cuda:4'), covar=tensor([0.1401, 0.0877, 0.2247, 0.1332, 0.1078, 0.1283, 0.1433, 0.2320], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0245, 0.0139, 0.0121, 0.0132, 0.0153, 0.0118, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 05:59:14,784 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.9366, 2.8391, 2.2990, 3.3281, 2.9389, 2.8823, 1.0863, 2.9280], device='cuda:4'), covar=tensor([0.2191, 0.1676, 0.3002, 0.2815, 0.3379, 0.2170, 0.6159, 0.2677], device='cuda:4'), in_proj_covar=tensor([0.0246, 0.0215, 0.0250, 0.0304, 0.0298, 0.0248, 0.0272, 0.0273], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 06:00:00,328 INFO [finetune.py:976] (4/7) Epoch 14, batch 4050, loss[loss=0.1902, simple_loss=0.2613, pruned_loss=0.05955, over 4903.00 frames. ], tot_loss[loss=0.1821, simple_loss=0.2502, pruned_loss=0.05703, over 952795.12 frames. ], batch size: 43, lr: 3.53e-03, grad_scale: 32.0 2023-04-27 06:00:19,676 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=78521.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:00:43,534 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.54 vs. limit=2.0 2023-04-27 06:00:51,058 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.238e+02 1.700e+02 2.035e+02 2.375e+02 5.106e+02, threshold=4.071e+02, percent-clipped=4.0 2023-04-27 06:01:03,043 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=78556.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:01:06,003 INFO [finetune.py:976] (4/7) Epoch 14, batch 4100, loss[loss=0.1834, simple_loss=0.2544, pruned_loss=0.05616, over 4775.00 frames. ], tot_loss[loss=0.1837, simple_loss=0.2523, pruned_loss=0.05753, over 953299.71 frames. ], batch size: 29, lr: 3.53e-03, grad_scale: 32.0 2023-04-27 06:01:27,541 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.7239, 1.8209, 1.7963, 1.4075, 1.9193, 1.5857, 2.3336, 1.4838], device='cuda:4'), covar=tensor([0.3502, 0.1645, 0.4456, 0.2515, 0.1361, 0.2311, 0.1506, 0.4826], device='cuda:4'), in_proj_covar=tensor([0.0346, 0.0349, 0.0430, 0.0358, 0.0384, 0.0386, 0.0376, 0.0423], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 06:01:45,287 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=78588.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 06:02:05,270 INFO [finetune.py:976] (4/7) Epoch 14, batch 4150, loss[loss=0.1893, simple_loss=0.2705, pruned_loss=0.05408, over 4868.00 frames. ], tot_loss[loss=0.1851, simple_loss=0.254, pruned_loss=0.0581, over 953235.01 frames. ], batch size: 34, lr: 3.53e-03, grad_scale: 32.0 2023-04-27 06:02:09,278 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-27 06:02:09,553 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=78617.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:02:29,495 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.351e+02 1.709e+02 1.960e+02 2.352e+02 3.930e+02, threshold=3.920e+02, percent-clipped=0.0 2023-04-27 06:02:38,718 INFO [finetune.py:976] (4/7) Epoch 14, batch 4200, loss[loss=0.1387, simple_loss=0.2091, pruned_loss=0.03412, over 4755.00 frames. ], tot_loss[loss=0.1859, simple_loss=0.2554, pruned_loss=0.05816, over 953344.67 frames. ], batch size: 26, lr: 3.53e-03, grad_scale: 32.0 2023-04-27 06:03:12,005 INFO [finetune.py:976] (4/7) Epoch 14, batch 4250, loss[loss=0.1672, simple_loss=0.2338, pruned_loss=0.05026, over 4815.00 frames. ], tot_loss[loss=0.1848, simple_loss=0.2541, pruned_loss=0.05773, over 953684.02 frames. ], batch size: 25, lr: 3.53e-03, grad_scale: 32.0 2023-04-27 06:03:36,178 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.058e+02 1.588e+02 1.905e+02 2.240e+02 4.270e+02, threshold=3.810e+02, percent-clipped=2.0 2023-04-27 06:03:45,440 INFO [finetune.py:976] (4/7) Epoch 14, batch 4300, loss[loss=0.1579, simple_loss=0.2341, pruned_loss=0.04081, over 4777.00 frames. ], tot_loss[loss=0.1836, simple_loss=0.2522, pruned_loss=0.05752, over 955356.38 frames. ], batch size: 29, lr: 3.53e-03, grad_scale: 32.0 2023-04-27 06:04:14,037 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-27 06:04:19,294 INFO [finetune.py:976] (4/7) Epoch 14, batch 4350, loss[loss=0.1455, simple_loss=0.2225, pruned_loss=0.03421, over 4786.00 frames. ], tot_loss[loss=0.1801, simple_loss=0.248, pruned_loss=0.05607, over 954970.41 frames. ], batch size: 29, lr: 3.53e-03, grad_scale: 32.0 2023-04-27 06:04:20,036 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4891, 1.4402, 1.8468, 1.8210, 1.3440, 1.1589, 1.5021, 1.0103], device='cuda:4'), covar=tensor([0.0563, 0.0595, 0.0439, 0.0509, 0.0774, 0.1141, 0.0632, 0.0640], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0069, 0.0069, 0.0067, 0.0075, 0.0096, 0.0075, 0.0069], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 06:04:22,429 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=78816.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:04:31,018 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-27 06:04:43,555 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.082e+02 1.581e+02 1.858e+02 2.218e+02 4.471e+02, threshold=3.716e+02, percent-clipped=1.0 2023-04-27 06:05:04,245 INFO [finetune.py:976] (4/7) Epoch 14, batch 4400, loss[loss=0.1686, simple_loss=0.234, pruned_loss=0.05162, over 4320.00 frames. ], tot_loss[loss=0.1806, simple_loss=0.2487, pruned_loss=0.0563, over 954015.59 frames. ], batch size: 18, lr: 3.53e-03, grad_scale: 32.0 2023-04-27 06:05:38,841 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=78888.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:05:57,976 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-04-27 06:06:11,504 INFO [finetune.py:976] (4/7) Epoch 14, batch 4450, loss[loss=0.2262, simple_loss=0.296, pruned_loss=0.07816, over 4822.00 frames. ], tot_loss[loss=0.1825, simple_loss=0.2507, pruned_loss=0.05717, over 954822.42 frames. ], batch size: 38, lr: 3.53e-03, grad_scale: 32.0 2023-04-27 06:06:12,194 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=78912.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:06:44,184 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=78936.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:07:03,710 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.730e+02 2.046e+02 2.407e+02 5.260e+02, threshold=4.092e+02, percent-clipped=6.0 2023-04-27 06:07:18,901 INFO [finetune.py:976] (4/7) Epoch 14, batch 4500, loss[loss=0.1429, simple_loss=0.2108, pruned_loss=0.03751, over 4809.00 frames. ], tot_loss[loss=0.1834, simple_loss=0.2521, pruned_loss=0.05738, over 954357.56 frames. ], batch size: 33, lr: 3.53e-03, grad_scale: 32.0 2023-04-27 06:07:21,404 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=78965.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:07:52,197 INFO [finetune.py:976] (4/7) Epoch 14, batch 4550, loss[loss=0.1767, simple_loss=0.2537, pruned_loss=0.04984, over 4931.00 frames. ], tot_loss[loss=0.1852, simple_loss=0.2538, pruned_loss=0.05831, over 954317.40 frames. ], batch size: 33, lr: 3.53e-03, grad_scale: 32.0 2023-04-27 06:08:01,470 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=79026.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:08:04,475 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5016, 1.7722, 1.8761, 1.9654, 1.7828, 1.8255, 1.9431, 1.9141], device='cuda:4'), covar=tensor([0.4139, 0.6145, 0.5351, 0.4901, 0.6017, 0.8197, 0.5869, 0.5515], device='cuda:4'), in_proj_covar=tensor([0.0331, 0.0375, 0.0316, 0.0330, 0.0340, 0.0397, 0.0353, 0.0325], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 06:08:14,994 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.614e+01 1.580e+02 1.894e+02 2.383e+02 3.819e+02, threshold=3.787e+02, percent-clipped=0.0 2023-04-27 06:08:17,273 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7235, 3.5776, 1.0124, 1.7334, 2.0135, 2.5871, 1.9773, 0.9848], device='cuda:4'), covar=tensor([0.1366, 0.0993, 0.1925, 0.1479, 0.1045, 0.0970, 0.1543, 0.1974], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0247, 0.0139, 0.0122, 0.0133, 0.0154, 0.0119, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 06:08:26,105 INFO [finetune.py:976] (4/7) Epoch 14, batch 4600, loss[loss=0.1841, simple_loss=0.2479, pruned_loss=0.06016, over 4894.00 frames. ], tot_loss[loss=0.1844, simple_loss=0.253, pruned_loss=0.0579, over 955161.91 frames. ], batch size: 37, lr: 3.53e-03, grad_scale: 32.0 2023-04-27 06:08:58,132 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2687, 2.9388, 2.5114, 2.3028, 1.5417, 1.6079, 2.7238, 1.6838], device='cuda:4'), covar=tensor([0.1492, 0.1344, 0.1179, 0.1531, 0.2175, 0.1757, 0.0797, 0.1872], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0214, 0.0169, 0.0205, 0.0202, 0.0185, 0.0158, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 06:08:59,225 INFO [finetune.py:976] (4/7) Epoch 14, batch 4650, loss[loss=0.2222, simple_loss=0.2857, pruned_loss=0.07939, over 4926.00 frames. ], tot_loss[loss=0.1823, simple_loss=0.2504, pruned_loss=0.05709, over 955099.57 frames. ], batch size: 38, lr: 3.53e-03, grad_scale: 32.0 2023-04-27 06:09:00,392 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.47 vs. limit=2.0 2023-04-27 06:09:01,748 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.8334, 3.8076, 2.8514, 4.5225, 3.9117, 3.8435, 1.9650, 3.7782], device='cuda:4'), covar=tensor([0.1800, 0.1276, 0.3065, 0.1374, 0.3000, 0.2051, 0.5509, 0.2470], device='cuda:4'), in_proj_covar=tensor([0.0247, 0.0216, 0.0250, 0.0304, 0.0298, 0.0249, 0.0272, 0.0273], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 06:09:02,360 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=79116.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:09:21,596 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.186e+02 1.677e+02 1.942e+02 2.274e+02 5.469e+02, threshold=3.883e+02, percent-clipped=3.0 2023-04-27 06:09:29,212 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9030, 2.2598, 0.9252, 1.2682, 1.6388, 1.1561, 2.4786, 1.4090], device='cuda:4'), covar=tensor([0.0691, 0.0653, 0.0680, 0.1168, 0.0422, 0.1007, 0.0353, 0.0674], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0049, 0.0047, 0.0050, 0.0052, 0.0076, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 06:09:32,661 INFO [finetune.py:976] (4/7) Epoch 14, batch 4700, loss[loss=0.201, simple_loss=0.2655, pruned_loss=0.06825, over 4918.00 frames. ], tot_loss[loss=0.1817, simple_loss=0.2494, pruned_loss=0.05703, over 954048.37 frames. ], batch size: 37, lr: 3.53e-03, grad_scale: 32.0 2023-04-27 06:09:33,406 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5284, 1.9800, 1.7228, 1.8687, 1.5280, 1.5615, 1.6549, 1.3297], device='cuda:4'), covar=tensor([0.1850, 0.1423, 0.0932, 0.1283, 0.3582, 0.1475, 0.1960, 0.2584], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0310, 0.0224, 0.0281, 0.0312, 0.0265, 0.0254, 0.0270], device='cuda:4'), out_proj_covar=tensor([1.1692e-04, 1.2325e-04, 8.9118e-05, 1.1200e-04, 1.2703e-04, 1.0588e-04, 1.0259e-04, 1.0759e-04], device='cuda:4') 2023-04-27 06:09:34,546 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=79164.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:10:05,832 INFO [finetune.py:976] (4/7) Epoch 14, batch 4750, loss[loss=0.1996, simple_loss=0.2711, pruned_loss=0.06406, over 4774.00 frames. ], tot_loss[loss=0.1799, simple_loss=0.2471, pruned_loss=0.05633, over 953440.03 frames. ], batch size: 28, lr: 3.53e-03, grad_scale: 32.0 2023-04-27 06:10:06,561 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=79212.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:10:37,980 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.049e+02 1.597e+02 1.966e+02 2.340e+02 3.997e+02, threshold=3.932e+02, percent-clipped=2.0 2023-04-27 06:10:41,101 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1006, 2.5131, 1.0183, 1.4009, 1.8626, 1.2904, 3.3029, 1.7883], device='cuda:4'), covar=tensor([0.0664, 0.0617, 0.0808, 0.1274, 0.0503, 0.1032, 0.0307, 0.0631], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0049, 0.0047, 0.0050, 0.0052, 0.0076, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 06:10:58,673 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=79260.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:10:59,216 INFO [finetune.py:976] (4/7) Epoch 14, batch 4800, loss[loss=0.1807, simple_loss=0.2539, pruned_loss=0.05372, over 4927.00 frames. ], tot_loss[loss=0.1809, simple_loss=0.249, pruned_loss=0.05645, over 954105.88 frames. ], batch size: 33, lr: 3.53e-03, grad_scale: 32.0 2023-04-27 06:11:25,261 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=79285.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 06:11:55,459 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0926, 0.7067, 0.9699, 0.7800, 1.2254, 0.9819, 0.8633, 0.9883], device='cuda:4'), covar=tensor([0.1669, 0.1504, 0.2113, 0.1720, 0.1027, 0.1427, 0.1708, 0.2197], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0320, 0.0357, 0.0293, 0.0331, 0.0318, 0.0306, 0.0364], device='cuda:4'), out_proj_covar=tensor([6.4458e-05, 6.7154e-05, 7.6698e-05, 5.9830e-05, 6.8932e-05, 6.7343e-05, 6.4653e-05, 7.7785e-05], device='cuda:4') 2023-04-27 06:11:58,653 INFO [finetune.py:976] (4/7) Epoch 14, batch 4850, loss[loss=0.1538, simple_loss=0.2142, pruned_loss=0.04669, over 4725.00 frames. ], tot_loss[loss=0.1843, simple_loss=0.2527, pruned_loss=0.05793, over 955100.60 frames. ], batch size: 23, lr: 3.53e-03, grad_scale: 32.0 2023-04-27 06:12:05,526 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.37 vs. limit=5.0 2023-04-27 06:12:10,416 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=79321.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:12:12,861 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6390, 3.0039, 0.9671, 1.7976, 2.3129, 1.7410, 4.5128, 2.3273], device='cuda:4'), covar=tensor([0.0605, 0.0847, 0.0926, 0.1276, 0.0540, 0.0951, 0.0219, 0.0596], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0067, 0.0049, 0.0047, 0.0050, 0.0052, 0.0076, 0.0052], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 06:12:23,700 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=79333.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:12:36,866 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.159e+02 1.635e+02 2.018e+02 2.486e+02 3.725e+02, threshold=4.037e+02, percent-clipped=0.0 2023-04-27 06:12:42,251 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=79346.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 06:12:56,923 INFO [finetune.py:976] (4/7) Epoch 14, batch 4900, loss[loss=0.1601, simple_loss=0.2349, pruned_loss=0.04262, over 4820.00 frames. ], tot_loss[loss=0.1851, simple_loss=0.254, pruned_loss=0.05805, over 953956.81 frames. ], batch size: 25, lr: 3.53e-03, grad_scale: 32.0 2023-04-27 06:13:00,247 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=79364.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:13:07,486 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=79375.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 06:13:19,225 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=79394.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:13:30,000 INFO [finetune.py:976] (4/7) Epoch 14, batch 4950, loss[loss=0.1505, simple_loss=0.2279, pruned_loss=0.03655, over 4815.00 frames. ], tot_loss[loss=0.186, simple_loss=0.255, pruned_loss=0.05844, over 954133.65 frames. ], batch size: 33, lr: 3.52e-03, grad_scale: 32.0 2023-04-27 06:13:40,500 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=79425.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:13:47,631 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=79436.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 06:13:53,548 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.062e+02 1.668e+02 1.946e+02 2.385e+02 4.906e+02, threshold=3.893e+02, percent-clipped=2.0 2023-04-27 06:14:03,215 INFO [finetune.py:976] (4/7) Epoch 14, batch 5000, loss[loss=0.1288, simple_loss=0.1997, pruned_loss=0.02897, over 4778.00 frames. ], tot_loss[loss=0.1856, simple_loss=0.2544, pruned_loss=0.05845, over 952392.02 frames. ], batch size: 25, lr: 3.52e-03, grad_scale: 32.0 2023-04-27 06:14:30,832 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6069, 2.5944, 2.2307, 2.3156, 2.8009, 2.2074, 3.6095, 2.0334], device='cuda:4'), covar=tensor([0.4285, 0.2821, 0.4541, 0.4245, 0.2150, 0.3300, 0.1783, 0.4634], device='cuda:4'), in_proj_covar=tensor([0.0344, 0.0350, 0.0432, 0.0357, 0.0385, 0.0384, 0.0376, 0.0422], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 06:14:33,274 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5611, 2.0006, 2.4095, 3.0389, 2.4019, 1.8213, 1.8472, 2.4799], device='cuda:4'), covar=tensor([0.3520, 0.3354, 0.1795, 0.2580, 0.2715, 0.3043, 0.3786, 0.2113], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0248, 0.0224, 0.0317, 0.0216, 0.0231, 0.0230, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 06:14:35,554 INFO [finetune.py:976] (4/7) Epoch 14, batch 5050, loss[loss=0.1729, simple_loss=0.2492, pruned_loss=0.0483, over 4827.00 frames. ], tot_loss[loss=0.1841, simple_loss=0.2526, pruned_loss=0.05786, over 952879.14 frames. ], batch size: 33, lr: 3.52e-03, grad_scale: 16.0 2023-04-27 06:14:59,517 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.023e+02 1.519e+02 1.891e+02 2.332e+02 5.445e+02, threshold=3.783e+02, percent-clipped=1.0 2023-04-27 06:15:08,088 INFO [finetune.py:976] (4/7) Epoch 14, batch 5100, loss[loss=0.1499, simple_loss=0.2133, pruned_loss=0.04327, over 4868.00 frames. ], tot_loss[loss=0.1813, simple_loss=0.2492, pruned_loss=0.05675, over 952272.82 frames. ], batch size: 44, lr: 3.52e-03, grad_scale: 16.0 2023-04-27 06:15:25,158 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=79585.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:15:41,435 INFO [finetune.py:976] (4/7) Epoch 14, batch 5150, loss[loss=0.178, simple_loss=0.2474, pruned_loss=0.05431, over 4808.00 frames. ], tot_loss[loss=0.1806, simple_loss=0.2488, pruned_loss=0.05619, over 954545.62 frames. ], batch size: 41, lr: 3.52e-03, grad_scale: 16.0 2023-04-27 06:15:48,513 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=79621.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:16:02,441 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=79641.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 06:16:06,015 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=79646.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:16:06,494 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.050e+02 1.596e+02 1.950e+02 2.248e+02 3.363e+02, threshold=3.899e+02, percent-clipped=0.0 2023-04-27 06:16:12,026 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7936, 2.2998, 1.8809, 2.0857, 1.5898, 1.8321, 1.8458, 1.4781], device='cuda:4'), covar=tensor([0.1789, 0.1211, 0.0770, 0.1213, 0.3061, 0.1145, 0.1758, 0.2383], device='cuda:4'), in_proj_covar=tensor([0.0289, 0.0309, 0.0223, 0.0281, 0.0311, 0.0263, 0.0253, 0.0268], device='cuda:4'), out_proj_covar=tensor([1.1662e-04, 1.2320e-04, 8.8784e-05, 1.1193e-04, 1.2656e-04, 1.0498e-04, 1.0239e-04, 1.0686e-04], device='cuda:4') 2023-04-27 06:16:12,609 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=79657.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:16:14,973 INFO [finetune.py:976] (4/7) Epoch 14, batch 5200, loss[loss=0.255, simple_loss=0.3211, pruned_loss=0.09445, over 4092.00 frames. ], tot_loss[loss=0.1819, simple_loss=0.2505, pruned_loss=0.05665, over 950956.64 frames. ], batch size: 65, lr: 3.52e-03, grad_scale: 16.0 2023-04-27 06:16:19,854 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=79669.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:16:31,047 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.7121, 2.0014, 1.8886, 2.0886, 1.8397, 2.0035, 1.9730, 1.9184], device='cuda:4'), covar=tensor([0.4793, 0.7032, 0.5612, 0.4904, 0.6523, 0.8043, 0.7123, 0.6622], device='cuda:4'), in_proj_covar=tensor([0.0333, 0.0377, 0.0319, 0.0333, 0.0342, 0.0400, 0.0357, 0.0328], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 06:16:34,496 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=79689.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:16:54,349 INFO [finetune.py:976] (4/7) Epoch 14, batch 5250, loss[loss=0.1842, simple_loss=0.2656, pruned_loss=0.05144, over 4818.00 frames. ], tot_loss[loss=0.1832, simple_loss=0.2529, pruned_loss=0.05676, over 953537.80 frames. ], batch size: 40, lr: 3.52e-03, grad_scale: 16.0 2023-04-27 06:17:03,902 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=79718.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:17:05,052 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=79720.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:17:19,833 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=79731.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 06:17:36,846 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.087e+02 1.641e+02 2.072e+02 2.610e+02 5.253e+02, threshold=4.143e+02, percent-clipped=2.0 2023-04-27 06:17:45,879 INFO [finetune.py:976] (4/7) Epoch 14, batch 5300, loss[loss=0.2164, simple_loss=0.2892, pruned_loss=0.0718, over 4811.00 frames. ], tot_loss[loss=0.186, simple_loss=0.2556, pruned_loss=0.05824, over 952844.20 frames. ], batch size: 40, lr: 3.52e-03, grad_scale: 16.0 2023-04-27 06:17:46,596 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=79762.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:18:38,536 INFO [finetune.py:976] (4/7) Epoch 14, batch 5350, loss[loss=0.2069, simple_loss=0.2726, pruned_loss=0.07062, over 4733.00 frames. ], tot_loss[loss=0.1854, simple_loss=0.2553, pruned_loss=0.05775, over 952057.24 frames. ], batch size: 59, lr: 3.52e-03, grad_scale: 16.0 2023-04-27 06:18:45,992 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=79823.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:18:46,827 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.14 vs. limit=5.0 2023-04-27 06:19:00,020 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5100, 1.2546, 4.3924, 4.0719, 3.8236, 4.1247, 4.0495, 3.8724], device='cuda:4'), covar=tensor([0.7170, 0.6058, 0.1052, 0.1774, 0.1119, 0.1717, 0.1525, 0.1549], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0303, 0.0400, 0.0401, 0.0346, 0.0402, 0.0309, 0.0361], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 06:19:02,343 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.086e+02 1.530e+02 1.860e+02 2.217e+02 4.388e+02, threshold=3.721e+02, percent-clipped=2.0 2023-04-27 06:19:11,411 INFO [finetune.py:976] (4/7) Epoch 14, batch 5400, loss[loss=0.2099, simple_loss=0.2593, pruned_loss=0.08025, over 4722.00 frames. ], tot_loss[loss=0.1845, simple_loss=0.2535, pruned_loss=0.0578, over 952551.01 frames. ], batch size: 23, lr: 3.52e-03, grad_scale: 16.0 2023-04-27 06:19:23,237 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.13 vs. limit=5.0 2023-04-27 06:19:45,249 INFO [finetune.py:976] (4/7) Epoch 14, batch 5450, loss[loss=0.1595, simple_loss=0.2222, pruned_loss=0.04845, over 4829.00 frames. ], tot_loss[loss=0.1817, simple_loss=0.2502, pruned_loss=0.05661, over 955915.31 frames. ], batch size: 33, lr: 3.52e-03, grad_scale: 16.0 2023-04-27 06:19:50,246 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8892, 4.0634, 0.8012, 2.2235, 2.1713, 2.7015, 2.3106, 0.9920], device='cuda:4'), covar=tensor([0.1391, 0.0999, 0.2306, 0.1328, 0.1084, 0.1144, 0.1561, 0.2170], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0248, 0.0140, 0.0123, 0.0133, 0.0155, 0.0120, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 06:20:04,595 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=79941.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:20:04,613 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=79941.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 06:20:09,137 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.082e+02 1.493e+02 1.714e+02 2.129e+02 3.310e+02, threshold=3.428e+02, percent-clipped=0.0 2023-04-27 06:20:18,681 INFO [finetune.py:976] (4/7) Epoch 14, batch 5500, loss[loss=0.1262, simple_loss=0.2022, pruned_loss=0.02513, over 4777.00 frames. ], tot_loss[loss=0.1792, simple_loss=0.2476, pruned_loss=0.05545, over 954662.67 frames. ], batch size: 28, lr: 3.52e-03, grad_scale: 16.0 2023-04-27 06:20:36,321 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=79989.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 06:20:36,339 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=79989.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:20:53,197 INFO [finetune.py:976] (4/7) Epoch 14, batch 5550, loss[loss=0.1557, simple_loss=0.2177, pruned_loss=0.04684, over 3891.00 frames. ], tot_loss[loss=0.1813, simple_loss=0.2496, pruned_loss=0.05653, over 953741.11 frames. ], batch size: 17, lr: 3.52e-03, grad_scale: 16.0 2023-04-27 06:20:54,495 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=80013.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:20:59,232 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=80020.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:21:04,359 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-04-27 06:21:06,031 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=80031.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 06:21:09,683 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=80037.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:21:16,570 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.402e+01 1.777e+02 2.079e+02 2.504e+02 5.110e+02, threshold=4.158e+02, percent-clipped=3.0 2023-04-27 06:21:24,325 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2186, 2.1053, 1.6927, 1.7100, 2.3046, 1.8244, 2.5988, 1.5050], device='cuda:4'), covar=tensor([0.3960, 0.2058, 0.5245, 0.3114, 0.1673, 0.2575, 0.1732, 0.5070], device='cuda:4'), in_proj_covar=tensor([0.0343, 0.0347, 0.0428, 0.0355, 0.0383, 0.0381, 0.0372, 0.0420], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 06:21:24,824 INFO [finetune.py:976] (4/7) Epoch 14, batch 5600, loss[loss=0.1768, simple_loss=0.2506, pruned_loss=0.05156, over 4867.00 frames. ], tot_loss[loss=0.1832, simple_loss=0.2524, pruned_loss=0.05702, over 952766.27 frames. ], batch size: 44, lr: 3.52e-03, grad_scale: 16.0 2023-04-27 06:21:28,975 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=80068.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:21:35,802 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=80079.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 06:22:01,083 INFO [finetune.py:976] (4/7) Epoch 14, batch 5650, loss[loss=0.2159, simple_loss=0.2888, pruned_loss=0.07155, over 4913.00 frames. ], tot_loss[loss=0.1868, simple_loss=0.2563, pruned_loss=0.05859, over 954443.60 frames. ], batch size: 36, lr: 3.52e-03, grad_scale: 16.0 2023-04-27 06:22:11,109 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=80118.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:22:28,366 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.156e+02 1.596e+02 1.966e+02 2.394e+02 5.142e+02, threshold=3.932e+02, percent-clipped=2.0 2023-04-27 06:22:47,951 INFO [finetune.py:976] (4/7) Epoch 14, batch 5700, loss[loss=0.15, simple_loss=0.2068, pruned_loss=0.04664, over 4255.00 frames. ], tot_loss[loss=0.1862, simple_loss=0.2545, pruned_loss=0.05896, over 936074.83 frames. ], batch size: 18, lr: 3.52e-03, grad_scale: 16.0 2023-04-27 06:22:52,272 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1534, 1.4552, 1.7489, 1.8677, 1.7967, 1.9034, 1.7822, 1.7479], device='cuda:4'), covar=tensor([0.3610, 0.5316, 0.5016, 0.4770, 0.5760, 0.7236, 0.5268, 0.4928], device='cuda:4'), in_proj_covar=tensor([0.0334, 0.0378, 0.0321, 0.0334, 0.0344, 0.0403, 0.0357, 0.0329], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 06:22:57,988 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4564, 2.1938, 2.5767, 2.9248, 2.4855, 2.2365, 2.3861, 2.2786], device='cuda:4'), covar=tensor([0.4711, 0.7123, 0.7415, 0.6092, 0.6031, 0.8861, 0.8903, 0.9518], device='cuda:4'), in_proj_covar=tensor([0.0414, 0.0406, 0.0494, 0.0510, 0.0444, 0.0465, 0.0472, 0.0475], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 06:23:10,468 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-04-27 06:23:36,604 INFO [finetune.py:976] (4/7) Epoch 15, batch 0, loss[loss=0.1522, simple_loss=0.2275, pruned_loss=0.03845, over 4868.00 frames. ], tot_loss[loss=0.1522, simple_loss=0.2275, pruned_loss=0.03845, over 4868.00 frames. ], batch size: 32, lr: 3.52e-03, grad_scale: 16.0 2023-04-27 06:23:36,604 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 06:23:39,650 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9338, 1.8614, 1.6972, 1.3608, 1.9525, 1.5848, 2.2290, 1.5550], device='cuda:4'), covar=tensor([0.3263, 0.1414, 0.4435, 0.2563, 0.1230, 0.1949, 0.1581, 0.4106], device='cuda:4'), in_proj_covar=tensor([0.0343, 0.0346, 0.0427, 0.0355, 0.0381, 0.0381, 0.0372, 0.0419], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 06:23:53,230 INFO [finetune.py:1010] (4/7) Epoch 15, validation: loss=0.1536, simple_loss=0.2258, pruned_loss=0.04063, over 2265189.00 frames. 2023-04-27 06:23:53,231 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-27 06:23:53,377 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9191, 1.4534, 1.7251, 1.7524, 1.7188, 1.4128, 0.7787, 1.3889], device='cuda:4'), covar=tensor([0.3797, 0.3733, 0.2033, 0.2519, 0.2725, 0.2863, 0.4549, 0.2455], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0249, 0.0225, 0.0318, 0.0216, 0.0231, 0.0231, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 06:24:48,179 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4106, 2.8478, 1.0529, 1.5071, 2.3073, 1.4913, 4.1017, 2.0131], device='cuda:4'), covar=tensor([0.0630, 0.0862, 0.0860, 0.1247, 0.0531, 0.0984, 0.0244, 0.0626], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0049, 0.0047, 0.0050, 0.0052, 0.0076, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 06:24:56,494 INFO [finetune.py:976] (4/7) Epoch 15, batch 50, loss[loss=0.2427, simple_loss=0.308, pruned_loss=0.08873, over 4128.00 frames. ], tot_loss[loss=0.1861, simple_loss=0.2575, pruned_loss=0.05732, over 216272.83 frames. ], batch size: 66, lr: 3.52e-03, grad_scale: 16.0 2023-04-27 06:25:05,177 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=80241.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:25:05,231 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0483, 1.4321, 1.8130, 2.0742, 1.8945, 1.4700, 0.9900, 1.5834], device='cuda:4'), covar=tensor([0.3854, 0.4227, 0.2100, 0.2754, 0.2928, 0.3018, 0.4664, 0.2555], device='cuda:4'), in_proj_covar=tensor([0.0289, 0.0249, 0.0225, 0.0317, 0.0216, 0.0230, 0.0231, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 06:25:08,722 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.035e+02 1.528e+02 1.806e+02 2.187e+02 6.322e+02, threshold=3.611e+02, percent-clipped=3.0 2023-04-27 06:25:08,843 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4773, 1.2074, 0.3204, 1.2284, 1.1047, 1.3517, 1.2826, 1.2628], device='cuda:4'), covar=tensor([0.0538, 0.0391, 0.0448, 0.0578, 0.0301, 0.0535, 0.0520, 0.0588], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0024, 0.0023, 0.0029, 0.0020, 0.0028, 0.0028, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0044, 0.0038, 0.0050, 0.0038, 0.0049, 0.0048, 0.0050], device='cuda:4') 2023-04-27 06:25:29,974 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.13 vs. limit=2.0 2023-04-27 06:25:36,888 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=80283.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:25:40,295 INFO [finetune.py:976] (4/7) Epoch 15, batch 100, loss[loss=0.1561, simple_loss=0.2276, pruned_loss=0.04225, over 4820.00 frames. ], tot_loss[loss=0.1809, simple_loss=0.2502, pruned_loss=0.05578, over 380064.91 frames. ], batch size: 41, lr: 3.52e-03, grad_scale: 16.0 2023-04-27 06:25:41,452 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=80289.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:25:55,292 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.63 vs. limit=2.0 2023-04-27 06:25:56,995 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=80313.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:26:12,988 INFO [finetune.py:976] (4/7) Epoch 15, batch 150, loss[loss=0.1772, simple_loss=0.2481, pruned_loss=0.05317, over 4753.00 frames. ], tot_loss[loss=0.1773, simple_loss=0.2452, pruned_loss=0.0547, over 507332.85 frames. ], batch size: 27, lr: 3.51e-03, grad_scale: 16.0 2023-04-27 06:26:13,100 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9072, 1.9902, 0.9658, 1.5645, 2.3577, 1.6842, 1.6124, 1.6980], device='cuda:4'), covar=tensor([0.0505, 0.0366, 0.0315, 0.0551, 0.0225, 0.0498, 0.0496, 0.0576], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0024, 0.0023, 0.0029, 0.0020, 0.0029, 0.0028, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0045, 0.0038, 0.0050, 0.0038, 0.0049, 0.0048, 0.0050], device='cuda:4') 2023-04-27 06:26:18,204 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=80344.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:26:20,364 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.178e+02 1.629e+02 1.838e+02 2.299e+02 3.632e+02, threshold=3.676e+02, percent-clipped=1.0 2023-04-27 06:26:20,502 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0422, 1.9382, 1.7379, 1.5714, 2.2087, 1.6634, 2.5459, 1.5669], device='cuda:4'), covar=tensor([0.3862, 0.1992, 0.4404, 0.3102, 0.1680, 0.2509, 0.1504, 0.4416], device='cuda:4'), in_proj_covar=tensor([0.0343, 0.0348, 0.0429, 0.0357, 0.0383, 0.0382, 0.0373, 0.0422], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 06:26:22,792 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-27 06:26:28,925 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=80361.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:26:30,178 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6800, 1.5358, 0.6581, 1.3307, 1.7369, 1.4811, 1.4034, 1.4768], device='cuda:4'), covar=tensor([0.0493, 0.0376, 0.0381, 0.0561, 0.0278, 0.0490, 0.0486, 0.0565], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0025, 0.0023, 0.0029, 0.0020, 0.0029, 0.0028, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0045, 0.0038, 0.0050, 0.0038, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 06:26:46,348 INFO [finetune.py:976] (4/7) Epoch 15, batch 200, loss[loss=0.2088, simple_loss=0.2772, pruned_loss=0.07016, over 4854.00 frames. ], tot_loss[loss=0.1771, simple_loss=0.2445, pruned_loss=0.05482, over 607120.03 frames. ], batch size: 44, lr: 3.51e-03, grad_scale: 16.0 2023-04-27 06:26:50,420 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=80393.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:27:06,657 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=80418.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:27:18,650 INFO [finetune.py:976] (4/7) Epoch 15, batch 250, loss[loss=0.1784, simple_loss=0.2349, pruned_loss=0.06093, over 4712.00 frames. ], tot_loss[loss=0.1786, simple_loss=0.2466, pruned_loss=0.05535, over 682913.26 frames. ], batch size: 23, lr: 3.51e-03, grad_scale: 16.0 2023-04-27 06:27:25,587 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.144e+02 1.693e+02 2.067e+02 2.411e+02 4.714e+02, threshold=4.133e+02, percent-clipped=3.0 2023-04-27 06:27:31,402 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=80454.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:27:36,892 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=80463.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:27:38,599 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=80466.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:27:51,746 INFO [finetune.py:976] (4/7) Epoch 15, batch 300, loss[loss=0.1418, simple_loss=0.2236, pruned_loss=0.02999, over 4757.00 frames. ], tot_loss[loss=0.1815, simple_loss=0.2501, pruned_loss=0.0564, over 742640.95 frames. ], batch size: 54, lr: 3.51e-03, grad_scale: 16.0 2023-04-27 06:27:56,476 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=80494.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:28:14,216 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1600, 1.8960, 1.6729, 1.5840, 2.0704, 1.6737, 2.2876, 1.5193], device='cuda:4'), covar=tensor([0.3412, 0.1542, 0.3915, 0.2638, 0.1409, 0.1985, 0.1515, 0.4130], device='cuda:4'), in_proj_covar=tensor([0.0345, 0.0350, 0.0432, 0.0357, 0.0385, 0.0384, 0.0375, 0.0424], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 06:28:27,728 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=80524.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:28:28,338 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3913, 3.0353, 2.3346, 2.8456, 2.0736, 2.6036, 2.5712, 1.9306], device='cuda:4'), covar=tensor([0.1911, 0.0997, 0.0822, 0.1104, 0.2992, 0.1111, 0.1864, 0.2777], device='cuda:4'), in_proj_covar=tensor([0.0288, 0.0308, 0.0222, 0.0279, 0.0310, 0.0261, 0.0252, 0.0267], device='cuda:4'), out_proj_covar=tensor([1.1605e-04, 1.2258e-04, 8.8420e-05, 1.1089e-04, 1.2615e-04, 1.0438e-04, 1.0201e-04, 1.0631e-04], device='cuda:4') 2023-04-27 06:28:36,208 INFO [finetune.py:976] (4/7) Epoch 15, batch 350, loss[loss=0.1952, simple_loss=0.2668, pruned_loss=0.06175, over 4914.00 frames. ], tot_loss[loss=0.1827, simple_loss=0.252, pruned_loss=0.05671, over 789851.78 frames. ], batch size: 36, lr: 3.51e-03, grad_scale: 16.0 2023-04-27 06:28:37,554 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0902, 2.5860, 2.1018, 2.4735, 1.8494, 2.1980, 2.0980, 1.6456], device='cuda:4'), covar=tensor([0.1752, 0.1202, 0.0774, 0.1077, 0.2753, 0.0964, 0.1767, 0.2362], device='cuda:4'), in_proj_covar=tensor([0.0287, 0.0307, 0.0221, 0.0278, 0.0309, 0.0261, 0.0252, 0.0266], device='cuda:4'), out_proj_covar=tensor([1.1582e-04, 1.2234e-04, 8.8270e-05, 1.1076e-04, 1.2598e-04, 1.0423e-04, 1.0192e-04, 1.0616e-04], device='cuda:4') 2023-04-27 06:28:42,201 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.160e+02 1.655e+02 1.989e+02 2.509e+02 3.787e+02, threshold=3.978e+02, percent-clipped=0.0 2023-04-27 06:28:48,613 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=80555.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:28:55,269 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-27 06:29:15,647 INFO [finetune.py:976] (4/7) Epoch 15, batch 400, loss[loss=0.1512, simple_loss=0.2205, pruned_loss=0.04088, over 4802.00 frames. ], tot_loss[loss=0.181, simple_loss=0.251, pruned_loss=0.05544, over 826000.05 frames. ], batch size: 25, lr: 3.51e-03, grad_scale: 16.0 2023-04-27 06:30:07,685 INFO [finetune.py:976] (4/7) Epoch 15, batch 450, loss[loss=0.1737, simple_loss=0.239, pruned_loss=0.05425, over 4868.00 frames. ], tot_loss[loss=0.1799, simple_loss=0.2496, pruned_loss=0.05508, over 853615.35 frames. ], batch size: 31, lr: 3.51e-03, grad_scale: 16.0 2023-04-27 06:30:13,733 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=80639.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:30:18,557 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.255e+02 1.612e+02 1.974e+02 2.306e+02 4.694e+02, threshold=3.949e+02, percent-clipped=1.0 2023-04-27 06:31:13,678 INFO [finetune.py:976] (4/7) Epoch 15, batch 500, loss[loss=0.2149, simple_loss=0.2721, pruned_loss=0.07881, over 4901.00 frames. ], tot_loss[loss=0.1783, simple_loss=0.2477, pruned_loss=0.05442, over 877814.46 frames. ], batch size: 43, lr: 3.51e-03, grad_scale: 16.0 2023-04-27 06:31:52,661 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3117, 1.2124, 1.5466, 1.4595, 1.2431, 1.0701, 1.2372, 0.9019], device='cuda:4'), covar=tensor([0.0522, 0.0607, 0.0396, 0.0440, 0.0644, 0.0983, 0.0490, 0.0579], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0070, 0.0069, 0.0068, 0.0075, 0.0096, 0.0075, 0.0069], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 06:31:57,443 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5815, 1.2530, 4.2573, 4.0129, 3.7597, 3.9425, 3.9128, 3.8020], device='cuda:4'), covar=tensor([0.6818, 0.5957, 0.0889, 0.1525, 0.1065, 0.1628, 0.1680, 0.1343], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0307, 0.0401, 0.0403, 0.0349, 0.0404, 0.0312, 0.0361], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 06:32:06,675 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=80736.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:32:07,804 INFO [finetune.py:976] (4/7) Epoch 15, batch 550, loss[loss=0.1974, simple_loss=0.2467, pruned_loss=0.07402, over 4713.00 frames. ], tot_loss[loss=0.177, simple_loss=0.2455, pruned_loss=0.05423, over 895369.37 frames. ], batch size: 23, lr: 3.51e-03, grad_scale: 16.0 2023-04-27 06:32:09,156 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7557, 2.0980, 1.6962, 1.3320, 1.3117, 1.2893, 1.7272, 1.2519], device='cuda:4'), covar=tensor([0.1662, 0.1260, 0.1490, 0.1889, 0.2353, 0.2033, 0.1056, 0.2046], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0212, 0.0168, 0.0203, 0.0200, 0.0184, 0.0156, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 06:32:13,248 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.128e+02 1.610e+02 1.836e+02 2.201e+02 4.321e+02, threshold=3.672e+02, percent-clipped=1.0 2023-04-27 06:32:14,547 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=80749.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:32:32,589 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=80773.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:32:41,567 INFO [finetune.py:976] (4/7) Epoch 15, batch 600, loss[loss=0.1778, simple_loss=0.2432, pruned_loss=0.05619, over 4940.00 frames. ], tot_loss[loss=0.1775, simple_loss=0.2459, pruned_loss=0.05457, over 909422.23 frames. ], batch size: 38, lr: 3.51e-03, grad_scale: 16.0 2023-04-27 06:32:47,160 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=80797.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:32:59,002 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3663, 3.0593, 2.4447, 2.8112, 2.2120, 2.6383, 2.6289, 1.9921], device='cuda:4'), covar=tensor([0.2448, 0.1322, 0.0867, 0.1294, 0.3374, 0.1240, 0.2243, 0.2987], device='cuda:4'), in_proj_covar=tensor([0.0291, 0.0312, 0.0224, 0.0281, 0.0312, 0.0263, 0.0254, 0.0270], device='cuda:4'), out_proj_covar=tensor([1.1722e-04, 1.2427e-04, 8.9326e-05, 1.1186e-04, 1.2717e-04, 1.0513e-04, 1.0298e-04, 1.0766e-04], device='cuda:4') 2023-04-27 06:33:03,035 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=80819.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:33:12,689 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=80834.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:33:15,003 INFO [finetune.py:976] (4/7) Epoch 15, batch 650, loss[loss=0.1815, simple_loss=0.2585, pruned_loss=0.05226, over 4744.00 frames. ], tot_loss[loss=0.1804, simple_loss=0.2495, pruned_loss=0.05567, over 919849.70 frames. ], batch size: 27, lr: 3.51e-03, grad_scale: 16.0 2023-04-27 06:33:19,960 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8941, 2.7356, 1.9237, 2.1182, 1.4502, 1.4302, 2.0764, 1.3924], device='cuda:4'), covar=tensor([0.1720, 0.1465, 0.1541, 0.1837, 0.2475, 0.2111, 0.1065, 0.2184], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0213, 0.0169, 0.0204, 0.0201, 0.0185, 0.0157, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 06:33:20,403 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.224e+02 1.693e+02 1.997e+02 2.392e+02 3.553e+02, threshold=3.994e+02, percent-clipped=0.0 2023-04-27 06:33:22,321 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=80850.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:33:48,173 INFO [finetune.py:976] (4/7) Epoch 15, batch 700, loss[loss=0.2445, simple_loss=0.2938, pruned_loss=0.09756, over 4737.00 frames. ], tot_loss[loss=0.1824, simple_loss=0.2512, pruned_loss=0.0568, over 925696.37 frames. ], batch size: 59, lr: 3.51e-03, grad_scale: 16.0 2023-04-27 06:34:21,750 INFO [finetune.py:976] (4/7) Epoch 15, batch 750, loss[loss=0.2057, simple_loss=0.279, pruned_loss=0.06624, over 4905.00 frames. ], tot_loss[loss=0.1841, simple_loss=0.253, pruned_loss=0.05757, over 931754.87 frames. ], batch size: 37, lr: 3.51e-03, grad_scale: 16.0 2023-04-27 06:34:22,440 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=80939.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:34:27,259 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.065e+02 1.649e+02 1.974e+02 2.385e+02 5.578e+02, threshold=3.948e+02, percent-clipped=5.0 2023-04-27 06:34:54,975 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=80987.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:34:55,529 INFO [finetune.py:976] (4/7) Epoch 15, batch 800, loss[loss=0.1994, simple_loss=0.2697, pruned_loss=0.0645, over 4907.00 frames. ], tot_loss[loss=0.1841, simple_loss=0.2537, pruned_loss=0.05723, over 937581.93 frames. ], batch size: 37, lr: 3.51e-03, grad_scale: 16.0 2023-04-27 06:35:35,157 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1646, 1.9765, 2.2618, 2.5857, 2.6130, 2.1542, 1.8192, 2.3561], device='cuda:4'), covar=tensor([0.0910, 0.1082, 0.0598, 0.0576, 0.0553, 0.0769, 0.0738, 0.0543], device='cuda:4'), in_proj_covar=tensor([0.0193, 0.0205, 0.0185, 0.0174, 0.0180, 0.0185, 0.0156, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 06:35:35,655 INFO [finetune.py:976] (4/7) Epoch 15, batch 850, loss[loss=0.2002, simple_loss=0.2588, pruned_loss=0.07077, over 4764.00 frames. ], tot_loss[loss=0.1818, simple_loss=0.2509, pruned_loss=0.05641, over 942880.79 frames. ], batch size: 27, lr: 3.51e-03, grad_scale: 16.0 2023-04-27 06:35:46,468 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.971e+01 1.581e+02 1.945e+02 2.309e+02 3.593e+02, threshold=3.889e+02, percent-clipped=0.0 2023-04-27 06:35:47,793 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=81049.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:36:41,755 INFO [finetune.py:976] (4/7) Epoch 15, batch 900, loss[loss=0.1661, simple_loss=0.2396, pruned_loss=0.04631, over 4729.00 frames. ], tot_loss[loss=0.1802, simple_loss=0.2489, pruned_loss=0.05575, over 946578.34 frames. ], batch size: 23, lr: 3.51e-03, grad_scale: 16.0 2023-04-27 06:36:49,897 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=81092.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:36:52,939 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=81097.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:37:15,233 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=81116.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:37:22,537 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=81119.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:37:34,999 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=81129.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:37:48,250 INFO [finetune.py:976] (4/7) Epoch 15, batch 950, loss[loss=0.1731, simple_loss=0.2416, pruned_loss=0.05227, over 4737.00 frames. ], tot_loss[loss=0.1787, simple_loss=0.2469, pruned_loss=0.05527, over 949901.02 frames. ], batch size: 59, lr: 3.51e-03, grad_scale: 16.0 2023-04-27 06:38:04,672 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.127e+02 1.625e+02 2.019e+02 2.368e+02 4.092e+02, threshold=4.038e+02, percent-clipped=2.0 2023-04-27 06:38:06,677 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=81150.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:38:17,044 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=81167.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:38:23,605 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=81177.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:38:32,221 INFO [finetune.py:976] (4/7) Epoch 15, batch 1000, loss[loss=0.1567, simple_loss=0.2326, pruned_loss=0.04042, over 4783.00 frames. ], tot_loss[loss=0.1796, simple_loss=0.2479, pruned_loss=0.0557, over 950290.92 frames. ], batch size: 28, lr: 3.51e-03, grad_scale: 16.0 2023-04-27 06:38:38,869 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=81198.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:38:39,523 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9188, 2.4818, 0.9861, 1.3434, 1.8932, 1.2429, 3.1648, 1.5487], device='cuda:4'), covar=tensor([0.0694, 0.0620, 0.0787, 0.1302, 0.0515, 0.1019, 0.0254, 0.0706], device='cuda:4'), in_proj_covar=tensor([0.0052, 0.0067, 0.0049, 0.0047, 0.0050, 0.0053, 0.0077, 0.0052], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 06:39:03,039 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4375, 1.6421, 1.6984, 1.8955, 1.7139, 1.8329, 1.7692, 1.7431], device='cuda:4'), covar=tensor([0.3866, 0.5757, 0.4597, 0.4691, 0.5941, 0.7974, 0.5851, 0.5092], device='cuda:4'), in_proj_covar=tensor([0.0329, 0.0374, 0.0317, 0.0331, 0.0341, 0.0399, 0.0355, 0.0326], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 06:39:05,803 INFO [finetune.py:976] (4/7) Epoch 15, batch 1050, loss[loss=0.1512, simple_loss=0.2203, pruned_loss=0.0411, over 4852.00 frames. ], tot_loss[loss=0.1807, simple_loss=0.2502, pruned_loss=0.0556, over 951975.94 frames. ], batch size: 31, lr: 3.51e-03, grad_scale: 16.0 2023-04-27 06:39:11,771 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.267e+02 1.726e+02 1.946e+02 2.343e+02 3.308e+02, threshold=3.891e+02, percent-clipped=0.0 2023-04-27 06:39:38,633 INFO [finetune.py:976] (4/7) Epoch 15, batch 1100, loss[loss=0.1657, simple_loss=0.245, pruned_loss=0.04324, over 4859.00 frames. ], tot_loss[loss=0.1823, simple_loss=0.2526, pruned_loss=0.05595, over 954235.15 frames. ], batch size: 44, lr: 3.51e-03, grad_scale: 16.0 2023-04-27 06:39:46,742 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=81299.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:39:47,928 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=81301.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:39:53,398 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0883, 1.7408, 2.0051, 2.4759, 2.4922, 1.9803, 1.6572, 1.9895], device='cuda:4'), covar=tensor([0.0818, 0.1057, 0.0698, 0.0521, 0.0579, 0.0774, 0.0766, 0.0588], device='cuda:4'), in_proj_covar=tensor([0.0192, 0.0204, 0.0184, 0.0173, 0.0180, 0.0184, 0.0155, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 06:40:11,235 INFO [finetune.py:976] (4/7) Epoch 15, batch 1150, loss[loss=0.1711, simple_loss=0.2326, pruned_loss=0.05474, over 4665.00 frames. ], tot_loss[loss=0.1833, simple_loss=0.254, pruned_loss=0.05636, over 953868.87 frames. ], batch size: 23, lr: 3.51e-03, grad_scale: 16.0 2023-04-27 06:40:18,548 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.629e+02 1.924e+02 2.399e+02 3.865e+02, threshold=3.848e+02, percent-clipped=0.0 2023-04-27 06:40:27,088 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=81360.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:40:28,313 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=81362.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:40:44,974 INFO [finetune.py:976] (4/7) Epoch 15, batch 1200, loss[loss=0.1746, simple_loss=0.2301, pruned_loss=0.05958, over 4715.00 frames. ], tot_loss[loss=0.1823, simple_loss=0.2524, pruned_loss=0.05606, over 953899.20 frames. ], batch size: 59, lr: 3.51e-03, grad_scale: 16.0 2023-04-27 06:40:48,495 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=81392.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:40:53,698 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1502, 1.6596, 2.0295, 2.4538, 2.0467, 1.5823, 1.2209, 1.8233], device='cuda:4'), covar=tensor([0.2931, 0.3062, 0.1563, 0.2049, 0.2417, 0.2619, 0.4255, 0.1917], device='cuda:4'), in_proj_covar=tensor([0.0287, 0.0245, 0.0222, 0.0314, 0.0213, 0.0229, 0.0228, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 06:41:12,546 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=81429.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:41:18,367 INFO [finetune.py:976] (4/7) Epoch 15, batch 1250, loss[loss=0.1541, simple_loss=0.2092, pruned_loss=0.04954, over 3976.00 frames. ], tot_loss[loss=0.1801, simple_loss=0.2496, pruned_loss=0.05532, over 954317.99 frames. ], batch size: 17, lr: 3.51e-03, grad_scale: 16.0 2023-04-27 06:41:20,060 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=81440.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:41:31,409 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.048e+02 1.659e+02 1.910e+02 2.320e+02 4.141e+02, threshold=3.819e+02, percent-clipped=2.0 2023-04-27 06:42:05,045 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=81472.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:42:08,091 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=81477.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:42:20,541 INFO [finetune.py:976] (4/7) Epoch 15, batch 1300, loss[loss=0.1891, simple_loss=0.2489, pruned_loss=0.06472, over 4342.00 frames. ], tot_loss[loss=0.1781, simple_loss=0.2468, pruned_loss=0.05464, over 954868.91 frames. ], batch size: 65, lr: 3.50e-03, grad_scale: 32.0 2023-04-27 06:43:12,244 INFO [finetune.py:976] (4/7) Epoch 15, batch 1350, loss[loss=0.2039, simple_loss=0.2805, pruned_loss=0.06366, over 4866.00 frames. ], tot_loss[loss=0.1778, simple_loss=0.2469, pruned_loss=0.0544, over 953285.60 frames. ], batch size: 44, lr: 3.50e-03, grad_scale: 32.0 2023-04-27 06:43:24,613 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.020e+02 1.644e+02 2.016e+02 2.436e+02 4.078e+02, threshold=4.033e+02, percent-clipped=1.0 2023-04-27 06:44:07,695 INFO [finetune.py:976] (4/7) Epoch 15, batch 1400, loss[loss=0.1973, simple_loss=0.2775, pruned_loss=0.05855, over 4903.00 frames. ], tot_loss[loss=0.1793, simple_loss=0.2491, pruned_loss=0.05478, over 953955.51 frames. ], batch size: 43, lr: 3.50e-03, grad_scale: 32.0 2023-04-27 06:44:10,757 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7941, 2.2948, 1.0237, 1.2592, 1.5984, 1.1972, 2.4671, 1.3485], device='cuda:4'), covar=tensor([0.0711, 0.0622, 0.0610, 0.1227, 0.0427, 0.1005, 0.0290, 0.0700], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0048, 0.0047, 0.0050, 0.0052, 0.0076, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 06:44:27,735 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-27 06:44:37,186 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.36 vs. limit=5.0 2023-04-27 06:44:41,266 INFO [finetune.py:976] (4/7) Epoch 15, batch 1450, loss[loss=0.2119, simple_loss=0.2821, pruned_loss=0.07088, over 4858.00 frames. ], tot_loss[loss=0.1809, simple_loss=0.2512, pruned_loss=0.05532, over 954342.14 frames. ], batch size: 44, lr: 3.50e-03, grad_scale: 32.0 2023-04-27 06:44:47,198 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.153e+02 1.626e+02 2.006e+02 2.422e+02 4.010e+02, threshold=4.013e+02, percent-clipped=0.0 2023-04-27 06:44:53,611 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=81655.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:44:55,285 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=81657.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:45:04,224 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=81670.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:45:15,084 INFO [finetune.py:976] (4/7) Epoch 15, batch 1500, loss[loss=0.169, simple_loss=0.2478, pruned_loss=0.04516, over 4787.00 frames. ], tot_loss[loss=0.181, simple_loss=0.2516, pruned_loss=0.05514, over 956217.25 frames. ], batch size: 29, lr: 3.50e-03, grad_scale: 32.0 2023-04-27 06:45:17,600 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9809, 2.6418, 2.0132, 2.1257, 1.4329, 1.4489, 2.2059, 1.4379], device='cuda:4'), covar=tensor([0.1582, 0.1423, 0.1447, 0.1639, 0.2348, 0.1925, 0.0928, 0.2035], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0212, 0.0168, 0.0203, 0.0199, 0.0183, 0.0155, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 06:45:44,065 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=81731.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 06:45:48,176 INFO [finetune.py:976] (4/7) Epoch 15, batch 1550, loss[loss=0.1907, simple_loss=0.2598, pruned_loss=0.06082, over 4919.00 frames. ], tot_loss[loss=0.181, simple_loss=0.2518, pruned_loss=0.05504, over 954705.54 frames. ], batch size: 38, lr: 3.50e-03, grad_scale: 32.0 2023-04-27 06:45:53,673 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.064e+02 1.553e+02 1.872e+02 2.288e+02 6.577e+02, threshold=3.744e+02, percent-clipped=2.0 2023-04-27 06:46:04,669 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0658, 1.5644, 1.5297, 1.9537, 2.2356, 1.8979, 1.7903, 1.5562], device='cuda:4'), covar=tensor([0.1697, 0.1813, 0.1530, 0.1780, 0.1050, 0.2001, 0.2040, 0.1859], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0318, 0.0353, 0.0292, 0.0330, 0.0314, 0.0304, 0.0362], device='cuda:4'), out_proj_covar=tensor([6.4331e-05, 6.6512e-05, 7.5552e-05, 5.9593e-05, 6.8614e-05, 6.6450e-05, 6.4297e-05, 7.7448e-05], device='cuda:4') 2023-04-27 06:46:11,793 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=81772.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:46:21,441 INFO [finetune.py:976] (4/7) Epoch 15, batch 1600, loss[loss=0.1718, simple_loss=0.2415, pruned_loss=0.05106, over 4903.00 frames. ], tot_loss[loss=0.1796, simple_loss=0.2496, pruned_loss=0.05478, over 956306.94 frames. ], batch size: 36, lr: 3.50e-03, grad_scale: 32.0 2023-04-27 06:46:30,145 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7143, 1.3819, 1.8970, 2.0659, 1.7873, 1.7281, 1.8157, 1.8197], device='cuda:4'), covar=tensor([0.6033, 0.8141, 0.8164, 0.9028, 0.7539, 1.0172, 0.9940, 0.9467], device='cuda:4'), in_proj_covar=tensor([0.0415, 0.0407, 0.0494, 0.0507, 0.0443, 0.0466, 0.0473, 0.0475], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 06:46:42,457 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=81820.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:46:53,853 INFO [finetune.py:976] (4/7) Epoch 15, batch 1650, loss[loss=0.1677, simple_loss=0.2296, pruned_loss=0.05286, over 4933.00 frames. ], tot_loss[loss=0.1766, simple_loss=0.2461, pruned_loss=0.05356, over 956664.23 frames. ], batch size: 38, lr: 3.50e-03, grad_scale: 32.0 2023-04-27 06:47:04,624 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.076e+02 1.625e+02 1.858e+02 2.341e+02 4.893e+02, threshold=3.717e+02, percent-clipped=2.0 2023-04-27 06:47:37,585 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.32 vs. limit=5.0 2023-04-27 06:47:59,314 INFO [finetune.py:976] (4/7) Epoch 15, batch 1700, loss[loss=0.1581, simple_loss=0.2215, pruned_loss=0.04739, over 4183.00 frames. ], tot_loss[loss=0.1748, simple_loss=0.2439, pruned_loss=0.05287, over 955345.49 frames. ], batch size: 18, lr: 3.50e-03, grad_scale: 32.0 2023-04-27 06:49:06,084 INFO [finetune.py:976] (4/7) Epoch 15, batch 1750, loss[loss=0.1669, simple_loss=0.2293, pruned_loss=0.0522, over 4057.00 frames. ], tot_loss[loss=0.178, simple_loss=0.2471, pruned_loss=0.05445, over 953849.64 frames. ], batch size: 17, lr: 3.50e-03, grad_scale: 32.0 2023-04-27 06:49:16,875 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.080e+01 1.647e+02 2.011e+02 2.347e+02 5.729e+02, threshold=4.021e+02, percent-clipped=2.0 2023-04-27 06:49:26,285 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=81955.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:49:27,546 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=81957.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:49:27,595 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6948, 1.2940, 1.8488, 2.1774, 1.8012, 1.6789, 1.7443, 1.6949], device='cuda:4'), covar=tensor([0.4988, 0.7558, 0.7278, 0.6696, 0.6374, 0.8961, 0.9172, 0.9790], device='cuda:4'), in_proj_covar=tensor([0.0415, 0.0407, 0.0494, 0.0507, 0.0443, 0.0466, 0.0473, 0.0475], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 06:49:31,747 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5497, 1.4624, 0.6281, 1.2187, 1.3991, 1.4253, 1.3208, 1.3133], device='cuda:4'), covar=tensor([0.0482, 0.0363, 0.0385, 0.0551, 0.0298, 0.0496, 0.0487, 0.0564], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0024, 0.0023, 0.0029, 0.0020, 0.0029, 0.0028, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0045, 0.0038, 0.0051, 0.0038, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 06:49:41,242 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3335, 1.6436, 1.7749, 1.8650, 1.7298, 1.8356, 1.8695, 1.8352], device='cuda:4'), covar=tensor([0.4687, 0.6113, 0.5064, 0.4941, 0.5994, 0.8032, 0.5931, 0.5603], device='cuda:4'), in_proj_covar=tensor([0.0329, 0.0375, 0.0319, 0.0334, 0.0343, 0.0400, 0.0354, 0.0327], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 06:49:48,405 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6046, 1.4145, 4.4599, 4.2100, 3.8992, 4.2135, 4.1039, 3.9383], device='cuda:4'), covar=tensor([0.7510, 0.5866, 0.0985, 0.1567, 0.1155, 0.1522, 0.1324, 0.1500], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0308, 0.0403, 0.0405, 0.0349, 0.0405, 0.0315, 0.0363], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 06:49:48,433 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=81987.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:49:48,925 INFO [finetune.py:976] (4/7) Epoch 15, batch 1800, loss[loss=0.189, simple_loss=0.2591, pruned_loss=0.05938, over 4916.00 frames. ], tot_loss[loss=0.1811, simple_loss=0.2513, pruned_loss=0.05544, over 954896.78 frames. ], batch size: 37, lr: 3.50e-03, grad_scale: 32.0 2023-04-27 06:49:59,485 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=82003.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:50:00,739 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=82005.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:50:16,417 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=82026.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 06:50:24,045 INFO [finetune.py:976] (4/7) Epoch 15, batch 1850, loss[loss=0.206, simple_loss=0.2771, pruned_loss=0.06748, over 4698.00 frames. ], tot_loss[loss=0.1826, simple_loss=0.2531, pruned_loss=0.05608, over 954395.04 frames. ], batch size: 59, lr: 3.50e-03, grad_scale: 32.0 2023-04-27 06:50:26,603 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4003, 1.7720, 2.2280, 2.8994, 2.2830, 1.7538, 1.7231, 2.0902], device='cuda:4'), covar=tensor([0.2957, 0.3310, 0.1594, 0.2437, 0.2700, 0.2702, 0.3942, 0.2078], device='cuda:4'), in_proj_covar=tensor([0.0287, 0.0244, 0.0222, 0.0313, 0.0214, 0.0228, 0.0228, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 06:50:29,487 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.115e+02 1.760e+02 1.989e+02 2.247e+02 4.902e+02, threshold=3.979e+02, percent-clipped=1.0 2023-04-27 06:50:30,238 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=82048.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:50:34,527 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=82055.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:50:56,617 INFO [finetune.py:976] (4/7) Epoch 15, batch 1900, loss[loss=0.1777, simple_loss=0.2473, pruned_loss=0.05404, over 4739.00 frames. ], tot_loss[loss=0.1826, simple_loss=0.2532, pruned_loss=0.05598, over 954505.48 frames. ], batch size: 27, lr: 3.50e-03, grad_scale: 32.0 2023-04-27 06:51:14,317 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=82116.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:51:22,456 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.75 vs. limit=2.0 2023-04-27 06:51:30,519 INFO [finetune.py:976] (4/7) Epoch 15, batch 1950, loss[loss=0.2034, simple_loss=0.2662, pruned_loss=0.07027, over 4872.00 frames. ], tot_loss[loss=0.1817, simple_loss=0.252, pruned_loss=0.05573, over 954232.57 frames. ], batch size: 34, lr: 3.50e-03, grad_scale: 32.0 2023-04-27 06:51:36,462 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.029e+02 1.535e+02 1.831e+02 2.341e+02 3.747e+02, threshold=3.661e+02, percent-clipped=0.0 2023-04-27 06:51:39,240 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.67 vs. limit=2.0 2023-04-27 06:51:44,956 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7274, 1.6033, 1.7720, 2.0561, 2.1324, 1.7152, 1.3314, 1.8911], device='cuda:4'), covar=tensor([0.0835, 0.1201, 0.0745, 0.0608, 0.0582, 0.0800, 0.0806, 0.0572], device='cuda:4'), in_proj_covar=tensor([0.0192, 0.0203, 0.0183, 0.0174, 0.0179, 0.0183, 0.0155, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 06:52:04,206 INFO [finetune.py:976] (4/7) Epoch 15, batch 2000, loss[loss=0.1504, simple_loss=0.2072, pruned_loss=0.04679, over 4712.00 frames. ], tot_loss[loss=0.18, simple_loss=0.2497, pruned_loss=0.05518, over 954549.30 frames. ], batch size: 23, lr: 3.50e-03, grad_scale: 32.0 2023-04-27 06:52:06,805 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8330, 1.2686, 4.8615, 4.5420, 4.2030, 4.6384, 4.2633, 4.2788], device='cuda:4'), covar=tensor([0.7729, 0.6523, 0.0904, 0.1643, 0.1106, 0.1249, 0.1951, 0.1526], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0308, 0.0400, 0.0404, 0.0347, 0.0405, 0.0313, 0.0361], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 06:52:23,971 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5948, 1.8053, 1.8931, 2.0371, 1.7960, 1.9498, 2.0676, 2.0227], device='cuda:4'), covar=tensor([0.3673, 0.5539, 0.4965, 0.4675, 0.5977, 0.7628, 0.5399, 0.5240], device='cuda:4'), in_proj_covar=tensor([0.0328, 0.0374, 0.0318, 0.0334, 0.0343, 0.0400, 0.0353, 0.0327], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 06:52:33,079 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-27 06:52:37,368 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3401, 2.0933, 1.8679, 1.7963, 2.1674, 1.7737, 2.3966, 1.7108], device='cuda:4'), covar=tensor([0.3423, 0.1605, 0.3801, 0.2788, 0.1731, 0.2213, 0.1953, 0.3726], device='cuda:4'), in_proj_covar=tensor([0.0343, 0.0348, 0.0427, 0.0355, 0.0383, 0.0383, 0.0374, 0.0418], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 06:52:37,842 INFO [finetune.py:976] (4/7) Epoch 15, batch 2050, loss[loss=0.1608, simple_loss=0.2299, pruned_loss=0.04586, over 4254.00 frames. ], tot_loss[loss=0.1787, simple_loss=0.2476, pruned_loss=0.05486, over 954220.49 frames. ], batch size: 65, lr: 3.50e-03, grad_scale: 32.0 2023-04-27 06:52:43,310 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.160e+02 1.629e+02 1.931e+02 2.371e+02 3.947e+02, threshold=3.861e+02, percent-clipped=2.0 2023-04-27 06:52:54,920 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0648, 1.0339, 1.2534, 1.1567, 1.0511, 0.9401, 0.9831, 0.5485], device='cuda:4'), covar=tensor([0.0603, 0.0563, 0.0500, 0.0553, 0.0740, 0.1199, 0.0494, 0.0785], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0069, 0.0069, 0.0067, 0.0075, 0.0095, 0.0074, 0.0068], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 06:53:21,146 INFO [finetune.py:976] (4/7) Epoch 15, batch 2100, loss[loss=0.1959, simple_loss=0.2528, pruned_loss=0.0695, over 4897.00 frames. ], tot_loss[loss=0.1773, simple_loss=0.246, pruned_loss=0.05425, over 954728.69 frames. ], batch size: 32, lr: 3.50e-03, grad_scale: 32.0 2023-04-27 06:54:04,937 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=82326.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:54:13,976 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6233, 1.2012, 1.6641, 2.0550, 1.6826, 1.5688, 1.6867, 1.6239], device='cuda:4'), covar=tensor([0.4831, 0.6679, 0.6833, 0.6409, 0.6180, 0.8390, 0.8510, 0.8699], device='cuda:4'), in_proj_covar=tensor([0.0417, 0.0407, 0.0498, 0.0511, 0.0445, 0.0468, 0.0476, 0.0479], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 06:54:24,279 INFO [finetune.py:976] (4/7) Epoch 15, batch 2150, loss[loss=0.1911, simple_loss=0.2621, pruned_loss=0.06006, over 4841.00 frames. ], tot_loss[loss=0.1798, simple_loss=0.2492, pruned_loss=0.05513, over 954532.90 frames. ], batch size: 44, lr: 3.50e-03, grad_scale: 32.0 2023-04-27 06:54:33,154 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=82343.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:54:35,048 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=82346.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:54:35,530 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.316e+02 1.732e+02 1.970e+02 2.465e+02 9.359e+02, threshold=3.940e+02, percent-clipped=1.0 2023-04-27 06:54:37,270 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-27 06:54:46,140 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4458, 3.3476, 0.8318, 1.7392, 1.9295, 2.3397, 1.8681, 1.0822], device='cuda:4'), covar=tensor([0.1519, 0.1217, 0.2125, 0.1317, 0.1063, 0.1096, 0.1706, 0.1891], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0246, 0.0137, 0.0121, 0.0132, 0.0154, 0.0118, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 06:54:58,341 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=82374.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:55:13,104 INFO [finetune.py:976] (4/7) Epoch 15, batch 2200, loss[loss=0.1865, simple_loss=0.2606, pruned_loss=0.0562, over 4751.00 frames. ], tot_loss[loss=0.1802, simple_loss=0.2501, pruned_loss=0.05515, over 954796.86 frames. ], batch size: 54, lr: 3.50e-03, grad_scale: 32.0 2023-04-27 06:55:42,648 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=82407.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:55:45,053 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=82411.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:55:54,021 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9707, 1.4445, 1.8458, 2.1683, 1.8220, 1.4800, 1.0120, 1.6003], device='cuda:4'), covar=tensor([0.3215, 0.3297, 0.1661, 0.2249, 0.2477, 0.2644, 0.4552, 0.2085], device='cuda:4'), in_proj_covar=tensor([0.0288, 0.0245, 0.0223, 0.0315, 0.0214, 0.0228, 0.0229, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 06:56:12,257 INFO [finetune.py:976] (4/7) Epoch 15, batch 2250, loss[loss=0.2183, simple_loss=0.2844, pruned_loss=0.07613, over 4888.00 frames. ], tot_loss[loss=0.182, simple_loss=0.2522, pruned_loss=0.05591, over 954891.78 frames. ], batch size: 35, lr: 3.50e-03, grad_scale: 32.0 2023-04-27 06:56:19,196 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.169e+02 1.596e+02 1.887e+02 2.124e+02 5.729e+02, threshold=3.773e+02, percent-clipped=1.0 2023-04-27 06:56:28,081 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5403, 2.0816, 1.7004, 1.9457, 1.5882, 1.6663, 1.6461, 1.2555], device='cuda:4'), covar=tensor([0.1967, 0.1398, 0.1006, 0.1405, 0.3351, 0.1409, 0.2174, 0.2996], device='cuda:4'), in_proj_covar=tensor([0.0289, 0.0311, 0.0223, 0.0284, 0.0313, 0.0266, 0.0255, 0.0269], device='cuda:4'), out_proj_covar=tensor([1.1637e-04, 1.2348e-04, 8.8891e-05, 1.1298e-04, 1.2754e-04, 1.0607e-04, 1.0326e-04, 1.0707e-04], device='cuda:4') 2023-04-27 06:56:45,463 INFO [finetune.py:976] (4/7) Epoch 15, batch 2300, loss[loss=0.1696, simple_loss=0.2424, pruned_loss=0.04844, over 4722.00 frames. ], tot_loss[loss=0.1825, simple_loss=0.2528, pruned_loss=0.05604, over 953584.07 frames. ], batch size: 59, lr: 3.50e-03, grad_scale: 32.0 2023-04-27 06:56:49,021 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=82493.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 06:57:00,256 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0919, 0.7540, 0.9432, 0.8220, 1.2441, 0.9728, 0.8325, 0.9705], device='cuda:4'), covar=tensor([0.1685, 0.1411, 0.1945, 0.1497, 0.0976, 0.1388, 0.1555, 0.2228], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0318, 0.0353, 0.0293, 0.0331, 0.0315, 0.0304, 0.0362], device='cuda:4'), out_proj_covar=tensor([6.4100e-05, 6.6549e-05, 7.5616e-05, 5.9970e-05, 6.8880e-05, 6.6744e-05, 6.4416e-05, 7.7380e-05], device='cuda:4') 2023-04-27 06:57:12,851 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9401, 1.9223, 1.7111, 1.6198, 2.0388, 1.6790, 2.4212, 1.5779], device='cuda:4'), covar=tensor([0.3436, 0.1788, 0.4466, 0.2562, 0.1500, 0.2185, 0.1549, 0.3971], device='cuda:4'), in_proj_covar=tensor([0.0345, 0.0349, 0.0430, 0.0357, 0.0384, 0.0386, 0.0375, 0.0422], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 06:57:18,140 INFO [finetune.py:976] (4/7) Epoch 15, batch 2350, loss[loss=0.2035, simple_loss=0.2582, pruned_loss=0.07438, over 4219.00 frames. ], tot_loss[loss=0.1815, simple_loss=0.251, pruned_loss=0.05605, over 954274.99 frames. ], batch size: 66, lr: 3.50e-03, grad_scale: 32.0 2023-04-27 06:57:25,044 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.877e+01 1.633e+02 1.887e+02 2.224e+02 3.641e+02, threshold=3.774e+02, percent-clipped=0.0 2023-04-27 06:57:30,338 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=82554.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 06:57:38,099 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.28 vs. limit=5.0 2023-04-27 06:57:39,322 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-27 06:57:40,468 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=82569.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:57:51,900 INFO [finetune.py:976] (4/7) Epoch 15, batch 2400, loss[loss=0.1879, simple_loss=0.2578, pruned_loss=0.05901, over 4915.00 frames. ], tot_loss[loss=0.1791, simple_loss=0.2481, pruned_loss=0.05505, over 954253.27 frames. ], batch size: 46, lr: 3.50e-03, grad_scale: 32.0 2023-04-27 06:57:52,659 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4062, 1.8761, 2.4298, 2.9265, 2.3081, 1.9241, 1.8123, 2.3876], device='cuda:4'), covar=tensor([0.3487, 0.3385, 0.1669, 0.2963, 0.2914, 0.2852, 0.4285, 0.2327], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0249, 0.0225, 0.0318, 0.0217, 0.0231, 0.0232, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 06:58:10,800 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9167, 1.6982, 1.8436, 2.2856, 2.3011, 1.8450, 1.5264, 2.0912], device='cuda:4'), covar=tensor([0.0862, 0.1180, 0.0713, 0.0577, 0.0584, 0.0799, 0.0815, 0.0533], device='cuda:4'), in_proj_covar=tensor([0.0190, 0.0202, 0.0181, 0.0172, 0.0177, 0.0182, 0.0154, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 06:58:21,020 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=82630.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:58:25,779 INFO [finetune.py:976] (4/7) Epoch 15, batch 2450, loss[loss=0.1972, simple_loss=0.2614, pruned_loss=0.06652, over 4905.00 frames. ], tot_loss[loss=0.1777, simple_loss=0.2459, pruned_loss=0.05476, over 954252.83 frames. ], batch size: 36, lr: 3.49e-03, grad_scale: 32.0 2023-04-27 06:58:28,856 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=82643.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:58:31,168 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.750e+01 1.533e+02 1.809e+02 2.233e+02 3.484e+02, threshold=3.618e+02, percent-clipped=0.0 2023-04-27 06:58:50,612 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2995, 1.4804, 1.7133, 1.8636, 1.7150, 1.8047, 1.8294, 1.7723], device='cuda:4'), covar=tensor([0.4296, 0.5877, 0.4681, 0.4485, 0.6047, 0.7600, 0.5244, 0.5042], device='cuda:4'), in_proj_covar=tensor([0.0326, 0.0370, 0.0315, 0.0331, 0.0339, 0.0396, 0.0351, 0.0323], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 06:59:04,415 INFO [finetune.py:976] (4/7) Epoch 15, batch 2500, loss[loss=0.1581, simple_loss=0.219, pruned_loss=0.0486, over 4753.00 frames. ], tot_loss[loss=0.1797, simple_loss=0.2481, pruned_loss=0.05559, over 954936.39 frames. ], batch size: 23, lr: 3.49e-03, grad_scale: 32.0 2023-04-27 06:59:11,879 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=82691.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:59:25,007 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=82702.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 06:59:37,992 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=82711.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:00:09,107 INFO [finetune.py:976] (4/7) Epoch 15, batch 2550, loss[loss=0.2159, simple_loss=0.2824, pruned_loss=0.07471, over 4834.00 frames. ], tot_loss[loss=0.1798, simple_loss=0.2492, pruned_loss=0.05517, over 955658.18 frames. ], batch size: 30, lr: 3.49e-03, grad_scale: 32.0 2023-04-27 07:00:14,538 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.081e+02 1.661e+02 1.980e+02 2.344e+02 4.065e+02, threshold=3.961e+02, percent-clipped=3.0 2023-04-27 07:00:24,401 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=82759.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:00:47,935 INFO [finetune.py:976] (4/7) Epoch 15, batch 2600, loss[loss=0.2243, simple_loss=0.2858, pruned_loss=0.08138, over 4759.00 frames. ], tot_loss[loss=0.1821, simple_loss=0.2517, pruned_loss=0.05625, over 956040.37 frames. ], batch size: 27, lr: 3.49e-03, grad_scale: 32.0 2023-04-27 07:01:20,494 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-27 07:01:33,095 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.7505, 2.4117, 2.6633, 3.3501, 3.0560, 2.5434, 2.2762, 2.9182], device='cuda:4'), covar=tensor([0.0764, 0.0991, 0.0587, 0.0477, 0.0601, 0.0751, 0.0696, 0.0447], device='cuda:4'), in_proj_covar=tensor([0.0191, 0.0204, 0.0183, 0.0173, 0.0179, 0.0184, 0.0155, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 07:01:40,271 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7835, 3.7596, 0.8208, 2.0290, 2.1436, 2.5464, 2.0550, 0.9614], device='cuda:4'), covar=tensor([0.1323, 0.1070, 0.2210, 0.1266, 0.1046, 0.1144, 0.1645, 0.2040], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0244, 0.0136, 0.0120, 0.0131, 0.0152, 0.0117, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 07:01:43,236 INFO [finetune.py:976] (4/7) Epoch 15, batch 2650, loss[loss=0.1556, simple_loss=0.2356, pruned_loss=0.03786, over 4856.00 frames. ], tot_loss[loss=0.1823, simple_loss=0.2522, pruned_loss=0.05615, over 956595.45 frames. ], batch size: 44, lr: 3.49e-03, grad_scale: 32.0 2023-04-27 07:01:47,656 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.81 vs. limit=2.0 2023-04-27 07:01:48,667 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.169e+02 1.684e+02 1.897e+02 2.238e+02 3.554e+02, threshold=3.793e+02, percent-clipped=0.0 2023-04-27 07:01:49,961 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=82849.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 07:01:59,042 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-27 07:02:33,354 INFO [finetune.py:976] (4/7) Epoch 15, batch 2700, loss[loss=0.178, simple_loss=0.2559, pruned_loss=0.05011, over 4858.00 frames. ], tot_loss[loss=0.1822, simple_loss=0.252, pruned_loss=0.05617, over 953890.12 frames. ], batch size: 31, lr: 3.49e-03, grad_scale: 32.0 2023-04-27 07:03:04,631 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=82925.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:03:12,267 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.90 vs. limit=5.0 2023-04-27 07:03:12,571 INFO [finetune.py:976] (4/7) Epoch 15, batch 2750, loss[loss=0.1598, simple_loss=0.2292, pruned_loss=0.04522, over 4817.00 frames. ], tot_loss[loss=0.1798, simple_loss=0.2494, pruned_loss=0.05509, over 955470.24 frames. ], batch size: 40, lr: 3.49e-03, grad_scale: 32.0 2023-04-27 07:03:18,530 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.092e+02 1.634e+02 2.005e+02 2.343e+02 4.529e+02, threshold=4.010e+02, percent-clipped=3.0 2023-04-27 07:03:28,557 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=82963.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:03:46,119 INFO [finetune.py:976] (4/7) Epoch 15, batch 2800, loss[loss=0.1692, simple_loss=0.2352, pruned_loss=0.05155, over 4925.00 frames. ], tot_loss[loss=0.1769, simple_loss=0.2461, pruned_loss=0.05387, over 957018.50 frames. ], batch size: 37, lr: 3.49e-03, grad_scale: 32.0 2023-04-27 07:03:52,207 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7441, 1.7865, 0.8658, 1.3525, 1.7943, 1.6253, 1.4603, 1.5061], device='cuda:4'), covar=tensor([0.0515, 0.0365, 0.0353, 0.0569, 0.0282, 0.0534, 0.0523, 0.0576], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0024, 0.0023, 0.0029, 0.0020, 0.0029, 0.0028, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0045, 0.0038, 0.0051, 0.0038, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 07:03:55,379 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=83002.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:04:06,871 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3941, 3.2782, 0.9026, 1.9011, 1.8988, 2.3562, 1.9675, 0.9891], device='cuda:4'), covar=tensor([0.1472, 0.1119, 0.2026, 0.1253, 0.1032, 0.1054, 0.1414, 0.1865], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0246, 0.0137, 0.0121, 0.0131, 0.0153, 0.0118, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 07:04:09,807 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=83024.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:04:19,801 INFO [finetune.py:976] (4/7) Epoch 15, batch 2850, loss[loss=0.158, simple_loss=0.2283, pruned_loss=0.04384, over 4857.00 frames. ], tot_loss[loss=0.1761, simple_loss=0.2449, pruned_loss=0.05368, over 954987.44 frames. ], batch size: 31, lr: 3.49e-03, grad_scale: 32.0 2023-04-27 07:04:25,721 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.603e+01 1.615e+02 1.794e+02 2.149e+02 4.463e+02, threshold=3.588e+02, percent-clipped=2.0 2023-04-27 07:04:27,623 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=83050.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:05:03,759 INFO [finetune.py:976] (4/7) Epoch 15, batch 2900, loss[loss=0.2136, simple_loss=0.2854, pruned_loss=0.07094, over 4786.00 frames. ], tot_loss[loss=0.1788, simple_loss=0.2478, pruned_loss=0.05489, over 954088.19 frames. ], batch size: 54, lr: 3.49e-03, grad_scale: 32.0 2023-04-27 07:05:57,665 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.62 vs. limit=2.0 2023-04-27 07:06:07,041 INFO [finetune.py:976] (4/7) Epoch 15, batch 2950, loss[loss=0.181, simple_loss=0.2484, pruned_loss=0.05677, over 4758.00 frames. ], tot_loss[loss=0.1812, simple_loss=0.2507, pruned_loss=0.05583, over 955334.96 frames. ], batch size: 23, lr: 3.49e-03, grad_scale: 32.0 2023-04-27 07:06:09,001 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.4385, 1.3215, 1.3432, 1.0273, 1.4055, 1.1768, 1.8018, 1.2258], device='cuda:4'), covar=tensor([0.3629, 0.1803, 0.5151, 0.2758, 0.1489, 0.2218, 0.1514, 0.4820], device='cuda:4'), in_proj_covar=tensor([0.0345, 0.0349, 0.0428, 0.0357, 0.0385, 0.0384, 0.0374, 0.0422], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 07:06:18,120 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.270e+02 1.793e+02 2.066e+02 2.627e+02 6.571e+02, threshold=4.132e+02, percent-clipped=5.0 2023-04-27 07:06:19,457 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=83149.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 07:06:40,599 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7493, 1.3624, 1.8927, 2.2261, 1.8998, 1.7346, 1.8424, 1.8203], device='cuda:4'), covar=tensor([0.4923, 0.7011, 0.6404, 0.6139, 0.5877, 0.8702, 0.8072, 0.8022], device='cuda:4'), in_proj_covar=tensor([0.0414, 0.0405, 0.0493, 0.0507, 0.0444, 0.0466, 0.0473, 0.0477], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 07:06:56,294 INFO [finetune.py:976] (4/7) Epoch 15, batch 3000, loss[loss=0.1515, simple_loss=0.2212, pruned_loss=0.04093, over 4869.00 frames. ], tot_loss[loss=0.1829, simple_loss=0.2524, pruned_loss=0.05673, over 955295.86 frames. ], batch size: 34, lr: 3.49e-03, grad_scale: 32.0 2023-04-27 07:06:56,294 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 07:07:03,598 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3517, 1.2604, 3.8193, 3.5208, 3.4589, 3.6549, 3.7587, 3.3923], device='cuda:4'), covar=tensor([0.7264, 0.5628, 0.1258, 0.2018, 0.1319, 0.1485, 0.0757, 0.1797], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0307, 0.0401, 0.0406, 0.0348, 0.0404, 0.0312, 0.0360], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 07:07:05,619 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3802, 1.3271, 3.9016, 3.5912, 3.5098, 3.7385, 3.8189, 3.4306], device='cuda:4'), covar=tensor([0.7153, 0.5548, 0.1328, 0.2091, 0.1382, 0.1540, 0.0772, 0.1706], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0307, 0.0401, 0.0406, 0.0348, 0.0404, 0.0312, 0.0360], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 07:07:05,860 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2663, 1.5347, 1.3899, 1.7443, 1.6502, 1.6962, 1.4235, 3.0550], device='cuda:4'), covar=tensor([0.0624, 0.0777, 0.0777, 0.1262, 0.0612, 0.0474, 0.0749, 0.0187], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0044, 0.0040, 0.0038, 0.0038, 0.0057], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 07:07:06,876 INFO [finetune.py:1010] (4/7) Epoch 15, validation: loss=0.1516, simple_loss=0.2237, pruned_loss=0.03975, over 2265189.00 frames. 2023-04-27 07:07:06,876 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-27 07:07:08,954 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.53 vs. limit=5.0 2023-04-27 07:07:12,341 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=83196.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:07:12,902 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=83197.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 07:07:30,742 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=83225.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:07:35,708 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.65 vs. limit=2.0 2023-04-27 07:07:38,608 INFO [finetune.py:976] (4/7) Epoch 15, batch 3050, loss[loss=0.1919, simple_loss=0.2678, pruned_loss=0.05797, over 4770.00 frames. ], tot_loss[loss=0.183, simple_loss=0.2529, pruned_loss=0.05656, over 955936.39 frames. ], batch size: 28, lr: 3.49e-03, grad_scale: 32.0 2023-04-27 07:07:50,805 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.173e+02 1.583e+02 1.890e+02 2.106e+02 3.614e+02, threshold=3.781e+02, percent-clipped=0.0 2023-04-27 07:08:02,518 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=83257.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:08:24,058 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=83273.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:08:36,021 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=83284.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:08:42,492 INFO [finetune.py:976] (4/7) Epoch 15, batch 3100, loss[loss=0.1709, simple_loss=0.2319, pruned_loss=0.05489, over 4869.00 frames. ], tot_loss[loss=0.1817, simple_loss=0.2516, pruned_loss=0.05594, over 955261.63 frames. ], batch size: 34, lr: 3.49e-03, grad_scale: 32.0 2023-04-27 07:09:04,916 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1422, 1.6204, 1.9743, 2.5461, 1.9981, 1.5707, 1.3585, 1.8154], device='cuda:4'), covar=tensor([0.3264, 0.3517, 0.1866, 0.2098, 0.2588, 0.2819, 0.4441, 0.2272], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0246, 0.0223, 0.0318, 0.0216, 0.0229, 0.0230, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 07:09:10,986 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7049, 1.6042, 2.0090, 2.0158, 1.5241, 1.4026, 1.6597, 1.1501], device='cuda:4'), covar=tensor([0.0528, 0.0669, 0.0396, 0.0595, 0.0827, 0.1169, 0.0719, 0.0716], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0070, 0.0069, 0.0068, 0.0075, 0.0096, 0.0075, 0.0069], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 07:09:15,044 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=83319.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:09:26,584 INFO [finetune.py:976] (4/7) Epoch 15, batch 3150, loss[loss=0.1848, simple_loss=0.2522, pruned_loss=0.05871, over 4742.00 frames. ], tot_loss[loss=0.1807, simple_loss=0.2499, pruned_loss=0.05568, over 958098.90 frames. ], batch size: 26, lr: 3.49e-03, grad_scale: 32.0 2023-04-27 07:09:31,512 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=83345.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:09:32,641 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.218e+02 1.641e+02 1.865e+02 2.202e+02 3.570e+02, threshold=3.730e+02, percent-clipped=0.0 2023-04-27 07:09:44,460 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6173, 1.1740, 1.3565, 1.2053, 1.7621, 1.4131, 1.1565, 1.3136], device='cuda:4'), covar=tensor([0.1679, 0.1344, 0.2407, 0.1548, 0.0874, 0.1556, 0.1768, 0.1996], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0320, 0.0357, 0.0295, 0.0334, 0.0317, 0.0308, 0.0367], device='cuda:4'), out_proj_covar=tensor([6.4401e-05, 6.6859e-05, 7.6365e-05, 6.0325e-05, 6.9527e-05, 6.6939e-05, 6.5160e-05, 7.8316e-05], device='cuda:4') 2023-04-27 07:09:48,706 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8653, 1.5017, 1.4632, 1.5371, 2.0713, 1.6502, 1.3396, 1.3850], device='cuda:4'), covar=tensor([0.1419, 0.1250, 0.1781, 0.1465, 0.0716, 0.1349, 0.1783, 0.1996], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0319, 0.0356, 0.0295, 0.0334, 0.0316, 0.0308, 0.0367], device='cuda:4'), out_proj_covar=tensor([6.4329e-05, 6.6816e-05, 7.6316e-05, 6.0266e-05, 6.9484e-05, 6.6873e-05, 6.5122e-05, 7.8271e-05], device='cuda:4') 2023-04-27 07:09:55,128 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7294, 2.1414, 1.9237, 2.0522, 1.7185, 1.8043, 1.7918, 1.4735], device='cuda:4'), covar=tensor([0.1506, 0.1151, 0.0785, 0.1149, 0.3000, 0.1190, 0.1738, 0.2338], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0312, 0.0223, 0.0284, 0.0315, 0.0267, 0.0255, 0.0271], device='cuda:4'), out_proj_covar=tensor([1.1691e-04, 1.2402e-04, 8.9021e-05, 1.1309e-04, 1.2841e-04, 1.0654e-04, 1.0312e-04, 1.0793e-04], device='cuda:4') 2023-04-27 07:09:59,394 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5702, 1.1663, 1.3260, 1.2274, 1.6813, 1.3334, 1.0852, 1.2808], device='cuda:4'), covar=tensor([0.1587, 0.1407, 0.2010, 0.1478, 0.0981, 0.1534, 0.2061, 0.2127], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0319, 0.0356, 0.0294, 0.0333, 0.0316, 0.0307, 0.0366], device='cuda:4'), out_proj_covar=tensor([6.4271e-05, 6.6761e-05, 7.6234e-05, 6.0181e-05, 6.9373e-05, 6.6830e-05, 6.5055e-05, 7.8220e-05], device='cuda:4') 2023-04-27 07:09:59,877 INFO [finetune.py:976] (4/7) Epoch 15, batch 3200, loss[loss=0.1652, simple_loss=0.2286, pruned_loss=0.05092, over 4824.00 frames. ], tot_loss[loss=0.1781, simple_loss=0.2466, pruned_loss=0.05481, over 956077.99 frames. ], batch size: 30, lr: 3.49e-03, grad_scale: 32.0 2023-04-27 07:10:04,790 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6640, 2.0827, 1.8828, 1.9663, 1.6242, 1.8045, 1.7292, 1.3344], device='cuda:4'), covar=tensor([0.1503, 0.1015, 0.0683, 0.0951, 0.2680, 0.0870, 0.1572, 0.2234], device='cuda:4'), in_proj_covar=tensor([0.0289, 0.0311, 0.0223, 0.0283, 0.0314, 0.0266, 0.0254, 0.0270], device='cuda:4'), out_proj_covar=tensor([1.1647e-04, 1.2358e-04, 8.8675e-05, 1.1267e-04, 1.2791e-04, 1.0614e-04, 1.0263e-04, 1.0746e-04], device='cuda:4') 2023-04-27 07:10:44,125 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3644, 1.3279, 1.5929, 1.6162, 1.2984, 1.0944, 1.3631, 0.9638], device='cuda:4'), covar=tensor([0.0697, 0.0552, 0.0440, 0.0516, 0.0728, 0.1007, 0.0578, 0.0631], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0070, 0.0069, 0.0068, 0.0075, 0.0096, 0.0075, 0.0069], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 07:10:45,254 INFO [finetune.py:976] (4/7) Epoch 15, batch 3250, loss[loss=0.1496, simple_loss=0.2318, pruned_loss=0.03366, over 4757.00 frames. ], tot_loss[loss=0.178, simple_loss=0.2466, pruned_loss=0.05469, over 957064.37 frames. ], batch size: 28, lr: 3.49e-03, grad_scale: 32.0 2023-04-27 07:10:56,135 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.105e+02 1.623e+02 1.932e+02 2.416e+02 4.902e+02, threshold=3.863e+02, percent-clipped=4.0 2023-04-27 07:11:47,249 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-27 07:11:50,145 INFO [finetune.py:976] (4/7) Epoch 15, batch 3300, loss[loss=0.2712, simple_loss=0.3379, pruned_loss=0.1022, over 4900.00 frames. ], tot_loss[loss=0.1827, simple_loss=0.2515, pruned_loss=0.05695, over 955316.30 frames. ], batch size: 43, lr: 3.49e-03, grad_scale: 64.0 2023-04-27 07:12:23,031 INFO [finetune.py:976] (4/7) Epoch 15, batch 3350, loss[loss=0.1602, simple_loss=0.2411, pruned_loss=0.0396, over 4863.00 frames. ], tot_loss[loss=0.1851, simple_loss=0.2543, pruned_loss=0.05794, over 955066.31 frames. ], batch size: 34, lr: 3.49e-03, grad_scale: 64.0 2023-04-27 07:12:28,446 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.001e+02 1.803e+02 2.175e+02 2.734e+02 4.804e+02, threshold=4.350e+02, percent-clipped=5.0 2023-04-27 07:12:31,595 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=83552.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:12:56,225 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3229, 1.5908, 1.4503, 1.7294, 1.6958, 1.9357, 1.4451, 3.3400], device='cuda:4'), covar=tensor([0.0553, 0.0700, 0.0714, 0.1128, 0.0578, 0.0497, 0.0700, 0.0159], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0044, 0.0040, 0.0038, 0.0038, 0.0057], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 07:12:56,720 INFO [finetune.py:976] (4/7) Epoch 15, batch 3400, loss[loss=0.163, simple_loss=0.2395, pruned_loss=0.04326, over 4795.00 frames. ], tot_loss[loss=0.1852, simple_loss=0.2547, pruned_loss=0.05781, over 955507.19 frames. ], batch size: 51, lr: 3.49e-03, grad_scale: 64.0 2023-04-27 07:13:09,191 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-27 07:13:18,196 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=83619.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:13:19,397 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2341, 1.5318, 1.3964, 1.7085, 1.6741, 1.8793, 1.3869, 3.5252], device='cuda:4'), covar=tensor([0.0596, 0.0793, 0.0802, 0.1213, 0.0621, 0.0488, 0.0724, 0.0143], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0044, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 07:13:30,155 INFO [finetune.py:976] (4/7) Epoch 15, batch 3450, loss[loss=0.192, simple_loss=0.2595, pruned_loss=0.06223, over 4820.00 frames. ], tot_loss[loss=0.1851, simple_loss=0.2547, pruned_loss=0.05781, over 954847.99 frames. ], batch size: 39, lr: 3.49e-03, grad_scale: 64.0 2023-04-27 07:13:31,412 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=83640.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:13:35,583 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.177e+02 1.561e+02 1.911e+02 2.345e+02 4.135e+02, threshold=3.821e+02, percent-clipped=0.0 2023-04-27 07:13:49,254 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=83667.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:14:13,281 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.94 vs. limit=2.0 2023-04-27 07:14:13,519 INFO [finetune.py:976] (4/7) Epoch 15, batch 3500, loss[loss=0.1932, simple_loss=0.2554, pruned_loss=0.06549, over 4833.00 frames. ], tot_loss[loss=0.1834, simple_loss=0.2523, pruned_loss=0.05721, over 954845.17 frames. ], batch size: 39, lr: 3.49e-03, grad_scale: 64.0 2023-04-27 07:15:15,482 INFO [finetune.py:976] (4/7) Epoch 15, batch 3550, loss[loss=0.1704, simple_loss=0.2444, pruned_loss=0.04817, over 4803.00 frames. ], tot_loss[loss=0.1799, simple_loss=0.2484, pruned_loss=0.05569, over 955607.57 frames. ], batch size: 51, lr: 3.49e-03, grad_scale: 64.0 2023-04-27 07:15:21,402 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.049e+02 1.464e+02 1.732e+02 2.080e+02 4.644e+02, threshold=3.463e+02, percent-clipped=1.0 2023-04-27 07:15:29,340 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=83760.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:15:35,281 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.3159, 3.2294, 2.4376, 3.8446, 3.2784, 3.3394, 1.5481, 3.3302], device='cuda:4'), covar=tensor([0.1885, 0.1453, 0.3515, 0.2154, 0.4011, 0.1896, 0.5828, 0.2473], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0215, 0.0251, 0.0302, 0.0298, 0.0247, 0.0270, 0.0270], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 07:15:49,271 INFO [finetune.py:976] (4/7) Epoch 15, batch 3600, loss[loss=0.1877, simple_loss=0.2574, pruned_loss=0.05893, over 4811.00 frames. ], tot_loss[loss=0.1789, simple_loss=0.2469, pruned_loss=0.05546, over 955495.68 frames. ], batch size: 38, lr: 3.49e-03, grad_scale: 64.0 2023-04-27 07:16:03,915 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0288, 2.6026, 1.0936, 1.4288, 1.9354, 1.2175, 3.1134, 1.5775], device='cuda:4'), covar=tensor([0.0676, 0.0583, 0.0793, 0.1084, 0.0456, 0.0940, 0.0215, 0.0623], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0065, 0.0048, 0.0046, 0.0049, 0.0051, 0.0075, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0007, 0.0008, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-27 07:16:32,023 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=83821.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:16:46,117 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-27 07:16:55,918 INFO [finetune.py:976] (4/7) Epoch 15, batch 3650, loss[loss=0.1525, simple_loss=0.2245, pruned_loss=0.04027, over 4763.00 frames. ], tot_loss[loss=0.1804, simple_loss=0.2489, pruned_loss=0.05602, over 955576.70 frames. ], batch size: 28, lr: 3.48e-03, grad_scale: 64.0 2023-04-27 07:17:01,429 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.661e+01 1.633e+02 1.914e+02 2.329e+02 4.921e+02, threshold=3.827e+02, percent-clipped=4.0 2023-04-27 07:17:05,009 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=83852.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:17:29,863 INFO [finetune.py:976] (4/7) Epoch 15, batch 3700, loss[loss=0.1783, simple_loss=0.2496, pruned_loss=0.05346, over 4909.00 frames. ], tot_loss[loss=0.1835, simple_loss=0.2529, pruned_loss=0.05702, over 955919.66 frames. ], batch size: 35, lr: 3.48e-03, grad_scale: 32.0 2023-04-27 07:17:37,201 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=83900.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:18:03,516 INFO [finetune.py:976] (4/7) Epoch 15, batch 3750, loss[loss=0.2012, simple_loss=0.2672, pruned_loss=0.06762, over 4831.00 frames. ], tot_loss[loss=0.185, simple_loss=0.2549, pruned_loss=0.05753, over 955926.75 frames. ], batch size: 49, lr: 3.48e-03, grad_scale: 32.0 2023-04-27 07:18:04,867 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=83940.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:18:09,576 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.106e+02 1.640e+02 2.069e+02 2.492e+02 6.161e+02, threshold=4.139e+02, percent-clipped=2.0 2023-04-27 07:18:10,283 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=83949.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:18:34,074 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2030, 4.5043, 0.9638, 2.2643, 2.9619, 2.8826, 2.6667, 1.0921], device='cuda:4'), covar=tensor([0.1254, 0.1140, 0.2201, 0.1365, 0.0888, 0.1187, 0.1309, 0.2099], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0245, 0.0137, 0.0121, 0.0130, 0.0152, 0.0118, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 07:18:34,104 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=83985.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:18:36,294 INFO [finetune.py:976] (4/7) Epoch 15, batch 3800, loss[loss=0.2074, simple_loss=0.2791, pruned_loss=0.06791, over 4866.00 frames. ], tot_loss[loss=0.184, simple_loss=0.2542, pruned_loss=0.05687, over 955951.53 frames. ], batch size: 34, lr: 3.48e-03, grad_scale: 32.0 2023-04-27 07:18:36,355 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=83988.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:18:51,947 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=84010.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 07:19:10,673 INFO [finetune.py:976] (4/7) Epoch 15, batch 3850, loss[loss=0.1652, simple_loss=0.2344, pruned_loss=0.04801, over 4754.00 frames. ], tot_loss[loss=0.1823, simple_loss=0.2525, pruned_loss=0.05612, over 955467.96 frames. ], batch size: 54, lr: 3.48e-03, grad_scale: 32.0 2023-04-27 07:19:21,669 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1334, 1.3314, 1.6613, 1.7146, 2.1378, 1.8090, 1.5207, 1.4992], device='cuda:4'), covar=tensor([0.1562, 0.1512, 0.1810, 0.1431, 0.0942, 0.1517, 0.1994, 0.2026], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0317, 0.0352, 0.0292, 0.0330, 0.0314, 0.0304, 0.0363], device='cuda:4'), out_proj_covar=tensor([6.3989e-05, 6.6255e-05, 7.5227e-05, 5.9621e-05, 6.8690e-05, 6.6339e-05, 6.4357e-05, 7.7427e-05], device='cuda:4') 2023-04-27 07:19:21,671 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=84046.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:19:22,757 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.590e+02 1.914e+02 2.294e+02 5.295e+02, threshold=3.828e+02, percent-clipped=1.0 2023-04-27 07:19:35,244 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1358, 1.4497, 1.3547, 1.6485, 1.6093, 1.5412, 1.3522, 2.4588], device='cuda:4'), covar=tensor([0.0578, 0.0787, 0.0797, 0.1218, 0.0606, 0.0463, 0.0739, 0.0235], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0044, 0.0040, 0.0038, 0.0038, 0.0057], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 07:20:15,362 INFO [finetune.py:976] (4/7) Epoch 15, batch 3900, loss[loss=0.1975, simple_loss=0.2583, pruned_loss=0.06832, over 4888.00 frames. ], tot_loss[loss=0.1813, simple_loss=0.2502, pruned_loss=0.05616, over 956720.07 frames. ], batch size: 32, lr: 3.48e-03, grad_scale: 32.0 2023-04-27 07:20:16,659 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.67 vs. limit=2.0 2023-04-27 07:20:27,927 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=84097.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:20:50,572 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=84116.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:21:09,133 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5963, 1.3427, 4.3321, 4.1013, 3.8346, 4.0976, 3.9765, 3.8256], device='cuda:4'), covar=tensor([0.7088, 0.5955, 0.1034, 0.1411, 0.0985, 0.1570, 0.1741, 0.1347], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0307, 0.0402, 0.0405, 0.0349, 0.0404, 0.0312, 0.0361], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 07:21:20,860 INFO [finetune.py:976] (4/7) Epoch 15, batch 3950, loss[loss=0.2009, simple_loss=0.2544, pruned_loss=0.07369, over 4823.00 frames. ], tot_loss[loss=0.1788, simple_loss=0.2473, pruned_loss=0.05517, over 957122.78 frames. ], batch size: 40, lr: 3.48e-03, grad_scale: 32.0 2023-04-27 07:21:34,733 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.122e+01 1.668e+02 1.985e+02 2.633e+02 4.581e+02, threshold=3.971e+02, percent-clipped=4.0 2023-04-27 07:21:41,490 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=84158.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:21:48,082 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5856, 1.3065, 3.9943, 3.4937, 3.5945, 3.7533, 3.5950, 3.3807], device='cuda:4'), covar=tensor([0.9188, 0.8259, 0.1907, 0.3250, 0.2259, 0.4411, 0.4267, 0.3320], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0306, 0.0402, 0.0404, 0.0348, 0.0403, 0.0312, 0.0360], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 07:22:12,111 INFO [finetune.py:976] (4/7) Epoch 15, batch 4000, loss[loss=0.2199, simple_loss=0.2867, pruned_loss=0.07651, over 4814.00 frames. ], tot_loss[loss=0.1785, simple_loss=0.2469, pruned_loss=0.05503, over 956367.69 frames. ], batch size: 41, lr: 3.48e-03, grad_scale: 32.0 2023-04-27 07:22:31,715 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=84199.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:23:00,613 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=84236.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:23:02,207 INFO [finetune.py:976] (4/7) Epoch 15, batch 4050, loss[loss=0.1993, simple_loss=0.2848, pruned_loss=0.05689, over 4729.00 frames. ], tot_loss[loss=0.1822, simple_loss=0.2517, pruned_loss=0.05642, over 955509.91 frames. ], batch size: 59, lr: 3.48e-03, grad_scale: 32.0 2023-04-27 07:23:09,757 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.868e+01 1.686e+02 2.049e+02 2.449e+02 3.463e+02, threshold=4.099e+02, percent-clipped=0.0 2023-04-27 07:23:18,193 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=84260.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:23:19,936 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.8965, 0.9984, 3.1572, 2.7484, 2.9219, 2.9740, 2.9963, 2.6578], device='cuda:4'), covar=tensor([0.9669, 0.8009, 0.2576, 0.4009, 0.2831, 0.3973, 0.3707, 0.3813], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0306, 0.0402, 0.0405, 0.0348, 0.0404, 0.0312, 0.0361], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 07:23:35,687 INFO [finetune.py:976] (4/7) Epoch 15, batch 4100, loss[loss=0.1778, simple_loss=0.2469, pruned_loss=0.05436, over 4806.00 frames. ], tot_loss[loss=0.185, simple_loss=0.2552, pruned_loss=0.05736, over 955914.62 frames. ], batch size: 41, lr: 3.48e-03, grad_scale: 32.0 2023-04-27 07:23:42,198 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=84297.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:23:48,049 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=84305.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 07:23:57,591 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7953, 2.4419, 1.7858, 1.8311, 1.3327, 1.3411, 1.8799, 1.2905], device='cuda:4'), covar=tensor([0.1622, 0.1228, 0.1362, 0.1523, 0.2286, 0.1914, 0.0913, 0.1968], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0215, 0.0170, 0.0205, 0.0202, 0.0185, 0.0157, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 07:24:08,385 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=84336.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:24:09,504 INFO [finetune.py:976] (4/7) Epoch 15, batch 4150, loss[loss=0.1984, simple_loss=0.275, pruned_loss=0.06093, over 4935.00 frames. ], tot_loss[loss=0.1859, simple_loss=0.2561, pruned_loss=0.05786, over 953548.94 frames. ], batch size: 42, lr: 3.48e-03, grad_scale: 32.0 2023-04-27 07:24:11,378 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=84341.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:24:16,004 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.201e+02 1.658e+02 1.986e+02 2.406e+02 7.031e+02, threshold=3.971e+02, percent-clipped=3.0 2023-04-27 07:24:20,278 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.53 vs. limit=5.0 2023-04-27 07:24:36,624 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-04-27 07:24:43,283 INFO [finetune.py:976] (4/7) Epoch 15, batch 4200, loss[loss=0.1505, simple_loss=0.225, pruned_loss=0.03801, over 4762.00 frames. ], tot_loss[loss=0.184, simple_loss=0.2546, pruned_loss=0.05673, over 954204.29 frames. ], batch size: 28, lr: 3.48e-03, grad_scale: 32.0 2023-04-27 07:24:43,401 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6082, 1.4464, 1.8073, 1.7776, 1.3859, 1.2661, 1.4875, 1.0570], device='cuda:4'), covar=tensor([0.0522, 0.0739, 0.0432, 0.0723, 0.0793, 0.1140, 0.0649, 0.0669], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0070, 0.0070, 0.0068, 0.0076, 0.0097, 0.0076, 0.0069], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 07:24:49,255 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=84397.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:25:03,267 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=84416.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:25:22,954 INFO [finetune.py:976] (4/7) Epoch 15, batch 4250, loss[loss=0.1491, simple_loss=0.2233, pruned_loss=0.03748, over 4816.00 frames. ], tot_loss[loss=0.1812, simple_loss=0.2513, pruned_loss=0.05551, over 955064.13 frames. ], batch size: 39, lr: 3.48e-03, grad_scale: 32.0 2023-04-27 07:25:32,383 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.77 vs. limit=5.0 2023-04-27 07:25:33,112 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.033e+02 1.587e+02 1.989e+02 2.520e+02 5.183e+02, threshold=3.979e+02, percent-clipped=3.0 2023-04-27 07:25:42,001 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=84453.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:25:52,502 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2014, 1.6845, 2.0814, 2.6019, 2.0371, 1.6148, 1.3016, 1.8513], device='cuda:4'), covar=tensor([0.3743, 0.3585, 0.1987, 0.2672, 0.3243, 0.3199, 0.4716, 0.2317], device='cuda:4'), in_proj_covar=tensor([0.0291, 0.0248, 0.0225, 0.0317, 0.0216, 0.0229, 0.0230, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 07:25:56,125 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=84464.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:26:28,013 INFO [finetune.py:976] (4/7) Epoch 15, batch 4300, loss[loss=0.1792, simple_loss=0.253, pruned_loss=0.05273, over 4801.00 frames. ], tot_loss[loss=0.18, simple_loss=0.2493, pruned_loss=0.05535, over 957067.73 frames. ], batch size: 51, lr: 3.48e-03, grad_scale: 32.0 2023-04-27 07:27:19,049 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=84525.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:27:20,281 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9298, 2.5182, 1.9614, 2.3398, 1.7042, 2.0368, 2.1127, 1.5287], device='cuda:4'), covar=tensor([0.1997, 0.1316, 0.0937, 0.1226, 0.3366, 0.1280, 0.1850, 0.2794], device='cuda:4'), in_proj_covar=tensor([0.0287, 0.0308, 0.0219, 0.0281, 0.0311, 0.0264, 0.0251, 0.0266], device='cuda:4'), out_proj_covar=tensor([1.1564e-04, 1.2273e-04, 8.7459e-05, 1.1192e-04, 1.2672e-04, 1.0533e-04, 1.0165e-04, 1.0588e-04], device='cuda:4') 2023-04-27 07:27:32,535 INFO [finetune.py:976] (4/7) Epoch 15, batch 4350, loss[loss=0.1583, simple_loss=0.2273, pruned_loss=0.04464, over 4816.00 frames. ], tot_loss[loss=0.1773, simple_loss=0.2456, pruned_loss=0.05451, over 956235.21 frames. ], batch size: 38, lr: 3.48e-03, grad_scale: 32.0 2023-04-27 07:27:44,744 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.129e+02 1.607e+02 1.900e+02 2.324e+02 5.232e+02, threshold=3.801e+02, percent-clipped=2.0 2023-04-27 07:27:54,631 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=84555.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:28:26,842 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=84586.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:28:27,954 INFO [finetune.py:976] (4/7) Epoch 15, batch 4400, loss[loss=0.1928, simple_loss=0.261, pruned_loss=0.06233, over 4890.00 frames. ], tot_loss[loss=0.1776, simple_loss=0.2461, pruned_loss=0.05454, over 955587.75 frames. ], batch size: 32, lr: 3.48e-03, grad_scale: 32.0 2023-04-27 07:28:30,484 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=84592.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:28:39,051 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=84605.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:28:59,176 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=84633.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:29:02,150 INFO [finetune.py:976] (4/7) Epoch 15, batch 4450, loss[loss=0.1763, simple_loss=0.253, pruned_loss=0.04981, over 4800.00 frames. ], tot_loss[loss=0.1802, simple_loss=0.2495, pruned_loss=0.05543, over 954924.32 frames. ], batch size: 51, lr: 3.48e-03, grad_scale: 32.0 2023-04-27 07:29:04,129 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=84641.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:29:08,290 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.195e+02 1.631e+02 1.866e+02 2.330e+02 4.316e+02, threshold=3.732e+02, percent-clipped=2.0 2023-04-27 07:29:11,996 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=84653.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:29:36,308 INFO [finetune.py:976] (4/7) Epoch 15, batch 4500, loss[loss=0.1743, simple_loss=0.2549, pruned_loss=0.04685, over 4810.00 frames. ], tot_loss[loss=0.1822, simple_loss=0.2521, pruned_loss=0.05619, over 956527.44 frames. ], batch size: 41, lr: 3.48e-03, grad_scale: 32.0 2023-04-27 07:29:37,004 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=84689.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:29:38,835 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=84692.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:29:40,110 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=84694.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:30:09,977 INFO [finetune.py:976] (4/7) Epoch 15, batch 4550, loss[loss=0.1742, simple_loss=0.2545, pruned_loss=0.04693, over 4838.00 frames. ], tot_loss[loss=0.1834, simple_loss=0.2535, pruned_loss=0.05665, over 952393.02 frames. ], batch size: 49, lr: 3.48e-03, grad_scale: 32.0 2023-04-27 07:30:16,092 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.003e+02 1.700e+02 1.997e+02 2.528e+02 3.461e+02, threshold=3.994e+02, percent-clipped=0.0 2023-04-27 07:30:19,282 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=84753.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:30:43,752 INFO [finetune.py:976] (4/7) Epoch 15, batch 4600, loss[loss=0.1882, simple_loss=0.2549, pruned_loss=0.06078, over 4860.00 frames. ], tot_loss[loss=0.1834, simple_loss=0.2536, pruned_loss=0.05663, over 951501.28 frames. ], batch size: 31, lr: 3.48e-03, grad_scale: 32.0 2023-04-27 07:30:51,269 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=84800.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:30:51,853 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=84801.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:31:33,985 INFO [finetune.py:976] (4/7) Epoch 15, batch 4650, loss[loss=0.1606, simple_loss=0.2303, pruned_loss=0.04544, over 4755.00 frames. ], tot_loss[loss=0.182, simple_loss=0.2519, pruned_loss=0.056, over 953121.20 frames. ], batch size: 54, lr: 3.48e-03, grad_scale: 32.0 2023-04-27 07:31:45,945 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.111e+02 1.628e+02 1.889e+02 2.261e+02 4.327e+02, threshold=3.778e+02, percent-clipped=1.0 2023-04-27 07:31:47,588 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.59 vs. limit=5.0 2023-04-27 07:31:55,709 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=84855.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:31:56,404 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8529, 1.3704, 1.9104, 2.3023, 1.9346, 1.7972, 1.8860, 1.8343], device='cuda:4'), covar=tensor([0.4633, 0.6437, 0.6353, 0.5613, 0.5975, 0.7502, 0.7730, 0.7959], device='cuda:4'), in_proj_covar=tensor([0.0415, 0.0404, 0.0491, 0.0504, 0.0443, 0.0465, 0.0473, 0.0476], device='cuda:4'), out_proj_covar=tensor([1.0012e-04, 9.9708e-05, 1.1051e-04, 1.2015e-04, 1.0654e-04, 1.1203e-04, 1.1260e-04, 1.1317e-04], device='cuda:4') 2023-04-27 07:32:05,699 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=84861.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:32:30,054 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=84881.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:32:39,700 INFO [finetune.py:976] (4/7) Epoch 15, batch 4700, loss[loss=0.1523, simple_loss=0.2231, pruned_loss=0.04077, over 4789.00 frames. ], tot_loss[loss=0.1791, simple_loss=0.2487, pruned_loss=0.05475, over 955329.99 frames. ], batch size: 29, lr: 3.48e-03, grad_scale: 32.0 2023-04-27 07:32:48,400 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=84892.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:33:00,658 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=84903.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:33:46,042 INFO [finetune.py:976] (4/7) Epoch 15, batch 4750, loss[loss=0.1826, simple_loss=0.2457, pruned_loss=0.05971, over 4807.00 frames. ], tot_loss[loss=0.1764, simple_loss=0.2459, pruned_loss=0.05342, over 955356.60 frames. ], batch size: 45, lr: 3.48e-03, grad_scale: 32.0 2023-04-27 07:33:47,829 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=84940.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:33:53,645 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.954e+01 1.625e+02 2.007e+02 2.365e+02 3.914e+02, threshold=4.014e+02, percent-clipped=2.0 2023-04-27 07:33:55,680 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=84951.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:34:19,538 INFO [finetune.py:976] (4/7) Epoch 15, batch 4800, loss[loss=0.1755, simple_loss=0.2525, pruned_loss=0.04924, over 4895.00 frames. ], tot_loss[loss=0.1799, simple_loss=0.2493, pruned_loss=0.05524, over 955839.63 frames. ], batch size: 35, lr: 3.47e-03, grad_scale: 32.0 2023-04-27 07:34:20,210 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=84989.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:34:20,284 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0918, 1.5265, 1.8903, 2.3341, 1.9573, 1.5000, 1.1874, 1.7037], device='cuda:4'), covar=tensor([0.3403, 0.3570, 0.1837, 0.2513, 0.2712, 0.2834, 0.4628, 0.2258], device='cuda:4'), in_proj_covar=tensor([0.0289, 0.0246, 0.0223, 0.0314, 0.0215, 0.0229, 0.0229, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 07:34:23,032 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=84992.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:34:33,977 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6637, 1.9267, 1.6579, 1.4244, 1.1812, 1.2432, 1.7620, 1.1742], device='cuda:4'), covar=tensor([0.1821, 0.1509, 0.1464, 0.1836, 0.2452, 0.2052, 0.0976, 0.2118], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0213, 0.0169, 0.0204, 0.0201, 0.0184, 0.0155, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 07:34:36,394 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=85012.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:34:40,015 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8636, 2.5072, 2.0200, 2.4054, 1.7167, 1.9868, 2.0547, 1.6514], device='cuda:4'), covar=tensor([0.2119, 0.1173, 0.0872, 0.1106, 0.3258, 0.1311, 0.1821, 0.2318], device='cuda:4'), in_proj_covar=tensor([0.0288, 0.0309, 0.0220, 0.0282, 0.0313, 0.0264, 0.0251, 0.0266], device='cuda:4'), out_proj_covar=tensor([1.1581e-04, 1.2295e-04, 8.7920e-05, 1.1197e-04, 1.2751e-04, 1.0545e-04, 1.0169e-04, 1.0609e-04], device='cuda:4') 2023-04-27 07:34:53,702 INFO [finetune.py:976] (4/7) Epoch 15, batch 4850, loss[loss=0.1778, simple_loss=0.2428, pruned_loss=0.05644, over 4783.00 frames. ], tot_loss[loss=0.1833, simple_loss=0.2533, pruned_loss=0.0566, over 955242.30 frames. ], batch size: 54, lr: 3.47e-03, grad_scale: 32.0 2023-04-27 07:34:54,927 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=85040.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:35:01,198 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.085e+02 1.698e+02 1.957e+02 2.327e+02 4.552e+02, threshold=3.914e+02, percent-clipped=3.0 2023-04-27 07:35:08,455 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7644, 3.7486, 0.7822, 1.8594, 2.2522, 2.5977, 2.0986, 1.0543], device='cuda:4'), covar=tensor([0.1274, 0.0835, 0.2225, 0.1291, 0.0936, 0.1021, 0.1693, 0.2010], device='cuda:4'), in_proj_covar=tensor([0.0115, 0.0240, 0.0135, 0.0119, 0.0129, 0.0149, 0.0116, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 07:35:27,218 INFO [finetune.py:976] (4/7) Epoch 15, batch 4900, loss[loss=0.1739, simple_loss=0.2529, pruned_loss=0.04741, over 4795.00 frames. ], tot_loss[loss=0.184, simple_loss=0.2542, pruned_loss=0.05689, over 954431.74 frames. ], batch size: 51, lr: 3.47e-03, grad_scale: 32.0 2023-04-27 07:35:40,334 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=85107.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:35:45,189 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=85115.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:36:00,017 INFO [finetune.py:976] (4/7) Epoch 15, batch 4950, loss[loss=0.1735, simple_loss=0.2447, pruned_loss=0.05108, over 4862.00 frames. ], tot_loss[loss=0.1846, simple_loss=0.2552, pruned_loss=0.05699, over 955373.51 frames. ], batch size: 34, lr: 3.47e-03, grad_scale: 32.0 2023-04-27 07:36:07,086 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.205e+02 1.668e+02 1.905e+02 2.304e+02 4.878e+02, threshold=3.810e+02, percent-clipped=1.0 2023-04-27 07:36:12,984 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=85156.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:36:20,320 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=85168.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 07:36:25,199 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=85176.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:36:26,485 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-04-27 07:36:28,191 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=85181.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:36:33,335 INFO [finetune.py:976] (4/7) Epoch 15, batch 5000, loss[loss=0.1544, simple_loss=0.2189, pruned_loss=0.04497, over 4808.00 frames. ], tot_loss[loss=0.1824, simple_loss=0.253, pruned_loss=0.05588, over 955146.63 frames. ], batch size: 25, lr: 3.47e-03, grad_scale: 32.0 2023-04-27 07:37:00,636 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=85229.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:37:12,538 INFO [finetune.py:976] (4/7) Epoch 15, batch 5050, loss[loss=0.1761, simple_loss=0.2454, pruned_loss=0.05344, over 4819.00 frames. ], tot_loss[loss=0.1805, simple_loss=0.25, pruned_loss=0.0555, over 954285.42 frames. ], batch size: 25, lr: 3.47e-03, grad_scale: 32.0 2023-04-27 07:37:25,140 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.163e+02 1.655e+02 2.048e+02 2.440e+02 4.868e+02, threshold=4.096e+02, percent-clipped=4.0 2023-04-27 07:38:17,994 INFO [finetune.py:976] (4/7) Epoch 15, batch 5100, loss[loss=0.1716, simple_loss=0.2408, pruned_loss=0.0512, over 4866.00 frames. ], tot_loss[loss=0.1785, simple_loss=0.2475, pruned_loss=0.05475, over 953518.85 frames. ], batch size: 31, lr: 3.47e-03, grad_scale: 32.0 2023-04-27 07:38:18,666 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=85289.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:38:19,337 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0386, 2.2988, 1.9731, 1.8813, 2.1992, 1.8654, 2.7558, 1.5016], device='cuda:4'), covar=tensor([0.3665, 0.1489, 0.4170, 0.2740, 0.1676, 0.2330, 0.1205, 0.4560], device='cuda:4'), in_proj_covar=tensor([0.0342, 0.0349, 0.0427, 0.0355, 0.0384, 0.0380, 0.0372, 0.0420], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 07:38:41,085 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=85307.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:39:06,642 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.77 vs. limit=5.0 2023-04-27 07:39:06,780 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=85337.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:39:07,341 INFO [finetune.py:976] (4/7) Epoch 15, batch 5150, loss[loss=0.1895, simple_loss=0.2669, pruned_loss=0.05601, over 4788.00 frames. ], tot_loss[loss=0.1782, simple_loss=0.247, pruned_loss=0.0547, over 952766.89 frames. ], batch size: 51, lr: 3.47e-03, grad_scale: 32.0 2023-04-27 07:39:20,282 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.171e+02 1.606e+02 1.874e+02 2.297e+02 4.541e+02, threshold=3.747e+02, percent-clipped=1.0 2023-04-27 07:39:47,577 INFO [finetune.py:976] (4/7) Epoch 15, batch 5200, loss[loss=0.2199, simple_loss=0.2886, pruned_loss=0.07566, over 4809.00 frames. ], tot_loss[loss=0.1814, simple_loss=0.2507, pruned_loss=0.05602, over 953428.65 frames. ], batch size: 45, lr: 3.47e-03, grad_scale: 32.0 2023-04-27 07:39:52,601 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1302, 2.7317, 2.0605, 2.1459, 1.5073, 1.5255, 2.1982, 1.4806], device='cuda:4'), covar=tensor([0.1787, 0.1370, 0.1501, 0.1718, 0.2527, 0.2055, 0.1041, 0.2178], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0212, 0.0168, 0.0203, 0.0200, 0.0183, 0.0155, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 07:40:06,485 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9401, 1.4408, 1.7511, 1.6649, 1.7473, 1.4463, 0.7410, 1.3646], device='cuda:4'), covar=tensor([0.3730, 0.3537, 0.1787, 0.2532, 0.2689, 0.2710, 0.4578, 0.2265], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0246, 0.0223, 0.0315, 0.0216, 0.0230, 0.0229, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 07:40:15,420 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=85428.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:40:21,370 INFO [finetune.py:976] (4/7) Epoch 15, batch 5250, loss[loss=0.1395, simple_loss=0.2231, pruned_loss=0.02796, over 4754.00 frames. ], tot_loss[loss=0.181, simple_loss=0.2511, pruned_loss=0.05543, over 953879.78 frames. ], batch size: 28, lr: 3.47e-03, grad_scale: 32.0 2023-04-27 07:40:27,422 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.183e+02 1.584e+02 2.032e+02 2.482e+02 6.006e+02, threshold=4.063e+02, percent-clipped=4.0 2023-04-27 07:40:33,768 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=85456.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:40:38,408 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=85463.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 07:40:40,834 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-27 07:40:42,996 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6723, 1.5126, 1.6808, 2.0049, 2.0592, 1.6665, 1.2512, 1.9118], device='cuda:4'), covar=tensor([0.0858, 0.1192, 0.0777, 0.0606, 0.0601, 0.0765, 0.0891, 0.0561], device='cuda:4'), in_proj_covar=tensor([0.0193, 0.0207, 0.0184, 0.0177, 0.0181, 0.0186, 0.0157, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 07:40:44,661 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=85471.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:40:51,999 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=85483.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:40:54,957 INFO [finetune.py:976] (4/7) Epoch 15, batch 5300, loss[loss=0.1832, simple_loss=0.2582, pruned_loss=0.05413, over 4713.00 frames. ], tot_loss[loss=0.1829, simple_loss=0.253, pruned_loss=0.05641, over 954634.50 frames. ], batch size: 59, lr: 3.47e-03, grad_scale: 32.0 2023-04-27 07:40:55,690 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=85489.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:41:04,900 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=85504.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:41:23,626 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=85530.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:41:28,361 INFO [finetune.py:976] (4/7) Epoch 15, batch 5350, loss[loss=0.17, simple_loss=0.2356, pruned_loss=0.05222, over 4718.00 frames. ], tot_loss[loss=0.1818, simple_loss=0.2522, pruned_loss=0.05566, over 954989.01 frames. ], batch size: 59, lr: 3.47e-03, grad_scale: 32.0 2023-04-27 07:41:32,141 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=85544.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:41:34,433 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.243e+02 1.620e+02 1.904e+02 2.320e+02 4.507e+02, threshold=3.809e+02, percent-clipped=1.0 2023-04-27 07:41:49,853 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1487, 2.6658, 1.0228, 1.5560, 2.1313, 1.3493, 3.5031, 1.9035], device='cuda:4'), covar=tensor([0.0655, 0.0589, 0.0848, 0.1236, 0.0497, 0.0981, 0.0195, 0.0612], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0048, 0.0047, 0.0050, 0.0052, 0.0075, 0.0052], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 07:42:02,178 INFO [finetune.py:976] (4/7) Epoch 15, batch 5400, loss[loss=0.2111, simple_loss=0.2742, pruned_loss=0.07395, over 4828.00 frames. ], tot_loss[loss=0.1811, simple_loss=0.2513, pruned_loss=0.05541, over 956550.12 frames. ], batch size: 33, lr: 3.47e-03, grad_scale: 32.0 2023-04-27 07:42:04,161 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=85591.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:42:13,950 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=85607.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:42:27,859 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=85625.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:42:41,671 INFO [finetune.py:976] (4/7) Epoch 15, batch 5450, loss[loss=0.1883, simple_loss=0.2546, pruned_loss=0.06096, over 4945.00 frames. ], tot_loss[loss=0.1787, simple_loss=0.2481, pruned_loss=0.05468, over 957250.86 frames. ], batch size: 42, lr: 3.47e-03, grad_scale: 32.0 2023-04-27 07:42:46,155 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.8199, 2.0617, 2.0023, 2.2340, 1.9271, 2.0852, 2.0379, 2.0098], device='cuda:4'), covar=tensor([0.4002, 0.6897, 0.5458, 0.4849, 0.6585, 0.8238, 0.6788, 0.6639], device='cuda:4'), in_proj_covar=tensor([0.0329, 0.0371, 0.0317, 0.0331, 0.0342, 0.0396, 0.0352, 0.0324], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 07:42:53,051 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.070e+02 1.619e+02 1.980e+02 2.331e+02 6.511e+02, threshold=3.961e+02, percent-clipped=1.0 2023-04-27 07:43:02,875 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=85655.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:43:04,198 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.3841, 2.8155, 2.6079, 2.8381, 2.5877, 2.8035, 2.7816, 2.6736], device='cuda:4'), covar=tensor([0.3320, 0.5773, 0.4565, 0.3950, 0.5558, 0.6805, 0.5302, 0.4876], device='cuda:4'), in_proj_covar=tensor([0.0329, 0.0371, 0.0317, 0.0331, 0.0341, 0.0396, 0.0351, 0.0323], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 07:43:05,703 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-27 07:43:45,382 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=85686.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 07:43:46,492 INFO [finetune.py:976] (4/7) Epoch 15, batch 5500, loss[loss=0.1435, simple_loss=0.2109, pruned_loss=0.03802, over 4775.00 frames. ], tot_loss[loss=0.1757, simple_loss=0.2448, pruned_loss=0.05329, over 957093.03 frames. ], batch size: 26, lr: 3.47e-03, grad_scale: 32.0 2023-04-27 07:44:51,824 INFO [finetune.py:976] (4/7) Epoch 15, batch 5550, loss[loss=0.22, simple_loss=0.2899, pruned_loss=0.07504, over 4825.00 frames. ], tot_loss[loss=0.1771, simple_loss=0.2464, pruned_loss=0.05389, over 957516.68 frames. ], batch size: 39, lr: 3.47e-03, grad_scale: 32.0 2023-04-27 07:45:02,123 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.106e+02 1.632e+02 1.848e+02 2.307e+02 4.705e+02, threshold=3.696e+02, percent-clipped=4.0 2023-04-27 07:45:05,989 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5679, 1.7985, 1.8758, 1.9927, 1.7930, 1.9564, 1.9611, 1.8849], device='cuda:4'), covar=tensor([0.4022, 0.6403, 0.4722, 0.4686, 0.6201, 0.8062, 0.5797, 0.5818], device='cuda:4'), in_proj_covar=tensor([0.0332, 0.0375, 0.0320, 0.0334, 0.0345, 0.0401, 0.0356, 0.0327], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 07:45:11,452 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=85763.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:45:16,810 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=85771.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:45:21,451 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8013, 2.2534, 1.8300, 2.0739, 1.4817, 1.8613, 1.9880, 1.4495], device='cuda:4'), covar=tensor([0.2268, 0.1599, 0.1229, 0.1678, 0.3683, 0.1684, 0.2042, 0.2869], device='cuda:4'), in_proj_covar=tensor([0.0282, 0.0305, 0.0218, 0.0278, 0.0308, 0.0260, 0.0248, 0.0265], device='cuda:4'), out_proj_covar=tensor([1.1356e-04, 1.2151e-04, 8.6798e-05, 1.1048e-04, 1.2519e-04, 1.0389e-04, 1.0023e-04, 1.0544e-04], device='cuda:4') 2023-04-27 07:45:24,799 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=85784.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:45:27,127 INFO [finetune.py:976] (4/7) Epoch 15, batch 5600, loss[loss=0.176, simple_loss=0.2499, pruned_loss=0.05112, over 4795.00 frames. ], tot_loss[loss=0.1794, simple_loss=0.2493, pruned_loss=0.05474, over 955818.25 frames. ], batch size: 25, lr: 3.47e-03, grad_scale: 32.0 2023-04-27 07:45:40,675 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=85811.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:45:45,405 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=85819.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:45:57,426 INFO [finetune.py:976] (4/7) Epoch 15, batch 5650, loss[loss=0.1778, simple_loss=0.2536, pruned_loss=0.05104, over 4816.00 frames. ], tot_loss[loss=0.1808, simple_loss=0.2512, pruned_loss=0.05518, over 953986.25 frames. ], batch size: 51, lr: 3.47e-03, grad_scale: 32.0 2023-04-27 07:45:58,081 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=85839.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:46:03,465 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.035e+02 1.549e+02 1.787e+02 2.141e+02 3.216e+02, threshold=3.573e+02, percent-clipped=0.0 2023-04-27 07:46:11,598 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.14 vs. limit=5.0 2023-04-27 07:46:19,897 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=85875.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:46:26,480 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=85886.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:46:27,653 INFO [finetune.py:976] (4/7) Epoch 15, batch 5700, loss[loss=0.1294, simple_loss=0.1817, pruned_loss=0.03852, over 3840.00 frames. ], tot_loss[loss=0.1783, simple_loss=0.2478, pruned_loss=0.05443, over 938031.53 frames. ], batch size: 16, lr: 3.47e-03, grad_scale: 64.0 2023-04-27 07:46:30,806 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3823, 1.4653, 1.8998, 2.0239, 1.9123, 2.0723, 1.9344, 1.9392], device='cuda:4'), covar=tensor([0.3204, 0.4991, 0.4656, 0.4271, 0.5047, 0.6570, 0.4824, 0.4458], device='cuda:4'), in_proj_covar=tensor([0.0332, 0.0375, 0.0320, 0.0334, 0.0345, 0.0400, 0.0355, 0.0327], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 07:46:35,002 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.8435, 3.4439, 2.7131, 3.0003, 2.1019, 2.1378, 2.9884, 2.1856], device='cuda:4'), covar=tensor([0.1419, 0.1295, 0.1242, 0.1365, 0.2060, 0.1806, 0.0801, 0.1710], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0212, 0.0168, 0.0204, 0.0200, 0.0183, 0.0155, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 07:46:59,231 INFO [finetune.py:976] (4/7) Epoch 16, batch 0, loss[loss=0.2207, simple_loss=0.2931, pruned_loss=0.07415, over 4732.00 frames. ], tot_loss[loss=0.2207, simple_loss=0.2931, pruned_loss=0.07415, over 4732.00 frames. ], batch size: 59, lr: 3.46e-03, grad_scale: 64.0 2023-04-27 07:46:59,231 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 07:47:07,011 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1722, 2.5302, 0.9698, 1.4074, 1.9550, 1.3276, 3.0145, 1.7260], device='cuda:4'), covar=tensor([0.0596, 0.0532, 0.0778, 0.1248, 0.0446, 0.0943, 0.0279, 0.0606], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0048, 0.0046, 0.0050, 0.0052, 0.0075, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 07:47:15,738 INFO [finetune.py:1010] (4/7) Epoch 16, validation: loss=0.1534, simple_loss=0.2252, pruned_loss=0.04076, over 2265189.00 frames. 2023-04-27 07:47:15,738 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-27 07:47:30,500 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=85930.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 07:47:34,158 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=85936.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:47:41,351 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.024e+02 1.546e+02 1.841e+02 2.213e+02 4.481e+02, threshold=3.682e+02, percent-clipped=3.0 2023-04-27 07:47:43,316 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6793, 3.5100, 0.8642, 1.8517, 2.0453, 2.3795, 2.0218, 0.9573], device='cuda:4'), covar=tensor([0.1236, 0.0716, 0.2041, 0.1200, 0.0875, 0.0987, 0.1364, 0.1947], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0244, 0.0136, 0.0121, 0.0131, 0.0152, 0.0118, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 07:47:46,977 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=85957.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:47:53,858 INFO [finetune.py:976] (4/7) Epoch 16, batch 50, loss[loss=0.1567, simple_loss=0.2271, pruned_loss=0.04317, over 4904.00 frames. ], tot_loss[loss=0.1832, simple_loss=0.2546, pruned_loss=0.05587, over 217832.83 frames. ], batch size: 36, lr: 3.46e-03, grad_scale: 64.0 2023-04-27 07:48:04,120 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=85981.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 07:48:10,282 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=85991.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 07:48:23,356 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=86002.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:48:44,016 INFO [finetune.py:976] (4/7) Epoch 16, batch 100, loss[loss=0.1235, simple_loss=0.2059, pruned_loss=0.02056, over 4758.00 frames. ], tot_loss[loss=0.1797, simple_loss=0.2493, pruned_loss=0.05508, over 382103.84 frames. ], batch size: 27, lr: 3.46e-03, grad_scale: 64.0 2023-04-27 07:48:45,853 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=86018.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:49:24,775 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5360, 3.3003, 0.8325, 1.6878, 1.9651, 2.2516, 1.9315, 0.9499], device='cuda:4'), covar=tensor([0.1340, 0.1010, 0.1980, 0.1364, 0.0986, 0.1098, 0.1458, 0.1971], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0244, 0.0136, 0.0121, 0.0131, 0.0152, 0.0118, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 07:49:27,107 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.062e+02 1.526e+02 1.932e+02 2.355e+02 3.544e+02, threshold=3.863e+02, percent-clipped=0.0 2023-04-27 07:49:34,861 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=86052.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:49:47,406 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=86063.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:49:49,098 INFO [finetune.py:976] (4/7) Epoch 16, batch 150, loss[loss=0.1476, simple_loss=0.2086, pruned_loss=0.04331, over 4814.00 frames. ], tot_loss[loss=0.1773, simple_loss=0.2457, pruned_loss=0.05442, over 510095.80 frames. ], batch size: 25, lr: 3.46e-03, grad_scale: 64.0 2023-04-27 07:49:52,567 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=86069.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:50:02,670 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=86084.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:50:13,598 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8770, 2.0206, 1.9304, 2.2307, 2.1792, 2.4976, 1.8318, 3.9785], device='cuda:4'), covar=tensor([0.0499, 0.0684, 0.0715, 0.1011, 0.0551, 0.0379, 0.0658, 0.0175], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0040, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 07:50:37,979 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=86113.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:50:45,303 INFO [finetune.py:976] (4/7) Epoch 16, batch 200, loss[loss=0.208, simple_loss=0.2569, pruned_loss=0.07952, over 4209.00 frames. ], tot_loss[loss=0.1751, simple_loss=0.2422, pruned_loss=0.05394, over 606674.19 frames. ], batch size: 65, lr: 3.46e-03, grad_scale: 64.0 2023-04-27 07:51:06,756 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=86130.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:51:07,904 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=86132.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:51:16,787 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=86139.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:51:22,243 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.150e+02 1.564e+02 1.820e+02 2.219e+02 4.567e+02, threshold=3.639e+02, percent-clipped=1.0 2023-04-27 07:51:33,767 INFO [finetune.py:976] (4/7) Epoch 16, batch 250, loss[loss=0.1611, simple_loss=0.2349, pruned_loss=0.04358, over 4754.00 frames. ], tot_loss[loss=0.1757, simple_loss=0.2437, pruned_loss=0.05381, over 684011.57 frames. ], batch size: 27, lr: 3.46e-03, grad_scale: 64.0 2023-04-27 07:51:48,353 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=86186.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:51:48,949 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=86187.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:52:06,460 INFO [finetune.py:976] (4/7) Epoch 16, batch 300, loss[loss=0.1876, simple_loss=0.2431, pruned_loss=0.06603, over 4810.00 frames. ], tot_loss[loss=0.1809, simple_loss=0.2491, pruned_loss=0.0563, over 744071.59 frames. ], batch size: 40, lr: 3.46e-03, grad_scale: 64.0 2023-04-27 07:52:10,739 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.10 vs. limit=2.0 2023-04-27 07:52:17,999 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=86231.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:52:19,820 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=86234.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:52:19,910 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.9132, 2.1119, 2.0764, 2.2703, 1.9294, 2.1467, 2.1526, 2.0959], device='cuda:4'), covar=tensor([0.3894, 0.6442, 0.5086, 0.4603, 0.5806, 0.7644, 0.6576, 0.5785], device='cuda:4'), in_proj_covar=tensor([0.0331, 0.0373, 0.0319, 0.0333, 0.0345, 0.0400, 0.0354, 0.0326], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 07:52:27,197 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=86246.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:52:28,286 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.076e+02 1.717e+02 1.966e+02 2.307e+02 3.761e+02, threshold=3.931e+02, percent-clipped=1.0 2023-04-27 07:52:31,453 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6264, 1.3707, 1.5637, 1.9630, 1.9183, 1.5858, 1.4051, 1.7986], device='cuda:4'), covar=tensor([0.0757, 0.1194, 0.0662, 0.0486, 0.0543, 0.0755, 0.0747, 0.0499], device='cuda:4'), in_proj_covar=tensor([0.0192, 0.0205, 0.0183, 0.0176, 0.0179, 0.0186, 0.0156, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 07:52:37,031 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.55 vs. limit=5.0 2023-04-27 07:52:39,055 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.62 vs. limit=2.0 2023-04-27 07:52:39,296 INFO [finetune.py:976] (4/7) Epoch 16, batch 350, loss[loss=0.1774, simple_loss=0.2498, pruned_loss=0.05248, over 4859.00 frames. ], tot_loss[loss=0.182, simple_loss=0.2507, pruned_loss=0.0567, over 791306.11 frames. ], batch size: 44, lr: 3.46e-03, grad_scale: 32.0 2023-04-27 07:52:50,051 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=86281.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:52:54,332 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=86286.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 07:53:07,230 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=86307.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:53:10,811 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=86313.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:53:12,611 INFO [finetune.py:976] (4/7) Epoch 16, batch 400, loss[loss=0.177, simple_loss=0.2567, pruned_loss=0.04864, over 4736.00 frames. ], tot_loss[loss=0.1811, simple_loss=0.251, pruned_loss=0.05561, over 828983.23 frames. ], batch size: 27, lr: 3.46e-03, grad_scale: 32.0 2023-04-27 07:53:21,591 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=86329.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:53:35,566 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.079e+02 1.581e+02 1.858e+02 2.136e+02 3.923e+02, threshold=3.716e+02, percent-clipped=0.0 2023-04-27 07:53:41,146 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=86358.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:53:42,075 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.72 vs. limit=2.0 2023-04-27 07:53:45,899 INFO [finetune.py:976] (4/7) Epoch 16, batch 450, loss[loss=0.1932, simple_loss=0.2529, pruned_loss=0.06675, over 4825.00 frames. ], tot_loss[loss=0.1796, simple_loss=0.2495, pruned_loss=0.05488, over 857384.93 frames. ], batch size: 38, lr: 3.46e-03, grad_scale: 32.0 2023-04-27 07:54:08,696 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8594, 2.4084, 1.8509, 1.8774, 1.4174, 1.4340, 1.9982, 1.4027], device='cuda:4'), covar=tensor([0.1622, 0.1486, 0.1460, 0.1733, 0.2420, 0.1881, 0.1037, 0.1984], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0214, 0.0170, 0.0206, 0.0202, 0.0185, 0.0156, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 07:54:14,885 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=86408.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:54:19,720 INFO [finetune.py:976] (4/7) Epoch 16, batch 500, loss[loss=0.2045, simple_loss=0.2641, pruned_loss=0.07244, over 4902.00 frames. ], tot_loss[loss=0.1785, simple_loss=0.2475, pruned_loss=0.05476, over 879201.22 frames. ], batch size: 35, lr: 3.46e-03, grad_scale: 32.0 2023-04-27 07:54:25,197 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=86425.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:54:34,599 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2205, 1.7755, 2.1481, 2.4677, 2.0782, 1.7006, 1.3512, 1.8311], device='cuda:4'), covar=tensor([0.3401, 0.3327, 0.1793, 0.2319, 0.3005, 0.2664, 0.4305, 0.2203], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0245, 0.0223, 0.0315, 0.0216, 0.0229, 0.0228, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 07:54:42,233 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=86440.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:54:45,737 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=86445.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:54:52,712 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.729e+01 1.556e+02 1.970e+02 2.447e+02 5.301e+02, threshold=3.940e+02, percent-clipped=3.0 2023-04-27 07:55:03,733 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-27 07:55:08,757 INFO [finetune.py:976] (4/7) Epoch 16, batch 550, loss[loss=0.1906, simple_loss=0.2525, pruned_loss=0.06434, over 4444.00 frames. ], tot_loss[loss=0.1758, simple_loss=0.2445, pruned_loss=0.05348, over 897075.29 frames. ], batch size: 19, lr: 3.46e-03, grad_scale: 32.0 2023-04-27 07:55:27,655 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-27 07:55:29,841 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.11 vs. limit=2.0 2023-04-27 07:56:00,657 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=86501.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:56:09,267 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.43 vs. limit=2.0 2023-04-27 07:56:09,557 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=86506.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 07:56:20,906 INFO [finetune.py:976] (4/7) Epoch 16, batch 600, loss[loss=0.1955, simple_loss=0.2684, pruned_loss=0.06135, over 4834.00 frames. ], tot_loss[loss=0.177, simple_loss=0.2456, pruned_loss=0.05419, over 911331.64 frames. ], batch size: 33, lr: 3.46e-03, grad_scale: 32.0 2023-04-27 07:56:32,808 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7901, 1.4108, 1.3319, 1.5770, 1.9776, 1.5604, 1.3225, 1.2574], device='cuda:4'), covar=tensor([0.1494, 0.1594, 0.2078, 0.1315, 0.0860, 0.1678, 0.1949, 0.2264], device='cuda:4'), in_proj_covar=tensor([0.0306, 0.0312, 0.0350, 0.0289, 0.0329, 0.0311, 0.0299, 0.0360], device='cuda:4'), out_proj_covar=tensor([6.3267e-05, 6.5197e-05, 7.4795e-05, 5.9126e-05, 6.8614e-05, 6.5673e-05, 6.3203e-05, 7.6881e-05], device='cuda:4') 2023-04-27 07:56:35,204 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=86531.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:56:45,565 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3938, 1.9675, 2.3048, 2.7177, 2.5654, 2.2297, 1.8360, 2.4841], device='cuda:4'), covar=tensor([0.0655, 0.1025, 0.0578, 0.0478, 0.0564, 0.0675, 0.0687, 0.0438], device='cuda:4'), in_proj_covar=tensor([0.0190, 0.0203, 0.0182, 0.0173, 0.0177, 0.0184, 0.0154, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 07:56:54,677 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.044e+02 1.714e+02 1.969e+02 2.421e+02 4.960e+02, threshold=3.938e+02, percent-clipped=3.0 2023-04-27 07:57:05,048 INFO [finetune.py:976] (4/7) Epoch 16, batch 650, loss[loss=0.195, simple_loss=0.2794, pruned_loss=0.05533, over 4816.00 frames. ], tot_loss[loss=0.1781, simple_loss=0.2476, pruned_loss=0.05436, over 922003.80 frames. ], batch size: 51, lr: 3.46e-03, grad_scale: 32.0 2023-04-27 07:57:12,975 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=86579.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:57:13,632 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=86580.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:57:18,207 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=86586.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 07:57:20,573 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3478, 1.5815, 1.3745, 1.5115, 1.3052, 1.4121, 1.4172, 1.1142], device='cuda:4'), covar=tensor([0.1637, 0.1233, 0.0936, 0.1110, 0.3379, 0.1062, 0.1589, 0.2205], device='cuda:4'), in_proj_covar=tensor([0.0289, 0.0310, 0.0222, 0.0282, 0.0315, 0.0265, 0.0252, 0.0268], device='cuda:4'), out_proj_covar=tensor([1.1635e-04, 1.2343e-04, 8.8465e-05, 1.1218e-04, 1.2824e-04, 1.0553e-04, 1.0205e-04, 1.0663e-04], device='cuda:4') 2023-04-27 07:57:29,905 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=86602.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:57:36,591 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=86613.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:57:38,332 INFO [finetune.py:976] (4/7) Epoch 16, batch 700, loss[loss=0.2012, simple_loss=0.2726, pruned_loss=0.06492, over 4872.00 frames. ], tot_loss[loss=0.1808, simple_loss=0.2503, pruned_loss=0.05566, over 927758.15 frames. ], batch size: 31, lr: 3.46e-03, grad_scale: 32.0 2023-04-27 07:57:49,360 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=86634.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 07:57:54,604 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=86641.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:57:55,230 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8499, 1.0537, 1.5738, 1.6936, 1.6551, 1.7517, 1.6004, 1.5695], device='cuda:4'), covar=tensor([0.4192, 0.5487, 0.4606, 0.4697, 0.5409, 0.6880, 0.5017, 0.4706], device='cuda:4'), in_proj_covar=tensor([0.0330, 0.0371, 0.0318, 0.0332, 0.0344, 0.0397, 0.0352, 0.0325], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 07:58:00,294 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.046e+02 1.513e+02 1.841e+02 2.164e+02 3.453e+02, threshold=3.681e+02, percent-clipped=0.0 2023-04-27 07:58:06,398 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=86658.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:58:08,180 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=86661.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:58:11,213 INFO [finetune.py:976] (4/7) Epoch 16, batch 750, loss[loss=0.1571, simple_loss=0.2524, pruned_loss=0.03089, over 4841.00 frames. ], tot_loss[loss=0.1822, simple_loss=0.252, pruned_loss=0.05624, over 932874.94 frames. ], batch size: 49, lr: 3.46e-03, grad_scale: 32.0 2023-04-27 07:58:15,646 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-27 07:58:33,234 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1370, 1.2998, 1.5767, 1.7314, 1.6176, 1.7567, 1.6265, 1.6325], device='cuda:4'), covar=tensor([0.3911, 0.5359, 0.4538, 0.4245, 0.5429, 0.7496, 0.4861, 0.4338], device='cuda:4'), in_proj_covar=tensor([0.0330, 0.0371, 0.0318, 0.0333, 0.0344, 0.0397, 0.0352, 0.0325], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 07:58:35,082 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2394, 1.4374, 1.7907, 1.9016, 1.8484, 1.9264, 1.8145, 1.8124], device='cuda:4'), covar=tensor([0.3851, 0.5613, 0.4957, 0.5038, 0.5601, 0.7848, 0.5342, 0.5173], device='cuda:4'), in_proj_covar=tensor([0.0330, 0.0371, 0.0318, 0.0333, 0.0344, 0.0397, 0.0352, 0.0325], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 07:58:38,828 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=86706.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:58:38,861 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4118, 3.2373, 0.7719, 1.7099, 1.8686, 2.2811, 1.8071, 0.8917], device='cuda:4'), covar=tensor([0.1440, 0.0889, 0.2144, 0.1304, 0.1063, 0.1036, 0.1513, 0.2051], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0244, 0.0136, 0.0120, 0.0131, 0.0152, 0.0119, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 07:58:40,066 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=86708.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:58:44,816 INFO [finetune.py:976] (4/7) Epoch 16, batch 800, loss[loss=0.1573, simple_loss=0.2268, pruned_loss=0.04393, over 4823.00 frames. ], tot_loss[loss=0.181, simple_loss=0.2511, pruned_loss=0.05547, over 938472.24 frames. ], batch size: 38, lr: 3.46e-03, grad_scale: 32.0 2023-04-27 07:58:50,358 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=86725.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:58:50,407 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8365, 1.4342, 1.9196, 2.3251, 1.9292, 1.7691, 1.8671, 1.8647], device='cuda:4'), covar=tensor([0.4563, 0.6685, 0.6365, 0.5665, 0.5902, 0.8045, 0.7850, 0.8679], device='cuda:4'), in_proj_covar=tensor([0.0418, 0.0408, 0.0496, 0.0506, 0.0447, 0.0470, 0.0477, 0.0480], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 07:58:58,283 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4611, 1.3855, 1.6880, 1.6920, 1.3186, 1.1801, 1.4354, 0.8712], device='cuda:4'), covar=tensor([0.0624, 0.0647, 0.0389, 0.0493, 0.0725, 0.1022, 0.0549, 0.0652], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0070, 0.0070, 0.0068, 0.0076, 0.0097, 0.0076, 0.0069], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 07:59:00,611 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9689, 2.5367, 1.9639, 1.8135, 1.4395, 1.5047, 2.0003, 1.4856], device='cuda:4'), covar=tensor([0.1693, 0.1329, 0.1507, 0.1799, 0.2379, 0.1894, 0.0977, 0.2001], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0212, 0.0168, 0.0203, 0.0199, 0.0183, 0.0155, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 07:59:05,803 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.056e+02 1.560e+02 1.911e+02 2.310e+02 3.714e+02, threshold=3.821e+02, percent-clipped=1.0 2023-04-27 07:59:11,623 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=86756.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:59:17,588 INFO [finetune.py:976] (4/7) Epoch 16, batch 850, loss[loss=0.2158, simple_loss=0.2872, pruned_loss=0.07218, over 4704.00 frames. ], tot_loss[loss=0.1802, simple_loss=0.2493, pruned_loss=0.05552, over 942248.47 frames. ], batch size: 59, lr: 3.46e-03, grad_scale: 32.0 2023-04-27 07:59:21,928 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=86773.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:59:36,380 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=86796.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:59:39,983 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=86801.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 07:59:49,824 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=86814.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 07:59:50,492 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7047, 1.2108, 1.7625, 2.2178, 1.8183, 1.6856, 1.7686, 1.7102], device='cuda:4'), covar=tensor([0.4320, 0.6173, 0.5512, 0.5482, 0.5299, 0.7230, 0.6938, 0.8097], device='cuda:4'), in_proj_covar=tensor([0.0419, 0.0409, 0.0496, 0.0507, 0.0448, 0.0471, 0.0478, 0.0481], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 07:59:50,969 INFO [finetune.py:976] (4/7) Epoch 16, batch 900, loss[loss=0.1815, simple_loss=0.2441, pruned_loss=0.0594, over 4843.00 frames. ], tot_loss[loss=0.1771, simple_loss=0.2458, pruned_loss=0.05422, over 943871.31 frames. ], batch size: 49, lr: 3.46e-03, grad_scale: 32.0 2023-04-27 07:59:58,453 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2055, 1.8630, 2.1195, 2.6247, 2.1834, 1.7134, 1.6988, 1.9648], device='cuda:4'), covar=tensor([0.2750, 0.3078, 0.1527, 0.1973, 0.2387, 0.2385, 0.4049, 0.2217], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0249, 0.0225, 0.0318, 0.0219, 0.0231, 0.0230, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 08:00:22,181 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.095e+02 1.518e+02 1.851e+02 2.220e+02 4.721e+02, threshold=3.701e+02, percent-clipped=1.0 2023-04-27 08:00:52,315 INFO [finetune.py:976] (4/7) Epoch 16, batch 950, loss[loss=0.1789, simple_loss=0.2601, pruned_loss=0.04888, over 4817.00 frames. ], tot_loss[loss=0.176, simple_loss=0.2443, pruned_loss=0.05381, over 945939.34 frames. ], batch size: 41, lr: 3.46e-03, grad_scale: 32.0 2023-04-27 08:01:03,389 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=86875.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:01:05,152 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=86878.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:01:05,186 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7173, 2.3699, 1.7106, 1.8012, 1.2898, 1.3070, 1.8470, 1.2256], device='cuda:4'), covar=tensor([0.1639, 0.1268, 0.1430, 0.1599, 0.2296, 0.1914, 0.0975, 0.2029], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0212, 0.0168, 0.0203, 0.0199, 0.0183, 0.0155, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 08:01:37,258 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=86902.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:01:58,434 INFO [finetune.py:976] (4/7) Epoch 16, batch 1000, loss[loss=0.1279, simple_loss=0.1947, pruned_loss=0.03053, over 4734.00 frames. ], tot_loss[loss=0.1758, simple_loss=0.2446, pruned_loss=0.05348, over 946218.56 frames. ], batch size: 23, lr: 3.46e-03, grad_scale: 32.0 2023-04-27 08:02:18,019 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=86936.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:02:19,922 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=86939.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:02:31,464 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.164e+02 1.765e+02 1.981e+02 2.453e+02 9.285e+02, threshold=3.962e+02, percent-clipped=3.0 2023-04-27 08:02:37,036 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=86950.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:02:59,852 INFO [finetune.py:976] (4/7) Epoch 16, batch 1050, loss[loss=0.1899, simple_loss=0.2673, pruned_loss=0.05628, over 4929.00 frames. ], tot_loss[loss=0.1777, simple_loss=0.2475, pruned_loss=0.05389, over 948212.72 frames. ], batch size: 33, lr: 3.46e-03, grad_scale: 32.0 2023-04-27 08:03:05,915 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2577, 1.5410, 1.4300, 1.4845, 1.2922, 1.3889, 1.3857, 1.0829], device='cuda:4'), covar=tensor([0.1673, 0.1179, 0.0837, 0.1179, 0.3064, 0.1085, 0.1567, 0.1932], device='cuda:4'), in_proj_covar=tensor([0.0291, 0.0315, 0.0224, 0.0285, 0.0317, 0.0268, 0.0255, 0.0271], device='cuda:4'), out_proj_covar=tensor([1.1730e-04, 1.2519e-04, 8.9292e-05, 1.1343e-04, 1.2901e-04, 1.0679e-04, 1.0303e-04, 1.0780e-04], device='cuda:4') 2023-04-27 08:03:13,849 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6279, 1.4669, 0.7720, 1.2354, 1.8022, 1.4942, 1.3628, 1.3619], device='cuda:4'), covar=tensor([0.0500, 0.0400, 0.0365, 0.0596, 0.0269, 0.0544, 0.0503, 0.0599], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0024, 0.0023, 0.0029, 0.0020, 0.0028, 0.0028, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0045, 0.0038, 0.0051, 0.0038, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 08:03:28,654 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=87002.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 08:03:43,716 INFO [finetune.py:976] (4/7) Epoch 16, batch 1100, loss[loss=0.2253, simple_loss=0.2963, pruned_loss=0.07717, over 4898.00 frames. ], tot_loss[loss=0.1786, simple_loss=0.249, pruned_loss=0.05407, over 951753.09 frames. ], batch size: 43, lr: 3.45e-03, grad_scale: 32.0 2023-04-27 08:04:06,136 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.082e+02 1.660e+02 1.994e+02 2.342e+02 3.600e+02, threshold=3.988e+02, percent-clipped=0.0 2023-04-27 08:04:15,691 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=87063.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 08:04:17,376 INFO [finetune.py:976] (4/7) Epoch 16, batch 1150, loss[loss=0.1712, simple_loss=0.2412, pruned_loss=0.05062, over 4896.00 frames. ], tot_loss[loss=0.1799, simple_loss=0.2504, pruned_loss=0.05472, over 950556.24 frames. ], batch size: 46, lr: 3.45e-03, grad_scale: 32.0 2023-04-27 08:04:37,606 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=87096.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:04:40,688 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=87101.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 08:04:50,576 INFO [finetune.py:976] (4/7) Epoch 16, batch 1200, loss[loss=0.1967, simple_loss=0.2586, pruned_loss=0.0674, over 4912.00 frames. ], tot_loss[loss=0.1797, simple_loss=0.2501, pruned_loss=0.0547, over 952304.66 frames. ], batch size: 37, lr: 3.45e-03, grad_scale: 32.0 2023-04-27 08:05:10,106 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=87144.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:05:13,049 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.133e+02 1.748e+02 2.017e+02 2.271e+02 5.228e+02, threshold=4.034e+02, percent-clipped=1.0 2023-04-27 08:05:13,120 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=87149.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:05:24,041 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3405, 1.7330, 2.2468, 2.7007, 2.1915, 1.6921, 1.4725, 1.9989], device='cuda:4'), covar=tensor([0.3104, 0.3245, 0.1517, 0.2417, 0.2555, 0.2729, 0.4110, 0.2062], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0247, 0.0224, 0.0316, 0.0217, 0.0230, 0.0228, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 08:05:24,514 INFO [finetune.py:976] (4/7) Epoch 16, batch 1250, loss[loss=0.1568, simple_loss=0.2252, pruned_loss=0.04421, over 4932.00 frames. ], tot_loss[loss=0.1777, simple_loss=0.2477, pruned_loss=0.05381, over 953196.57 frames. ], batch size: 33, lr: 3.45e-03, grad_scale: 32.0 2023-04-27 08:05:27,050 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=87170.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:05:28,864 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-27 08:05:36,827 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5825, 3.4685, 2.7200, 3.1118, 2.6359, 2.7519, 2.9988, 2.2966], device='cuda:4'), covar=tensor([0.1924, 0.1001, 0.0819, 0.1294, 0.2565, 0.1305, 0.1721, 0.2557], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0312, 0.0223, 0.0283, 0.0315, 0.0267, 0.0254, 0.0269], device='cuda:4'), out_proj_covar=tensor([1.1672e-04, 1.2420e-04, 8.8793e-05, 1.1272e-04, 1.2822e-04, 1.0634e-04, 1.0261e-04, 1.0727e-04], device='cuda:4') 2023-04-27 08:05:44,984 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.58 vs. limit=2.0 2023-04-27 08:05:58,111 INFO [finetune.py:976] (4/7) Epoch 16, batch 1300, loss[loss=0.1907, simple_loss=0.2573, pruned_loss=0.062, over 4747.00 frames. ], tot_loss[loss=0.1748, simple_loss=0.2442, pruned_loss=0.05272, over 954009.14 frames. ], batch size: 26, lr: 3.45e-03, grad_scale: 32.0 2023-04-27 08:06:09,228 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=87230.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:06:12,088 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=87234.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:06:13,313 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=87236.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:06:21,119 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.142e+02 1.625e+02 1.883e+02 2.267e+02 3.477e+02, threshold=3.766e+02, percent-clipped=0.0 2023-04-27 08:06:31,008 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.8517, 1.8911, 1.7313, 1.5869, 2.1197, 1.8671, 2.5107, 1.4816], device='cuda:4'), covar=tensor([0.3545, 0.1835, 0.4747, 0.2708, 0.1544, 0.1962, 0.1464, 0.4647], device='cuda:4'), in_proj_covar=tensor([0.0347, 0.0354, 0.0433, 0.0364, 0.0391, 0.0387, 0.0378, 0.0427], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 08:06:31,975 INFO [finetune.py:976] (4/7) Epoch 16, batch 1350, loss[loss=0.1778, simple_loss=0.2566, pruned_loss=0.04949, over 4912.00 frames. ], tot_loss[loss=0.1765, simple_loss=0.2454, pruned_loss=0.05378, over 953697.14 frames. ], batch size: 38, lr: 3.45e-03, grad_scale: 32.0 2023-04-27 08:06:45,593 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=87284.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:06:50,334 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=87291.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:07:16,734 INFO [finetune.py:976] (4/7) Epoch 16, batch 1400, loss[loss=0.2455, simple_loss=0.3074, pruned_loss=0.09176, over 4809.00 frames. ], tot_loss[loss=0.1806, simple_loss=0.2504, pruned_loss=0.05539, over 953361.92 frames. ], batch size: 39, lr: 3.45e-03, grad_scale: 32.0 2023-04-27 08:07:19,006 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.77 vs. limit=2.0 2023-04-27 08:07:47,239 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.81 vs. limit=5.0 2023-04-27 08:07:58,926 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.057e+02 1.760e+02 2.100e+02 2.306e+02 6.240e+02, threshold=4.200e+02, percent-clipped=3.0 2023-04-27 08:08:08,071 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9535, 1.4401, 1.8209, 1.7251, 1.7541, 1.4745, 0.8412, 1.4200], device='cuda:4'), covar=tensor([0.3415, 0.3260, 0.1684, 0.2259, 0.2508, 0.2531, 0.4060, 0.2015], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0246, 0.0223, 0.0315, 0.0217, 0.0229, 0.0228, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 08:08:15,360 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=87358.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 08:08:17,817 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=87362.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:08:20,154 INFO [finetune.py:976] (4/7) Epoch 16, batch 1450, loss[loss=0.1772, simple_loss=0.2451, pruned_loss=0.05469, over 4891.00 frames. ], tot_loss[loss=0.1824, simple_loss=0.2524, pruned_loss=0.05623, over 955459.94 frames. ], batch size: 35, lr: 3.45e-03, grad_scale: 32.0 2023-04-27 08:09:08,677 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3799, 2.2250, 2.4436, 2.8284, 2.6532, 2.1374, 1.9703, 2.2881], device='cuda:4'), covar=tensor([0.0869, 0.0979, 0.0705, 0.0577, 0.0673, 0.0913, 0.0806, 0.0640], device='cuda:4'), in_proj_covar=tensor([0.0192, 0.0205, 0.0185, 0.0175, 0.0179, 0.0185, 0.0156, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 08:09:14,068 INFO [finetune.py:976] (4/7) Epoch 16, batch 1500, loss[loss=0.1771, simple_loss=0.2489, pruned_loss=0.05269, over 4830.00 frames. ], tot_loss[loss=0.183, simple_loss=0.2535, pruned_loss=0.05623, over 956645.00 frames. ], batch size: 47, lr: 3.45e-03, grad_scale: 32.0 2023-04-27 08:09:18,467 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=87423.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:09:36,462 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.033e+02 1.658e+02 1.977e+02 2.415e+02 4.579e+02, threshold=3.953e+02, percent-clipped=1.0 2023-04-27 08:09:43,862 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=87461.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:09:46,778 INFO [finetune.py:976] (4/7) Epoch 16, batch 1550, loss[loss=0.1802, simple_loss=0.2539, pruned_loss=0.05328, over 4778.00 frames. ], tot_loss[loss=0.1815, simple_loss=0.2521, pruned_loss=0.0555, over 955942.96 frames. ], batch size: 51, lr: 3.45e-03, grad_scale: 32.0 2023-04-27 08:09:49,332 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=87470.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:10:30,187 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=87505.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:10:42,329 INFO [finetune.py:976] (4/7) Epoch 16, batch 1600, loss[loss=0.1624, simple_loss=0.2276, pruned_loss=0.04855, over 4805.00 frames. ], tot_loss[loss=0.18, simple_loss=0.2503, pruned_loss=0.05482, over 957460.63 frames. ], batch size: 25, lr: 3.45e-03, grad_scale: 32.0 2023-04-27 08:10:43,625 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=87518.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:10:46,103 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=87522.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:10:54,389 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=87534.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:11:05,883 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.006e+02 1.660e+02 1.902e+02 2.277e+02 5.320e+02, threshold=3.803e+02, percent-clipped=1.0 2023-04-27 08:11:16,206 INFO [finetune.py:976] (4/7) Epoch 16, batch 1650, loss[loss=0.1823, simple_loss=0.2493, pruned_loss=0.05762, over 4726.00 frames. ], tot_loss[loss=0.177, simple_loss=0.247, pruned_loss=0.05352, over 956352.92 frames. ], batch size: 59, lr: 3.45e-03, grad_scale: 32.0 2023-04-27 08:11:16,326 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=87566.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 08:11:26,482 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=87582.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:11:28,266 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=87584.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:11:29,409 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=87586.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:11:31,297 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=87589.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:11:43,105 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=87605.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:11:49,704 INFO [finetune.py:976] (4/7) Epoch 16, batch 1700, loss[loss=0.1957, simple_loss=0.2788, pruned_loss=0.05626, over 4852.00 frames. ], tot_loss[loss=0.1754, simple_loss=0.245, pruned_loss=0.05289, over 956039.41 frames. ], batch size: 44, lr: 3.45e-03, grad_scale: 32.0 2023-04-27 08:12:06,655 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.77 vs. limit=2.0 2023-04-27 08:12:08,462 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=87645.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 08:12:12,238 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.051e+02 1.609e+02 2.053e+02 2.465e+02 3.849e+02, threshold=4.106e+02, percent-clipped=1.0 2023-04-27 08:12:13,454 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=87650.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:12:16,727 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-27 08:12:18,811 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=87658.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 08:12:20,045 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9125, 2.3549, 1.9138, 2.2206, 1.6858, 1.9740, 2.0053, 1.4179], device='cuda:4'), covar=tensor([0.1996, 0.1381, 0.0974, 0.1378, 0.3169, 0.1309, 0.1873, 0.2864], device='cuda:4'), in_proj_covar=tensor([0.0287, 0.0310, 0.0221, 0.0282, 0.0314, 0.0263, 0.0252, 0.0267], device='cuda:4'), out_proj_covar=tensor([1.1543e-04, 1.2345e-04, 8.8110e-05, 1.1216e-04, 1.2775e-04, 1.0489e-04, 1.0200e-04, 1.0625e-04], device='cuda:4') 2023-04-27 08:12:23,601 INFO [finetune.py:976] (4/7) Epoch 16, batch 1750, loss[loss=0.1571, simple_loss=0.2293, pruned_loss=0.0425, over 4897.00 frames. ], tot_loss[loss=0.1779, simple_loss=0.2477, pruned_loss=0.05404, over 954685.66 frames. ], batch size: 35, lr: 3.45e-03, grad_scale: 32.0 2023-04-27 08:12:23,725 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=87666.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 08:12:39,362 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-27 08:12:55,815 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=87706.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 08:13:07,610 INFO [finetune.py:976] (4/7) Epoch 16, batch 1800, loss[loss=0.1961, simple_loss=0.2731, pruned_loss=0.05958, over 4841.00 frames. ], tot_loss[loss=0.1803, simple_loss=0.2505, pruned_loss=0.05501, over 955490.68 frames. ], batch size: 47, lr: 3.45e-03, grad_scale: 32.0 2023-04-27 08:13:14,245 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=87718.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:13:48,825 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.59 vs. limit=2.0 2023-04-27 08:13:49,464 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.133e+02 1.657e+02 1.996e+02 2.458e+02 4.294e+02, threshold=3.992e+02, percent-clipped=1.0 2023-04-27 08:14:08,974 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7812, 1.2689, 1.8678, 2.2920, 1.8645, 1.7265, 1.7856, 1.7565], device='cuda:4'), covar=tensor([0.4693, 0.6781, 0.5907, 0.6075, 0.6265, 0.7801, 0.7996, 0.7872], device='cuda:4'), in_proj_covar=tensor([0.0417, 0.0406, 0.0494, 0.0505, 0.0445, 0.0470, 0.0476, 0.0479], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 08:14:12,970 INFO [finetune.py:976] (4/7) Epoch 16, batch 1850, loss[loss=0.1643, simple_loss=0.2441, pruned_loss=0.04224, over 4785.00 frames. ], tot_loss[loss=0.1831, simple_loss=0.2533, pruned_loss=0.0565, over 956408.89 frames. ], batch size: 51, lr: 3.45e-03, grad_scale: 32.0 2023-04-27 08:15:07,317 INFO [finetune.py:976] (4/7) Epoch 16, batch 1900, loss[loss=0.1728, simple_loss=0.248, pruned_loss=0.04876, over 4905.00 frames. ], tot_loss[loss=0.1829, simple_loss=0.2538, pruned_loss=0.05603, over 957408.01 frames. ], batch size: 37, lr: 3.45e-03, grad_scale: 32.0 2023-04-27 08:15:08,022 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=87817.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:15:28,299 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.071e+02 1.628e+02 1.844e+02 2.203e+02 4.331e+02, threshold=3.688e+02, percent-clipped=1.0 2023-04-27 08:15:37,121 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=87861.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 08:15:40,075 INFO [finetune.py:976] (4/7) Epoch 16, batch 1950, loss[loss=0.1369, simple_loss=0.2016, pruned_loss=0.03605, over 3947.00 frames. ], tot_loss[loss=0.1801, simple_loss=0.2507, pruned_loss=0.05475, over 957591.17 frames. ], batch size: 17, lr: 3.45e-03, grad_scale: 32.0 2023-04-27 08:16:03,453 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.6940, 1.7287, 1.7808, 1.3537, 1.8958, 1.5574, 2.3454, 1.5198], device='cuda:4'), covar=tensor([0.3665, 0.1763, 0.4391, 0.2814, 0.1446, 0.2321, 0.1330, 0.4587], device='cuda:4'), in_proj_covar=tensor([0.0342, 0.0350, 0.0428, 0.0358, 0.0385, 0.0383, 0.0373, 0.0422], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 08:16:10,454 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=87886.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:16:47,482 INFO [finetune.py:976] (4/7) Epoch 16, batch 2000, loss[loss=0.1841, simple_loss=0.2547, pruned_loss=0.05679, over 4851.00 frames. ], tot_loss[loss=0.178, simple_loss=0.2481, pruned_loss=0.05391, over 956774.79 frames. ], batch size: 49, lr: 3.45e-03, grad_scale: 32.0 2023-04-27 08:16:49,946 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0759, 2.5199, 2.1591, 2.2890, 1.7187, 2.1914, 2.1977, 1.5623], device='cuda:4'), covar=tensor([0.1683, 0.0992, 0.0760, 0.1296, 0.2934, 0.1152, 0.1740, 0.2533], device='cuda:4'), in_proj_covar=tensor([0.0286, 0.0309, 0.0220, 0.0281, 0.0312, 0.0261, 0.0251, 0.0265], device='cuda:4'), out_proj_covar=tensor([1.1485e-04, 1.2315e-04, 8.7792e-05, 1.1173e-04, 1.2706e-04, 1.0418e-04, 1.0151e-04, 1.0564e-04], device='cuda:4') 2023-04-27 08:16:49,986 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7374, 1.3156, 1.7874, 2.2876, 1.8416, 1.6853, 1.7751, 1.7112], device='cuda:4'), covar=tensor([0.4579, 0.6644, 0.6652, 0.5634, 0.5650, 0.7763, 0.7913, 0.8406], device='cuda:4'), in_proj_covar=tensor([0.0416, 0.0405, 0.0493, 0.0504, 0.0444, 0.0469, 0.0474, 0.0478], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 08:16:59,000 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=87934.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:17:02,680 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=87940.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 08:17:05,749 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=87945.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:17:08,486 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.110e+02 1.621e+02 1.906e+02 2.380e+02 5.072e+02, threshold=3.811e+02, percent-clipped=4.0 2023-04-27 08:17:17,324 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=87961.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 08:17:21,297 INFO [finetune.py:976] (4/7) Epoch 16, batch 2050, loss[loss=0.1416, simple_loss=0.2186, pruned_loss=0.0323, over 4028.00 frames. ], tot_loss[loss=0.1762, simple_loss=0.2455, pruned_loss=0.05344, over 957601.94 frames. ], batch size: 17, lr: 3.45e-03, grad_scale: 32.0 2023-04-27 08:17:21,734 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.43 vs. limit=2.0 2023-04-27 08:17:55,291 INFO [finetune.py:976] (4/7) Epoch 16, batch 2100, loss[loss=0.1827, simple_loss=0.2604, pruned_loss=0.05247, over 4902.00 frames. ], tot_loss[loss=0.1763, simple_loss=0.2454, pruned_loss=0.05359, over 958114.15 frames. ], batch size: 35, lr: 3.45e-03, grad_scale: 32.0 2023-04-27 08:17:57,101 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=88018.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:18:15,155 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9701, 2.4107, 1.0277, 1.3581, 1.9292, 1.2281, 3.1492, 1.5781], device='cuda:4'), covar=tensor([0.0686, 0.0689, 0.0755, 0.1202, 0.0468, 0.0958, 0.0206, 0.0617], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0048, 0.0046, 0.0050, 0.0052, 0.0075, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0007, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 08:18:16,282 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.262e+02 1.786e+02 2.026e+02 2.569e+02 5.831e+02, threshold=4.052e+02, percent-clipped=4.0 2023-04-27 08:18:28,508 INFO [finetune.py:976] (4/7) Epoch 16, batch 2150, loss[loss=0.2202, simple_loss=0.2836, pruned_loss=0.07842, over 4821.00 frames. ], tot_loss[loss=0.1801, simple_loss=0.2498, pruned_loss=0.05525, over 957390.56 frames. ], batch size: 33, lr: 3.45e-03, grad_scale: 32.0 2023-04-27 08:18:29,059 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=88066.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:18:29,947 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.91 vs. limit=2.0 2023-04-27 08:18:50,533 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=88100.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:18:53,519 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.55 vs. limit=2.0 2023-04-27 08:19:01,084 INFO [finetune.py:976] (4/7) Epoch 16, batch 2200, loss[loss=0.1907, simple_loss=0.2594, pruned_loss=0.06096, over 4294.00 frames. ], tot_loss[loss=0.1818, simple_loss=0.252, pruned_loss=0.05579, over 955772.48 frames. ], batch size: 66, lr: 3.45e-03, grad_scale: 32.0 2023-04-27 08:19:02,312 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=88117.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:19:11,116 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0501, 0.6243, 0.9031, 0.7344, 1.2346, 0.9906, 0.8226, 0.9373], device='cuda:4'), covar=tensor([0.1651, 0.1653, 0.2191, 0.1695, 0.1052, 0.1445, 0.1763, 0.2376], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0314, 0.0354, 0.0292, 0.0332, 0.0313, 0.0303, 0.0364], device='cuda:4'), out_proj_covar=tensor([6.3787e-05, 6.5453e-05, 7.5584e-05, 5.9601e-05, 6.9119e-05, 6.6156e-05, 6.3967e-05, 7.7718e-05], device='cuda:4') 2023-04-27 08:19:34,551 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.043e+02 1.597e+02 1.844e+02 2.202e+02 4.029e+02, threshold=3.688e+02, percent-clipped=0.0 2023-04-27 08:19:47,615 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=88161.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 08:19:47,633 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=88161.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:19:55,583 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=88165.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:19:56,111 INFO [finetune.py:976] (4/7) Epoch 16, batch 2250, loss[loss=0.1809, simple_loss=0.2537, pruned_loss=0.0541, over 4841.00 frames. ], tot_loss[loss=0.1825, simple_loss=0.2529, pruned_loss=0.05608, over 955746.12 frames. ], batch size: 49, lr: 3.45e-03, grad_scale: 32.0 2023-04-27 08:20:09,345 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.51 vs. limit=5.0 2023-04-27 08:20:50,810 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=88209.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:21:01,646 INFO [finetune.py:976] (4/7) Epoch 16, batch 2300, loss[loss=0.1827, simple_loss=0.2538, pruned_loss=0.05581, over 4780.00 frames. ], tot_loss[loss=0.1812, simple_loss=0.2519, pruned_loss=0.05529, over 955732.89 frames. ], batch size: 26, lr: 3.44e-03, grad_scale: 32.0 2023-04-27 08:21:07,856 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-27 08:21:18,860 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=88240.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:21:21,932 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=88245.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:21:24,254 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.059e+01 1.606e+02 1.877e+02 2.265e+02 8.419e+02, threshold=3.754e+02, percent-clipped=4.0 2023-04-27 08:21:32,159 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=88261.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 08:21:41,586 INFO [finetune.py:976] (4/7) Epoch 16, batch 2350, loss[loss=0.2057, simple_loss=0.2717, pruned_loss=0.06985, over 4920.00 frames. ], tot_loss[loss=0.179, simple_loss=0.2495, pruned_loss=0.05429, over 957840.98 frames. ], batch size: 37, lr: 3.44e-03, grad_scale: 64.0 2023-04-27 08:22:14,473 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=88288.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:22:16,365 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=88291.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 08:22:17,547 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=88293.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:22:19,434 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6163, 1.6991, 0.7552, 1.3147, 1.9287, 1.4971, 1.4119, 1.4873], device='cuda:4'), covar=tensor([0.0500, 0.0365, 0.0366, 0.0556, 0.0243, 0.0519, 0.0479, 0.0580], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0024, 0.0022, 0.0029, 0.0020, 0.0028, 0.0028, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0044, 0.0037, 0.0051, 0.0038, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 08:22:38,196 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=88309.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:22:48,484 INFO [finetune.py:976] (4/7) Epoch 16, batch 2400, loss[loss=0.1816, simple_loss=0.2482, pruned_loss=0.05744, over 4863.00 frames. ], tot_loss[loss=0.177, simple_loss=0.2466, pruned_loss=0.05365, over 958262.05 frames. ], batch size: 44, lr: 3.44e-03, grad_scale: 64.0 2023-04-27 08:23:00,221 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5404, 1.4368, 1.8161, 1.8860, 1.4385, 1.2971, 1.5371, 0.9120], device='cuda:4'), covar=tensor([0.0585, 0.0637, 0.0419, 0.0572, 0.0827, 0.1193, 0.0646, 0.0743], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0070, 0.0069, 0.0067, 0.0075, 0.0096, 0.0075, 0.0068], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 08:23:05,267 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-27 08:23:16,408 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-04-27 08:23:19,575 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-27 08:23:23,057 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.024e+02 1.498e+02 1.832e+02 2.156e+02 3.450e+02, threshold=3.664e+02, percent-clipped=0.0 2023-04-27 08:23:25,028 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=88352.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 08:23:33,494 INFO [finetune.py:976] (4/7) Epoch 16, batch 2450, loss[loss=0.2213, simple_loss=0.2684, pruned_loss=0.08712, over 4830.00 frames. ], tot_loss[loss=0.1744, simple_loss=0.2434, pruned_loss=0.05263, over 958112.96 frames. ], batch size: 30, lr: 3.44e-03, grad_scale: 64.0 2023-04-27 08:23:52,925 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4343, 2.9263, 0.8705, 1.7841, 2.2927, 1.5844, 4.0648, 2.0678], device='cuda:4'), covar=tensor([0.0662, 0.0835, 0.0935, 0.1232, 0.0497, 0.0927, 0.0212, 0.0582], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0049, 0.0046, 0.0050, 0.0052, 0.0075, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0007, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 08:24:30,565 INFO [finetune.py:976] (4/7) Epoch 16, batch 2500, loss[loss=0.2084, simple_loss=0.2935, pruned_loss=0.06167, over 4858.00 frames. ], tot_loss[loss=0.1762, simple_loss=0.2457, pruned_loss=0.05338, over 958094.92 frames. ], batch size: 44, lr: 3.44e-03, grad_scale: 32.0 2023-04-27 08:24:35,459 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.98 vs. limit=5.0 2023-04-27 08:24:54,799 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.230e+02 1.712e+02 2.054e+02 2.521e+02 4.262e+02, threshold=4.109e+02, percent-clipped=2.0 2023-04-27 08:24:58,641 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=88456.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:25:04,747 INFO [finetune.py:976] (4/7) Epoch 16, batch 2550, loss[loss=0.1377, simple_loss=0.2162, pruned_loss=0.02957, over 4789.00 frames. ], tot_loss[loss=0.1788, simple_loss=0.249, pruned_loss=0.05433, over 954519.72 frames. ], batch size: 26, lr: 3.44e-03, grad_scale: 32.0 2023-04-27 08:25:35,701 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8572, 1.4386, 1.9888, 2.3680, 1.9693, 1.8425, 1.8952, 1.8763], device='cuda:4'), covar=tensor([0.5268, 0.7425, 0.6738, 0.6350, 0.6916, 0.9354, 0.8779, 0.9149], device='cuda:4'), in_proj_covar=tensor([0.0419, 0.0408, 0.0495, 0.0507, 0.0448, 0.0473, 0.0478, 0.0481], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 08:25:38,575 INFO [finetune.py:976] (4/7) Epoch 16, batch 2600, loss[loss=0.1744, simple_loss=0.2488, pruned_loss=0.05, over 4809.00 frames. ], tot_loss[loss=0.1799, simple_loss=0.2508, pruned_loss=0.05449, over 956542.94 frames. ], batch size: 38, lr: 3.44e-03, grad_scale: 32.0 2023-04-27 08:26:18,045 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.329e+02 1.735e+02 2.047e+02 2.467e+02 4.454e+02, threshold=4.094e+02, percent-clipped=2.0 2023-04-27 08:26:30,454 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7338, 1.7509, 0.8700, 1.4039, 1.8836, 1.6031, 1.4722, 1.5637], device='cuda:4'), covar=tensor([0.0522, 0.0364, 0.0339, 0.0568, 0.0272, 0.0537, 0.0522, 0.0567], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0024, 0.0022, 0.0029, 0.0020, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0044, 0.0038, 0.0051, 0.0038, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 08:26:39,229 INFO [finetune.py:976] (4/7) Epoch 16, batch 2650, loss[loss=0.2455, simple_loss=0.3081, pruned_loss=0.09141, over 4813.00 frames. ], tot_loss[loss=0.1801, simple_loss=0.2515, pruned_loss=0.05439, over 956527.64 frames. ], batch size: 40, lr: 3.44e-03, grad_scale: 32.0 2023-04-27 08:27:20,832 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8265, 1.5025, 1.9194, 2.2709, 1.9814, 1.8198, 1.9248, 1.8011], device='cuda:4'), covar=tensor([0.4213, 0.6086, 0.5535, 0.5073, 0.5293, 0.7320, 0.7327, 0.7780], device='cuda:4'), in_proj_covar=tensor([0.0420, 0.0410, 0.0496, 0.0509, 0.0449, 0.0474, 0.0480, 0.0482], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 08:27:29,037 INFO [finetune.py:976] (4/7) Epoch 16, batch 2700, loss[loss=0.187, simple_loss=0.268, pruned_loss=0.05306, over 4809.00 frames. ], tot_loss[loss=0.1796, simple_loss=0.2509, pruned_loss=0.05416, over 955277.58 frames. ], batch size: 40, lr: 3.44e-03, grad_scale: 32.0 2023-04-27 08:28:11,602 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=88647.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 08:28:13,350 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.055e+02 1.464e+02 1.764e+02 2.078e+02 3.620e+02, threshold=3.528e+02, percent-clipped=0.0 2023-04-27 08:28:36,532 INFO [finetune.py:976] (4/7) Epoch 16, batch 2750, loss[loss=0.1706, simple_loss=0.2376, pruned_loss=0.05173, over 4897.00 frames. ], tot_loss[loss=0.178, simple_loss=0.2483, pruned_loss=0.0538, over 955409.19 frames. ], batch size: 35, lr: 3.44e-03, grad_scale: 32.0 2023-04-27 08:28:43,420 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.8427, 3.7575, 2.7845, 4.4081, 3.8844, 3.8756, 1.8094, 3.7821], device='cuda:4'), covar=tensor([0.2071, 0.1335, 0.3324, 0.1791, 0.3280, 0.1924, 0.5845, 0.2638], device='cuda:4'), in_proj_covar=tensor([0.0241, 0.0211, 0.0251, 0.0301, 0.0296, 0.0246, 0.0268, 0.0270], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 08:29:21,110 INFO [finetune.py:976] (4/7) Epoch 16, batch 2800, loss[loss=0.201, simple_loss=0.2634, pruned_loss=0.06926, over 4793.00 frames. ], tot_loss[loss=0.1761, simple_loss=0.2455, pruned_loss=0.05332, over 953999.42 frames. ], batch size: 26, lr: 3.44e-03, grad_scale: 32.0 2023-04-27 08:29:23,070 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1515, 1.9339, 2.0680, 2.5890, 2.5008, 2.0602, 1.6665, 2.2482], device='cuda:4'), covar=tensor([0.0772, 0.1018, 0.0669, 0.0487, 0.0554, 0.0824, 0.0807, 0.0504], device='cuda:4'), in_proj_covar=tensor([0.0192, 0.0204, 0.0185, 0.0176, 0.0179, 0.0184, 0.0156, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 08:29:33,974 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6009, 1.4641, 1.7958, 1.8486, 1.4815, 1.2873, 1.4764, 0.8650], device='cuda:4'), covar=tensor([0.0549, 0.0790, 0.0444, 0.0698, 0.0781, 0.1125, 0.0776, 0.0812], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0070, 0.0070, 0.0068, 0.0076, 0.0097, 0.0076, 0.0069], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 08:29:40,443 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6583, 1.3170, 1.7818, 2.1343, 1.8291, 1.6341, 1.7656, 1.6732], device='cuda:4'), covar=tensor([0.4317, 0.6654, 0.5910, 0.5431, 0.5671, 0.7517, 0.7247, 0.8282], device='cuda:4'), in_proj_covar=tensor([0.0420, 0.0410, 0.0497, 0.0509, 0.0449, 0.0475, 0.0481, 0.0483], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 08:29:42,736 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.076e+02 1.582e+02 1.928e+02 2.328e+02 4.353e+02, threshold=3.856e+02, percent-clipped=4.0 2023-04-27 08:29:47,979 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=88756.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:29:54,458 INFO [finetune.py:976] (4/7) Epoch 16, batch 2850, loss[loss=0.175, simple_loss=0.2453, pruned_loss=0.0524, over 4787.00 frames. ], tot_loss[loss=0.1766, simple_loss=0.2455, pruned_loss=0.05383, over 951865.78 frames. ], batch size: 29, lr: 3.44e-03, grad_scale: 32.0 2023-04-27 08:30:13,720 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4943, 1.7047, 1.8284, 1.9540, 1.7564, 1.8786, 1.9605, 1.8732], device='cuda:4'), covar=tensor([0.3832, 0.6024, 0.4967, 0.4752, 0.5620, 0.7594, 0.5312, 0.5039], device='cuda:4'), in_proj_covar=tensor([0.0330, 0.0371, 0.0316, 0.0332, 0.0342, 0.0394, 0.0351, 0.0323], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 08:30:20,171 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=88804.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:30:28,461 INFO [finetune.py:976] (4/7) Epoch 16, batch 2900, loss[loss=0.1629, simple_loss=0.2314, pruned_loss=0.04721, over 4787.00 frames. ], tot_loss[loss=0.1795, simple_loss=0.2484, pruned_loss=0.05528, over 954236.87 frames. ], batch size: 25, lr: 3.44e-03, grad_scale: 32.0 2023-04-27 08:30:43,085 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1592, 1.5375, 1.4092, 1.6641, 1.6116, 1.7442, 1.3753, 3.3202], device='cuda:4'), covar=tensor([0.0600, 0.0785, 0.0780, 0.1216, 0.0614, 0.0554, 0.0743, 0.0154], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0044, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 08:30:50,708 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.204e+02 1.633e+02 2.026e+02 2.312e+02 4.286e+02, threshold=4.053e+02, percent-clipped=1.0 2023-04-27 08:31:02,490 INFO [finetune.py:976] (4/7) Epoch 16, batch 2950, loss[loss=0.2019, simple_loss=0.273, pruned_loss=0.06541, over 4811.00 frames. ], tot_loss[loss=0.1817, simple_loss=0.2514, pruned_loss=0.05603, over 953725.14 frames. ], batch size: 40, lr: 3.44e-03, grad_scale: 32.0 2023-04-27 08:31:11,700 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6256, 2.7129, 2.1012, 2.4602, 2.8465, 2.2867, 3.6608, 1.8209], device='cuda:4'), covar=tensor([0.3885, 0.2147, 0.4666, 0.3286, 0.1726, 0.2698, 0.1154, 0.4760], device='cuda:4'), in_proj_covar=tensor([0.0344, 0.0350, 0.0429, 0.0359, 0.0386, 0.0385, 0.0374, 0.0422], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 08:31:58,218 INFO [finetune.py:976] (4/7) Epoch 16, batch 3000, loss[loss=0.207, simple_loss=0.2717, pruned_loss=0.0712, over 4801.00 frames. ], tot_loss[loss=0.1833, simple_loss=0.2535, pruned_loss=0.0566, over 955218.79 frames. ], batch size: 51, lr: 3.44e-03, grad_scale: 32.0 2023-04-27 08:31:58,218 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 08:32:14,557 INFO [finetune.py:1010] (4/7) Epoch 16, validation: loss=0.1523, simple_loss=0.2234, pruned_loss=0.04062, over 2265189.00 frames. 2023-04-27 08:32:14,557 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-27 08:32:46,710 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8156, 1.6188, 1.7693, 2.1740, 2.1603, 1.7494, 1.5264, 1.9604], device='cuda:4'), covar=tensor([0.0852, 0.1252, 0.0767, 0.0585, 0.0617, 0.0885, 0.0806, 0.0546], device='cuda:4'), in_proj_covar=tensor([0.0192, 0.0205, 0.0185, 0.0176, 0.0180, 0.0184, 0.0157, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 08:32:46,737 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7206, 2.3436, 1.6683, 1.6593, 1.2249, 1.2650, 1.7474, 1.1868], device='cuda:4'), covar=tensor([0.1618, 0.1222, 0.1492, 0.1710, 0.2451, 0.1979, 0.1048, 0.2123], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0213, 0.0168, 0.0205, 0.0200, 0.0183, 0.0156, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 08:32:47,286 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=88947.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 08:32:49,013 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.654e+02 1.995e+02 2.342e+02 3.743e+02, threshold=3.990e+02, percent-clipped=0.0 2023-04-27 08:32:59,787 INFO [finetune.py:976] (4/7) Epoch 16, batch 3050, loss[loss=0.1902, simple_loss=0.2637, pruned_loss=0.05832, over 4790.00 frames. ], tot_loss[loss=0.1826, simple_loss=0.2534, pruned_loss=0.0559, over 954496.15 frames. ], batch size: 45, lr: 3.44e-03, grad_scale: 32.0 2023-04-27 08:33:30,460 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=88995.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 08:33:55,682 INFO [finetune.py:976] (4/7) Epoch 16, batch 3100, loss[loss=0.1837, simple_loss=0.2606, pruned_loss=0.05343, over 4903.00 frames. ], tot_loss[loss=0.1799, simple_loss=0.2506, pruned_loss=0.05458, over 955410.71 frames. ], batch size: 36, lr: 3.44e-03, grad_scale: 32.0 2023-04-27 08:34:13,191 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=89027.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:34:38,477 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.056e+02 1.571e+02 1.837e+02 2.281e+02 4.804e+02, threshold=3.674e+02, percent-clipped=2.0 2023-04-27 08:34:59,021 INFO [finetune.py:976] (4/7) Epoch 16, batch 3150, loss[loss=0.1634, simple_loss=0.2314, pruned_loss=0.0477, over 4901.00 frames. ], tot_loss[loss=0.179, simple_loss=0.2486, pruned_loss=0.05465, over 957489.35 frames. ], batch size: 32, lr: 3.44e-03, grad_scale: 32.0 2023-04-27 08:35:14,701 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=89088.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:35:37,804 INFO [finetune.py:976] (4/7) Epoch 16, batch 3200, loss[loss=0.1764, simple_loss=0.2414, pruned_loss=0.05569, over 4821.00 frames. ], tot_loss[loss=0.1759, simple_loss=0.2453, pruned_loss=0.05325, over 956933.70 frames. ], batch size: 41, lr: 3.44e-03, grad_scale: 32.0 2023-04-27 08:36:09,670 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-04-27 08:36:10,705 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.6913, 4.5193, 3.2659, 5.3109, 4.5962, 4.6360, 1.7898, 4.6089], device='cuda:4'), covar=tensor([0.1610, 0.0869, 0.3144, 0.0925, 0.3976, 0.1635, 0.6280, 0.2118], device='cuda:4'), in_proj_covar=tensor([0.0241, 0.0211, 0.0251, 0.0301, 0.0294, 0.0246, 0.0269, 0.0269], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 08:36:23,960 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.134e+02 1.556e+02 1.903e+02 2.206e+02 4.255e+02, threshold=3.807e+02, percent-clipped=1.0 2023-04-27 08:36:42,704 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.29 vs. limit=5.0 2023-04-27 08:36:44,940 INFO [finetune.py:976] (4/7) Epoch 16, batch 3250, loss[loss=0.2419, simple_loss=0.3077, pruned_loss=0.08802, over 4824.00 frames. ], tot_loss[loss=0.1769, simple_loss=0.2461, pruned_loss=0.05383, over 954900.53 frames. ], batch size: 45, lr: 3.44e-03, grad_scale: 32.0 2023-04-27 08:36:46,915 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=89169.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:36:55,133 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-27 08:37:46,007 INFO [finetune.py:976] (4/7) Epoch 16, batch 3300, loss[loss=0.1775, simple_loss=0.2532, pruned_loss=0.05094, over 4774.00 frames. ], tot_loss[loss=0.179, simple_loss=0.2489, pruned_loss=0.0546, over 956036.18 frames. ], batch size: 54, lr: 3.44e-03, grad_scale: 32.0 2023-04-27 08:37:47,350 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0257, 2.2779, 1.4211, 1.6796, 2.4469, 1.8859, 1.8191, 1.9327], device='cuda:4'), covar=tensor([0.0494, 0.0338, 0.0279, 0.0540, 0.0206, 0.0507, 0.0500, 0.0549], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0024, 0.0022, 0.0029, 0.0020, 0.0028, 0.0028, 0.0029], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0044, 0.0038, 0.0051, 0.0038, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 08:38:07,665 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=89230.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:38:08,267 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1017, 2.5223, 0.8632, 1.4738, 1.5284, 1.9558, 1.5727, 0.7869], device='cuda:4'), covar=tensor([0.1569, 0.1113, 0.1776, 0.1317, 0.1138, 0.0941, 0.1550, 0.1746], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0245, 0.0138, 0.0121, 0.0131, 0.0153, 0.0118, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 08:38:16,971 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8175, 2.4040, 1.9005, 1.8764, 1.4059, 1.4059, 1.9571, 1.3562], device='cuda:4'), covar=tensor([0.1488, 0.1395, 0.1307, 0.1618, 0.2252, 0.1906, 0.0884, 0.1950], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0213, 0.0168, 0.0204, 0.0200, 0.0184, 0.0155, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 08:38:21,744 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.219e+02 1.680e+02 1.944e+02 2.414e+02 4.518e+02, threshold=3.887e+02, percent-clipped=2.0 2023-04-27 08:38:43,925 INFO [finetune.py:976] (4/7) Epoch 16, batch 3350, loss[loss=0.1857, simple_loss=0.2581, pruned_loss=0.05661, over 4779.00 frames. ], tot_loss[loss=0.1811, simple_loss=0.2513, pruned_loss=0.05545, over 956327.88 frames. ], batch size: 51, lr: 3.44e-03, grad_scale: 32.0 2023-04-27 08:38:50,226 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.57 vs. limit=2.0 2023-04-27 08:39:17,176 INFO [finetune.py:976] (4/7) Epoch 16, batch 3400, loss[loss=0.207, simple_loss=0.2763, pruned_loss=0.06887, over 4822.00 frames. ], tot_loss[loss=0.1815, simple_loss=0.2519, pruned_loss=0.0555, over 955143.89 frames. ], batch size: 33, lr: 3.44e-03, grad_scale: 32.0 2023-04-27 08:39:40,163 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.038e+02 1.554e+02 1.949e+02 2.299e+02 3.513e+02, threshold=3.898e+02, percent-clipped=0.0 2023-04-27 08:39:52,543 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4746, 1.7668, 1.8581, 2.0088, 1.8434, 1.9274, 1.9373, 1.8918], device='cuda:4'), covar=tensor([0.4245, 0.6038, 0.4789, 0.4790, 0.5798, 0.7773, 0.5918, 0.5591], device='cuda:4'), in_proj_covar=tensor([0.0333, 0.0374, 0.0318, 0.0334, 0.0345, 0.0396, 0.0354, 0.0326], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 08:39:53,662 INFO [finetune.py:976] (4/7) Epoch 16, batch 3450, loss[loss=0.1579, simple_loss=0.2274, pruned_loss=0.04416, over 4273.00 frames. ], tot_loss[loss=0.18, simple_loss=0.2505, pruned_loss=0.05478, over 954483.93 frames. ], batch size: 65, lr: 3.43e-03, grad_scale: 32.0 2023-04-27 08:40:04,298 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=89374.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:40:15,630 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0205, 2.4564, 0.8471, 1.3293, 1.4227, 1.8728, 1.5517, 0.7538], device='cuda:4'), covar=tensor([0.1848, 0.1413, 0.1997, 0.1604, 0.1322, 0.1118, 0.1795, 0.1953], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0246, 0.0138, 0.0121, 0.0131, 0.0154, 0.0119, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 08:40:16,190 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=89383.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:40:54,851 INFO [finetune.py:976] (4/7) Epoch 16, batch 3500, loss[loss=0.148, simple_loss=0.2237, pruned_loss=0.03622, over 4758.00 frames. ], tot_loss[loss=0.1787, simple_loss=0.2483, pruned_loss=0.05455, over 954967.22 frames. ], batch size: 27, lr: 3.43e-03, grad_scale: 32.0 2023-04-27 08:41:00,486 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=89425.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:41:07,533 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=89435.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:41:09,951 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=89439.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 08:41:18,078 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.096e+02 1.586e+02 1.877e+02 2.284e+02 3.930e+02, threshold=3.755e+02, percent-clipped=1.0 2023-04-27 08:41:21,838 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2030, 2.5889, 0.8445, 1.5455, 1.5765, 1.9352, 1.6556, 0.8291], device='cuda:4'), covar=tensor([0.1431, 0.1146, 0.1711, 0.1232, 0.1038, 0.0882, 0.1510, 0.1647], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0247, 0.0139, 0.0122, 0.0132, 0.0154, 0.0119, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 08:41:34,097 INFO [finetune.py:976] (4/7) Epoch 16, batch 3550, loss[loss=0.135, simple_loss=0.1985, pruned_loss=0.03568, over 4791.00 frames. ], tot_loss[loss=0.1771, simple_loss=0.2461, pruned_loss=0.05398, over 955374.68 frames. ], batch size: 26, lr: 3.43e-03, grad_scale: 32.0 2023-04-27 08:41:52,565 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-04-27 08:41:58,256 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=89486.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:41:58,647 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=2.01 vs. limit=2.0 2023-04-27 08:42:08,143 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=89500.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 08:42:20,189 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=89509.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:42:29,609 INFO [finetune.py:976] (4/7) Epoch 16, batch 3600, loss[loss=0.182, simple_loss=0.2489, pruned_loss=0.05752, over 4101.00 frames. ], tot_loss[loss=0.1755, simple_loss=0.2442, pruned_loss=0.05343, over 952880.46 frames. ], batch size: 65, lr: 3.43e-03, grad_scale: 32.0 2023-04-27 08:42:30,369 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5795, 1.4078, 0.7206, 1.2328, 1.4439, 1.4651, 1.2956, 1.3389], device='cuda:4'), covar=tensor([0.0479, 0.0369, 0.0381, 0.0556, 0.0288, 0.0518, 0.0508, 0.0544], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0024, 0.0022, 0.0029, 0.0020, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0044, 0.0037, 0.0051, 0.0038, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 08:42:41,044 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=89525.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:43:13,813 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.211e+02 1.668e+02 1.907e+02 2.305e+02 5.911e+02, threshold=3.815e+02, percent-clipped=3.0 2023-04-27 08:43:13,952 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3888, 1.4050, 1.6443, 1.6992, 1.3296, 1.1356, 1.4975, 1.0598], device='cuda:4'), covar=tensor([0.0665, 0.0609, 0.0467, 0.0566, 0.0718, 0.1137, 0.0565, 0.0623], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0071, 0.0070, 0.0069, 0.0077, 0.0098, 0.0076, 0.0069], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 08:43:36,162 INFO [finetune.py:976] (4/7) Epoch 16, batch 3650, loss[loss=0.1953, simple_loss=0.2724, pruned_loss=0.05913, over 4912.00 frames. ], tot_loss[loss=0.1781, simple_loss=0.2474, pruned_loss=0.0544, over 953421.37 frames. ], batch size: 43, lr: 3.43e-03, grad_scale: 32.0 2023-04-27 08:43:38,726 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=89570.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:43:45,521 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.06 vs. limit=5.0 2023-04-27 08:44:12,687 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9132, 1.6159, 2.0402, 2.4188, 2.0046, 1.7899, 1.9198, 1.9229], device='cuda:4'), covar=tensor([0.5159, 0.7629, 0.7711, 0.6235, 0.6363, 0.9915, 0.9315, 0.9805], device='cuda:4'), in_proj_covar=tensor([0.0419, 0.0408, 0.0495, 0.0507, 0.0449, 0.0473, 0.0478, 0.0481], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 08:44:23,115 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.12 vs. limit=2.0 2023-04-27 08:44:42,832 INFO [finetune.py:976] (4/7) Epoch 16, batch 3700, loss[loss=0.1889, simple_loss=0.26, pruned_loss=0.05893, over 4827.00 frames. ], tot_loss[loss=0.1797, simple_loss=0.2498, pruned_loss=0.05483, over 953671.97 frames. ], batch size: 47, lr: 3.43e-03, grad_scale: 32.0 2023-04-27 08:45:26,139 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.109e+02 1.724e+02 2.041e+02 2.428e+02 5.061e+02, threshold=4.082e+02, percent-clipped=2.0 2023-04-27 08:45:48,271 INFO [finetune.py:976] (4/7) Epoch 16, batch 3750, loss[loss=0.1458, simple_loss=0.2157, pruned_loss=0.03792, over 4698.00 frames. ], tot_loss[loss=0.1805, simple_loss=0.251, pruned_loss=0.05502, over 954095.27 frames. ], batch size: 23, lr: 3.43e-03, grad_scale: 32.0 2023-04-27 08:46:02,543 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=89680.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:46:10,266 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=89683.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:46:24,604 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0846, 1.5585, 1.9745, 2.1375, 1.9163, 1.5287, 1.0552, 1.6453], device='cuda:4'), covar=tensor([0.3496, 0.3511, 0.1712, 0.2437, 0.2822, 0.2814, 0.4337, 0.2156], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0247, 0.0224, 0.0315, 0.0217, 0.0230, 0.0227, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 08:46:49,772 INFO [finetune.py:976] (4/7) Epoch 16, batch 3800, loss[loss=0.1847, simple_loss=0.2628, pruned_loss=0.0533, over 4715.00 frames. ], tot_loss[loss=0.1819, simple_loss=0.2527, pruned_loss=0.05555, over 954525.97 frames. ], batch size: 59, lr: 3.43e-03, grad_scale: 32.0 2023-04-27 08:47:09,440 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=89730.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:47:10,029 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=89731.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:47:21,766 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=89741.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:47:30,411 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0536, 1.8456, 2.2938, 2.5418, 2.1079, 1.9994, 2.1216, 2.1107], device='cuda:4'), covar=tensor([0.5220, 0.7649, 0.7887, 0.6235, 0.6906, 0.9238, 0.9859, 0.9825], device='cuda:4'), in_proj_covar=tensor([0.0419, 0.0408, 0.0493, 0.0507, 0.0447, 0.0473, 0.0477, 0.0480], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 08:47:31,967 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.072e+02 1.563e+02 1.915e+02 2.320e+02 3.976e+02, threshold=3.830e+02, percent-clipped=0.0 2023-04-27 08:47:43,647 INFO [finetune.py:976] (4/7) Epoch 16, batch 3850, loss[loss=0.159, simple_loss=0.2197, pruned_loss=0.04911, over 4762.00 frames. ], tot_loss[loss=0.1807, simple_loss=0.2509, pruned_loss=0.05523, over 953215.13 frames. ], batch size: 27, lr: 3.43e-03, grad_scale: 32.0 2023-04-27 08:47:58,307 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=89781.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:48:12,383 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=89795.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 08:48:17,985 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8214, 2.1337, 1.7473, 1.3450, 1.3186, 1.3410, 1.8187, 1.2925], device='cuda:4'), covar=tensor([0.1835, 0.1383, 0.1499, 0.1971, 0.2450, 0.2098, 0.1032, 0.2152], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0212, 0.0167, 0.0204, 0.0199, 0.0183, 0.0155, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 08:48:26,725 INFO [finetune.py:976] (4/7) Epoch 16, batch 3900, loss[loss=0.2091, simple_loss=0.2681, pruned_loss=0.07507, over 4845.00 frames. ], tot_loss[loss=0.1791, simple_loss=0.2489, pruned_loss=0.0547, over 952674.51 frames. ], batch size: 47, lr: 3.43e-03, grad_scale: 32.0 2023-04-27 08:48:33,336 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=89825.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:48:33,934 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2465, 1.5551, 1.3783, 1.8243, 1.7295, 2.0085, 1.4575, 3.7345], device='cuda:4'), covar=tensor([0.0592, 0.0795, 0.0821, 0.1203, 0.0593, 0.0435, 0.0723, 0.0134], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0044, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 08:48:48,865 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.776e+01 1.757e+02 2.140e+02 2.503e+02 6.967e+02, threshold=4.281e+02, percent-clipped=3.0 2023-04-27 08:48:59,106 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=89865.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:48:59,136 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=89865.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:48:59,623 INFO [finetune.py:976] (4/7) Epoch 16, batch 3950, loss[loss=0.1589, simple_loss=0.2261, pruned_loss=0.04587, over 4914.00 frames. ], tot_loss[loss=0.1758, simple_loss=0.245, pruned_loss=0.05329, over 951715.65 frames. ], batch size: 36, lr: 3.43e-03, grad_scale: 32.0 2023-04-27 08:49:05,392 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=89873.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:49:23,984 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.4674, 4.4562, 2.9995, 5.1762, 4.5249, 4.4817, 1.8408, 4.3443], device='cuda:4'), covar=tensor([0.1535, 0.0801, 0.3693, 0.0857, 0.2872, 0.1796, 0.5927, 0.2257], device='cuda:4'), in_proj_covar=tensor([0.0247, 0.0215, 0.0254, 0.0306, 0.0301, 0.0251, 0.0275, 0.0275], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 08:49:27,775 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8281, 1.3833, 1.8909, 2.3152, 1.9425, 1.8095, 1.8958, 1.8704], device='cuda:4'), covar=tensor([0.4961, 0.6835, 0.6665, 0.5943, 0.6020, 0.8343, 0.8329, 0.8258], device='cuda:4'), in_proj_covar=tensor([0.0418, 0.0406, 0.0492, 0.0506, 0.0446, 0.0471, 0.0476, 0.0480], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 08:49:31,256 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7595, 1.9305, 1.1381, 1.5129, 2.1237, 1.6458, 1.5503, 1.7076], device='cuda:4'), covar=tensor([0.0494, 0.0365, 0.0302, 0.0559, 0.0260, 0.0514, 0.0501, 0.0554], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0024, 0.0022, 0.0029, 0.0020, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0045, 0.0038, 0.0051, 0.0038, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 08:49:33,479 INFO [finetune.py:976] (4/7) Epoch 16, batch 4000, loss[loss=0.1753, simple_loss=0.2428, pruned_loss=0.05394, over 4904.00 frames. ], tot_loss[loss=0.1756, simple_loss=0.2445, pruned_loss=0.05335, over 953105.65 frames. ], batch size: 36, lr: 3.43e-03, grad_scale: 32.0 2023-04-27 08:49:38,290 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.7801, 2.0714, 1.9983, 2.1722, 1.9096, 1.9925, 2.0777, 2.0574], device='cuda:4'), covar=tensor([0.4106, 0.6588, 0.5449, 0.4583, 0.5926, 0.7559, 0.6433, 0.6015], device='cuda:4'), in_proj_covar=tensor([0.0331, 0.0373, 0.0317, 0.0333, 0.0344, 0.0395, 0.0353, 0.0325], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 08:49:47,303 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=89926.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 08:50:10,126 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5728, 1.6120, 0.8258, 1.3054, 1.6333, 1.4670, 1.3286, 1.4132], device='cuda:4'), covar=tensor([0.0463, 0.0353, 0.0360, 0.0522, 0.0290, 0.0476, 0.0458, 0.0531], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0024, 0.0022, 0.0029, 0.0020, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0044, 0.0037, 0.0051, 0.0038, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 08:50:11,836 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.112e+02 1.722e+02 2.012e+02 2.405e+02 4.800e+02, threshold=4.024e+02, percent-clipped=2.0 2023-04-27 08:50:21,844 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.77 vs. limit=5.0 2023-04-27 08:50:33,226 INFO [finetune.py:976] (4/7) Epoch 16, batch 4050, loss[loss=0.1869, simple_loss=0.2492, pruned_loss=0.0623, over 4769.00 frames. ], tot_loss[loss=0.179, simple_loss=0.2482, pruned_loss=0.05488, over 950895.27 frames. ], batch size: 26, lr: 3.43e-03, grad_scale: 32.0 2023-04-27 08:50:34,471 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9046, 1.2347, 1.6581, 2.0232, 1.6719, 1.3432, 0.9752, 1.3746], device='cuda:4'), covar=tensor([0.3737, 0.4265, 0.2122, 0.2807, 0.3046, 0.3040, 0.5005, 0.2542], device='cuda:4'), in_proj_covar=tensor([0.0291, 0.0248, 0.0226, 0.0317, 0.0218, 0.0231, 0.0229, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 08:51:40,301 INFO [finetune.py:976] (4/7) Epoch 16, batch 4100, loss[loss=0.2003, simple_loss=0.2378, pruned_loss=0.0814, over 4360.00 frames. ], tot_loss[loss=0.1816, simple_loss=0.2514, pruned_loss=0.05593, over 950615.30 frames. ], batch size: 18, lr: 3.43e-03, grad_scale: 32.0 2023-04-27 08:52:02,658 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=90030.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:52:12,518 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=90036.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:52:27,090 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.030e+02 1.746e+02 2.004e+02 2.346e+02 5.282e+02, threshold=4.007e+02, percent-clipped=2.0 2023-04-27 08:52:48,658 INFO [finetune.py:976] (4/7) Epoch 16, batch 4150, loss[loss=0.2278, simple_loss=0.2894, pruned_loss=0.08305, over 4812.00 frames. ], tot_loss[loss=0.182, simple_loss=0.252, pruned_loss=0.05598, over 950478.61 frames. ], batch size: 47, lr: 3.43e-03, grad_scale: 32.0 2023-04-27 08:53:06,919 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=90078.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:53:09,298 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=90081.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:53:26,048 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=90095.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 08:53:39,082 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2081, 1.2397, 3.8002, 3.5604, 3.3650, 3.6308, 3.6175, 3.3118], device='cuda:4'), covar=tensor([0.7670, 0.5559, 0.1183, 0.1745, 0.1361, 0.1939, 0.1555, 0.1614], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0301, 0.0399, 0.0399, 0.0345, 0.0405, 0.0307, 0.0361], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 08:53:50,594 INFO [finetune.py:976] (4/7) Epoch 16, batch 4200, loss[loss=0.1567, simple_loss=0.2248, pruned_loss=0.04435, over 4837.00 frames. ], tot_loss[loss=0.1827, simple_loss=0.2528, pruned_loss=0.05629, over 950122.47 frames. ], batch size: 47, lr: 3.43e-03, grad_scale: 32.0 2023-04-27 08:53:59,900 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=90129.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:54:01,076 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=90130.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:54:14,874 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=90143.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 08:54:24,333 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.851e+01 1.635e+02 1.941e+02 2.339e+02 3.883e+02, threshold=3.882e+02, percent-clipped=0.0 2023-04-27 08:54:34,089 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=90165.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:54:34,611 INFO [finetune.py:976] (4/7) Epoch 16, batch 4250, loss[loss=0.1554, simple_loss=0.2235, pruned_loss=0.0436, over 4867.00 frames. ], tot_loss[loss=0.1791, simple_loss=0.2491, pruned_loss=0.05452, over 952406.79 frames. ], batch size: 34, lr: 3.43e-03, grad_scale: 32.0 2023-04-27 08:54:38,888 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7110, 1.7170, 0.7975, 1.4449, 1.8156, 1.5925, 1.4619, 1.5473], device='cuda:4'), covar=tensor([0.0489, 0.0357, 0.0350, 0.0530, 0.0271, 0.0494, 0.0473, 0.0531], device='cuda:4'), in_proj_covar=tensor([0.0029, 0.0024, 0.0022, 0.0029, 0.0020, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0045, 0.0038, 0.0051, 0.0038, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 08:54:52,884 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=90191.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:55:06,210 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=90213.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:55:08,444 INFO [finetune.py:976] (4/7) Epoch 16, batch 4300, loss[loss=0.2222, simple_loss=0.2799, pruned_loss=0.0822, over 4743.00 frames. ], tot_loss[loss=0.1779, simple_loss=0.2475, pruned_loss=0.0541, over 952946.12 frames. ], batch size: 27, lr: 3.43e-03, grad_scale: 32.0 2023-04-27 08:55:11,531 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=90221.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 08:55:19,285 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=90232.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:55:28,016 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1369, 2.2917, 2.1992, 1.8867, 2.2935, 1.9253, 2.9133, 1.8391], device='cuda:4'), covar=tensor([0.3240, 0.1418, 0.3376, 0.2474, 0.1409, 0.2016, 0.1139, 0.3613], device='cuda:4'), in_proj_covar=tensor([0.0347, 0.0350, 0.0434, 0.0359, 0.0388, 0.0386, 0.0375, 0.0424], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 08:55:32,137 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.154e+02 1.571e+02 1.935e+02 2.216e+02 3.879e+02, threshold=3.870e+02, percent-clipped=0.0 2023-04-27 08:55:41,845 INFO [finetune.py:976] (4/7) Epoch 16, batch 4350, loss[loss=0.2057, simple_loss=0.2773, pruned_loss=0.06711, over 4902.00 frames. ], tot_loss[loss=0.1764, simple_loss=0.2454, pruned_loss=0.0537, over 951411.08 frames. ], batch size: 37, lr: 3.43e-03, grad_scale: 32.0 2023-04-27 08:55:57,632 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.8651, 3.7946, 2.7933, 4.5476, 3.9362, 3.9303, 1.7594, 3.8669], device='cuda:4'), covar=tensor([0.1873, 0.1250, 0.3358, 0.1457, 0.3800, 0.1811, 0.6050, 0.2604], device='cuda:4'), in_proj_covar=tensor([0.0246, 0.0213, 0.0252, 0.0303, 0.0299, 0.0250, 0.0273, 0.0274], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 08:56:00,600 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=90293.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:56:15,530 INFO [finetune.py:976] (4/7) Epoch 16, batch 4400, loss[loss=0.1794, simple_loss=0.2519, pruned_loss=0.05348, over 4840.00 frames. ], tot_loss[loss=0.1765, simple_loss=0.2461, pruned_loss=0.05348, over 951713.39 frames. ], batch size: 30, lr: 3.43e-03, grad_scale: 32.0 2023-04-27 08:56:34,938 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=90336.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:56:55,341 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.166e+02 1.655e+02 1.875e+02 2.353e+02 3.798e+02, threshold=3.751e+02, percent-clipped=0.0 2023-04-27 08:56:56,143 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=90351.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:57:10,216 INFO [finetune.py:976] (4/7) Epoch 16, batch 4450, loss[loss=0.2801, simple_loss=0.3327, pruned_loss=0.1137, over 4805.00 frames. ], tot_loss[loss=0.1789, simple_loss=0.2493, pruned_loss=0.0542, over 952199.45 frames. ], batch size: 51, lr: 3.43e-03, grad_scale: 32.0 2023-04-27 08:57:32,200 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-27 08:57:33,655 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=90384.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:57:58,252 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-27 08:58:04,789 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=90412.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:58:06,670 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1279, 1.6358, 2.0794, 2.5743, 2.0307, 1.5626, 1.4137, 1.8706], device='cuda:4'), covar=tensor([0.3319, 0.3177, 0.1641, 0.2191, 0.2571, 0.2652, 0.3891, 0.2082], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0249, 0.0227, 0.0318, 0.0219, 0.0232, 0.0230, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 08:58:07,141 INFO [finetune.py:976] (4/7) Epoch 16, batch 4500, loss[loss=0.1921, simple_loss=0.2639, pruned_loss=0.06012, over 4844.00 frames. ], tot_loss[loss=0.1791, simple_loss=0.2497, pruned_loss=0.0542, over 952407.28 frames. ], batch size: 49, lr: 3.43e-03, grad_scale: 64.0 2023-04-27 08:58:32,110 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0004, 2.4876, 1.0538, 1.4099, 1.9106, 1.2501, 3.2979, 1.6523], device='cuda:4'), covar=tensor([0.0723, 0.0634, 0.0823, 0.1287, 0.0512, 0.1047, 0.0226, 0.0643], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0048, 0.0046, 0.0050, 0.0052, 0.0075, 0.0052], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0007, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 08:58:38,162 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5316, 1.9434, 1.3796, 1.2956, 1.1232, 1.1944, 1.4139, 1.0190], device='cuda:4'), covar=tensor([0.1968, 0.1467, 0.1813, 0.2024, 0.2721, 0.2474, 0.1182, 0.2415], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0213, 0.0169, 0.0205, 0.0201, 0.0184, 0.0156, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 08:58:52,081 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.119e+02 1.589e+02 1.862e+02 2.271e+02 3.876e+02, threshold=3.723e+02, percent-clipped=1.0 2023-04-27 08:59:04,462 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=90459.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:59:14,176 INFO [finetune.py:976] (4/7) Epoch 16, batch 4550, loss[loss=0.1762, simple_loss=0.2465, pruned_loss=0.05288, over 4841.00 frames. ], tot_loss[loss=0.1792, simple_loss=0.2507, pruned_loss=0.05388, over 954042.64 frames. ], batch size: 30, lr: 3.43e-03, grad_scale: 64.0 2023-04-27 08:59:43,240 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=90486.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 08:59:55,956 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=90496.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:00:09,154 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2047, 1.6408, 1.4260, 2.0414, 2.2670, 1.9076, 1.8602, 1.5836], device='cuda:4'), covar=tensor([0.1950, 0.1807, 0.1780, 0.1776, 0.1261, 0.2026, 0.2047, 0.2427], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0313, 0.0353, 0.0292, 0.0329, 0.0313, 0.0303, 0.0367], device='cuda:4'), out_proj_covar=tensor([6.3796e-05, 6.5165e-05, 7.5312e-05, 5.9493e-05, 6.8517e-05, 6.5895e-05, 6.4040e-05, 7.8396e-05], device='cuda:4') 2023-04-27 09:00:15,725 INFO [finetune.py:976] (4/7) Epoch 16, batch 4600, loss[loss=0.1819, simple_loss=0.2498, pruned_loss=0.05701, over 4826.00 frames. ], tot_loss[loss=0.1787, simple_loss=0.2501, pruned_loss=0.05361, over 953417.65 frames. ], batch size: 30, lr: 3.43e-03, grad_scale: 64.0 2023-04-27 09:00:18,338 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=90520.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:00:18,919 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=90521.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 09:00:37,736 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.078e+02 1.675e+02 1.952e+02 2.329e+02 6.121e+02, threshold=3.905e+02, percent-clipped=3.0 2023-04-27 09:00:43,137 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=90557.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:00:49,343 INFO [finetune.py:976] (4/7) Epoch 16, batch 4650, loss[loss=0.1839, simple_loss=0.2396, pruned_loss=0.06407, over 4853.00 frames. ], tot_loss[loss=0.1783, simple_loss=0.2485, pruned_loss=0.05402, over 954523.33 frames. ], batch size: 47, lr: 3.42e-03, grad_scale: 64.0 2023-04-27 09:00:51,267 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=90569.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:00:56,118 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8389, 2.3818, 2.0940, 1.8196, 1.3190, 1.3896, 2.2848, 1.4087], device='cuda:4'), covar=tensor([0.1811, 0.1474, 0.1332, 0.1861, 0.2412, 0.2051, 0.0837, 0.2078], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0213, 0.0169, 0.0205, 0.0201, 0.0184, 0.0156, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 09:01:03,237 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=90588.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:01:13,198 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=5.00 vs. limit=5.0 2023-04-27 09:01:23,251 INFO [finetune.py:976] (4/7) Epoch 16, batch 4700, loss[loss=0.1378, simple_loss=0.2183, pruned_loss=0.02866, over 4757.00 frames. ], tot_loss[loss=0.1758, simple_loss=0.2458, pruned_loss=0.05293, over 956210.96 frames. ], batch size: 28, lr: 3.42e-03, grad_scale: 64.0 2023-04-27 09:01:29,505 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-27 09:01:45,414 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.002e+02 1.583e+02 1.845e+02 2.338e+02 5.124e+02, threshold=3.690e+02, percent-clipped=1.0 2023-04-27 09:02:01,941 INFO [finetune.py:976] (4/7) Epoch 16, batch 4750, loss[loss=0.2082, simple_loss=0.2716, pruned_loss=0.07245, over 4832.00 frames. ], tot_loss[loss=0.1754, simple_loss=0.2449, pruned_loss=0.05302, over 956973.08 frames. ], batch size: 33, lr: 3.42e-03, grad_scale: 32.0 2023-04-27 09:02:03,396 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-27 09:02:57,063 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=90707.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:03:07,869 INFO [finetune.py:976] (4/7) Epoch 16, batch 4800, loss[loss=0.2598, simple_loss=0.3288, pruned_loss=0.09537, over 4777.00 frames. ], tot_loss[loss=0.1773, simple_loss=0.2471, pruned_loss=0.05376, over 954873.95 frames. ], batch size: 59, lr: 3.42e-03, grad_scale: 32.0 2023-04-27 09:03:36,200 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.106e+02 1.708e+02 2.005e+02 2.486e+02 4.014e+02, threshold=4.009e+02, percent-clipped=3.0 2023-04-27 09:03:38,051 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7377, 2.2028, 1.6882, 1.4106, 1.2895, 1.3131, 1.7235, 1.1845], device='cuda:4'), covar=tensor([0.1884, 0.1391, 0.1668, 0.2026, 0.2699, 0.2270, 0.1131, 0.2322], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0213, 0.0168, 0.0205, 0.0201, 0.0184, 0.0156, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 09:03:47,213 INFO [finetune.py:976] (4/7) Epoch 16, batch 4850, loss[loss=0.2137, simple_loss=0.2769, pruned_loss=0.07521, over 4737.00 frames. ], tot_loss[loss=0.1804, simple_loss=0.2511, pruned_loss=0.05483, over 956134.68 frames. ], batch size: 59, lr: 3.42e-03, grad_scale: 32.0 2023-04-27 09:04:00,377 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=90786.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:04:36,234 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=90815.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:04:36,791 INFO [finetune.py:976] (4/7) Epoch 16, batch 4900, loss[loss=0.1574, simple_loss=0.2083, pruned_loss=0.05324, over 4317.00 frames. ], tot_loss[loss=0.181, simple_loss=0.2516, pruned_loss=0.05523, over 953210.48 frames. ], batch size: 18, lr: 3.42e-03, grad_scale: 32.0 2023-04-27 09:05:01,143 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=90834.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:05:21,482 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=90849.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:05:22,614 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.194e+02 1.643e+02 1.978e+02 2.332e+02 3.633e+02, threshold=3.955e+02, percent-clipped=0.0 2023-04-27 09:05:23,821 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=90852.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:05:38,744 INFO [finetune.py:976] (4/7) Epoch 16, batch 4950, loss[loss=0.1979, simple_loss=0.2635, pruned_loss=0.06618, over 4920.00 frames. ], tot_loss[loss=0.1831, simple_loss=0.2541, pruned_loss=0.05607, over 955239.37 frames. ], batch size: 33, lr: 3.42e-03, grad_scale: 32.0 2023-04-27 09:05:51,847 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3292, 1.6299, 1.4034, 1.5478, 1.4383, 1.3355, 1.3801, 1.0980], device='cuda:4'), covar=tensor([0.1748, 0.1252, 0.0976, 0.1210, 0.3523, 0.1188, 0.1674, 0.2408], device='cuda:4'), in_proj_covar=tensor([0.0286, 0.0308, 0.0222, 0.0280, 0.0312, 0.0261, 0.0250, 0.0265], device='cuda:4'), out_proj_covar=tensor([1.1491e-04, 1.2252e-04, 8.8201e-05, 1.1122e-04, 1.2711e-04, 1.0387e-04, 1.0111e-04, 1.0554e-04], device='cuda:4') 2023-04-27 09:05:54,256 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=90888.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:06:08,379 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=90910.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:06:12,378 INFO [finetune.py:976] (4/7) Epoch 16, batch 5000, loss[loss=0.1829, simple_loss=0.2505, pruned_loss=0.05766, over 4816.00 frames. ], tot_loss[loss=0.1813, simple_loss=0.2519, pruned_loss=0.05541, over 952791.23 frames. ], batch size: 38, lr: 3.42e-03, grad_scale: 32.0 2023-04-27 09:06:27,004 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=90936.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:06:36,102 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.088e+02 1.563e+02 1.794e+02 2.191e+02 3.401e+02, threshold=3.587e+02, percent-clipped=0.0 2023-04-27 09:06:46,205 INFO [finetune.py:976] (4/7) Epoch 16, batch 5050, loss[loss=0.1516, simple_loss=0.2204, pruned_loss=0.04137, over 4740.00 frames. ], tot_loss[loss=0.1786, simple_loss=0.2485, pruned_loss=0.05439, over 953169.42 frames. ], batch size: 59, lr: 3.42e-03, grad_scale: 32.0 2023-04-27 09:07:04,421 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-27 09:07:07,638 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9214, 2.5266, 2.0527, 2.3026, 1.7484, 2.0767, 2.1098, 1.5026], device='cuda:4'), covar=tensor([0.2139, 0.1022, 0.0893, 0.1308, 0.3143, 0.1213, 0.1860, 0.2809], device='cuda:4'), in_proj_covar=tensor([0.0283, 0.0305, 0.0220, 0.0278, 0.0310, 0.0259, 0.0249, 0.0264], device='cuda:4'), out_proj_covar=tensor([1.1388e-04, 1.2145e-04, 8.7414e-05, 1.1030e-04, 1.2602e-04, 1.0295e-04, 1.0053e-04, 1.0488e-04], device='cuda:4') 2023-04-27 09:07:13,894 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=91007.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:07:19,797 INFO [finetune.py:976] (4/7) Epoch 16, batch 5100, loss[loss=0.1816, simple_loss=0.2442, pruned_loss=0.05948, over 4822.00 frames. ], tot_loss[loss=0.176, simple_loss=0.2453, pruned_loss=0.05334, over 952278.48 frames. ], batch size: 39, lr: 3.42e-03, grad_scale: 32.0 2023-04-27 09:07:24,068 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8303, 1.1162, 1.6782, 2.2312, 1.8765, 1.7153, 1.7298, 1.7296], device='cuda:4'), covar=tensor([0.4621, 0.6467, 0.6453, 0.5989, 0.5723, 0.7628, 0.7713, 0.8121], device='cuda:4'), in_proj_covar=tensor([0.0420, 0.0405, 0.0494, 0.0506, 0.0447, 0.0472, 0.0478, 0.0483], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 09:07:31,395 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.66 vs. limit=5.0 2023-04-27 09:07:38,557 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7141, 1.2853, 1.3587, 1.4673, 1.9250, 1.5343, 1.3498, 1.2978], device='cuda:4'), covar=tensor([0.1303, 0.1386, 0.1878, 0.1256, 0.0799, 0.1529, 0.1915, 0.2129], device='cuda:4'), in_proj_covar=tensor([0.0303, 0.0308, 0.0347, 0.0286, 0.0323, 0.0309, 0.0299, 0.0360], device='cuda:4'), out_proj_covar=tensor([6.2562e-05, 6.4186e-05, 7.4121e-05, 5.8124e-05, 6.7335e-05, 6.5072e-05, 6.3123e-05, 7.6988e-05], device='cuda:4') 2023-04-27 09:07:43,900 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.024e+02 1.518e+02 1.785e+02 2.206e+02 4.900e+02, threshold=3.570e+02, percent-clipped=2.0 2023-04-27 09:07:46,425 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=91055.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:07:53,571 INFO [finetune.py:976] (4/7) Epoch 16, batch 5150, loss[loss=0.2138, simple_loss=0.2796, pruned_loss=0.07396, over 4838.00 frames. ], tot_loss[loss=0.1766, simple_loss=0.2457, pruned_loss=0.05374, over 953231.37 frames. ], batch size: 47, lr: 3.42e-03, grad_scale: 32.0 2023-04-27 09:08:25,149 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-04-27 09:08:48,472 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=91115.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:08:48,976 INFO [finetune.py:976] (4/7) Epoch 16, batch 5200, loss[loss=0.1681, simple_loss=0.2362, pruned_loss=0.04996, over 4875.00 frames. ], tot_loss[loss=0.1796, simple_loss=0.2492, pruned_loss=0.05501, over 952651.64 frames. ], batch size: 31, lr: 3.42e-03, grad_scale: 32.0 2023-04-27 09:09:38,602 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.048e+02 1.654e+02 2.065e+02 2.493e+02 5.806e+02, threshold=4.130e+02, percent-clipped=4.0 2023-04-27 09:09:39,330 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=91152.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:09:51,505 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=91163.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:09:53,296 INFO [finetune.py:976] (4/7) Epoch 16, batch 5250, loss[loss=0.1966, simple_loss=0.2773, pruned_loss=0.05796, over 4719.00 frames. ], tot_loss[loss=0.1814, simple_loss=0.2516, pruned_loss=0.05563, over 951987.94 frames. ], batch size: 59, lr: 3.42e-03, grad_scale: 32.0 2023-04-27 09:10:44,541 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=91200.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:10:47,138 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-04-27 09:10:47,648 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=91205.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:11:00,426 INFO [finetune.py:976] (4/7) Epoch 16, batch 5300, loss[loss=0.1582, simple_loss=0.2354, pruned_loss=0.04051, over 4784.00 frames. ], tot_loss[loss=0.1825, simple_loss=0.253, pruned_loss=0.05603, over 952469.79 frames. ], batch size: 29, lr: 3.42e-03, grad_scale: 32.0 2023-04-27 09:11:25,291 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.127e+02 1.661e+02 1.899e+02 2.252e+02 3.532e+02, threshold=3.799e+02, percent-clipped=0.0 2023-04-27 09:11:25,435 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6880, 1.5909, 2.0157, 2.0430, 1.5844, 1.3048, 1.7643, 1.1915], device='cuda:4'), covar=tensor([0.0745, 0.0617, 0.0496, 0.0726, 0.0780, 0.1107, 0.0703, 0.0728], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0070, 0.0069, 0.0068, 0.0075, 0.0096, 0.0075, 0.0067], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 09:11:29,891 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.88 vs. limit=5.0 2023-04-27 09:11:34,435 INFO [finetune.py:976] (4/7) Epoch 16, batch 5350, loss[loss=0.1938, simple_loss=0.2553, pruned_loss=0.06616, over 4761.00 frames. ], tot_loss[loss=0.1822, simple_loss=0.2532, pruned_loss=0.05559, over 954300.89 frames. ], batch size: 27, lr: 3.42e-03, grad_scale: 32.0 2023-04-27 09:11:39,476 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1591, 2.8049, 2.1606, 2.1821, 1.5632, 1.5326, 2.1837, 1.4246], device='cuda:4'), covar=tensor([0.1680, 0.1410, 0.1453, 0.1682, 0.2383, 0.1957, 0.1000, 0.2013], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0212, 0.0168, 0.0204, 0.0201, 0.0184, 0.0156, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 09:11:47,131 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=91285.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:12:08,306 INFO [finetune.py:976] (4/7) Epoch 16, batch 5400, loss[loss=0.1741, simple_loss=0.2408, pruned_loss=0.0537, over 4184.00 frames. ], tot_loss[loss=0.1808, simple_loss=0.2513, pruned_loss=0.05513, over 955033.30 frames. ], batch size: 65, lr: 3.42e-03, grad_scale: 32.0 2023-04-27 09:12:19,397 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0102, 0.6957, 0.8286, 0.6773, 1.1174, 0.9193, 0.8376, 0.8247], device='cuda:4'), covar=tensor([0.1204, 0.1281, 0.1530, 0.1435, 0.0759, 0.1229, 0.1221, 0.1719], device='cuda:4'), in_proj_covar=tensor([0.0305, 0.0310, 0.0349, 0.0286, 0.0325, 0.0310, 0.0300, 0.0362], device='cuda:4'), out_proj_covar=tensor([6.3005e-05, 6.4635e-05, 7.4464e-05, 5.8181e-05, 6.7542e-05, 6.5408e-05, 6.3306e-05, 7.7441e-05], device='cuda:4') 2023-04-27 09:12:20,675 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-27 09:12:28,679 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=91346.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:12:32,096 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.075e+02 1.605e+02 2.003e+02 2.320e+02 6.253e+02, threshold=4.007e+02, percent-clipped=1.0 2023-04-27 09:12:42,146 INFO [finetune.py:976] (4/7) Epoch 16, batch 5450, loss[loss=0.1489, simple_loss=0.2207, pruned_loss=0.0385, over 4739.00 frames. ], tot_loss[loss=0.1776, simple_loss=0.2474, pruned_loss=0.05392, over 952847.19 frames. ], batch size: 23, lr: 3.42e-03, grad_scale: 32.0 2023-04-27 09:12:42,816 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0295, 2.3645, 2.3851, 2.7842, 2.5158, 2.6714, 2.1947, 4.9447], device='cuda:4'), covar=tensor([0.0495, 0.0688, 0.0674, 0.0959, 0.0555, 0.0432, 0.0644, 0.0081], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0040, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 09:13:15,978 INFO [finetune.py:976] (4/7) Epoch 16, batch 5500, loss[loss=0.2219, simple_loss=0.2862, pruned_loss=0.07879, over 4736.00 frames. ], tot_loss[loss=0.1756, simple_loss=0.2448, pruned_loss=0.05321, over 952310.79 frames. ], batch size: 54, lr: 3.42e-03, grad_scale: 32.0 2023-04-27 09:13:38,282 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.525e+01 1.617e+02 1.882e+02 2.275e+02 4.330e+02, threshold=3.765e+02, percent-clipped=1.0 2023-04-27 09:13:49,965 INFO [finetune.py:976] (4/7) Epoch 16, batch 5550, loss[loss=0.1877, simple_loss=0.2556, pruned_loss=0.05988, over 4874.00 frames. ], tot_loss[loss=0.1777, simple_loss=0.2469, pruned_loss=0.05429, over 953230.66 frames. ], batch size: 34, lr: 3.42e-03, grad_scale: 32.0 2023-04-27 09:13:56,231 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3450, 1.8291, 2.2488, 2.6936, 2.2232, 1.7481, 1.4986, 2.0195], device='cuda:4'), covar=tensor([0.3652, 0.3323, 0.1804, 0.2530, 0.2993, 0.2915, 0.4255, 0.2122], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0247, 0.0228, 0.0316, 0.0218, 0.0232, 0.0229, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 09:14:12,398 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9473, 2.3990, 2.0257, 1.7752, 1.3964, 1.4782, 2.1928, 1.3908], device='cuda:4'), covar=tensor([0.1586, 0.1388, 0.1345, 0.1680, 0.2278, 0.1913, 0.0855, 0.1984], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0213, 0.0169, 0.0205, 0.0201, 0.0185, 0.0156, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 09:14:20,556 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=91505.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:14:24,095 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8284, 1.5532, 1.8214, 2.1875, 2.1251, 1.7714, 1.4724, 1.8933], device='cuda:4'), covar=tensor([0.0833, 0.1238, 0.0701, 0.0564, 0.0627, 0.0853, 0.0839, 0.0616], device='cuda:4'), in_proj_covar=tensor([0.0188, 0.0200, 0.0180, 0.0172, 0.0176, 0.0181, 0.0152, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 09:14:32,825 INFO [finetune.py:976] (4/7) Epoch 16, batch 5600, loss[loss=0.2475, simple_loss=0.3076, pruned_loss=0.09373, over 4862.00 frames. ], tot_loss[loss=0.1796, simple_loss=0.2499, pruned_loss=0.05461, over 955660.77 frames. ], batch size: 31, lr: 3.42e-03, grad_scale: 32.0 2023-04-27 09:14:44,132 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.52 vs. limit=5.0 2023-04-27 09:15:04,839 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=91542.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 09:15:15,398 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.016e+02 1.634e+02 1.984e+02 2.410e+02 6.572e+02, threshold=3.968e+02, percent-clipped=2.0 2023-04-27 09:15:16,633 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=91553.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:15:29,907 INFO [finetune.py:976] (4/7) Epoch 16, batch 5650, loss[loss=0.1715, simple_loss=0.2419, pruned_loss=0.0505, over 4824.00 frames. ], tot_loss[loss=0.1811, simple_loss=0.2523, pruned_loss=0.05499, over 956889.60 frames. ], batch size: 33, lr: 3.42e-03, grad_scale: 32.0 2023-04-27 09:16:20,698 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=91603.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 09:16:34,228 INFO [finetune.py:976] (4/7) Epoch 16, batch 5700, loss[loss=0.1541, simple_loss=0.2192, pruned_loss=0.04452, over 4063.00 frames. ], tot_loss[loss=0.179, simple_loss=0.2488, pruned_loss=0.05461, over 938106.96 frames. ], batch size: 17, lr: 3.42e-03, grad_scale: 32.0 2023-04-27 09:17:06,020 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=91641.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:17:19,302 INFO [finetune.py:976] (4/7) Epoch 17, batch 0, loss[loss=0.1783, simple_loss=0.2483, pruned_loss=0.05418, over 4905.00 frames. ], tot_loss[loss=0.1783, simple_loss=0.2483, pruned_loss=0.05418, over 4905.00 frames. ], batch size: 38, lr: 3.41e-03, grad_scale: 32.0 2023-04-27 09:17:19,302 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 09:17:22,555 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7851, 1.0071, 1.6530, 2.2492, 1.8782, 1.6985, 1.6861, 1.7217], device='cuda:4'), covar=tensor([0.5153, 0.7555, 0.6858, 0.6763, 0.6667, 0.8812, 0.9164, 0.8966], device='cuda:4'), in_proj_covar=tensor([0.0420, 0.0408, 0.0496, 0.0506, 0.0449, 0.0471, 0.0479, 0.0485], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 09:17:22,622 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4515, 1.4223, 3.8931, 3.6023, 3.5055, 3.6856, 3.8008, 3.5022], device='cuda:4'), covar=tensor([0.6857, 0.5197, 0.1262, 0.2137, 0.1282, 0.1623, 0.0805, 0.1643], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0303, 0.0400, 0.0400, 0.0345, 0.0408, 0.0307, 0.0361], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 09:17:23,142 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4607, 1.2955, 1.7059, 1.6526, 1.3175, 1.2549, 1.3436, 0.9047], device='cuda:4'), covar=tensor([0.0484, 0.0671, 0.0400, 0.0533, 0.0738, 0.1099, 0.0502, 0.0557], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0070, 0.0069, 0.0068, 0.0075, 0.0097, 0.0075, 0.0067], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 09:17:40,814 INFO [finetune.py:1010] (4/7) Epoch 17, validation: loss=0.1535, simple_loss=0.2247, pruned_loss=0.04111, over 2265189.00 frames. 2023-04-27 09:17:40,814 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-27 09:17:45,698 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.062e+02 1.575e+02 1.838e+02 2.202e+02 3.811e+02, threshold=3.676e+02, percent-clipped=0.0 2023-04-27 09:18:20,570 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.81 vs. limit=2.0 2023-04-27 09:18:24,711 INFO [finetune.py:976] (4/7) Epoch 17, batch 50, loss[loss=0.196, simple_loss=0.2755, pruned_loss=0.05825, over 4901.00 frames. ], tot_loss[loss=0.1765, simple_loss=0.2488, pruned_loss=0.05206, over 216416.67 frames. ], batch size: 46, lr: 3.41e-03, grad_scale: 32.0 2023-04-27 09:18:32,665 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1147, 2.2388, 1.9568, 1.8735, 2.2943, 1.7274, 2.8468, 1.7118], device='cuda:4'), covar=tensor([0.4172, 0.1862, 0.5072, 0.3079, 0.1967, 0.2864, 0.1335, 0.4807], device='cuda:4'), in_proj_covar=tensor([0.0343, 0.0348, 0.0432, 0.0359, 0.0385, 0.0383, 0.0374, 0.0424], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 09:18:39,438 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.53 vs. limit=2.0 2023-04-27 09:18:45,336 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5678, 3.6067, 1.2197, 1.7770, 2.0133, 2.4865, 2.1154, 1.2490], device='cuda:4'), covar=tensor([0.1872, 0.1532, 0.2298, 0.1819, 0.1334, 0.1347, 0.1685, 0.2009], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0243, 0.0137, 0.0121, 0.0132, 0.0153, 0.0117, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 09:18:57,080 INFO [finetune.py:976] (4/7) Epoch 17, batch 100, loss[loss=0.1643, simple_loss=0.2446, pruned_loss=0.04204, over 4793.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2427, pruned_loss=0.05067, over 379694.86 frames. ], batch size: 51, lr: 3.41e-03, grad_scale: 32.0 2023-04-27 09:19:02,435 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.149e+02 1.653e+02 1.974e+02 2.303e+02 4.426e+02, threshold=3.948e+02, percent-clipped=2.0 2023-04-27 09:19:04,169 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-27 09:19:30,053 INFO [finetune.py:976] (4/7) Epoch 17, batch 150, loss[loss=0.1751, simple_loss=0.2406, pruned_loss=0.05474, over 4821.00 frames. ], tot_loss[loss=0.1717, simple_loss=0.2412, pruned_loss=0.05116, over 506699.63 frames. ], batch size: 41, lr: 3.41e-03, grad_scale: 32.0 2023-04-27 09:20:03,534 INFO [finetune.py:976] (4/7) Epoch 17, batch 200, loss[loss=0.1848, simple_loss=0.2483, pruned_loss=0.06064, over 4827.00 frames. ], tot_loss[loss=0.1741, simple_loss=0.2423, pruned_loss=0.05298, over 605447.99 frames. ], batch size: 30, lr: 3.41e-03, grad_scale: 32.0 2023-04-27 09:20:04,229 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1618, 1.3603, 1.2443, 1.7064, 1.5187, 1.7224, 1.2974, 3.0847], device='cuda:4'), covar=tensor([0.0662, 0.0864, 0.0934, 0.1353, 0.0714, 0.0528, 0.0814, 0.0159], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 09:20:08,956 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.596e+01 1.677e+02 1.899e+02 2.270e+02 6.599e+02, threshold=3.797e+02, percent-clipped=2.0 2023-04-27 09:20:35,888 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=91891.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:20:36,990 INFO [finetune.py:976] (4/7) Epoch 17, batch 250, loss[loss=0.2156, simple_loss=0.2873, pruned_loss=0.07196, over 4813.00 frames. ], tot_loss[loss=0.1765, simple_loss=0.2458, pruned_loss=0.05358, over 684299.75 frames. ], batch size: 45, lr: 3.41e-03, grad_scale: 32.0 2023-04-27 09:20:40,162 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=91898.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 09:20:41,413 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7003, 1.6195, 2.0374, 2.1133, 1.5507, 1.4620, 1.6882, 1.0327], device='cuda:4'), covar=tensor([0.0506, 0.0719, 0.0413, 0.0579, 0.0748, 0.1185, 0.0681, 0.0710], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0069, 0.0068, 0.0067, 0.0075, 0.0096, 0.0074, 0.0067], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 09:21:03,341 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9540, 1.6090, 2.1627, 2.3735, 2.0531, 1.9462, 2.0728, 1.9847], device='cuda:4'), covar=tensor([0.4687, 0.6532, 0.6756, 0.5842, 0.6004, 0.7917, 0.8106, 0.8400], device='cuda:4'), in_proj_covar=tensor([0.0419, 0.0407, 0.0495, 0.0506, 0.0448, 0.0471, 0.0478, 0.0484], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 09:21:08,451 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=91941.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:21:08,493 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5401, 2.5759, 2.2115, 2.3165, 2.5476, 2.1423, 3.5594, 1.9655], device='cuda:4'), covar=tensor([0.3908, 0.2109, 0.4434, 0.3253, 0.2133, 0.2841, 0.1491, 0.4155], device='cuda:4'), in_proj_covar=tensor([0.0341, 0.0347, 0.0431, 0.0357, 0.0384, 0.0382, 0.0373, 0.0421], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 09:21:10,055 INFO [finetune.py:976] (4/7) Epoch 17, batch 300, loss[loss=0.1746, simple_loss=0.2508, pruned_loss=0.04919, over 4822.00 frames. ], tot_loss[loss=0.1765, simple_loss=0.2472, pruned_loss=0.05288, over 744624.26 frames. ], batch size: 33, lr: 3.41e-03, grad_scale: 32.0 2023-04-27 09:21:15,845 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.098e+02 1.651e+02 1.943e+02 2.275e+02 4.588e+02, threshold=3.886e+02, percent-clipped=2.0 2023-04-27 09:21:16,627 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=91952.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:21:18,456 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=91955.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:21:51,285 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=91981.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 09:21:56,126 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=91989.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:22:04,088 INFO [finetune.py:976] (4/7) Epoch 17, batch 350, loss[loss=0.1685, simple_loss=0.2383, pruned_loss=0.0494, over 4801.00 frames. ], tot_loss[loss=0.1775, simple_loss=0.2483, pruned_loss=0.05339, over 787811.52 frames. ], batch size: 39, lr: 3.41e-03, grad_scale: 32.0 2023-04-27 09:22:13,864 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2028, 1.8349, 2.1236, 2.5070, 2.4771, 2.0411, 1.7764, 2.2551], device='cuda:4'), covar=tensor([0.0748, 0.1016, 0.0645, 0.0482, 0.0515, 0.0778, 0.0704, 0.0511], device='cuda:4'), in_proj_covar=tensor([0.0190, 0.0201, 0.0182, 0.0172, 0.0176, 0.0181, 0.0152, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 09:22:36,923 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=92016.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:23:09,180 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=92042.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 09:23:09,647 INFO [finetune.py:976] (4/7) Epoch 17, batch 400, loss[loss=0.1969, simple_loss=0.2582, pruned_loss=0.06781, over 4859.00 frames. ], tot_loss[loss=0.179, simple_loss=0.2499, pruned_loss=0.05402, over 825981.50 frames. ], batch size: 34, lr: 3.41e-03, grad_scale: 32.0 2023-04-27 09:23:20,962 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.21 vs. limit=5.0 2023-04-27 09:23:21,385 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.614e+02 1.906e+02 2.282e+02 3.471e+02, threshold=3.811e+02, percent-clipped=0.0 2023-04-27 09:23:31,553 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-27 09:23:40,187 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4373, 3.5454, 1.0489, 1.7581, 2.0055, 2.4176, 1.9685, 0.9975], device='cuda:4'), covar=tensor([0.1699, 0.1526, 0.2387, 0.1681, 0.1258, 0.1376, 0.1770, 0.2207], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0243, 0.0138, 0.0121, 0.0132, 0.0153, 0.0118, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 09:24:03,174 INFO [finetune.py:976] (4/7) Epoch 17, batch 450, loss[loss=0.1697, simple_loss=0.2264, pruned_loss=0.05649, over 4826.00 frames. ], tot_loss[loss=0.1781, simple_loss=0.2484, pruned_loss=0.05383, over 854767.74 frames. ], batch size: 30, lr: 3.41e-03, grad_scale: 32.0 2023-04-27 09:24:28,165 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.79 vs. limit=2.0 2023-04-27 09:24:32,090 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6461, 1.3901, 1.7843, 1.8695, 1.4825, 1.1605, 1.4056, 0.8661], device='cuda:4'), covar=tensor([0.0497, 0.0676, 0.0461, 0.0533, 0.0712, 0.1587, 0.0683, 0.0767], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0069, 0.0068, 0.0067, 0.0075, 0.0096, 0.0074, 0.0067], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 09:24:36,847 INFO [finetune.py:976] (4/7) Epoch 17, batch 500, loss[loss=0.1745, simple_loss=0.2439, pruned_loss=0.0525, over 4818.00 frames. ], tot_loss[loss=0.1764, simple_loss=0.2464, pruned_loss=0.0532, over 878504.11 frames. ], batch size: 39, lr: 3.41e-03, grad_scale: 32.0 2023-04-27 09:24:42,152 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.066e+02 1.524e+02 1.904e+02 2.252e+02 3.809e+02, threshold=3.808e+02, percent-clipped=0.0 2023-04-27 09:24:51,470 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.80 vs. limit=5.0 2023-04-27 09:25:06,770 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8835, 2.4296, 1.8066, 1.8017, 1.3517, 1.4386, 1.8638, 1.2704], device='cuda:4'), covar=tensor([0.1799, 0.1326, 0.1578, 0.1839, 0.2576, 0.2145, 0.1112, 0.2231], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0214, 0.0169, 0.0205, 0.0202, 0.0185, 0.0156, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 09:25:09,769 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0745, 1.2675, 1.2101, 1.5390, 1.4244, 1.4115, 1.2460, 2.4453], device='cuda:4'), covar=tensor([0.0621, 0.0868, 0.0838, 0.1313, 0.0680, 0.0547, 0.0791, 0.0220], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0039, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0012, 0.0013, 0.0016], device='cuda:4') 2023-04-27 09:25:10,282 INFO [finetune.py:976] (4/7) Epoch 17, batch 550, loss[loss=0.2185, simple_loss=0.2851, pruned_loss=0.07596, over 4871.00 frames. ], tot_loss[loss=0.175, simple_loss=0.2444, pruned_loss=0.05276, over 893983.08 frames. ], batch size: 34, lr: 3.41e-03, grad_scale: 32.0 2023-04-27 09:25:13,461 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=92198.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 09:25:21,882 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.8309, 3.6474, 2.7621, 4.4340, 3.7782, 3.8472, 1.7408, 3.7946], device='cuda:4'), covar=tensor([0.1572, 0.1270, 0.3766, 0.1377, 0.3607, 0.1789, 0.5775, 0.2321], device='cuda:4'), in_proj_covar=tensor([0.0246, 0.0214, 0.0252, 0.0306, 0.0298, 0.0250, 0.0274, 0.0272], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 09:25:26,806 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2511, 1.5029, 1.6619, 1.7777, 1.7208, 1.8166, 1.7579, 1.7424], device='cuda:4'), covar=tensor([0.3983, 0.4856, 0.4304, 0.4440, 0.5246, 0.7106, 0.4898, 0.4574], device='cuda:4'), in_proj_covar=tensor([0.0333, 0.0373, 0.0318, 0.0333, 0.0346, 0.0397, 0.0355, 0.0325], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 09:25:44,129 INFO [finetune.py:976] (4/7) Epoch 17, batch 600, loss[loss=0.1889, simple_loss=0.26, pruned_loss=0.05888, over 4829.00 frames. ], tot_loss[loss=0.1759, simple_loss=0.2453, pruned_loss=0.05323, over 907046.73 frames. ], batch size: 33, lr: 3.41e-03, grad_scale: 32.0 2023-04-27 09:25:46,054 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=92246.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 09:25:46,682 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=92247.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:25:49,028 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.221e+02 1.690e+02 1.979e+02 2.477e+02 6.011e+02, threshold=3.959e+02, percent-clipped=1.0 2023-04-27 09:26:00,317 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-27 09:26:17,387 INFO [finetune.py:976] (4/7) Epoch 17, batch 650, loss[loss=0.192, simple_loss=0.2778, pruned_loss=0.0531, over 4835.00 frames. ], tot_loss[loss=0.178, simple_loss=0.2482, pruned_loss=0.05385, over 917926.33 frames. ], batch size: 47, lr: 3.41e-03, grad_scale: 32.0 2023-04-27 09:26:18,132 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6250, 1.7089, 1.5830, 1.2198, 1.2626, 1.2297, 1.5738, 1.1895], device='cuda:4'), covar=tensor([0.1812, 0.1493, 0.1459, 0.1646, 0.2403, 0.2015, 0.1087, 0.2041], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0215, 0.0170, 0.0206, 0.0203, 0.0186, 0.0157, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 09:26:29,964 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=92311.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:26:56,492 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=92337.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 09:27:00,060 INFO [finetune.py:976] (4/7) Epoch 17, batch 700, loss[loss=0.1578, simple_loss=0.2381, pruned_loss=0.03875, over 4742.00 frames. ], tot_loss[loss=0.18, simple_loss=0.2507, pruned_loss=0.05462, over 924907.38 frames. ], batch size: 54, lr: 3.41e-03, grad_scale: 32.0 2023-04-27 09:27:10,674 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.019e+02 1.654e+02 1.859e+02 2.105e+02 3.715e+02, threshold=3.718e+02, percent-clipped=1.0 2023-04-27 09:27:49,358 INFO [finetune.py:976] (4/7) Epoch 17, batch 750, loss[loss=0.1982, simple_loss=0.2588, pruned_loss=0.06877, over 4896.00 frames. ], tot_loss[loss=0.1806, simple_loss=0.252, pruned_loss=0.05463, over 930761.88 frames. ], batch size: 36, lr: 3.41e-03, grad_scale: 32.0 2023-04-27 09:28:44,566 INFO [finetune.py:976] (4/7) Epoch 17, batch 800, loss[loss=0.1759, simple_loss=0.2392, pruned_loss=0.05634, over 4920.00 frames. ], tot_loss[loss=0.1799, simple_loss=0.2512, pruned_loss=0.05433, over 935751.76 frames. ], batch size: 33, lr: 3.41e-03, grad_scale: 32.0 2023-04-27 09:28:54,789 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.154e+02 1.529e+02 1.784e+02 2.077e+02 4.006e+02, threshold=3.568e+02, percent-clipped=1.0 2023-04-27 09:29:04,670 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3001, 1.2613, 1.6129, 1.5254, 1.2657, 1.1315, 1.3014, 0.8989], device='cuda:4'), covar=tensor([0.0573, 0.0540, 0.0358, 0.0551, 0.0706, 0.1022, 0.0513, 0.0540], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0069, 0.0068, 0.0067, 0.0075, 0.0096, 0.0074, 0.0067], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 09:29:31,911 INFO [finetune.py:976] (4/7) Epoch 17, batch 850, loss[loss=0.1816, simple_loss=0.2529, pruned_loss=0.05508, over 4855.00 frames. ], tot_loss[loss=0.1784, simple_loss=0.2493, pruned_loss=0.05376, over 938499.99 frames. ], batch size: 49, lr: 3.41e-03, grad_scale: 32.0 2023-04-27 09:29:52,947 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5436, 1.4353, 4.2878, 4.0513, 3.7361, 4.0537, 3.8646, 3.7632], device='cuda:4'), covar=tensor([0.7437, 0.5874, 0.1002, 0.1535, 0.1069, 0.1572, 0.2037, 0.1529], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0303, 0.0398, 0.0402, 0.0344, 0.0406, 0.0306, 0.0361], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 09:30:05,439 INFO [finetune.py:976] (4/7) Epoch 17, batch 900, loss[loss=0.2152, simple_loss=0.2625, pruned_loss=0.08394, over 4725.00 frames. ], tot_loss[loss=0.1771, simple_loss=0.2473, pruned_loss=0.05347, over 943636.11 frames. ], batch size: 59, lr: 3.41e-03, grad_scale: 32.0 2023-04-27 09:30:07,972 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=92547.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:30:10,299 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.562e+01 1.556e+02 1.849e+02 2.275e+02 6.056e+02, threshold=3.698e+02, percent-clipped=3.0 2023-04-27 09:30:11,073 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.8626, 2.1565, 2.0259, 2.2412, 2.0720, 2.1763, 2.1128, 2.0936], device='cuda:4'), covar=tensor([0.4140, 0.5778, 0.5668, 0.4351, 0.5811, 0.7164, 0.6320, 0.5844], device='cuda:4'), in_proj_covar=tensor([0.0330, 0.0371, 0.0316, 0.0332, 0.0343, 0.0394, 0.0353, 0.0323], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 09:30:19,045 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7286, 1.5594, 1.9873, 2.0277, 1.5442, 1.4116, 1.6918, 0.9550], device='cuda:4'), covar=tensor([0.0532, 0.0691, 0.0363, 0.0626, 0.0865, 0.1255, 0.0646, 0.0734], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0070, 0.0069, 0.0068, 0.0075, 0.0096, 0.0075, 0.0068], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 09:30:36,119 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2668, 3.1928, 1.1726, 1.6977, 1.8161, 2.3648, 1.8245, 0.9923], device='cuda:4'), covar=tensor([0.1855, 0.1561, 0.2092, 0.1633, 0.1409, 0.1259, 0.1784, 0.2237], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0244, 0.0138, 0.0121, 0.0133, 0.0153, 0.0118, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 09:30:38,498 INFO [finetune.py:976] (4/7) Epoch 17, batch 950, loss[loss=0.2068, simple_loss=0.2707, pruned_loss=0.07146, over 4841.00 frames. ], tot_loss[loss=0.1763, simple_loss=0.2457, pruned_loss=0.0534, over 947458.56 frames. ], batch size: 44, lr: 3.40e-03, grad_scale: 32.0 2023-04-27 09:30:39,829 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=92595.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:30:49,703 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=92611.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:30:59,802 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=92625.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:31:07,668 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=92637.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 09:31:12,294 INFO [finetune.py:976] (4/7) Epoch 17, batch 1000, loss[loss=0.1534, simple_loss=0.2386, pruned_loss=0.03417, over 4926.00 frames. ], tot_loss[loss=0.1772, simple_loss=0.2471, pruned_loss=0.05366, over 950026.40 frames. ], batch size: 38, lr: 3.40e-03, grad_scale: 32.0 2023-04-27 09:31:17,228 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.750e+01 1.761e+02 2.088e+02 2.403e+02 5.880e+02, threshold=4.175e+02, percent-clipped=3.0 2023-04-27 09:31:22,263 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=92659.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:31:40,696 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=92685.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 09:31:41,333 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=92686.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:31:45,967 INFO [finetune.py:976] (4/7) Epoch 17, batch 1050, loss[loss=0.1887, simple_loss=0.2679, pruned_loss=0.05475, over 4895.00 frames. ], tot_loss[loss=0.1784, simple_loss=0.2495, pruned_loss=0.05361, over 952343.86 frames. ], batch size: 35, lr: 3.40e-03, grad_scale: 64.0 2023-04-27 09:31:48,380 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6202, 1.4968, 1.9962, 2.0024, 1.4807, 1.3517, 1.6353, 0.9814], device='cuda:4'), covar=tensor([0.0520, 0.0811, 0.0371, 0.0529, 0.0825, 0.1034, 0.0664, 0.0690], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0069, 0.0068, 0.0067, 0.0074, 0.0095, 0.0074, 0.0067], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 09:32:03,094 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=92720.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:32:30,324 INFO [finetune.py:976] (4/7) Epoch 17, batch 1100, loss[loss=0.1669, simple_loss=0.24, pruned_loss=0.04696, over 4877.00 frames. ], tot_loss[loss=0.1799, simple_loss=0.2513, pruned_loss=0.05429, over 954729.54 frames. ], batch size: 32, lr: 3.40e-03, grad_scale: 64.0 2023-04-27 09:32:36,187 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.169e+02 1.706e+02 1.976e+02 2.311e+02 4.775e+02, threshold=3.952e+02, percent-clipped=2.0 2023-04-27 09:33:11,522 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=92781.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:33:32,086 INFO [finetune.py:976] (4/7) Epoch 17, batch 1150, loss[loss=0.1737, simple_loss=0.2427, pruned_loss=0.05233, over 4831.00 frames. ], tot_loss[loss=0.1808, simple_loss=0.2524, pruned_loss=0.05462, over 956182.37 frames. ], batch size: 47, lr: 3.40e-03, grad_scale: 32.0 2023-04-27 09:33:32,493 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-04-27 09:34:20,065 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=92831.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:34:39,529 INFO [finetune.py:976] (4/7) Epoch 17, batch 1200, loss[loss=0.1509, simple_loss=0.2313, pruned_loss=0.03525, over 4808.00 frames. ], tot_loss[loss=0.1787, simple_loss=0.2502, pruned_loss=0.05362, over 957111.82 frames. ], batch size: 40, lr: 3.40e-03, grad_scale: 32.0 2023-04-27 09:34:50,260 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.173e+01 1.694e+02 1.853e+02 2.185e+02 5.052e+02, threshold=3.707e+02, percent-clipped=2.0 2023-04-27 09:35:44,789 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=92892.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:35:45,252 INFO [finetune.py:976] (4/7) Epoch 17, batch 1250, loss[loss=0.1517, simple_loss=0.2285, pruned_loss=0.03741, over 4916.00 frames. ], tot_loss[loss=0.1785, simple_loss=0.249, pruned_loss=0.05399, over 957605.80 frames. ], batch size: 37, lr: 3.40e-03, grad_scale: 32.0 2023-04-27 09:35:48,972 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=92899.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:36:06,411 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1788, 1.4742, 1.3161, 1.7345, 1.5440, 1.6339, 1.2893, 3.0585], device='cuda:4'), covar=tensor([0.0609, 0.0767, 0.0786, 0.1140, 0.0649, 0.0528, 0.0777, 0.0162], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0039, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 09:36:25,919 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=92931.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:36:45,900 INFO [finetune.py:976] (4/7) Epoch 17, batch 1300, loss[loss=0.1528, simple_loss=0.218, pruned_loss=0.04376, over 4911.00 frames. ], tot_loss[loss=0.1765, simple_loss=0.2463, pruned_loss=0.05338, over 958839.82 frames. ], batch size: 36, lr: 3.40e-03, grad_scale: 32.0 2023-04-27 09:36:56,601 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-04-27 09:36:57,404 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.209e+02 1.510e+02 1.807e+02 2.085e+02 3.413e+02, threshold=3.613e+02, percent-clipped=0.0 2023-04-27 09:37:08,815 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=92960.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:37:29,935 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=92975.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:37:32,389 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5199, 2.8684, 0.8513, 1.5499, 2.1301, 1.5650, 4.0569, 1.9828], device='cuda:4'), covar=tensor([0.0622, 0.0973, 0.0954, 0.1242, 0.0551, 0.0961, 0.0231, 0.0601], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0065, 0.0048, 0.0046, 0.0049, 0.0052, 0.0075, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0007, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 09:37:33,592 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=92981.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:37:33,645 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4393, 3.2415, 2.6428, 2.8280, 2.2549, 2.8889, 2.8255, 2.2048], device='cuda:4'), covar=tensor([0.2128, 0.1139, 0.0839, 0.1307, 0.3183, 0.1202, 0.1790, 0.2797], device='cuda:4'), in_proj_covar=tensor([0.0284, 0.0305, 0.0220, 0.0277, 0.0312, 0.0258, 0.0249, 0.0266], device='cuda:4'), out_proj_covar=tensor([1.1421e-04, 1.2137e-04, 8.7365e-05, 1.1002e-04, 1.2696e-04, 1.0281e-04, 1.0069e-04, 1.0563e-04], device='cuda:4') 2023-04-27 09:37:40,824 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=92992.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:37:41,319 INFO [finetune.py:976] (4/7) Epoch 17, batch 1350, loss[loss=0.1838, simple_loss=0.266, pruned_loss=0.05085, over 4836.00 frames. ], tot_loss[loss=0.1766, simple_loss=0.2462, pruned_loss=0.05349, over 956483.26 frames. ], batch size: 49, lr: 3.40e-03, grad_scale: 32.0 2023-04-27 09:37:49,814 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8847, 1.4403, 2.0059, 2.3748, 1.9769, 1.7973, 1.8742, 1.8511], device='cuda:4'), covar=tensor([0.5159, 0.7558, 0.6590, 0.6269, 0.6424, 0.8377, 0.8557, 0.9182], device='cuda:4'), in_proj_covar=tensor([0.0420, 0.0407, 0.0495, 0.0505, 0.0450, 0.0472, 0.0478, 0.0484], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 09:37:54,393 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=93010.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:38:16,148 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=93036.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:38:16,157 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=93036.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:38:26,353 INFO [finetune.py:976] (4/7) Epoch 17, batch 1400, loss[loss=0.164, simple_loss=0.2424, pruned_loss=0.04279, over 4819.00 frames. ], tot_loss[loss=0.1776, simple_loss=0.2479, pruned_loss=0.05368, over 956816.99 frames. ], batch size: 38, lr: 3.40e-03, grad_scale: 32.0 2023-04-27 09:38:44,540 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.119e+02 1.602e+02 1.954e+02 2.219e+02 5.640e+02, threshold=3.909e+02, percent-clipped=3.0 2023-04-27 09:39:08,831 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=93071.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:39:12,769 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=93076.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:39:28,194 INFO [finetune.py:976] (4/7) Epoch 17, batch 1450, loss[loss=0.1639, simple_loss=0.2356, pruned_loss=0.04609, over 4764.00 frames. ], tot_loss[loss=0.1786, simple_loss=0.2497, pruned_loss=0.05372, over 957688.06 frames. ], batch size: 28, lr: 3.40e-03, grad_scale: 32.0 2023-04-27 09:39:37,790 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=93097.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 09:40:01,746 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=93115.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:40:24,592 INFO [finetune.py:976] (4/7) Epoch 17, batch 1500, loss[loss=0.2117, simple_loss=0.2815, pruned_loss=0.07091, over 4815.00 frames. ], tot_loss[loss=0.1814, simple_loss=0.2523, pruned_loss=0.05526, over 954438.73 frames. ], batch size: 40, lr: 3.40e-03, grad_scale: 32.0 2023-04-27 09:40:31,495 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.190e+02 1.667e+02 1.959e+02 2.406e+02 7.498e+02, threshold=3.919e+02, percent-clipped=4.0 2023-04-27 09:40:54,001 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=93176.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:41:06,573 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=93187.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:41:15,562 INFO [finetune.py:976] (4/7) Epoch 17, batch 1550, loss[loss=0.1969, simple_loss=0.2656, pruned_loss=0.06413, over 4920.00 frames. ], tot_loss[loss=0.1798, simple_loss=0.2507, pruned_loss=0.05443, over 953801.22 frames. ], batch size: 38, lr: 3.40e-03, grad_scale: 32.0 2023-04-27 09:41:38,290 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6944, 1.5046, 1.3277, 1.4729, 1.9226, 1.5725, 1.2780, 1.2397], device='cuda:4'), covar=tensor([0.1383, 0.1064, 0.1653, 0.1277, 0.0633, 0.1328, 0.1829, 0.2035], device='cuda:4'), in_proj_covar=tensor([0.0305, 0.0311, 0.0351, 0.0289, 0.0327, 0.0311, 0.0301, 0.0366], device='cuda:4'), out_proj_covar=tensor([6.3031e-05, 6.4916e-05, 7.4822e-05, 5.8855e-05, 6.8030e-05, 6.5533e-05, 6.3558e-05, 7.8091e-05], device='cuda:4') 2023-04-27 09:41:49,255 INFO [finetune.py:976] (4/7) Epoch 17, batch 1600, loss[loss=0.1517, simple_loss=0.2271, pruned_loss=0.0382, over 4813.00 frames. ], tot_loss[loss=0.1785, simple_loss=0.2491, pruned_loss=0.05398, over 955125.84 frames. ], batch size: 40, lr: 3.40e-03, grad_scale: 32.0 2023-04-27 09:41:54,717 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.149e+02 1.590e+02 1.909e+02 2.207e+02 5.345e+02, threshold=3.818e+02, percent-clipped=1.0 2023-04-27 09:41:57,602 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=93255.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:42:06,779 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-04-27 09:42:15,843 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=93281.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:42:19,486 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=93287.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:42:23,065 INFO [finetune.py:976] (4/7) Epoch 17, batch 1650, loss[loss=0.1686, simple_loss=0.2418, pruned_loss=0.04765, over 4776.00 frames. ], tot_loss[loss=0.1763, simple_loss=0.2462, pruned_loss=0.05321, over 956158.07 frames. ], batch size: 27, lr: 3.40e-03, grad_scale: 32.0 2023-04-27 09:42:29,357 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5009, 1.5233, 1.5446, 2.2698, 2.2339, 1.8962, 1.8892, 1.6388], device='cuda:4'), covar=tensor([0.1953, 0.2428, 0.2520, 0.1940, 0.1681, 0.2559, 0.2716, 0.3112], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0312, 0.0352, 0.0290, 0.0328, 0.0311, 0.0301, 0.0366], device='cuda:4'), out_proj_covar=tensor([6.3348e-05, 6.5089e-05, 7.5041e-05, 5.9047e-05, 6.8241e-05, 6.5644e-05, 6.3595e-05, 7.8277e-05], device='cuda:4') 2023-04-27 09:42:32,308 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9834, 1.5600, 2.1306, 2.4184, 2.0476, 1.9638, 2.0341, 2.0162], device='cuda:4'), covar=tensor([0.4691, 0.6784, 0.6873, 0.5947, 0.6396, 0.8414, 0.8239, 0.7641], device='cuda:4'), in_proj_covar=tensor([0.0419, 0.0406, 0.0495, 0.0504, 0.0448, 0.0472, 0.0478, 0.0483], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 09:43:03,620 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=93329.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:43:04,870 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=93331.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:43:17,621 INFO [finetune.py:976] (4/7) Epoch 17, batch 1700, loss[loss=0.1757, simple_loss=0.2206, pruned_loss=0.06534, over 4193.00 frames. ], tot_loss[loss=0.1751, simple_loss=0.2442, pruned_loss=0.05303, over 955134.18 frames. ], batch size: 17, lr: 3.40e-03, grad_scale: 32.0 2023-04-27 09:43:22,700 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.47 vs. limit=5.0 2023-04-27 09:43:23,104 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.005e+02 1.559e+02 1.851e+02 2.220e+02 4.230e+02, threshold=3.703e+02, percent-clipped=1.0 2023-04-27 09:43:33,689 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=93366.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:43:41,238 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=93376.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:43:51,042 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=93392.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 09:43:51,575 INFO [finetune.py:976] (4/7) Epoch 17, batch 1750, loss[loss=0.186, simple_loss=0.2539, pruned_loss=0.05903, over 4764.00 frames. ], tot_loss[loss=0.1775, simple_loss=0.2467, pruned_loss=0.05411, over 955426.36 frames. ], batch size: 26, lr: 3.40e-03, grad_scale: 32.0 2023-04-27 09:43:52,950 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0630, 1.6740, 1.9078, 2.2931, 2.2289, 1.8120, 1.6372, 2.0299], device='cuda:4'), covar=tensor([0.0846, 0.1204, 0.0794, 0.0612, 0.0709, 0.0984, 0.0843, 0.0652], device='cuda:4'), in_proj_covar=tensor([0.0190, 0.0201, 0.0183, 0.0173, 0.0178, 0.0182, 0.0153, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 09:44:13,254 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=93424.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:44:16,140 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5345, 2.4697, 2.0989, 2.2681, 2.6033, 2.2758, 3.5628, 1.9682], device='cuda:4'), covar=tensor([0.3834, 0.2362, 0.4437, 0.3480, 0.1974, 0.2678, 0.1148, 0.4217], device='cuda:4'), in_proj_covar=tensor([0.0343, 0.0347, 0.0432, 0.0358, 0.0384, 0.0384, 0.0371, 0.0423], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 09:44:36,168 INFO [finetune.py:976] (4/7) Epoch 17, batch 1800, loss[loss=0.1995, simple_loss=0.2619, pruned_loss=0.06854, over 4881.00 frames. ], tot_loss[loss=0.1806, simple_loss=0.2508, pruned_loss=0.05518, over 954158.43 frames. ], batch size: 32, lr: 3.40e-03, grad_scale: 32.0 2023-04-27 09:44:47,780 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.257e+02 1.710e+02 1.900e+02 2.209e+02 3.616e+02, threshold=3.799e+02, percent-clipped=0.0 2023-04-27 09:45:05,914 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=93471.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:45:22,722 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=93487.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:45:26,248 INFO [finetune.py:976] (4/7) Epoch 17, batch 1850, loss[loss=0.1986, simple_loss=0.2704, pruned_loss=0.06345, over 4861.00 frames. ], tot_loss[loss=0.1818, simple_loss=0.2519, pruned_loss=0.0559, over 954954.15 frames. ], batch size: 34, lr: 3.40e-03, grad_scale: 32.0 2023-04-27 09:45:44,477 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6217, 1.1913, 1.7852, 2.0778, 1.6938, 1.6262, 1.6907, 1.6915], device='cuda:4'), covar=tensor([0.4171, 0.6015, 0.5526, 0.5192, 0.5275, 0.7195, 0.6717, 0.7801], device='cuda:4'), in_proj_covar=tensor([0.0420, 0.0406, 0.0495, 0.0505, 0.0448, 0.0473, 0.0478, 0.0483], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 09:46:27,192 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=93535.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:46:29,063 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5338, 1.8269, 1.6933, 2.2388, 1.9880, 2.0354, 1.6352, 4.6097], device='cuda:4'), covar=tensor([0.0564, 0.0753, 0.0815, 0.1119, 0.0617, 0.0522, 0.0752, 0.0094], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0012, 0.0013, 0.0016], device='cuda:4') 2023-04-27 09:46:37,236 INFO [finetune.py:976] (4/7) Epoch 17, batch 1900, loss[loss=0.1865, simple_loss=0.2555, pruned_loss=0.05879, over 4840.00 frames. ], tot_loss[loss=0.1813, simple_loss=0.2519, pruned_loss=0.05538, over 955936.78 frames. ], batch size: 44, lr: 3.40e-03, grad_scale: 32.0 2023-04-27 09:46:37,987 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.5495, 1.3427, 1.3305, 0.9301, 1.3446, 1.1380, 1.7330, 1.2699], device='cuda:4'), covar=tensor([0.3335, 0.1817, 0.5488, 0.2730, 0.1638, 0.2173, 0.1695, 0.4895], device='cuda:4'), in_proj_covar=tensor([0.0342, 0.0346, 0.0432, 0.0357, 0.0383, 0.0383, 0.0371, 0.0424], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 09:46:42,802 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.049e+02 1.573e+02 1.862e+02 2.216e+02 4.322e+02, threshold=3.725e+02, percent-clipped=2.0 2023-04-27 09:46:44,105 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1972, 2.1459, 1.8189, 1.8004, 2.2464, 1.7590, 2.8283, 1.6261], device='cuda:4'), covar=tensor([0.3568, 0.1974, 0.4861, 0.3329, 0.1760, 0.2787, 0.1446, 0.4545], device='cuda:4'), in_proj_covar=tensor([0.0342, 0.0346, 0.0432, 0.0357, 0.0383, 0.0383, 0.0371, 0.0424], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 09:46:44,734 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=93555.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:47:06,578 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=93587.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:47:10,125 INFO [finetune.py:976] (4/7) Epoch 17, batch 1950, loss[loss=0.1478, simple_loss=0.232, pruned_loss=0.03178, over 4761.00 frames. ], tot_loss[loss=0.1796, simple_loss=0.2502, pruned_loss=0.05448, over 953330.02 frames. ], batch size: 28, lr: 3.40e-03, grad_scale: 32.0 2023-04-27 09:47:16,344 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=93603.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:47:35,236 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=93631.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:47:37,612 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=93635.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:47:48,925 INFO [finetune.py:976] (4/7) Epoch 17, batch 2000, loss[loss=0.2054, simple_loss=0.2614, pruned_loss=0.0747, over 4903.00 frames. ], tot_loss[loss=0.1779, simple_loss=0.2482, pruned_loss=0.0538, over 954897.14 frames. ], batch size: 35, lr: 3.40e-03, grad_scale: 32.0 2023-04-27 09:47:54,466 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.097e+02 1.579e+02 1.885e+02 2.263e+02 4.038e+02, threshold=3.769e+02, percent-clipped=1.0 2023-04-27 09:48:03,176 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=93666.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:48:11,984 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=93679.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:48:17,872 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4962, 1.3484, 1.6940, 1.6542, 1.3933, 1.2837, 1.3085, 0.7361], device='cuda:4'), covar=tensor([0.0538, 0.1010, 0.0600, 0.0575, 0.0707, 0.1232, 0.0707, 0.0907], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0070, 0.0069, 0.0068, 0.0075, 0.0097, 0.0075, 0.0068], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 09:48:21,390 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=93692.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:48:21,911 INFO [finetune.py:976] (4/7) Epoch 17, batch 2050, loss[loss=0.1423, simple_loss=0.2149, pruned_loss=0.03486, over 4776.00 frames. ], tot_loss[loss=0.1745, simple_loss=0.2444, pruned_loss=0.0523, over 953397.70 frames. ], batch size: 28, lr: 3.40e-03, grad_scale: 32.0 2023-04-27 09:48:35,563 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=93714.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:48:53,239 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=93740.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:48:55,037 INFO [finetune.py:976] (4/7) Epoch 17, batch 2100, loss[loss=0.1817, simple_loss=0.249, pruned_loss=0.05717, over 4808.00 frames. ], tot_loss[loss=0.1756, simple_loss=0.2451, pruned_loss=0.05311, over 953261.09 frames. ], batch size: 38, lr: 3.40e-03, grad_scale: 32.0 2023-04-27 09:49:01,810 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.279e+02 1.585e+02 1.847e+02 2.242e+02 6.268e+02, threshold=3.694e+02, percent-clipped=2.0 2023-04-27 09:49:11,168 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1389, 1.5541, 1.4577, 1.9952, 2.1709, 1.8022, 1.7597, 1.5399], device='cuda:4'), covar=tensor([0.1813, 0.1880, 0.2001, 0.1763, 0.1242, 0.1853, 0.2413, 0.2314], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0315, 0.0352, 0.0292, 0.0330, 0.0312, 0.0303, 0.0368], device='cuda:4'), out_proj_covar=tensor([6.3856e-05, 6.5575e-05, 7.5120e-05, 5.9408e-05, 6.8680e-05, 6.5730e-05, 6.3855e-05, 7.8554e-05], device='cuda:4') 2023-04-27 09:49:12,967 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=93770.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:49:13,577 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=93771.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:49:17,569 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-27 09:49:30,256 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4959, 1.3667, 0.5045, 1.1867, 1.4558, 1.3735, 1.2575, 1.2738], device='cuda:4'), covar=tensor([0.0532, 0.0410, 0.0436, 0.0605, 0.0315, 0.0558, 0.0546, 0.0614], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0029, 0.0020, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0044, 0.0038, 0.0051, 0.0038, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 09:49:30,856 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6917, 1.7626, 0.7427, 1.3783, 1.8913, 1.5658, 1.4447, 1.4900], device='cuda:4'), covar=tensor([0.0531, 0.0393, 0.0380, 0.0576, 0.0282, 0.0509, 0.0552, 0.0595], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0029, 0.0020, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0044, 0.0038, 0.0051, 0.0038, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 09:49:33,703 INFO [finetune.py:976] (4/7) Epoch 17, batch 2150, loss[loss=0.2317, simple_loss=0.3016, pruned_loss=0.08086, over 4903.00 frames. ], tot_loss[loss=0.179, simple_loss=0.2488, pruned_loss=0.05462, over 951434.25 frames. ], batch size: 37, lr: 3.39e-03, grad_scale: 32.0 2023-04-27 09:50:04,672 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.55 vs. limit=5.0 2023-04-27 09:50:13,739 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=93819.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:50:26,741 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=93831.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:50:47,864 INFO [finetune.py:976] (4/7) Epoch 17, batch 2200, loss[loss=0.1926, simple_loss=0.2644, pruned_loss=0.06044, over 4843.00 frames. ], tot_loss[loss=0.1806, simple_loss=0.2509, pruned_loss=0.05512, over 952651.06 frames. ], batch size: 44, lr: 3.39e-03, grad_scale: 32.0 2023-04-27 09:50:59,349 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.152e+02 1.629e+02 1.916e+02 2.338e+02 4.475e+02, threshold=3.833e+02, percent-clipped=3.0 2023-04-27 09:51:41,202 INFO [finetune.py:976] (4/7) Epoch 17, batch 2250, loss[loss=0.1951, simple_loss=0.2661, pruned_loss=0.06209, over 4924.00 frames. ], tot_loss[loss=0.1814, simple_loss=0.2525, pruned_loss=0.05519, over 955464.04 frames. ], batch size: 33, lr: 3.39e-03, grad_scale: 32.0 2023-04-27 09:51:58,218 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.90 vs. limit=2.0 2023-04-27 09:52:14,468 INFO [finetune.py:976] (4/7) Epoch 17, batch 2300, loss[loss=0.1738, simple_loss=0.2514, pruned_loss=0.04808, over 4798.00 frames. ], tot_loss[loss=0.181, simple_loss=0.2525, pruned_loss=0.05474, over 953889.09 frames. ], batch size: 45, lr: 3.39e-03, grad_scale: 32.0 2023-04-27 09:52:20,959 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.059e+02 1.592e+02 1.958e+02 2.327e+02 4.753e+02, threshold=3.916e+02, percent-clipped=2.0 2023-04-27 09:52:40,179 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=93981.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:52:47,881 INFO [finetune.py:976] (4/7) Epoch 17, batch 2350, loss[loss=0.1683, simple_loss=0.2394, pruned_loss=0.04856, over 4824.00 frames. ], tot_loss[loss=0.1785, simple_loss=0.2494, pruned_loss=0.05382, over 954635.54 frames. ], batch size: 39, lr: 3.39e-03, grad_scale: 32.0 2023-04-27 09:53:40,622 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2081, 2.6216, 1.0197, 1.3778, 2.0781, 1.2336, 3.6515, 1.7332], device='cuda:4'), covar=tensor([0.0709, 0.0839, 0.0894, 0.1333, 0.0536, 0.1000, 0.0290, 0.0644], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0048, 0.0047, 0.0050, 0.0052, 0.0075, 0.0052], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 09:53:48,604 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=94042.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:53:49,096 INFO [finetune.py:976] (4/7) Epoch 17, batch 2400, loss[loss=0.1097, simple_loss=0.1808, pruned_loss=0.01928, over 4804.00 frames. ], tot_loss[loss=0.1748, simple_loss=0.2454, pruned_loss=0.05209, over 955277.51 frames. ], batch size: 25, lr: 3.39e-03, grad_scale: 32.0 2023-04-27 09:53:56,112 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.042e+02 1.528e+02 1.774e+02 2.113e+02 4.800e+02, threshold=3.549e+02, percent-clipped=1.0 2023-04-27 09:54:23,068 INFO [finetune.py:976] (4/7) Epoch 17, batch 2450, loss[loss=0.1785, simple_loss=0.2545, pruned_loss=0.05132, over 4836.00 frames. ], tot_loss[loss=0.1724, simple_loss=0.2426, pruned_loss=0.05112, over 956285.24 frames. ], batch size: 40, lr: 3.39e-03, grad_scale: 32.0 2023-04-27 09:54:46,628 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=94126.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:54:57,072 INFO [finetune.py:976] (4/7) Epoch 17, batch 2500, loss[loss=0.2017, simple_loss=0.2694, pruned_loss=0.06705, over 4809.00 frames. ], tot_loss[loss=0.1754, simple_loss=0.2457, pruned_loss=0.05254, over 957973.85 frames. ], batch size: 41, lr: 3.39e-03, grad_scale: 32.0 2023-04-27 09:55:03,648 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.026e+02 1.665e+02 1.997e+02 2.427e+02 4.291e+02, threshold=3.995e+02, percent-clipped=3.0 2023-04-27 09:55:06,702 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5926, 1.4027, 1.7899, 1.8350, 1.4248, 1.3179, 1.4859, 0.9024], device='cuda:4'), covar=tensor([0.0473, 0.0710, 0.0402, 0.0505, 0.0700, 0.1183, 0.0557, 0.0699], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0070, 0.0068, 0.0068, 0.0075, 0.0096, 0.0074, 0.0067], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 09:55:12,006 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.13 vs. limit=2.0 2023-04-27 09:55:31,032 INFO [finetune.py:976] (4/7) Epoch 17, batch 2550, loss[loss=0.1756, simple_loss=0.2557, pruned_loss=0.04777, over 4907.00 frames. ], tot_loss[loss=0.1787, simple_loss=0.2499, pruned_loss=0.05376, over 956910.14 frames. ], batch size: 36, lr: 3.39e-03, grad_scale: 32.0 2023-04-27 09:55:32,428 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-27 09:55:49,189 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1422, 1.3929, 1.2968, 1.7407, 1.5450, 1.8413, 1.3451, 3.3838], device='cuda:4'), covar=tensor([0.0702, 0.0947, 0.0951, 0.1287, 0.0721, 0.0557, 0.0874, 0.0187], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0039, 0.0037, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0012, 0.0013, 0.0016], device='cuda:4') 2023-04-27 09:55:52,787 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=94223.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 09:56:09,851 INFO [finetune.py:976] (4/7) Epoch 17, batch 2600, loss[loss=0.1304, simple_loss=0.1931, pruned_loss=0.03379, over 4494.00 frames. ], tot_loss[loss=0.1798, simple_loss=0.251, pruned_loss=0.05424, over 956747.90 frames. ], batch size: 20, lr: 3.39e-03, grad_scale: 32.0 2023-04-27 09:56:21,228 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.193e+02 1.596e+02 1.895e+02 2.398e+02 4.293e+02, threshold=3.790e+02, percent-clipped=2.0 2023-04-27 09:56:54,699 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0861, 1.7681, 1.9855, 2.3566, 2.4595, 1.8399, 1.6361, 2.1964], device='cuda:4'), covar=tensor([0.0722, 0.1059, 0.0675, 0.0518, 0.0507, 0.0833, 0.0715, 0.0490], device='cuda:4'), in_proj_covar=tensor([0.0188, 0.0199, 0.0180, 0.0170, 0.0174, 0.0179, 0.0151, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 09:57:00,078 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=94284.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 09:57:05,478 INFO [finetune.py:976] (4/7) Epoch 17, batch 2650, loss[loss=0.15, simple_loss=0.2301, pruned_loss=0.03495, over 4746.00 frames. ], tot_loss[loss=0.1815, simple_loss=0.2531, pruned_loss=0.05493, over 957044.01 frames. ], batch size: 27, lr: 3.39e-03, grad_scale: 32.0 2023-04-27 09:57:34,892 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=94337.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:57:38,467 INFO [finetune.py:976] (4/7) Epoch 17, batch 2700, loss[loss=0.1551, simple_loss=0.2299, pruned_loss=0.04014, over 4756.00 frames. ], tot_loss[loss=0.1799, simple_loss=0.2516, pruned_loss=0.05407, over 956250.15 frames. ], batch size: 27, lr: 3.39e-03, grad_scale: 32.0 2023-04-27 09:57:41,611 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=94348.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:57:43,966 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.591e+02 1.889e+02 2.175e+02 3.860e+02, threshold=3.779e+02, percent-clipped=1.0 2023-04-27 09:58:17,888 INFO [finetune.py:976] (4/7) Epoch 17, batch 2750, loss[loss=0.1433, simple_loss=0.2132, pruned_loss=0.03666, over 4892.00 frames. ], tot_loss[loss=0.1787, simple_loss=0.2496, pruned_loss=0.05392, over 956507.34 frames. ], batch size: 32, lr: 3.39e-03, grad_scale: 32.0 2023-04-27 09:58:38,894 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=94409.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 09:59:01,634 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=94426.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 09:59:02,306 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1119, 1.8239, 2.3288, 2.5929, 2.1400, 2.0232, 2.1109, 2.1549], device='cuda:4'), covar=tensor([0.5083, 0.7605, 0.7898, 0.6108, 0.6533, 0.8988, 1.0351, 0.9588], device='cuda:4'), in_proj_covar=tensor([0.0421, 0.0408, 0.0497, 0.0504, 0.0450, 0.0475, 0.0480, 0.0484], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 09:59:24,246 INFO [finetune.py:976] (4/7) Epoch 17, batch 2800, loss[loss=0.1957, simple_loss=0.2602, pruned_loss=0.06565, over 4827.00 frames. ], tot_loss[loss=0.1763, simple_loss=0.2461, pruned_loss=0.0533, over 957446.30 frames. ], batch size: 39, lr: 3.39e-03, grad_scale: 32.0 2023-04-27 09:59:35,022 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.123e+02 1.496e+02 1.786e+02 2.106e+02 4.249e+02, threshold=3.571e+02, percent-clipped=1.0 2023-04-27 09:59:46,817 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5609, 1.1936, 1.3507, 1.2387, 1.7231, 1.4515, 1.1379, 1.2788], device='cuda:4'), covar=tensor([0.1566, 0.1284, 0.1801, 0.1203, 0.0812, 0.1381, 0.1806, 0.2314], device='cuda:4'), in_proj_covar=tensor([0.0306, 0.0312, 0.0350, 0.0289, 0.0326, 0.0308, 0.0299, 0.0363], device='cuda:4'), out_proj_covar=tensor([6.3332e-05, 6.4967e-05, 7.4715e-05, 5.8649e-05, 6.7839e-05, 6.4970e-05, 6.3089e-05, 7.7466e-05], device='cuda:4') 2023-04-27 09:59:54,410 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=94474.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:00:01,047 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3289, 1.2968, 1.6512, 1.5868, 1.2516, 1.1428, 1.4158, 0.9364], device='cuda:4'), covar=tensor([0.0583, 0.0572, 0.0401, 0.0542, 0.0735, 0.1100, 0.0555, 0.0631], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0070, 0.0069, 0.0068, 0.0076, 0.0097, 0.0075, 0.0067], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 10:00:06,867 INFO [finetune.py:976] (4/7) Epoch 17, batch 2850, loss[loss=0.1747, simple_loss=0.2434, pruned_loss=0.05298, over 4819.00 frames. ], tot_loss[loss=0.175, simple_loss=0.2444, pruned_loss=0.05281, over 957991.37 frames. ], batch size: 45, lr: 3.39e-03, grad_scale: 32.0 2023-04-27 10:00:41,003 INFO [finetune.py:976] (4/7) Epoch 17, batch 2900, loss[loss=0.1798, simple_loss=0.2572, pruned_loss=0.05117, over 4925.00 frames. ], tot_loss[loss=0.1769, simple_loss=0.2471, pruned_loss=0.05341, over 959865.10 frames. ], batch size: 33, lr: 3.39e-03, grad_scale: 32.0 2023-04-27 10:00:46,390 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.024e+02 1.635e+02 2.008e+02 2.460e+02 5.439e+02, threshold=4.016e+02, percent-clipped=3.0 2023-04-27 10:01:04,315 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=94579.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 10:01:14,239 INFO [finetune.py:976] (4/7) Epoch 17, batch 2950, loss[loss=0.1922, simple_loss=0.2705, pruned_loss=0.05697, over 4746.00 frames. ], tot_loss[loss=0.1782, simple_loss=0.249, pruned_loss=0.05368, over 959266.21 frames. ], batch size: 54, lr: 3.39e-03, grad_scale: 32.0 2023-04-27 10:01:39,151 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.80 vs. limit=2.0 2023-04-27 10:01:48,975 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4649, 1.2013, 0.6407, 1.2054, 1.1604, 1.3494, 1.2680, 1.2807], device='cuda:4'), covar=tensor([0.0565, 0.0402, 0.0383, 0.0613, 0.0315, 0.0575, 0.0536, 0.0622], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0029, 0.0020, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0044, 0.0038, 0.0051, 0.0038, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 10:01:50,676 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0284, 1.6619, 1.8277, 2.1828, 2.2974, 1.8649, 1.7108, 2.0738], device='cuda:4'), covar=tensor([0.0682, 0.0990, 0.0606, 0.0507, 0.0492, 0.0700, 0.0697, 0.0470], device='cuda:4'), in_proj_covar=tensor([0.0189, 0.0201, 0.0182, 0.0171, 0.0176, 0.0180, 0.0152, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 10:02:00,093 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=94637.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:02:04,043 INFO [finetune.py:976] (4/7) Epoch 17, batch 3000, loss[loss=0.1524, simple_loss=0.2334, pruned_loss=0.03569, over 4770.00 frames. ], tot_loss[loss=0.1789, simple_loss=0.2499, pruned_loss=0.0539, over 957601.71 frames. ], batch size: 28, lr: 3.39e-03, grad_scale: 32.0 2023-04-27 10:02:04,043 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 10:02:27,438 INFO [finetune.py:1010] (4/7) Epoch 17, validation: loss=0.1526, simple_loss=0.2233, pruned_loss=0.04089, over 2265189.00 frames. 2023-04-27 10:02:27,439 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-27 10:02:36,293 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.43 vs. limit=2.0 2023-04-27 10:02:39,392 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.157e+02 1.688e+02 2.099e+02 2.458e+02 3.567e+02, threshold=4.198e+02, percent-clipped=0.0 2023-04-27 10:03:10,438 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=94685.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:03:15,761 INFO [finetune.py:976] (4/7) Epoch 17, batch 3050, loss[loss=0.1547, simple_loss=0.2403, pruned_loss=0.03449, over 4783.00 frames. ], tot_loss[loss=0.1789, simple_loss=0.2501, pruned_loss=0.05385, over 958683.17 frames. ], batch size: 29, lr: 3.39e-03, grad_scale: 32.0 2023-04-27 10:03:24,054 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=94704.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 10:03:47,899 INFO [finetune.py:976] (4/7) Epoch 17, batch 3100, loss[loss=0.1669, simple_loss=0.2357, pruned_loss=0.04906, over 4770.00 frames. ], tot_loss[loss=0.179, simple_loss=0.2495, pruned_loss=0.05425, over 957724.67 frames. ], batch size: 28, lr: 3.39e-03, grad_scale: 32.0 2023-04-27 10:03:55,871 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.099e+02 1.556e+02 1.873e+02 2.227e+02 4.960e+02, threshold=3.745e+02, percent-clipped=1.0 2023-04-27 10:04:21,244 INFO [finetune.py:976] (4/7) Epoch 17, batch 3150, loss[loss=0.1688, simple_loss=0.2341, pruned_loss=0.0518, over 4732.00 frames. ], tot_loss[loss=0.1774, simple_loss=0.2475, pruned_loss=0.05364, over 955923.94 frames. ], batch size: 54, lr: 3.39e-03, grad_scale: 64.0 2023-04-27 10:05:23,150 INFO [finetune.py:976] (4/7) Epoch 17, batch 3200, loss[loss=0.1801, simple_loss=0.2513, pruned_loss=0.05451, over 4839.00 frames. ], tot_loss[loss=0.1745, simple_loss=0.2442, pruned_loss=0.05237, over 957444.51 frames. ], batch size: 47, lr: 3.39e-03, grad_scale: 64.0 2023-04-27 10:05:34,601 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.148e+01 1.533e+02 1.770e+02 2.163e+02 3.313e+02, threshold=3.540e+02, percent-clipped=0.0 2023-04-27 10:05:53,672 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=94879.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 10:06:02,201 INFO [finetune.py:976] (4/7) Epoch 17, batch 3250, loss[loss=0.1376, simple_loss=0.2081, pruned_loss=0.03357, over 4765.00 frames. ], tot_loss[loss=0.175, simple_loss=0.2444, pruned_loss=0.05274, over 954391.39 frames. ], batch size: 28, lr: 3.39e-03, grad_scale: 64.0 2023-04-27 10:06:26,491 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=94927.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 10:06:36,232 INFO [finetune.py:976] (4/7) Epoch 17, batch 3300, loss[loss=0.1809, simple_loss=0.2314, pruned_loss=0.06517, over 4442.00 frames. ], tot_loss[loss=0.1781, simple_loss=0.2478, pruned_loss=0.05418, over 953123.94 frames. ], batch size: 19, lr: 3.38e-03, grad_scale: 64.0 2023-04-27 10:06:47,629 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.257e+02 1.598e+02 1.896e+02 2.226e+02 3.987e+02, threshold=3.793e+02, percent-clipped=1.0 2023-04-27 10:07:30,064 INFO [finetune.py:976] (4/7) Epoch 17, batch 3350, loss[loss=0.2275, simple_loss=0.302, pruned_loss=0.07652, over 4932.00 frames. ], tot_loss[loss=0.1795, simple_loss=0.2495, pruned_loss=0.05471, over 953468.58 frames. ], batch size: 41, lr: 3.38e-03, grad_scale: 64.0 2023-04-27 10:07:36,978 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=95004.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:08:03,817 INFO [finetune.py:976] (4/7) Epoch 17, batch 3400, loss[loss=0.2192, simple_loss=0.279, pruned_loss=0.07974, over 4862.00 frames. ], tot_loss[loss=0.1807, simple_loss=0.2508, pruned_loss=0.05529, over 954781.78 frames. ], batch size: 44, lr: 3.38e-03, grad_scale: 64.0 2023-04-27 10:08:09,294 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.057e+02 1.633e+02 1.895e+02 2.474e+02 5.536e+02, threshold=3.790e+02, percent-clipped=1.0 2023-04-27 10:08:09,366 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=95052.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:08:21,296 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.81 vs. limit=2.0 2023-04-27 10:08:23,995 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9504, 1.5713, 1.5503, 1.7592, 2.1602, 1.7167, 1.5257, 1.4876], device='cuda:4'), covar=tensor([0.1669, 0.1566, 0.1945, 0.1348, 0.0868, 0.1683, 0.2089, 0.2150], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0316, 0.0356, 0.0292, 0.0330, 0.0313, 0.0302, 0.0369], device='cuda:4'), out_proj_covar=tensor([6.4269e-05, 6.5963e-05, 7.5927e-05, 5.9235e-05, 6.8445e-05, 6.5940e-05, 6.3781e-05, 7.8775e-05], device='cuda:4') 2023-04-27 10:08:24,609 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2251, 1.4377, 1.6929, 1.8208, 1.7366, 1.8580, 1.7441, 1.7558], device='cuda:4'), covar=tensor([0.3877, 0.5290, 0.4647, 0.4415, 0.5460, 0.7289, 0.5305, 0.4653], device='cuda:4'), in_proj_covar=tensor([0.0329, 0.0371, 0.0317, 0.0332, 0.0342, 0.0394, 0.0353, 0.0323], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 10:08:37,248 INFO [finetune.py:976] (4/7) Epoch 17, batch 3450, loss[loss=0.2341, simple_loss=0.3002, pruned_loss=0.08402, over 4209.00 frames. ], tot_loss[loss=0.181, simple_loss=0.251, pruned_loss=0.05556, over 953988.04 frames. ], batch size: 65, lr: 3.38e-03, grad_scale: 64.0 2023-04-27 10:08:44,630 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=95105.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 10:09:11,015 INFO [finetune.py:976] (4/7) Epoch 17, batch 3500, loss[loss=0.1591, simple_loss=0.2367, pruned_loss=0.0408, over 4767.00 frames. ], tot_loss[loss=0.1784, simple_loss=0.2482, pruned_loss=0.05426, over 954010.75 frames. ], batch size: 28, lr: 3.38e-03, grad_scale: 64.0 2023-04-27 10:09:16,406 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.081e+02 1.623e+02 1.952e+02 2.265e+02 3.860e+02, threshold=3.904e+02, percent-clipped=1.0 2023-04-27 10:09:17,175 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1816, 1.3975, 2.0009, 2.5215, 2.0329, 1.5504, 1.3490, 1.7702], device='cuda:4'), covar=tensor([0.3344, 0.4023, 0.1893, 0.2632, 0.2707, 0.2798, 0.4500, 0.2290], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0246, 0.0226, 0.0315, 0.0217, 0.0230, 0.0228, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 10:09:26,087 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=95166.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 10:09:44,906 INFO [finetune.py:976] (4/7) Epoch 17, batch 3550, loss[loss=0.1498, simple_loss=0.2181, pruned_loss=0.04081, over 4830.00 frames. ], tot_loss[loss=0.1767, simple_loss=0.2458, pruned_loss=0.05378, over 954010.78 frames. ], batch size: 39, lr: 3.38e-03, grad_scale: 64.0 2023-04-27 10:10:15,819 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8632, 1.3409, 4.9037, 4.5961, 4.2905, 4.6311, 4.3700, 4.4155], device='cuda:4'), covar=tensor([0.6731, 0.5958, 0.0908, 0.1610, 0.1197, 0.1602, 0.1567, 0.1353], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0303, 0.0399, 0.0402, 0.0345, 0.0403, 0.0308, 0.0361], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 10:10:42,816 INFO [finetune.py:976] (4/7) Epoch 17, batch 3600, loss[loss=0.1718, simple_loss=0.2338, pruned_loss=0.05496, over 4255.00 frames. ], tot_loss[loss=0.1743, simple_loss=0.2433, pruned_loss=0.05262, over 954769.64 frames. ], batch size: 18, lr: 3.38e-03, grad_scale: 64.0 2023-04-27 10:10:42,942 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5731, 2.4154, 1.8718, 2.1119, 2.3383, 1.9690, 3.0220, 1.7900], device='cuda:4'), covar=tensor([0.3583, 0.2046, 0.4699, 0.3087, 0.1915, 0.2615, 0.2078, 0.4654], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0343, 0.0424, 0.0351, 0.0380, 0.0376, 0.0368, 0.0418], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 10:10:54,070 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.027e+02 1.577e+02 1.958e+02 2.437e+02 7.407e+02, threshold=3.916e+02, percent-clipped=3.0 2023-04-27 10:10:59,897 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=95260.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:11:10,464 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=95268.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:11:20,082 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.6895, 4.5634, 3.2380, 5.3709, 4.7024, 4.7070, 2.0265, 4.6889], device='cuda:4'), covar=tensor([0.1505, 0.0921, 0.2993, 0.0811, 0.4107, 0.1565, 0.5530, 0.1772], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0213, 0.0249, 0.0304, 0.0296, 0.0247, 0.0271, 0.0271], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 10:11:33,399 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.49 vs. limit=2.0 2023-04-27 10:11:34,140 INFO [finetune.py:976] (4/7) Epoch 17, batch 3650, loss[loss=0.2121, simple_loss=0.2808, pruned_loss=0.0717, over 4817.00 frames. ], tot_loss[loss=0.1758, simple_loss=0.2457, pruned_loss=0.05301, over 955586.75 frames. ], batch size: 39, lr: 3.38e-03, grad_scale: 64.0 2023-04-27 10:11:39,248 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2225, 1.3939, 1.7003, 1.8597, 1.7149, 1.8391, 1.7885, 1.7620], device='cuda:4'), covar=tensor([0.3677, 0.5168, 0.4230, 0.4338, 0.5408, 0.6969, 0.4879, 0.4719], device='cuda:4'), in_proj_covar=tensor([0.0329, 0.0370, 0.0316, 0.0331, 0.0342, 0.0393, 0.0353, 0.0323], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 10:11:51,438 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=95321.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:11:57,811 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=95329.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:12:05,325 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=95339.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:12:07,681 INFO [finetune.py:976] (4/7) Epoch 17, batch 3700, loss[loss=0.1995, simple_loss=0.2743, pruned_loss=0.06232, over 4817.00 frames. ], tot_loss[loss=0.1776, simple_loss=0.2485, pruned_loss=0.05331, over 956491.10 frames. ], batch size: 40, lr: 3.38e-03, grad_scale: 64.0 2023-04-27 10:12:13,173 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.977e+01 1.742e+02 1.979e+02 2.292e+02 4.077e+02, threshold=3.957e+02, percent-clipped=1.0 2023-04-27 10:12:41,445 INFO [finetune.py:976] (4/7) Epoch 17, batch 3750, loss[loss=0.2072, simple_loss=0.2688, pruned_loss=0.07287, over 4929.00 frames. ], tot_loss[loss=0.1781, simple_loss=0.2493, pruned_loss=0.05346, over 956141.44 frames. ], batch size: 33, lr: 3.38e-03, grad_scale: 32.0 2023-04-27 10:12:46,018 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=95400.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:13:06,437 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0867, 1.6471, 1.8938, 2.3731, 2.3445, 1.8680, 1.8110, 2.1289], device='cuda:4'), covar=tensor([0.0804, 0.1197, 0.0775, 0.0633, 0.0599, 0.0800, 0.0684, 0.0595], device='cuda:4'), in_proj_covar=tensor([0.0188, 0.0199, 0.0180, 0.0169, 0.0175, 0.0179, 0.0151, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 10:13:14,402 INFO [finetune.py:976] (4/7) Epoch 17, batch 3800, loss[loss=0.1274, simple_loss=0.2063, pruned_loss=0.02423, over 4753.00 frames. ], tot_loss[loss=0.1784, simple_loss=0.25, pruned_loss=0.0534, over 956557.71 frames. ], batch size: 27, lr: 3.38e-03, grad_scale: 32.0 2023-04-27 10:13:20,915 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.076e+02 1.628e+02 1.968e+02 2.270e+02 5.105e+02, threshold=3.935e+02, percent-clipped=1.0 2023-04-27 10:13:22,958 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.52 vs. limit=5.0 2023-04-27 10:13:25,884 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=95461.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 10:13:33,856 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.43 vs. limit=2.0 2023-04-27 10:13:48,141 INFO [finetune.py:976] (4/7) Epoch 17, batch 3850, loss[loss=0.1417, simple_loss=0.218, pruned_loss=0.03271, over 4773.00 frames. ], tot_loss[loss=0.1771, simple_loss=0.2485, pruned_loss=0.05279, over 955847.62 frames. ], batch size: 26, lr: 3.38e-03, grad_scale: 32.0 2023-04-27 10:14:20,427 INFO [finetune.py:976] (4/7) Epoch 17, batch 3900, loss[loss=0.1618, simple_loss=0.2327, pruned_loss=0.04544, over 4810.00 frames. ], tot_loss[loss=0.1763, simple_loss=0.2473, pruned_loss=0.05268, over 957704.98 frames. ], batch size: 51, lr: 3.38e-03, grad_scale: 32.0 2023-04-27 10:14:27,544 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.967e+01 1.562e+02 1.786e+02 2.211e+02 6.075e+02, threshold=3.573e+02, percent-clipped=1.0 2023-04-27 10:14:27,672 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1310, 2.5150, 2.5647, 2.8310, 2.7302, 2.6784, 2.3900, 4.8748], device='cuda:4'), covar=tensor([0.0404, 0.0551, 0.0556, 0.0834, 0.0426, 0.0356, 0.0515, 0.0120], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0039, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 10:14:31,990 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7238, 0.6868, 1.5558, 2.0707, 1.7762, 1.6124, 1.5572, 1.5959], device='cuda:4'), covar=tensor([0.4452, 0.5957, 0.5411, 0.5726, 0.5505, 0.6581, 0.6847, 0.6847], device='cuda:4'), in_proj_covar=tensor([0.0419, 0.0406, 0.0495, 0.0502, 0.0447, 0.0473, 0.0479, 0.0484], device='cuda:4'), out_proj_covar=tensor([1.0075e-04, 9.9850e-05, 1.1125e-04, 1.1972e-04, 1.0721e-04, 1.1354e-04, 1.1376e-04, 1.1434e-04], device='cuda:4') 2023-04-27 10:14:34,999 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1071, 2.4972, 0.9016, 1.3365, 1.7370, 1.2410, 3.0062, 1.7498], device='cuda:4'), covar=tensor([0.0636, 0.0497, 0.0733, 0.1217, 0.0512, 0.0955, 0.0264, 0.0587], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0065, 0.0048, 0.0046, 0.0050, 0.0052, 0.0075, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 10:14:52,351 INFO [finetune.py:976] (4/7) Epoch 17, batch 3950, loss[loss=0.1841, simple_loss=0.248, pruned_loss=0.06015, over 4821.00 frames. ], tot_loss[loss=0.175, simple_loss=0.2449, pruned_loss=0.05255, over 956611.84 frames. ], batch size: 40, lr: 3.38e-03, grad_scale: 32.0 2023-04-27 10:14:58,356 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4019, 1.6027, 1.8428, 1.9465, 1.8520, 1.9331, 1.8999, 1.9061], device='cuda:4'), covar=tensor([0.3862, 0.5634, 0.4942, 0.4651, 0.5813, 0.7485, 0.5609, 0.4821], device='cuda:4'), in_proj_covar=tensor([0.0330, 0.0371, 0.0317, 0.0332, 0.0344, 0.0394, 0.0354, 0.0324], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 10:15:08,540 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=95616.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:15:09,187 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=95617.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:15:10,982 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1758, 2.5308, 0.9417, 1.4036, 1.8700, 1.2793, 3.3676, 1.8520], device='cuda:4'), covar=tensor([0.0678, 0.0569, 0.0780, 0.1284, 0.0554, 0.1011, 0.0362, 0.0607], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0065, 0.0048, 0.0046, 0.0050, 0.0052, 0.0075, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 10:15:13,457 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=95624.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:15:31,265 INFO [finetune.py:976] (4/7) Epoch 17, batch 4000, loss[loss=0.1599, simple_loss=0.238, pruned_loss=0.04093, over 4810.00 frames. ], tot_loss[loss=0.1743, simple_loss=0.2438, pruned_loss=0.05242, over 954370.97 frames. ], batch size: 51, lr: 3.38e-03, grad_scale: 32.0 2023-04-27 10:15:43,871 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.207e+01 1.538e+02 1.849e+02 2.258e+02 4.125e+02, threshold=3.697e+02, percent-clipped=1.0 2023-04-27 10:16:16,168 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=95678.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:16:36,028 INFO [finetune.py:976] (4/7) Epoch 17, batch 4050, loss[loss=0.2066, simple_loss=0.2741, pruned_loss=0.06954, over 4784.00 frames. ], tot_loss[loss=0.1758, simple_loss=0.246, pruned_loss=0.05277, over 955022.54 frames. ], batch size: 29, lr: 3.38e-03, grad_scale: 32.0 2023-04-27 10:16:37,790 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=95695.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:17:36,636 INFO [finetune.py:976] (4/7) Epoch 17, batch 4100, loss[loss=0.2097, simple_loss=0.2677, pruned_loss=0.07582, over 4061.00 frames. ], tot_loss[loss=0.1788, simple_loss=0.2495, pruned_loss=0.05405, over 953336.36 frames. ], batch size: 65, lr: 3.38e-03, grad_scale: 32.0 2023-04-27 10:17:43,693 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.341e+02 1.651e+02 1.954e+02 2.364e+02 4.876e+02, threshold=3.909e+02, percent-clipped=1.0 2023-04-27 10:17:47,705 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3974, 3.2971, 0.8801, 1.7729, 1.8923, 2.2651, 1.9063, 1.0479], device='cuda:4'), covar=tensor([0.1422, 0.0779, 0.2003, 0.1257, 0.1019, 0.1113, 0.1506, 0.1830], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0242, 0.0137, 0.0119, 0.0132, 0.0152, 0.0117, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 10:17:50,088 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=95761.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 10:18:09,897 INFO [finetune.py:976] (4/7) Epoch 17, batch 4150, loss[loss=0.1942, simple_loss=0.2554, pruned_loss=0.06645, over 4881.00 frames. ], tot_loss[loss=0.1804, simple_loss=0.2517, pruned_loss=0.05457, over 955505.19 frames. ], batch size: 32, lr: 3.38e-03, grad_scale: 32.0 2023-04-27 10:18:21,838 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=95809.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 10:18:24,200 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=95812.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 10:18:43,507 INFO [finetune.py:976] (4/7) Epoch 17, batch 4200, loss[loss=0.1658, simple_loss=0.2363, pruned_loss=0.04768, over 4751.00 frames. ], tot_loss[loss=0.1802, simple_loss=0.2517, pruned_loss=0.05441, over 953075.19 frames. ], batch size: 26, lr: 3.38e-03, grad_scale: 32.0 2023-04-27 10:18:46,042 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=95847.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:18:49,668 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.908e+01 1.602e+02 1.963e+02 2.412e+02 3.992e+02, threshold=3.927e+02, percent-clipped=2.0 2023-04-27 10:18:57,714 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7536, 2.1350, 1.0088, 1.4453, 2.0694, 1.6633, 1.5324, 1.6477], device='cuda:4'), covar=tensor([0.0498, 0.0330, 0.0335, 0.0589, 0.0256, 0.0524, 0.0506, 0.0559], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0029, 0.0020, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0045, 0.0038, 0.0051, 0.0038, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 10:19:04,178 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=95873.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 10:19:16,331 INFO [finetune.py:976] (4/7) Epoch 17, batch 4250, loss[loss=0.1739, simple_loss=0.2523, pruned_loss=0.0477, over 4928.00 frames. ], tot_loss[loss=0.1787, simple_loss=0.2495, pruned_loss=0.05391, over 954680.16 frames. ], batch size: 38, lr: 3.38e-03, grad_scale: 32.0 2023-04-27 10:19:26,045 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=95908.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 10:19:31,924 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=95916.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:19:36,878 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=95924.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:19:38,514 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0678, 2.8408, 2.4043, 2.5316, 1.9628, 2.3529, 2.3883, 1.7481], device='cuda:4'), covar=tensor([0.2029, 0.1054, 0.0779, 0.1147, 0.3056, 0.1313, 0.1908, 0.2708], device='cuda:4'), in_proj_covar=tensor([0.0282, 0.0303, 0.0217, 0.0277, 0.0309, 0.0256, 0.0247, 0.0263], device='cuda:4'), out_proj_covar=tensor([1.1325e-04, 1.2065e-04, 8.6121e-05, 1.0972e-04, 1.2577e-04, 1.0192e-04, 9.9809e-05, 1.0456e-04], device='cuda:4') 2023-04-27 10:19:48,685 INFO [finetune.py:976] (4/7) Epoch 17, batch 4300, loss[loss=0.1652, simple_loss=0.2323, pruned_loss=0.04907, over 4866.00 frames. ], tot_loss[loss=0.1777, simple_loss=0.2478, pruned_loss=0.0538, over 954713.64 frames. ], batch size: 31, lr: 3.38e-03, grad_scale: 32.0 2023-04-27 10:19:49,430 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8812, 1.3652, 1.5860, 1.5239, 1.9820, 1.6651, 1.3498, 1.4944], device='cuda:4'), covar=tensor([0.1451, 0.1417, 0.1876, 0.1296, 0.0845, 0.1264, 0.1937, 0.1983], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0313, 0.0351, 0.0289, 0.0327, 0.0310, 0.0299, 0.0366], device='cuda:4'), out_proj_covar=tensor([6.3924e-05, 6.5329e-05, 7.4749e-05, 5.8637e-05, 6.7969e-05, 6.5345e-05, 6.3123e-05, 7.8001e-05], device='cuda:4') 2023-04-27 10:19:54,839 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.107e+02 1.623e+02 1.872e+02 2.208e+02 5.069e+02, threshold=3.743e+02, percent-clipped=1.0 2023-04-27 10:20:03,626 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=95964.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:20:09,301 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=95972.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:20:09,893 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=95973.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:20:14,646 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=95980.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:20:17,648 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.7559, 3.6993, 2.7924, 4.4264, 3.7405, 3.8203, 1.6142, 3.8232], device='cuda:4'), covar=tensor([0.1701, 0.1234, 0.3220, 0.1416, 0.3100, 0.1743, 0.5893, 0.2109], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0213, 0.0249, 0.0304, 0.0297, 0.0249, 0.0272, 0.0272], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 10:20:22,488 INFO [finetune.py:976] (4/7) Epoch 17, batch 4350, loss[loss=0.1886, simple_loss=0.2546, pruned_loss=0.06131, over 4860.00 frames. ], tot_loss[loss=0.1754, simple_loss=0.2448, pruned_loss=0.05296, over 955251.76 frames. ], batch size: 31, lr: 3.38e-03, grad_scale: 32.0 2023-04-27 10:20:23,767 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=95995.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:20:25,062 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6411, 3.2417, 2.7951, 2.6783, 2.0902, 2.1014, 2.9271, 2.1212], device='cuda:4'), covar=tensor([0.1669, 0.1514, 0.1203, 0.1589, 0.2344, 0.1861, 0.0842, 0.1858], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0212, 0.0167, 0.0204, 0.0200, 0.0184, 0.0155, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 10:21:06,651 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6185, 2.4765, 1.9618, 2.2944, 2.4675, 2.1255, 3.0427, 1.8448], device='cuda:4'), covar=tensor([0.3556, 0.1863, 0.4022, 0.2792, 0.1789, 0.2253, 0.2057, 0.3866], device='cuda:4'), in_proj_covar=tensor([0.0334, 0.0342, 0.0421, 0.0351, 0.0375, 0.0376, 0.0365, 0.0414], device='cuda:4'), out_proj_covar=tensor([9.9748e-05, 1.0300e-04, 1.2820e-04, 1.0623e-04, 1.1218e-04, 1.1281e-04, 1.0753e-04, 1.2556e-04], device='cuda:4') 2023-04-27 10:21:07,899 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=96041.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:21:08,979 INFO [finetune.py:976] (4/7) Epoch 17, batch 4400, loss[loss=0.202, simple_loss=0.2817, pruned_loss=0.06118, over 4870.00 frames. ], tot_loss[loss=0.1756, simple_loss=0.2453, pruned_loss=0.0529, over 954832.66 frames. ], batch size: 34, lr: 3.38e-03, grad_scale: 32.0 2023-04-27 10:21:09,043 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=96043.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:21:15,035 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.713e+01 1.567e+02 1.808e+02 2.181e+02 3.991e+02, threshold=3.617e+02, percent-clipped=1.0 2023-04-27 10:21:31,662 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0717, 2.3902, 1.0056, 1.3426, 1.7368, 1.2424, 3.2760, 1.7178], device='cuda:4'), covar=tensor([0.0694, 0.0721, 0.0820, 0.1305, 0.0554, 0.1055, 0.0242, 0.0664], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0065, 0.0048, 0.0046, 0.0049, 0.0052, 0.0074, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 10:21:34,766 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=96080.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:21:48,840 INFO [finetune.py:976] (4/7) Epoch 17, batch 4450, loss[loss=0.158, simple_loss=0.2426, pruned_loss=0.03665, over 4909.00 frames. ], tot_loss[loss=0.1771, simple_loss=0.2476, pruned_loss=0.05328, over 952315.72 frames. ], batch size: 37, lr: 3.38e-03, grad_scale: 32.0 2023-04-27 10:21:49,056 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-04-27 10:21:50,806 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6859, 1.3839, 1.6091, 2.0675, 1.9811, 1.6121, 1.2788, 1.8433], device='cuda:4'), covar=tensor([0.0798, 0.1220, 0.0838, 0.0491, 0.0585, 0.0776, 0.0812, 0.0534], device='cuda:4'), in_proj_covar=tensor([0.0189, 0.0201, 0.0181, 0.0171, 0.0176, 0.0180, 0.0152, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 10:22:00,476 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3834, 1.3657, 1.7389, 1.7054, 1.3429, 1.0916, 1.4826, 0.9149], device='cuda:4'), covar=tensor([0.0598, 0.0652, 0.0380, 0.0565, 0.0746, 0.1124, 0.0610, 0.0627], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0070, 0.0068, 0.0068, 0.0075, 0.0096, 0.0075, 0.0067], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 10:22:47,374 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0585, 1.0474, 1.2083, 1.2280, 1.0432, 0.9340, 1.0345, 0.6783], device='cuda:4'), covar=tensor([0.0542, 0.0579, 0.0472, 0.0509, 0.0651, 0.1252, 0.0425, 0.0649], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0070, 0.0068, 0.0068, 0.0075, 0.0096, 0.0074, 0.0067], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 10:22:53,792 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=96141.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:22:54,920 INFO [finetune.py:976] (4/7) Epoch 17, batch 4500, loss[loss=0.1906, simple_loss=0.2653, pruned_loss=0.058, over 4826.00 frames. ], tot_loss[loss=0.1799, simple_loss=0.2505, pruned_loss=0.05462, over 952143.46 frames. ], batch size: 47, lr: 3.37e-03, grad_scale: 32.0 2023-04-27 10:23:02,750 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.142e+02 1.688e+02 1.990e+02 2.416e+02 4.917e+02, threshold=3.981e+02, percent-clipped=4.0 2023-04-27 10:23:16,207 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=96165.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:23:17,977 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=96168.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 10:23:58,142 INFO [finetune.py:976] (4/7) Epoch 17, batch 4550, loss[loss=0.1461, simple_loss=0.2245, pruned_loss=0.03385, over 4768.00 frames. ], tot_loss[loss=0.1811, simple_loss=0.252, pruned_loss=0.05506, over 953105.51 frames. ], batch size: 28, lr: 3.37e-03, grad_scale: 32.0 2023-04-27 10:24:10,267 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.77 vs. limit=2.0 2023-04-27 10:24:10,602 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=96203.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 10:24:18,970 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0204, 1.0160, 1.2309, 1.1784, 1.0062, 0.9400, 0.9636, 0.5313], device='cuda:4'), covar=tensor([0.0606, 0.0582, 0.0509, 0.0549, 0.0811, 0.1167, 0.0519, 0.0744], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0070, 0.0068, 0.0068, 0.0075, 0.0096, 0.0074, 0.0067], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 10:24:42,299 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=96226.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:24:52,742 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1723, 1.5721, 1.3608, 1.8107, 1.7172, 1.7882, 1.4453, 3.0690], device='cuda:4'), covar=tensor([0.0643, 0.0780, 0.0821, 0.1193, 0.0603, 0.0478, 0.0748, 0.0175], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0039, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0012, 0.0013, 0.0016], device='cuda:4') 2023-04-27 10:24:54,976 INFO [finetune.py:976] (4/7) Epoch 17, batch 4600, loss[loss=0.1725, simple_loss=0.2456, pruned_loss=0.0497, over 4864.00 frames. ], tot_loss[loss=0.1795, simple_loss=0.2509, pruned_loss=0.05398, over 953288.37 frames. ], batch size: 34, lr: 3.37e-03, grad_scale: 32.0 2023-04-27 10:24:57,508 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0571, 1.1690, 5.0026, 4.7468, 4.4483, 4.7820, 4.4174, 4.4646], device='cuda:4'), covar=tensor([0.6606, 0.6363, 0.0965, 0.1610, 0.1074, 0.1332, 0.1525, 0.1426], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0306, 0.0401, 0.0404, 0.0348, 0.0404, 0.0310, 0.0364], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 10:25:01,049 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.153e+02 1.538e+02 1.819e+02 2.307e+02 5.540e+02, threshold=3.639e+02, percent-clipped=2.0 2023-04-27 10:25:13,318 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=96273.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:25:27,834 INFO [finetune.py:976] (4/7) Epoch 17, batch 4650, loss[loss=0.1452, simple_loss=0.2087, pruned_loss=0.04082, over 4807.00 frames. ], tot_loss[loss=0.1772, simple_loss=0.2477, pruned_loss=0.0533, over 953456.91 frames. ], batch size: 25, lr: 3.37e-03, grad_scale: 32.0 2023-04-27 10:25:39,948 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2403, 1.2972, 1.4012, 1.5791, 1.6374, 1.2549, 0.8760, 1.4283], device='cuda:4'), covar=tensor([0.0949, 0.1272, 0.0909, 0.0654, 0.0717, 0.0872, 0.0951, 0.0695], device='cuda:4'), in_proj_covar=tensor([0.0190, 0.0201, 0.0182, 0.0172, 0.0177, 0.0181, 0.0152, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 10:25:45,394 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=96321.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:25:47,251 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=96324.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:25:56,371 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=96336.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:25:57,049 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5347, 1.8620, 1.8130, 2.2459, 2.1107, 2.1770, 1.7717, 4.4976], device='cuda:4'), covar=tensor([0.0511, 0.0750, 0.0758, 0.1105, 0.0560, 0.0445, 0.0694, 0.0094], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0040, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 10:26:01,178 INFO [finetune.py:976] (4/7) Epoch 17, batch 4700, loss[loss=0.1965, simple_loss=0.2485, pruned_loss=0.07224, over 4824.00 frames. ], tot_loss[loss=0.1745, simple_loss=0.2444, pruned_loss=0.05228, over 955318.13 frames. ], batch size: 30, lr: 3.37e-03, grad_scale: 32.0 2023-04-27 10:26:07,580 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.173e+02 1.665e+02 1.923e+02 2.344e+02 4.487e+02, threshold=3.847e+02, percent-clipped=4.0 2023-04-27 10:26:43,950 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=96385.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:26:48,534 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=96390.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:26:55,841 INFO [finetune.py:976] (4/7) Epoch 17, batch 4750, loss[loss=0.1619, simple_loss=0.2371, pruned_loss=0.04338, over 4912.00 frames. ], tot_loss[loss=0.1754, simple_loss=0.2444, pruned_loss=0.05318, over 954238.53 frames. ], batch size: 36, lr: 3.37e-03, grad_scale: 32.0 2023-04-27 10:27:29,998 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.46 vs. limit=5.0 2023-04-27 10:27:44,225 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1188, 2.7025, 1.0135, 1.4260, 2.1882, 1.3348, 3.5928, 2.0130], device='cuda:4'), covar=tensor([0.0657, 0.0591, 0.0783, 0.1255, 0.0493, 0.1000, 0.0274, 0.0580], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0065, 0.0048, 0.0046, 0.0050, 0.0052, 0.0074, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 10:27:45,320 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=96436.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:27:50,990 INFO [finetune.py:976] (4/7) Epoch 17, batch 4800, loss[loss=0.1733, simple_loss=0.2364, pruned_loss=0.0551, over 4699.00 frames. ], tot_loss[loss=0.1785, simple_loss=0.2475, pruned_loss=0.05481, over 953130.93 frames. ], batch size: 23, lr: 3.37e-03, grad_scale: 32.0 2023-04-27 10:27:56,397 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=96451.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:27:57,983 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.109e+02 1.553e+02 2.009e+02 2.369e+02 4.346e+02, threshold=4.018e+02, percent-clipped=1.0 2023-04-27 10:28:02,358 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5142, 1.7088, 1.4534, 1.0464, 1.1699, 1.1526, 1.4159, 1.1368], device='cuda:4'), covar=tensor([0.1850, 0.1326, 0.1517, 0.1898, 0.2414, 0.2034, 0.1104, 0.2076], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0211, 0.0167, 0.0203, 0.0200, 0.0183, 0.0155, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 10:28:07,171 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=96468.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 10:28:19,401 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-27 10:28:34,489 INFO [finetune.py:976] (4/7) Epoch 17, batch 4850, loss[loss=0.1397, simple_loss=0.2059, pruned_loss=0.03671, over 4239.00 frames. ], tot_loss[loss=0.1785, simple_loss=0.2481, pruned_loss=0.05446, over 950618.64 frames. ], batch size: 18, lr: 3.37e-03, grad_scale: 32.0 2023-04-27 10:28:41,609 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=96503.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:28:49,917 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=96516.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 10:28:52,925 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=96521.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:29:07,105 INFO [finetune.py:976] (4/7) Epoch 17, batch 4900, loss[loss=0.2142, simple_loss=0.2791, pruned_loss=0.07468, over 4746.00 frames. ], tot_loss[loss=0.1807, simple_loss=0.2509, pruned_loss=0.05528, over 951847.68 frames. ], batch size: 27, lr: 3.37e-03, grad_scale: 32.0 2023-04-27 10:29:13,478 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=96551.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:29:15,080 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.788e+01 1.547e+02 1.882e+02 2.260e+02 4.956e+02, threshold=3.763e+02, percent-clipped=3.0 2023-04-27 10:29:37,119 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-04-27 10:29:39,379 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4462, 2.2230, 2.6619, 3.0722, 2.8759, 2.4250, 2.0150, 2.6039], device='cuda:4'), covar=tensor([0.0871, 0.0951, 0.0550, 0.0565, 0.0603, 0.0817, 0.0770, 0.0541], device='cuda:4'), in_proj_covar=tensor([0.0189, 0.0201, 0.0181, 0.0171, 0.0176, 0.0180, 0.0152, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 10:29:39,869 INFO [finetune.py:976] (4/7) Epoch 17, batch 4950, loss[loss=0.1644, simple_loss=0.2563, pruned_loss=0.03626, over 4914.00 frames. ], tot_loss[loss=0.1806, simple_loss=0.2512, pruned_loss=0.05493, over 951181.69 frames. ], batch size: 36, lr: 3.37e-03, grad_scale: 32.0 2023-04-27 10:30:09,384 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=96636.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:30:13,517 INFO [finetune.py:976] (4/7) Epoch 17, batch 5000, loss[loss=0.1651, simple_loss=0.2318, pruned_loss=0.04919, over 4789.00 frames. ], tot_loss[loss=0.1791, simple_loss=0.2498, pruned_loss=0.0542, over 953090.91 frames. ], batch size: 25, lr: 3.37e-03, grad_scale: 32.0 2023-04-27 10:30:21,475 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.257e+01 1.582e+02 1.781e+02 2.065e+02 3.108e+02, threshold=3.561e+02, percent-clipped=0.0 2023-04-27 10:30:25,186 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.24 vs. limit=5.0 2023-04-27 10:30:38,904 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=96680.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:30:41,299 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=96684.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:30:46,661 INFO [finetune.py:976] (4/7) Epoch 17, batch 5050, loss[loss=0.1632, simple_loss=0.2344, pruned_loss=0.04603, over 4865.00 frames. ], tot_loss[loss=0.1775, simple_loss=0.2476, pruned_loss=0.05365, over 953762.47 frames. ], batch size: 49, lr: 3.37e-03, grad_scale: 32.0 2023-04-27 10:30:49,197 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9867, 2.3603, 0.8698, 1.2447, 1.6131, 1.2423, 2.5120, 1.4196], device='cuda:4'), covar=tensor([0.0648, 0.0510, 0.0631, 0.1199, 0.0431, 0.0950, 0.0329, 0.0675], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0065, 0.0048, 0.0046, 0.0050, 0.0052, 0.0074, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0007, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 10:30:59,552 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0662, 2.5553, 2.0904, 2.3814, 1.8508, 2.2877, 2.2229, 1.5742], device='cuda:4'), covar=tensor([0.2101, 0.1304, 0.0989, 0.1331, 0.3416, 0.1230, 0.1947, 0.2877], device='cuda:4'), in_proj_covar=tensor([0.0283, 0.0305, 0.0218, 0.0279, 0.0311, 0.0258, 0.0248, 0.0265], device='cuda:4'), out_proj_covar=tensor([1.1361e-04, 1.2119e-04, 8.6847e-05, 1.1046e-04, 1.2630e-04, 1.0244e-04, 1.0042e-04, 1.0506e-04], device='cuda:4') 2023-04-27 10:31:15,682 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=96736.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:31:19,857 INFO [finetune.py:976] (4/7) Epoch 17, batch 5100, loss[loss=0.1554, simple_loss=0.2222, pruned_loss=0.04429, over 4922.00 frames. ], tot_loss[loss=0.1753, simple_loss=0.2449, pruned_loss=0.05286, over 954273.10 frames. ], batch size: 33, lr: 3.37e-03, grad_scale: 32.0 2023-04-27 10:31:21,739 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=96746.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:31:25,901 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.147e+02 1.614e+02 1.893e+02 2.231e+02 4.752e+02, threshold=3.786e+02, percent-clipped=4.0 2023-04-27 10:31:41,640 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5779, 1.6395, 1.4683, 1.0298, 1.2146, 1.1710, 1.4662, 1.1513], device='cuda:4'), covar=tensor([0.1871, 0.1435, 0.1630, 0.1991, 0.2434, 0.2050, 0.1134, 0.2200], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0213, 0.0168, 0.0204, 0.0201, 0.0184, 0.0156, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 10:31:54,155 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=96776.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:32:04,478 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=96784.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:32:15,346 INFO [finetune.py:976] (4/7) Epoch 17, batch 5150, loss[loss=0.1571, simple_loss=0.2162, pruned_loss=0.04897, over 4780.00 frames. ], tot_loss[loss=0.1753, simple_loss=0.2447, pruned_loss=0.05292, over 954332.00 frames. ], batch size: 28, lr: 3.37e-03, grad_scale: 32.0 2023-04-27 10:32:46,402 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=96821.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:32:56,533 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6639, 1.2273, 1.3778, 1.3668, 1.8846, 1.5149, 1.2199, 1.2675], device='cuda:4'), covar=tensor([0.1622, 0.1313, 0.1676, 0.1333, 0.0738, 0.1471, 0.2007, 0.2105], device='cuda:4'), in_proj_covar=tensor([0.0306, 0.0311, 0.0350, 0.0286, 0.0326, 0.0308, 0.0298, 0.0365], device='cuda:4'), out_proj_covar=tensor([6.3278e-05, 6.4771e-05, 7.4620e-05, 5.8146e-05, 6.7798e-05, 6.4821e-05, 6.2870e-05, 7.7811e-05], device='cuda:4') 2023-04-27 10:33:03,944 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=96837.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:33:12,846 INFO [finetune.py:976] (4/7) Epoch 17, batch 5200, loss[loss=0.1784, simple_loss=0.2513, pruned_loss=0.0528, over 4863.00 frames. ], tot_loss[loss=0.1779, simple_loss=0.2476, pruned_loss=0.05404, over 951572.79 frames. ], batch size: 31, lr: 3.37e-03, grad_scale: 32.0 2023-04-27 10:33:24,747 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.802e+01 1.740e+02 2.121e+02 2.596e+02 7.000e+02, threshold=4.242e+02, percent-clipped=4.0 2023-04-27 10:33:26,130 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4256, 3.3256, 1.1086, 1.8898, 1.8798, 2.4432, 1.8336, 1.1432], device='cuda:4'), covar=tensor([0.1350, 0.1034, 0.1773, 0.1131, 0.1056, 0.0940, 0.1473, 0.1792], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0244, 0.0138, 0.0121, 0.0133, 0.0154, 0.0119, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 10:33:47,098 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=96869.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:34:02,258 INFO [finetune.py:976] (4/7) Epoch 17, batch 5250, loss[loss=0.225, simple_loss=0.2888, pruned_loss=0.08055, over 4816.00 frames. ], tot_loss[loss=0.18, simple_loss=0.2507, pruned_loss=0.05463, over 952735.17 frames. ], batch size: 33, lr: 3.37e-03, grad_scale: 32.0 2023-04-27 10:34:16,395 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4897, 1.8000, 1.7489, 2.1882, 2.0234, 1.9224, 1.5970, 4.4805], device='cuda:4'), covar=tensor([0.0596, 0.0858, 0.0893, 0.1205, 0.0662, 0.0560, 0.0839, 0.0150], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0039, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 10:34:20,021 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2369, 1.2314, 3.7685, 3.4589, 3.3415, 3.6184, 3.6132, 3.2994], device='cuda:4'), covar=tensor([0.7507, 0.5977, 0.1293, 0.2165, 0.1247, 0.1763, 0.1775, 0.1732], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0306, 0.0403, 0.0406, 0.0350, 0.0406, 0.0312, 0.0367], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 10:34:35,559 INFO [finetune.py:976] (4/7) Epoch 17, batch 5300, loss[loss=0.1528, simple_loss=0.2151, pruned_loss=0.04526, over 4068.00 frames. ], tot_loss[loss=0.1814, simple_loss=0.2527, pruned_loss=0.05503, over 953184.85 frames. ], batch size: 17, lr: 3.37e-03, grad_scale: 32.0 2023-04-27 10:34:39,386 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9143, 1.3488, 1.7016, 1.6025, 1.6450, 1.3296, 0.7444, 1.3725], device='cuda:4'), covar=tensor([0.3251, 0.3159, 0.1753, 0.2335, 0.2470, 0.2666, 0.4418, 0.2008], device='cuda:4'), in_proj_covar=tensor([0.0291, 0.0247, 0.0227, 0.0316, 0.0218, 0.0231, 0.0228, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 10:34:41,677 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.055e+02 1.524e+02 1.845e+02 2.298e+02 4.151e+02, threshold=3.690e+02, percent-clipped=0.0 2023-04-27 10:34:43,098 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.68 vs. limit=2.0 2023-04-27 10:35:00,614 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=96980.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:35:08,783 INFO [finetune.py:976] (4/7) Epoch 17, batch 5350, loss[loss=0.1506, simple_loss=0.2216, pruned_loss=0.03976, over 4723.00 frames. ], tot_loss[loss=0.18, simple_loss=0.2516, pruned_loss=0.05422, over 952633.53 frames. ], batch size: 23, lr: 3.37e-03, grad_scale: 32.0 2023-04-27 10:35:13,334 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=97000.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:35:15,221 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-27 10:35:33,058 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=97028.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:35:42,644 INFO [finetune.py:976] (4/7) Epoch 17, batch 5400, loss[loss=0.1276, simple_loss=0.202, pruned_loss=0.02662, over 4795.00 frames. ], tot_loss[loss=0.1771, simple_loss=0.2482, pruned_loss=0.05302, over 953237.96 frames. ], batch size: 29, lr: 3.37e-03, grad_scale: 32.0 2023-04-27 10:35:44,603 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=97046.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:35:48,747 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.116e+02 1.600e+02 1.805e+02 2.105e+02 4.244e+02, threshold=3.609e+02, percent-clipped=1.0 2023-04-27 10:35:53,820 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=97061.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:36:03,177 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=5.70 vs. limit=5.0 2023-04-27 10:36:14,280 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2762, 1.6544, 1.4849, 1.8896, 1.8282, 1.9600, 1.5207, 3.5514], device='cuda:4'), covar=tensor([0.0589, 0.0748, 0.0816, 0.1082, 0.0600, 0.0454, 0.0713, 0.0141], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0039, 0.0037, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0012, 0.0013, 0.0016], device='cuda:4') 2023-04-27 10:36:15,420 INFO [finetune.py:976] (4/7) Epoch 17, batch 5450, loss[loss=0.1253, simple_loss=0.1995, pruned_loss=0.02555, over 4828.00 frames. ], tot_loss[loss=0.1748, simple_loss=0.2454, pruned_loss=0.05214, over 955054.59 frames. ], batch size: 25, lr: 3.37e-03, grad_scale: 32.0 2023-04-27 10:36:16,113 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=97094.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:36:21,101 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-27 10:36:38,896 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-27 10:36:40,623 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=97132.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:36:47,757 INFO [finetune.py:976] (4/7) Epoch 17, batch 5500, loss[loss=0.15, simple_loss=0.2151, pruned_loss=0.04244, over 4779.00 frames. ], tot_loss[loss=0.1726, simple_loss=0.2424, pruned_loss=0.05147, over 955638.20 frames. ], batch size: 26, lr: 3.37e-03, grad_scale: 32.0 2023-04-27 10:36:54,340 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.109e+02 1.627e+02 1.973e+02 2.284e+02 4.082e+02, threshold=3.945e+02, percent-clipped=1.0 2023-04-27 10:37:37,877 INFO [finetune.py:976] (4/7) Epoch 17, batch 5550, loss[loss=0.1641, simple_loss=0.2376, pruned_loss=0.04528, over 4800.00 frames. ], tot_loss[loss=0.1752, simple_loss=0.2446, pruned_loss=0.05294, over 956170.58 frames. ], batch size: 29, lr: 3.37e-03, grad_scale: 32.0 2023-04-27 10:38:15,509 INFO [finetune.py:976] (4/7) Epoch 17, batch 5600, loss[loss=0.1796, simple_loss=0.2674, pruned_loss=0.04592, over 4746.00 frames. ], tot_loss[loss=0.1776, simple_loss=0.2476, pruned_loss=0.05374, over 954981.04 frames. ], batch size: 54, lr: 3.37e-03, grad_scale: 32.0 2023-04-27 10:38:26,564 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.133e+02 1.758e+02 2.119e+02 2.585e+02 7.806e+02, threshold=4.239e+02, percent-clipped=5.0 2023-04-27 10:38:29,148 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.75 vs. limit=5.0 2023-04-27 10:38:37,076 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=97260.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:39:14,493 INFO [finetune.py:976] (4/7) Epoch 17, batch 5650, loss[loss=0.1676, simple_loss=0.2541, pruned_loss=0.04054, over 4819.00 frames. ], tot_loss[loss=0.1803, simple_loss=0.2517, pruned_loss=0.0545, over 954593.89 frames. ], batch size: 40, lr: 3.37e-03, grad_scale: 32.0 2023-04-27 10:39:34,148 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6357, 2.1325, 1.7001, 1.5259, 1.2212, 1.2321, 1.7856, 1.2111], device='cuda:4'), covar=tensor([0.1714, 0.1417, 0.1580, 0.1894, 0.2377, 0.1970, 0.0986, 0.2082], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0213, 0.0168, 0.0205, 0.0201, 0.0184, 0.0155, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 10:39:37,091 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2233, 2.9003, 2.4774, 2.6649, 2.1175, 2.5306, 2.4451, 1.8729], device='cuda:4'), covar=tensor([0.1659, 0.1150, 0.0717, 0.1072, 0.2820, 0.0981, 0.1651, 0.2395], device='cuda:4'), in_proj_covar=tensor([0.0282, 0.0303, 0.0217, 0.0277, 0.0309, 0.0256, 0.0247, 0.0264], device='cuda:4'), out_proj_covar=tensor([1.1310e-04, 1.2059e-04, 8.6223e-05, 1.0982e-04, 1.2540e-04, 1.0186e-04, 9.9894e-05, 1.0480e-04], device='cuda:4') 2023-04-27 10:39:47,542 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=97321.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:40:04,837 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-27 10:40:16,419 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=97342.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:40:16,914 INFO [finetune.py:976] (4/7) Epoch 17, batch 5700, loss[loss=0.1721, simple_loss=0.2314, pruned_loss=0.05642, over 4064.00 frames. ], tot_loss[loss=0.178, simple_loss=0.2477, pruned_loss=0.05413, over 932555.38 frames. ], batch size: 17, lr: 3.36e-03, grad_scale: 32.0 2023-04-27 10:40:19,359 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=97347.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:40:22,862 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.291e+01 1.370e+02 1.630e+02 1.988e+02 3.241e+02, threshold=3.261e+02, percent-clipped=0.0 2023-04-27 10:40:24,680 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=97356.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:40:26,464 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.1661, 3.0017, 2.4457, 3.6115, 3.1181, 3.1471, 1.4299, 3.0598], device='cuda:4'), covar=tensor([0.1941, 0.1362, 0.2767, 0.1960, 0.2862, 0.1841, 0.5628, 0.2419], device='cuda:4'), in_proj_covar=tensor([0.0242, 0.0212, 0.0248, 0.0304, 0.0297, 0.0249, 0.0273, 0.0270], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 10:40:46,576 INFO [finetune.py:976] (4/7) Epoch 18, batch 0, loss[loss=0.2102, simple_loss=0.2803, pruned_loss=0.07005, over 4813.00 frames. ], tot_loss[loss=0.2102, simple_loss=0.2803, pruned_loss=0.07005, over 4813.00 frames. ], batch size: 39, lr: 3.36e-03, grad_scale: 64.0 2023-04-27 10:40:46,576 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 10:41:03,167 INFO [finetune.py:1010] (4/7) Epoch 18, validation: loss=0.1537, simple_loss=0.225, pruned_loss=0.04121, over 2265189.00 frames. 2023-04-27 10:41:03,167 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-27 10:41:05,488 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=97374.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:41:28,961 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=97403.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 10:41:32,015 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=97408.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:41:40,832 INFO [finetune.py:976] (4/7) Epoch 18, batch 50, loss[loss=0.1814, simple_loss=0.2433, pruned_loss=0.05979, over 4823.00 frames. ], tot_loss[loss=0.1827, simple_loss=0.2536, pruned_loss=0.05588, over 216732.90 frames. ], batch size: 33, lr: 3.36e-03, grad_scale: 64.0 2023-04-27 10:41:49,628 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=97432.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:41:51,457 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=97435.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:42:02,223 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.680e+01 1.493e+02 1.785e+02 2.115e+02 3.636e+02, threshold=3.569e+02, percent-clipped=3.0 2023-04-27 10:42:14,117 INFO [finetune.py:976] (4/7) Epoch 18, batch 100, loss[loss=0.1384, simple_loss=0.2188, pruned_loss=0.02901, over 4794.00 frames. ], tot_loss[loss=0.1731, simple_loss=0.2435, pruned_loss=0.05132, over 381435.46 frames. ], batch size: 29, lr: 3.36e-03, grad_scale: 64.0 2023-04-27 10:42:22,015 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=97480.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:42:23,956 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=5.12 vs. limit=5.0 2023-04-27 10:42:25,737 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.69 vs. limit=5.0 2023-04-27 10:42:30,765 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-27 10:42:47,723 INFO [finetune.py:976] (4/7) Epoch 18, batch 150, loss[loss=0.156, simple_loss=0.2258, pruned_loss=0.0431, over 4759.00 frames. ], tot_loss[loss=0.1727, simple_loss=0.2414, pruned_loss=0.05201, over 509547.47 frames. ], batch size: 27, lr: 3.36e-03, grad_scale: 64.0 2023-04-27 10:43:03,121 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.4850, 4.3835, 3.0922, 5.1709, 4.5614, 4.4289, 1.8921, 4.4185], device='cuda:4'), covar=tensor([0.1749, 0.1072, 0.3349, 0.0969, 0.3771, 0.1655, 0.6202, 0.2239], device='cuda:4'), in_proj_covar=tensor([0.0243, 0.0212, 0.0247, 0.0303, 0.0297, 0.0249, 0.0272, 0.0271], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 10:43:08,513 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.039e+02 1.668e+02 1.928e+02 2.305e+02 4.422e+02, threshold=3.856e+02, percent-clipped=2.0 2023-04-27 10:43:09,853 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=97555.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:43:19,855 INFO [finetune.py:976] (4/7) Epoch 18, batch 200, loss[loss=0.1545, simple_loss=0.2269, pruned_loss=0.0411, over 4852.00 frames. ], tot_loss[loss=0.1726, simple_loss=0.2409, pruned_loss=0.05217, over 610116.20 frames. ], batch size: 49, lr: 3.36e-03, grad_scale: 64.0 2023-04-27 10:43:32,238 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5088, 2.9637, 0.9858, 1.8209, 2.3757, 1.5486, 4.2974, 2.2985], device='cuda:4'), covar=tensor([0.0649, 0.0846, 0.0985, 0.1283, 0.0536, 0.1041, 0.0321, 0.0612], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0049, 0.0047, 0.0050, 0.0052, 0.0074, 0.0052], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 10:43:55,466 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=97616.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:43:55,522 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=97616.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:43:58,891 INFO [finetune.py:976] (4/7) Epoch 18, batch 250, loss[loss=0.1571, simple_loss=0.2271, pruned_loss=0.04351, over 4935.00 frames. ], tot_loss[loss=0.1745, simple_loss=0.2437, pruned_loss=0.05266, over 688327.47 frames. ], batch size: 33, lr: 3.36e-03, grad_scale: 64.0 2023-04-27 10:44:27,934 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6537, 3.7185, 1.1866, 1.9461, 1.9506, 2.6434, 1.9248, 0.9979], device='cuda:4'), covar=tensor([0.1381, 0.0910, 0.1883, 0.1215, 0.1135, 0.0985, 0.1722, 0.2027], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0243, 0.0137, 0.0121, 0.0133, 0.0153, 0.0118, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 10:44:29,796 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0767, 2.5624, 1.0048, 1.4649, 2.0838, 1.2766, 3.4310, 1.8207], device='cuda:4'), covar=tensor([0.0656, 0.0634, 0.0791, 0.1254, 0.0466, 0.1015, 0.0362, 0.0613], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0049, 0.0046, 0.0050, 0.0052, 0.0074, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 10:44:42,422 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.134e+02 1.696e+02 2.035e+02 2.379e+02 5.416e+02, threshold=4.070e+02, percent-clipped=1.0 2023-04-27 10:44:49,678 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=97656.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:45:03,745 INFO [finetune.py:976] (4/7) Epoch 18, batch 300, loss[loss=0.1906, simple_loss=0.281, pruned_loss=0.05013, over 4812.00 frames. ], tot_loss[loss=0.1773, simple_loss=0.2478, pruned_loss=0.05336, over 749908.40 frames. ], batch size: 45, lr: 3.36e-03, grad_scale: 64.0 2023-04-27 10:45:09,979 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=97674.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 10:45:43,987 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=97698.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 10:45:47,055 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=97703.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:45:47,657 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=97704.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:46:07,662 INFO [finetune.py:976] (4/7) Epoch 18, batch 350, loss[loss=0.1666, simple_loss=0.245, pruned_loss=0.04407, over 4318.00 frames. ], tot_loss[loss=0.1781, simple_loss=0.2491, pruned_loss=0.0535, over 796483.01 frames. ], batch size: 65, lr: 3.36e-03, grad_scale: 64.0 2023-04-27 10:46:18,796 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=97730.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:46:28,150 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=97735.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 10:46:51,635 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.284e+02 1.687e+02 1.979e+02 2.374e+02 3.417e+02, threshold=3.957e+02, percent-clipped=0.0 2023-04-27 10:47:08,343 INFO [finetune.py:976] (4/7) Epoch 18, batch 400, loss[loss=0.1756, simple_loss=0.257, pruned_loss=0.04709, over 4912.00 frames. ], tot_loss[loss=0.1777, simple_loss=0.2492, pruned_loss=0.05311, over 829192.43 frames. ], batch size: 46, lr: 3.36e-03, grad_scale: 64.0 2023-04-27 10:47:17,150 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.69 vs. limit=2.0 2023-04-27 10:47:22,210 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=5.37 vs. limit=5.0 2023-04-27 10:47:32,322 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.9350, 3.8616, 2.7633, 4.5578, 4.0268, 3.9009, 1.5458, 4.0020], device='cuda:4'), covar=tensor([0.1739, 0.1067, 0.3035, 0.1259, 0.2724, 0.1713, 0.5791, 0.2032], device='cuda:4'), in_proj_covar=tensor([0.0236, 0.0207, 0.0242, 0.0296, 0.0290, 0.0243, 0.0265, 0.0264], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 10:47:42,004 INFO [finetune.py:976] (4/7) Epoch 18, batch 450, loss[loss=0.2407, simple_loss=0.2967, pruned_loss=0.09236, over 4797.00 frames. ], tot_loss[loss=0.178, simple_loss=0.249, pruned_loss=0.05352, over 858247.29 frames. ], batch size: 29, lr: 3.36e-03, grad_scale: 32.0 2023-04-27 10:47:42,720 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5349, 2.9550, 0.8413, 1.6808, 2.1384, 1.4494, 3.9955, 2.1818], device='cuda:4'), covar=tensor([0.0586, 0.0783, 0.0922, 0.1220, 0.0492, 0.0959, 0.0188, 0.0524], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0049, 0.0047, 0.0050, 0.0052, 0.0074, 0.0052], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 10:48:05,039 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.073e+02 1.551e+02 1.835e+02 2.204e+02 3.781e+02, threshold=3.670e+02, percent-clipped=0.0 2023-04-27 10:48:15,397 INFO [finetune.py:976] (4/7) Epoch 18, batch 500, loss[loss=0.1568, simple_loss=0.2313, pruned_loss=0.04115, over 4809.00 frames. ], tot_loss[loss=0.1753, simple_loss=0.2459, pruned_loss=0.05236, over 880356.01 frames. ], batch size: 51, lr: 3.36e-03, grad_scale: 32.0 2023-04-27 10:48:26,741 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=97888.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:48:42,857 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=97911.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:48:45,962 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=97916.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:48:48,923 INFO [finetune.py:976] (4/7) Epoch 18, batch 550, loss[loss=0.1688, simple_loss=0.2425, pruned_loss=0.04753, over 4855.00 frames. ], tot_loss[loss=0.1737, simple_loss=0.2438, pruned_loss=0.05177, over 896642.86 frames. ], batch size: 44, lr: 3.36e-03, grad_scale: 32.0 2023-04-27 10:48:59,978 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3642, 1.6492, 1.4280, 1.5816, 1.4511, 1.3817, 1.4709, 1.1528], device='cuda:4'), covar=tensor([0.1599, 0.1302, 0.0956, 0.1279, 0.3381, 0.1186, 0.1636, 0.2175], device='cuda:4'), in_proj_covar=tensor([0.0286, 0.0306, 0.0220, 0.0280, 0.0313, 0.0259, 0.0251, 0.0265], device='cuda:4'), out_proj_covar=tensor([1.1482e-04, 1.2144e-04, 8.7295e-05, 1.1117e-04, 1.2719e-04, 1.0311e-04, 1.0138e-04, 1.0540e-04], device='cuda:4') 2023-04-27 10:49:09,115 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4389, 1.0492, 0.5252, 1.1448, 1.1011, 1.3486, 1.2056, 1.2170], device='cuda:4'), covar=tensor([0.0475, 0.0382, 0.0391, 0.0539, 0.0301, 0.0472, 0.0457, 0.0531], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0029, 0.0020, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0045, 0.0038, 0.0051, 0.0038, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 10:49:09,141 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=97949.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:49:12,036 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.221e+01 1.571e+02 1.914e+02 2.299e+02 4.493e+02, threshold=3.828e+02, percent-clipped=2.0 2023-04-27 10:49:18,169 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=97964.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:49:22,483 INFO [finetune.py:976] (4/7) Epoch 18, batch 600, loss[loss=0.2409, simple_loss=0.308, pruned_loss=0.08696, over 4795.00 frames. ], tot_loss[loss=0.1763, simple_loss=0.246, pruned_loss=0.05327, over 911120.86 frames. ], batch size: 45, lr: 3.36e-03, grad_scale: 32.0 2023-04-27 10:49:32,837 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.7002, 3.0331, 0.8570, 1.6326, 2.2266, 1.5876, 4.2623, 2.1306], device='cuda:4'), covar=tensor([0.0615, 0.0909, 0.0929, 0.1452, 0.0561, 0.1061, 0.0420, 0.0654], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0049, 0.0047, 0.0050, 0.0052, 0.0074, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 10:49:40,416 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8066, 2.1631, 0.8628, 1.2399, 1.4664, 1.1814, 2.4804, 1.4594], device='cuda:4'), covar=tensor([0.0737, 0.0625, 0.0659, 0.1272, 0.0488, 0.1055, 0.0340, 0.0650], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0049, 0.0047, 0.0050, 0.0052, 0.0074, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 10:49:41,024 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=97998.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:49:46,899 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=98003.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:49:47,895 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.80 vs. limit=2.0 2023-04-27 10:49:55,558 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6936, 1.2815, 1.8479, 2.1711, 1.7836, 1.7105, 1.7229, 1.7495], device='cuda:4'), covar=tensor([0.4968, 0.7008, 0.6725, 0.6458, 0.6018, 0.8430, 0.8929, 0.9122], device='cuda:4'), in_proj_covar=tensor([0.0422, 0.0407, 0.0498, 0.0504, 0.0451, 0.0478, 0.0484, 0.0486], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 10:49:57,856 INFO [finetune.py:976] (4/7) Epoch 18, batch 650, loss[loss=0.1865, simple_loss=0.2655, pruned_loss=0.05372, over 4852.00 frames. ], tot_loss[loss=0.1785, simple_loss=0.2489, pruned_loss=0.05405, over 922494.23 frames. ], batch size: 44, lr: 3.36e-03, grad_scale: 32.0 2023-04-27 10:50:03,547 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=98030.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 10:50:03,561 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=98030.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:50:14,379 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=98046.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:50:18,390 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=98051.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:50:24,921 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8544, 2.2910, 1.2035, 1.5651, 2.2765, 1.7689, 1.5782, 1.8050], device='cuda:4'), covar=tensor([0.0520, 0.0354, 0.0309, 0.0564, 0.0248, 0.0515, 0.0511, 0.0570], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0029, 0.0020, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0045, 0.0038, 0.0051, 0.0038, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 10:50:26,536 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.176e+02 1.729e+02 2.021e+02 2.442e+02 4.739e+02, threshold=4.043e+02, percent-clipped=3.0 2023-04-27 10:50:48,619 INFO [finetune.py:976] (4/7) Epoch 18, batch 700, loss[loss=0.1354, simple_loss=0.2197, pruned_loss=0.02551, over 4823.00 frames. ], tot_loss[loss=0.1788, simple_loss=0.2496, pruned_loss=0.05395, over 932055.04 frames. ], batch size: 38, lr: 3.36e-03, grad_scale: 32.0 2023-04-27 10:50:58,807 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=98078.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:51:33,955 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=98106.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:51:45,522 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.3802, 3.3616, 2.4504, 3.9380, 3.4546, 3.3872, 1.4571, 3.2695], device='cuda:4'), covar=tensor([0.2089, 0.1406, 0.3082, 0.2201, 0.3186, 0.2179, 0.5814, 0.2990], device='cuda:4'), in_proj_covar=tensor([0.0239, 0.0210, 0.0245, 0.0301, 0.0294, 0.0246, 0.0268, 0.0267], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 10:51:53,724 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8953, 1.3317, 1.4618, 1.5351, 2.0300, 1.6245, 1.3254, 1.3589], device='cuda:4'), covar=tensor([0.1561, 0.1535, 0.1837, 0.1379, 0.0822, 0.1538, 0.1999, 0.2038], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0312, 0.0351, 0.0287, 0.0329, 0.0310, 0.0300, 0.0366], device='cuda:4'), out_proj_covar=tensor([6.3701e-05, 6.5053e-05, 7.4740e-05, 5.8196e-05, 6.8415e-05, 6.5210e-05, 6.3188e-05, 7.8106e-05], device='cuda:4') 2023-04-27 10:51:54,812 INFO [finetune.py:976] (4/7) Epoch 18, batch 750, loss[loss=0.1641, simple_loss=0.2399, pruned_loss=0.04416, over 4805.00 frames. ], tot_loss[loss=0.1788, simple_loss=0.25, pruned_loss=0.05377, over 937073.26 frames. ], batch size: 40, lr: 3.36e-03, grad_scale: 32.0 2023-04-27 10:52:31,991 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.015e+02 1.562e+02 1.753e+02 1.961e+02 3.537e+02, threshold=3.506e+02, percent-clipped=0.0 2023-04-27 10:52:42,179 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=98167.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:52:44,509 INFO [finetune.py:976] (4/7) Epoch 18, batch 800, loss[loss=0.1862, simple_loss=0.2522, pruned_loss=0.06013, over 4817.00 frames. ], tot_loss[loss=0.1781, simple_loss=0.2492, pruned_loss=0.05352, over 939001.37 frames. ], batch size: 33, lr: 3.36e-03, grad_scale: 32.0 2023-04-27 10:53:10,922 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=98211.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:53:17,964 INFO [finetune.py:976] (4/7) Epoch 18, batch 850, loss[loss=0.1624, simple_loss=0.2226, pruned_loss=0.05111, over 4915.00 frames. ], tot_loss[loss=0.1773, simple_loss=0.2477, pruned_loss=0.05344, over 942615.42 frames. ], batch size: 37, lr: 3.35e-03, grad_scale: 32.0 2023-04-27 10:53:32,039 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=98244.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:53:38,455 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.063e+02 1.514e+02 1.837e+02 2.117e+02 3.312e+02, threshold=3.674e+02, percent-clipped=0.0 2023-04-27 10:53:42,074 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=98259.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:53:46,225 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8195, 2.2546, 1.7878, 1.6637, 1.3742, 1.3899, 1.8949, 1.3265], device='cuda:4'), covar=tensor([0.1462, 0.1250, 0.1427, 0.1570, 0.1967, 0.1677, 0.0865, 0.1867], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0214, 0.0169, 0.0206, 0.0201, 0.0185, 0.0157, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 10:53:51,235 INFO [finetune.py:976] (4/7) Epoch 18, batch 900, loss[loss=0.1699, simple_loss=0.2245, pruned_loss=0.05762, over 4718.00 frames. ], tot_loss[loss=0.1756, simple_loss=0.2455, pruned_loss=0.05282, over 944880.38 frames. ], batch size: 23, lr: 3.35e-03, grad_scale: 32.0 2023-04-27 10:54:24,161 INFO [finetune.py:976] (4/7) Epoch 18, batch 950, loss[loss=0.1341, simple_loss=0.2094, pruned_loss=0.02935, over 4765.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.2437, pruned_loss=0.05205, over 947616.55 frames. ], batch size: 28, lr: 3.35e-03, grad_scale: 32.0 2023-04-27 10:54:30,300 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=98330.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 10:54:44,943 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.018e+02 1.562e+02 1.785e+02 2.020e+02 3.278e+02, threshold=3.570e+02, percent-clipped=0.0 2023-04-27 10:54:57,901 INFO [finetune.py:976] (4/7) Epoch 18, batch 1000, loss[loss=0.1306, simple_loss=0.203, pruned_loss=0.02912, over 4058.00 frames. ], tot_loss[loss=0.1742, simple_loss=0.245, pruned_loss=0.05175, over 950598.01 frames. ], batch size: 17, lr: 3.35e-03, grad_scale: 32.0 2023-04-27 10:55:02,792 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=98378.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 10:55:22,441 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=98410.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:55:25,921 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0874, 1.4367, 1.3664, 1.6640, 1.5726, 1.7275, 1.3303, 2.9467], device='cuda:4'), covar=tensor([0.0638, 0.0774, 0.0783, 0.1161, 0.0642, 0.0561, 0.0745, 0.0191], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 10:55:29,959 INFO [finetune.py:976] (4/7) Epoch 18, batch 1050, loss[loss=0.1499, simple_loss=0.2196, pruned_loss=0.04013, over 4775.00 frames. ], tot_loss[loss=0.1762, simple_loss=0.2475, pruned_loss=0.05245, over 949185.61 frames. ], batch size: 26, lr: 3.35e-03, grad_scale: 32.0 2023-04-27 10:55:40,009 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1463, 1.6043, 2.1105, 2.5592, 2.0368, 1.5803, 1.2415, 1.8814], device='cuda:4'), covar=tensor([0.3068, 0.3142, 0.1576, 0.2145, 0.2524, 0.2541, 0.4365, 0.2040], device='cuda:4'), in_proj_covar=tensor([0.0289, 0.0245, 0.0225, 0.0313, 0.0217, 0.0230, 0.0227, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 10:55:52,032 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.610e+02 1.888e+02 2.252e+02 4.818e+02, threshold=3.776e+02, percent-clipped=3.0 2023-04-27 10:55:56,970 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=98462.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:56:08,982 INFO [finetune.py:976] (4/7) Epoch 18, batch 1100, loss[loss=0.2091, simple_loss=0.2863, pruned_loss=0.06594, over 4923.00 frames. ], tot_loss[loss=0.1766, simple_loss=0.2482, pruned_loss=0.05251, over 951038.22 frames. ], batch size: 42, lr: 3.35e-03, grad_scale: 32.0 2023-04-27 10:56:09,116 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=98471.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:56:43,241 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3887, 1.2641, 1.6361, 1.6184, 1.2940, 1.1854, 1.3816, 0.8758], device='cuda:4'), covar=tensor([0.0592, 0.0613, 0.0407, 0.0477, 0.0767, 0.1118, 0.0508, 0.0617], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0069, 0.0068, 0.0067, 0.0075, 0.0095, 0.0074, 0.0067], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 10:56:54,545 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7968, 1.3592, 1.8181, 2.3919, 1.9000, 1.7381, 1.7966, 1.7644], device='cuda:4'), covar=tensor([0.4440, 0.6766, 0.6628, 0.5318, 0.5839, 0.7971, 0.7837, 0.9586], device='cuda:4'), in_proj_covar=tensor([0.0417, 0.0404, 0.0493, 0.0499, 0.0447, 0.0474, 0.0480, 0.0484], device='cuda:4'), out_proj_covar=tensor([1.0042e-04, 9.9326e-05, 1.1085e-04, 1.1927e-04, 1.0716e-04, 1.1388e-04, 1.1396e-04, 1.1425e-04], device='cuda:4') 2023-04-27 10:57:13,272 INFO [finetune.py:976] (4/7) Epoch 18, batch 1150, loss[loss=0.1759, simple_loss=0.2563, pruned_loss=0.04774, over 4861.00 frames. ], tot_loss[loss=0.178, simple_loss=0.2495, pruned_loss=0.05322, over 950643.13 frames. ], batch size: 34, lr: 3.35e-03, grad_scale: 32.0 2023-04-27 10:57:45,478 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=98543.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:57:46,072 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=98544.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:57:57,931 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.919e+01 1.656e+02 1.874e+02 2.209e+02 6.337e+02, threshold=3.749e+02, percent-clipped=3.0 2023-04-27 10:58:14,178 INFO [finetune.py:976] (4/7) Epoch 18, batch 1200, loss[loss=0.1604, simple_loss=0.2308, pruned_loss=0.04502, over 4797.00 frames. ], tot_loss[loss=0.1775, simple_loss=0.2488, pruned_loss=0.05313, over 950786.38 frames. ], batch size: 29, lr: 3.35e-03, grad_scale: 32.0 2023-04-27 10:58:15,587 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2671, 1.4281, 1.8556, 1.9571, 1.8549, 1.9701, 1.8842, 1.8865], device='cuda:4'), covar=tensor([0.3520, 0.5092, 0.4297, 0.4272, 0.5209, 0.6764, 0.4436, 0.4397], device='cuda:4'), in_proj_covar=tensor([0.0331, 0.0372, 0.0319, 0.0333, 0.0344, 0.0395, 0.0355, 0.0326], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 10:58:28,914 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=98592.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:58:36,400 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=98604.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 10:58:46,673 INFO [finetune.py:976] (4/7) Epoch 18, batch 1250, loss[loss=0.1579, simple_loss=0.2143, pruned_loss=0.05075, over 4731.00 frames. ], tot_loss[loss=0.1758, simple_loss=0.2464, pruned_loss=0.05254, over 951340.16 frames. ], batch size: 59, lr: 3.35e-03, grad_scale: 32.0 2023-04-27 10:59:10,194 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.095e+02 1.558e+02 1.895e+02 2.216e+02 4.222e+02, threshold=3.789e+02, percent-clipped=2.0 2023-04-27 10:59:20,472 INFO [finetune.py:976] (4/7) Epoch 18, batch 1300, loss[loss=0.1793, simple_loss=0.2478, pruned_loss=0.05538, over 4916.00 frames. ], tot_loss[loss=0.1734, simple_loss=0.2439, pruned_loss=0.05144, over 953751.36 frames. ], batch size: 37, lr: 3.35e-03, grad_scale: 32.0 2023-04-27 10:59:38,652 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.7210, 1.7089, 1.7614, 1.3243, 1.8336, 1.5297, 2.3376, 1.5384], device='cuda:4'), covar=tensor([0.3526, 0.1723, 0.4461, 0.2981, 0.1645, 0.2226, 0.1420, 0.4494], device='cuda:4'), in_proj_covar=tensor([0.0335, 0.0342, 0.0422, 0.0351, 0.0377, 0.0376, 0.0367, 0.0414], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 10:59:50,730 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0462, 1.7504, 1.8950, 2.2766, 2.2689, 1.8940, 1.6301, 2.1451], device='cuda:4'), covar=tensor([0.0756, 0.0998, 0.0699, 0.0540, 0.0586, 0.0797, 0.0730, 0.0531], device='cuda:4'), in_proj_covar=tensor([0.0186, 0.0198, 0.0179, 0.0170, 0.0175, 0.0179, 0.0149, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 10:59:53,702 INFO [finetune.py:976] (4/7) Epoch 18, batch 1350, loss[loss=0.2147, simple_loss=0.2822, pruned_loss=0.07359, over 4913.00 frames. ], tot_loss[loss=0.1738, simple_loss=0.2441, pruned_loss=0.05177, over 953416.48 frames. ], batch size: 37, lr: 3.35e-03, grad_scale: 32.0 2023-04-27 10:59:58,516 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5669, 1.3916, 0.5538, 1.2296, 1.3939, 1.4245, 1.2835, 1.3563], device='cuda:4'), covar=tensor([0.0520, 0.0415, 0.0405, 0.0606, 0.0297, 0.0546, 0.0503, 0.0621], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0029, 0.0020, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0045, 0.0038, 0.0051, 0.0038, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 11:00:17,035 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.199e+02 1.614e+02 2.005e+02 2.367e+02 4.015e+02, threshold=4.011e+02, percent-clipped=1.0 2023-04-27 11:00:22,050 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=98762.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:00:24,451 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=98766.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:00:24,480 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2058, 1.4087, 1.2772, 1.6545, 1.5210, 1.6578, 1.3037, 3.0386], device='cuda:4'), covar=tensor([0.0613, 0.0865, 0.0852, 0.1235, 0.0673, 0.0524, 0.0842, 0.0161], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0040, 0.0044, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 11:00:27,464 INFO [finetune.py:976] (4/7) Epoch 18, batch 1400, loss[loss=0.2058, simple_loss=0.2727, pruned_loss=0.06946, over 4805.00 frames. ], tot_loss[loss=0.1763, simple_loss=0.2475, pruned_loss=0.05249, over 954350.93 frames. ], batch size: 45, lr: 3.35e-03, grad_scale: 32.0 2023-04-27 11:00:28,864 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-27 11:00:47,552 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8645, 2.2165, 1.9791, 2.2052, 1.6037, 2.0110, 1.9246, 1.5092], device='cuda:4'), covar=tensor([0.1609, 0.1000, 0.0674, 0.1011, 0.2941, 0.0893, 0.1624, 0.2113], device='cuda:4'), in_proj_covar=tensor([0.0287, 0.0307, 0.0221, 0.0281, 0.0315, 0.0261, 0.0251, 0.0268], device='cuda:4'), out_proj_covar=tensor([1.1537e-04, 1.2227e-04, 8.7837e-05, 1.1143e-04, 1.2796e-04, 1.0380e-04, 1.0172e-04, 1.0636e-04], device='cuda:4') 2023-04-27 11:00:48,298 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3024, 1.2347, 1.6219, 1.6362, 1.2697, 1.1530, 1.3660, 0.8600], device='cuda:4'), covar=tensor([0.0631, 0.0656, 0.0393, 0.0571, 0.0792, 0.1053, 0.0566, 0.0651], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0069, 0.0068, 0.0068, 0.0075, 0.0096, 0.0074, 0.0068], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 11:00:54,325 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=98810.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:00:56,166 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0098, 0.9926, 1.2064, 1.1467, 0.9911, 0.9290, 0.9403, 0.6576], device='cuda:4'), covar=tensor([0.0582, 0.0565, 0.0442, 0.0596, 0.0719, 0.1123, 0.0475, 0.0675], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0069, 0.0068, 0.0068, 0.0075, 0.0096, 0.0074, 0.0067], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 11:01:00,975 INFO [finetune.py:976] (4/7) Epoch 18, batch 1450, loss[loss=0.1917, simple_loss=0.2727, pruned_loss=0.05541, over 4905.00 frames. ], tot_loss[loss=0.1772, simple_loss=0.2492, pruned_loss=0.05254, over 955845.70 frames. ], batch size: 37, lr: 3.35e-03, grad_scale: 32.0 2023-04-27 11:01:02,316 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1729, 1.6555, 2.1063, 2.3371, 2.0270, 1.6118, 1.3461, 1.8655], device='cuda:4'), covar=tensor([0.3094, 0.3067, 0.1467, 0.2440, 0.2548, 0.2588, 0.4167, 0.2020], device='cuda:4'), in_proj_covar=tensor([0.0289, 0.0245, 0.0226, 0.0315, 0.0218, 0.0230, 0.0227, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 11:01:24,437 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.190e+02 1.637e+02 1.928e+02 2.404e+02 4.442e+02, threshold=3.855e+02, percent-clipped=1.0 2023-04-27 11:01:34,709 INFO [finetune.py:976] (4/7) Epoch 18, batch 1500, loss[loss=0.1789, simple_loss=0.2541, pruned_loss=0.05187, over 4749.00 frames. ], tot_loss[loss=0.1793, simple_loss=0.2514, pruned_loss=0.05357, over 957049.47 frames. ], batch size: 54, lr: 3.35e-03, grad_scale: 32.0 2023-04-27 11:01:35,470 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7462, 2.2689, 1.7477, 1.7362, 1.2947, 1.3108, 1.8228, 1.2333], device='cuda:4'), covar=tensor([0.1777, 0.1446, 0.1412, 0.1740, 0.2403, 0.2001, 0.0999, 0.2128], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0212, 0.0168, 0.0204, 0.0199, 0.0184, 0.0156, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 11:01:40,266 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8116, 1.2301, 1.4318, 1.4849, 1.8685, 1.6049, 1.2559, 1.3960], device='cuda:4'), covar=tensor([0.1425, 0.1541, 0.1708, 0.1270, 0.0924, 0.1580, 0.1898, 0.2113], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0313, 0.0350, 0.0286, 0.0328, 0.0309, 0.0299, 0.0365], device='cuda:4'), out_proj_covar=tensor([6.3618e-05, 6.5122e-05, 7.4475e-05, 5.8086e-05, 6.8177e-05, 6.5049e-05, 6.3077e-05, 7.7820e-05], device='cuda:4') 2023-04-27 11:01:52,806 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=98890.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:02:02,813 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=98895.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:02:11,048 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=98899.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:02:35,632 INFO [finetune.py:976] (4/7) Epoch 18, batch 1550, loss[loss=0.2354, simple_loss=0.3054, pruned_loss=0.08269, over 4299.00 frames. ], tot_loss[loss=0.1786, simple_loss=0.251, pruned_loss=0.05311, over 957298.85 frames. ], batch size: 66, lr: 3.35e-03, grad_scale: 32.0 2023-04-27 11:03:17,443 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=98951.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:03:19,153 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.300e+01 1.636e+02 1.920e+02 2.366e+02 5.799e+02, threshold=3.840e+02, percent-clipped=2.0 2023-04-27 11:03:25,942 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=98956.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:03:41,676 INFO [finetune.py:976] (4/7) Epoch 18, batch 1600, loss[loss=0.197, simple_loss=0.2579, pruned_loss=0.06806, over 4852.00 frames. ], tot_loss[loss=0.1777, simple_loss=0.2492, pruned_loss=0.05305, over 954360.48 frames. ], batch size: 49, lr: 3.35e-03, grad_scale: 32.0 2023-04-27 11:03:41,817 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=98971.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:03:57,976 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-27 11:04:20,540 INFO [finetune.py:976] (4/7) Epoch 18, batch 1650, loss[loss=0.1477, simple_loss=0.1992, pruned_loss=0.04812, over 4203.00 frames. ], tot_loss[loss=0.1746, simple_loss=0.2456, pruned_loss=0.0518, over 953754.48 frames. ], batch size: 18, lr: 3.35e-03, grad_scale: 32.0 2023-04-27 11:04:27,352 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.8734, 1.9938, 1.9104, 1.5189, 2.0555, 1.6553, 2.6149, 1.6492], device='cuda:4'), covar=tensor([0.3395, 0.1687, 0.4568, 0.2678, 0.1521, 0.2291, 0.1264, 0.4273], device='cuda:4'), in_proj_covar=tensor([0.0338, 0.0346, 0.0428, 0.0353, 0.0380, 0.0379, 0.0369, 0.0419], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 11:04:27,363 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=99032.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:04:32,709 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7050, 0.7639, 1.5679, 2.0839, 1.7847, 1.6071, 1.5778, 1.5921], device='cuda:4'), covar=tensor([0.4396, 0.6219, 0.6594, 0.6244, 0.5985, 0.7607, 0.7108, 0.7587], device='cuda:4'), in_proj_covar=tensor([0.0422, 0.0407, 0.0497, 0.0503, 0.0451, 0.0478, 0.0484, 0.0489], device='cuda:4'), out_proj_covar=tensor([1.0164e-04, 9.9997e-05, 1.1178e-04, 1.2006e-04, 1.0811e-04, 1.1483e-04, 1.1481e-04, 1.1531e-04], device='cuda:4') 2023-04-27 11:04:40,863 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=99051.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:04:42,972 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.087e+02 1.609e+02 1.883e+02 2.336e+02 5.190e+02, threshold=3.766e+02, percent-clipped=4.0 2023-04-27 11:04:50,876 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=99066.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:04:53,813 INFO [finetune.py:976] (4/7) Epoch 18, batch 1700, loss[loss=0.153, simple_loss=0.2285, pruned_loss=0.03873, over 4822.00 frames. ], tot_loss[loss=0.1743, simple_loss=0.2443, pruned_loss=0.05208, over 953726.99 frames. ], batch size: 41, lr: 3.35e-03, grad_scale: 32.0 2023-04-27 11:05:01,797 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1224, 2.5330, 0.9652, 1.4773, 2.0567, 1.2948, 3.3247, 1.9540], device='cuda:4'), covar=tensor([0.0669, 0.0602, 0.0759, 0.1251, 0.0469, 0.0953, 0.0262, 0.0570], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0048, 0.0046, 0.0050, 0.0052, 0.0074, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 11:05:01,892 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-27 11:05:06,092 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7627, 1.4136, 1.3467, 1.6248, 1.9910, 1.5946, 1.3560, 1.3197], device='cuda:4'), covar=tensor([0.1613, 0.1421, 0.1880, 0.1332, 0.0820, 0.1673, 0.1929, 0.2133], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0311, 0.0350, 0.0286, 0.0327, 0.0308, 0.0299, 0.0365], device='cuda:4'), out_proj_covar=tensor([6.3458e-05, 6.4788e-05, 7.4384e-05, 5.8113e-05, 6.8059e-05, 6.4751e-05, 6.3007e-05, 7.7834e-05], device='cuda:4') 2023-04-27 11:05:08,637 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.47 vs. limit=5.0 2023-04-27 11:05:21,696 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=99112.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:05:22,862 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=99114.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:05:27,557 INFO [finetune.py:976] (4/7) Epoch 18, batch 1750, loss[loss=0.2073, simple_loss=0.2763, pruned_loss=0.06919, over 4836.00 frames. ], tot_loss[loss=0.177, simple_loss=0.2474, pruned_loss=0.05332, over 954486.21 frames. ], batch size: 47, lr: 3.35e-03, grad_scale: 32.0 2023-04-27 11:05:49,521 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7351, 2.0020, 1.0285, 1.4272, 2.2366, 1.6293, 1.5479, 1.5907], device='cuda:4'), covar=tensor([0.0471, 0.0332, 0.0294, 0.0521, 0.0237, 0.0500, 0.0485, 0.0557], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0029, 0.0019, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0044, 0.0038, 0.0051, 0.0038, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 11:05:50,008 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.166e+02 1.654e+02 1.955e+02 2.443e+02 4.969e+02, threshold=3.909e+02, percent-clipped=5.0 2023-04-27 11:06:01,215 INFO [finetune.py:976] (4/7) Epoch 18, batch 1800, loss[loss=0.198, simple_loss=0.2708, pruned_loss=0.06262, over 4818.00 frames. ], tot_loss[loss=0.178, simple_loss=0.2488, pruned_loss=0.05359, over 954165.00 frames. ], batch size: 40, lr: 3.35e-03, grad_scale: 32.0 2023-04-27 11:06:18,768 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=99199.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:06:20,638 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1632, 2.7725, 2.2216, 2.5604, 1.9581, 2.4825, 2.5404, 1.8298], device='cuda:4'), covar=tensor([0.2265, 0.1471, 0.0939, 0.1400, 0.3462, 0.1332, 0.2018, 0.2841], device='cuda:4'), in_proj_covar=tensor([0.0291, 0.0310, 0.0222, 0.0283, 0.0317, 0.0263, 0.0252, 0.0270], device='cuda:4'), out_proj_covar=tensor([1.1689e-04, 1.2329e-04, 8.8397e-05, 1.1247e-04, 1.2877e-04, 1.0451e-04, 1.0178e-04, 1.0742e-04], device='cuda:4') 2023-04-27 11:06:28,563 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-04-27 11:06:34,183 INFO [finetune.py:976] (4/7) Epoch 18, batch 1850, loss[loss=0.1789, simple_loss=0.2615, pruned_loss=0.04813, over 4902.00 frames. ], tot_loss[loss=0.1779, simple_loss=0.2488, pruned_loss=0.05345, over 955307.98 frames. ], batch size: 43, lr: 3.35e-03, grad_scale: 32.0 2023-04-27 11:06:39,642 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=99229.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:06:44,546 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.9371, 2.1201, 2.0961, 2.2271, 1.9921, 2.1156, 2.2935, 2.1477], device='cuda:4'), covar=tensor([0.4047, 0.6480, 0.5457, 0.4856, 0.5891, 0.7481, 0.6104, 0.5759], device='cuda:4'), in_proj_covar=tensor([0.0332, 0.0374, 0.0320, 0.0333, 0.0344, 0.0395, 0.0355, 0.0325], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 11:06:49,926 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=99246.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:06:50,533 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=99247.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:06:53,449 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=99251.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:06:55,164 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2887, 1.6933, 1.5450, 1.9487, 1.8548, 1.9323, 1.6257, 3.1807], device='cuda:4'), covar=tensor([0.0546, 0.0626, 0.0691, 0.0942, 0.0514, 0.0554, 0.0643, 0.0169], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0040, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 11:06:55,631 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.025e+02 1.735e+02 2.090e+02 2.546e+02 5.570e+02, threshold=4.180e+02, percent-clipped=4.0 2023-04-27 11:07:04,262 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.7287, 3.7203, 2.7187, 4.3370, 3.7114, 3.7170, 1.6569, 3.6635], device='cuda:4'), covar=tensor([0.1571, 0.1121, 0.2775, 0.1560, 0.2295, 0.1773, 0.5314, 0.2183], device='cuda:4'), in_proj_covar=tensor([0.0242, 0.0212, 0.0246, 0.0302, 0.0296, 0.0247, 0.0270, 0.0267], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 11:07:06,175 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=99268.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:07:07,879 INFO [finetune.py:976] (4/7) Epoch 18, batch 1900, loss[loss=0.1702, simple_loss=0.2545, pruned_loss=0.04301, over 4764.00 frames. ], tot_loss[loss=0.179, simple_loss=0.2505, pruned_loss=0.05372, over 954735.69 frames. ], batch size: 28, lr: 3.35e-03, grad_scale: 32.0 2023-04-27 11:07:20,149 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=99290.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:07:20,816 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.55 vs. limit=5.0 2023-04-27 11:07:30,991 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8421, 3.6718, 0.9152, 1.9530, 2.1681, 2.7442, 2.1175, 1.1104], device='cuda:4'), covar=tensor([0.1297, 0.0752, 0.2176, 0.1288, 0.1039, 0.0895, 0.1566, 0.1917], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0243, 0.0137, 0.0121, 0.0132, 0.0153, 0.0118, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 11:08:03,838 INFO [finetune.py:976] (4/7) Epoch 18, batch 1950, loss[loss=0.1609, simple_loss=0.2293, pruned_loss=0.04621, over 4749.00 frames. ], tot_loss[loss=0.1763, simple_loss=0.2479, pruned_loss=0.0523, over 955586.34 frames. ], batch size: 59, lr: 3.35e-03, grad_scale: 32.0 2023-04-27 11:08:05,836 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=99324.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:08:07,602 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=99327.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:08:09,364 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=99329.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:08:14,239 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9562, 1.4876, 1.5080, 1.7984, 2.0958, 1.7427, 1.5148, 1.4365], device='cuda:4'), covar=tensor([0.1628, 0.1461, 0.1845, 0.1187, 0.0964, 0.1463, 0.2178, 0.2197], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0313, 0.0351, 0.0288, 0.0329, 0.0309, 0.0301, 0.0368], device='cuda:4'), out_proj_covar=tensor([6.3980e-05, 6.5127e-05, 7.4634e-05, 5.8517e-05, 6.8367e-05, 6.5134e-05, 6.3461e-05, 7.8435e-05], device='cuda:4') 2023-04-27 11:08:14,307 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.58 vs. limit=2.0 2023-04-27 11:08:29,217 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-04-27 11:08:30,193 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.408e+01 1.522e+02 1.826e+02 2.209e+02 4.570e+02, threshold=3.652e+02, percent-clipped=1.0 2023-04-27 11:08:52,740 INFO [finetune.py:976] (4/7) Epoch 18, batch 2000, loss[loss=0.1601, simple_loss=0.2237, pruned_loss=0.04825, over 4905.00 frames. ], tot_loss[loss=0.1749, simple_loss=0.2457, pruned_loss=0.05205, over 954921.16 frames. ], batch size: 32, lr: 3.34e-03, grad_scale: 32.0 2023-04-27 11:09:12,997 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=99385.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:09:44,021 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=99407.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:09:54,332 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=99420.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:09:54,850 INFO [finetune.py:976] (4/7) Epoch 18, batch 2050, loss[loss=0.1862, simple_loss=0.2382, pruned_loss=0.06708, over 4731.00 frames. ], tot_loss[loss=0.1725, simple_loss=0.2424, pruned_loss=0.05131, over 955584.94 frames. ], batch size: 54, lr: 3.34e-03, grad_scale: 32.0 2023-04-27 11:10:00,697 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.79 vs. limit=2.0 2023-04-27 11:10:15,961 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.033e+02 1.635e+02 1.952e+02 2.334e+02 5.427e+02, threshold=3.904e+02, percent-clipped=2.0 2023-04-27 11:10:26,037 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=99468.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 11:10:28,726 INFO [finetune.py:976] (4/7) Epoch 18, batch 2100, loss[loss=0.1922, simple_loss=0.2667, pruned_loss=0.05886, over 4855.00 frames. ], tot_loss[loss=0.1737, simple_loss=0.2432, pruned_loss=0.05205, over 954707.67 frames. ], batch size: 44, lr: 3.34e-03, grad_scale: 32.0 2023-04-27 11:10:35,464 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=99481.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:10:49,277 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4776, 1.6170, 4.0415, 3.7408, 3.5582, 3.8403, 3.7997, 3.5780], device='cuda:4'), covar=tensor([0.7143, 0.5298, 0.1097, 0.1927, 0.1228, 0.1794, 0.1804, 0.1620], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0309, 0.0406, 0.0410, 0.0352, 0.0406, 0.0313, 0.0368], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 11:10:57,047 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8793, 1.0654, 1.5479, 1.7063, 1.6398, 1.7098, 1.5530, 1.5351], device='cuda:4'), covar=tensor([0.3625, 0.5167, 0.4209, 0.4198, 0.5255, 0.7107, 0.4747, 0.4645], device='cuda:4'), in_proj_covar=tensor([0.0330, 0.0372, 0.0320, 0.0331, 0.0342, 0.0393, 0.0354, 0.0324], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 11:11:02,726 INFO [finetune.py:976] (4/7) Epoch 18, batch 2150, loss[loss=0.2165, simple_loss=0.2948, pruned_loss=0.06907, over 4892.00 frames. ], tot_loss[loss=0.1772, simple_loss=0.2471, pruned_loss=0.0537, over 954437.84 frames. ], batch size: 43, lr: 3.34e-03, grad_scale: 32.0 2023-04-27 11:11:08,218 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=99529.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 11:11:18,931 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=99546.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:11:21,994 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=99551.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:11:23,740 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.043e+02 1.609e+02 1.810e+02 2.358e+02 5.458e+02, threshold=3.621e+02, percent-clipped=4.0 2023-04-27 11:11:32,861 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=99568.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:11:35,036 INFO [finetune.py:976] (4/7) Epoch 18, batch 2200, loss[loss=0.1894, simple_loss=0.2627, pruned_loss=0.05808, over 4867.00 frames. ], tot_loss[loss=0.1807, simple_loss=0.251, pruned_loss=0.05526, over 954301.79 frames. ], batch size: 31, lr: 3.34e-03, grad_scale: 32.0 2023-04-27 11:11:45,025 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=99585.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:11:47,517 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=99589.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:11:48,161 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4946, 1.9129, 1.8731, 2.0060, 1.8352, 1.9091, 1.9757, 1.9003], device='cuda:4'), covar=tensor([0.4104, 0.5545, 0.4647, 0.4534, 0.5543, 0.7024, 0.5028, 0.5088], device='cuda:4'), in_proj_covar=tensor([0.0330, 0.0371, 0.0320, 0.0331, 0.0343, 0.0393, 0.0354, 0.0325], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 11:11:51,020 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=99594.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:11:51,045 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2242, 1.3319, 4.9568, 4.5980, 4.2862, 4.6265, 4.3744, 4.3967], device='cuda:4'), covar=tensor([0.6289, 0.6101, 0.1101, 0.2228, 0.1045, 0.1273, 0.1557, 0.1567], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0308, 0.0403, 0.0407, 0.0349, 0.0404, 0.0311, 0.0366], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 11:11:54,027 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=99599.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:12:08,626 INFO [finetune.py:976] (4/7) Epoch 18, batch 2250, loss[loss=0.1258, simple_loss=0.1926, pruned_loss=0.02946, over 3909.00 frames. ], tot_loss[loss=0.1807, simple_loss=0.2515, pruned_loss=0.05498, over 953669.25 frames. ], batch size: 17, lr: 3.34e-03, grad_scale: 32.0 2023-04-27 11:12:11,482 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=99624.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:12:12,171 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9065, 1.3646, 1.6890, 1.7330, 1.6454, 1.3795, 0.7820, 1.3304], device='cuda:4'), covar=tensor([0.3155, 0.3325, 0.1742, 0.2290, 0.2390, 0.2609, 0.4553, 0.2075], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0245, 0.0226, 0.0314, 0.0217, 0.0230, 0.0228, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 11:12:13,846 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=99627.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:12:14,457 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.8659, 1.0684, 3.3039, 3.0470, 2.9408, 3.2119, 3.1832, 2.9458], device='cuda:4'), covar=tensor([0.6920, 0.5590, 0.1380, 0.2124, 0.1500, 0.1841, 0.1618, 0.1574], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0309, 0.0404, 0.0408, 0.0350, 0.0405, 0.0312, 0.0368], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 11:12:15,122 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=99629.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:12:21,645 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8148, 2.4903, 1.8317, 1.7246, 1.3269, 1.3805, 1.7709, 1.2675], device='cuda:4'), covar=tensor([0.1793, 0.1439, 0.1506, 0.1797, 0.2340, 0.2066, 0.1122, 0.2068], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0213, 0.0168, 0.0205, 0.0200, 0.0185, 0.0157, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 11:12:28,855 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=99650.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:12:31,121 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.508e+01 1.589e+02 1.867e+02 2.262e+02 4.430e+02, threshold=3.734e+02, percent-clipped=1.0 2023-04-27 11:12:37,322 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2473, 1.1297, 3.8632, 3.5893, 3.4338, 3.7176, 3.7173, 3.4217], device='cuda:4'), covar=tensor([0.6831, 0.5776, 0.1089, 0.1859, 0.1193, 0.1702, 0.1412, 0.1472], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0308, 0.0403, 0.0406, 0.0349, 0.0403, 0.0311, 0.0366], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 11:12:38,620 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-27 11:12:41,915 INFO [finetune.py:976] (4/7) Epoch 18, batch 2300, loss[loss=0.1485, simple_loss=0.2124, pruned_loss=0.04232, over 4932.00 frames. ], tot_loss[loss=0.1807, simple_loss=0.2519, pruned_loss=0.05477, over 954814.62 frames. ], batch size: 33, lr: 3.34e-03, grad_scale: 32.0 2023-04-27 11:12:45,387 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=99675.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:12:49,362 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=99680.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:12:49,389 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=99680.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:13:12,594 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=99707.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:13:26,445 INFO [finetune.py:976] (4/7) Epoch 18, batch 2350, loss[loss=0.1506, simple_loss=0.2263, pruned_loss=0.03744, over 4927.00 frames. ], tot_loss[loss=0.1775, simple_loss=0.2484, pruned_loss=0.0533, over 955591.10 frames. ], batch size: 38, lr: 3.34e-03, grad_scale: 32.0 2023-04-27 11:13:56,986 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=99741.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:14:16,630 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.069e+02 1.602e+02 1.875e+02 2.250e+02 5.277e+02, threshold=3.750e+02, percent-clipped=1.0 2023-04-27 11:14:17,850 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=99755.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:14:19,110 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3234, 2.8316, 2.4038, 2.7281, 2.0774, 2.5113, 2.6440, 1.9124], device='cuda:4'), covar=tensor([0.2006, 0.1137, 0.0816, 0.1143, 0.2988, 0.1054, 0.1773, 0.2914], device='cuda:4'), in_proj_covar=tensor([0.0289, 0.0308, 0.0221, 0.0283, 0.0315, 0.0262, 0.0251, 0.0269], device='cuda:4'), out_proj_covar=tensor([1.1615e-04, 1.2254e-04, 8.7809e-05, 1.1220e-04, 1.2778e-04, 1.0399e-04, 1.0159e-04, 1.0675e-04], device='cuda:4') 2023-04-27 11:14:33,760 INFO [finetune.py:976] (4/7) Epoch 18, batch 2400, loss[loss=0.1555, simple_loss=0.2152, pruned_loss=0.04793, over 4214.00 frames. ], tot_loss[loss=0.1738, simple_loss=0.2443, pruned_loss=0.05167, over 954467.72 frames. ], batch size: 18, lr: 3.34e-03, grad_scale: 32.0 2023-04-27 11:14:36,882 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=99776.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:14:58,096 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-27 11:15:26,254 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.3858, 1.3690, 1.4307, 1.0164, 1.3944, 1.2426, 1.7114, 1.3801], device='cuda:4'), covar=tensor([0.3759, 0.1923, 0.5448, 0.2757, 0.1570, 0.2287, 0.1616, 0.5107], device='cuda:4'), in_proj_covar=tensor([0.0338, 0.0345, 0.0425, 0.0352, 0.0380, 0.0377, 0.0368, 0.0417], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 11:15:35,839 INFO [finetune.py:976] (4/7) Epoch 18, batch 2450, loss[loss=0.1369, simple_loss=0.2047, pruned_loss=0.03455, over 3998.00 frames. ], tot_loss[loss=0.1723, simple_loss=0.2421, pruned_loss=0.05127, over 955141.42 frames. ], batch size: 17, lr: 3.34e-03, grad_scale: 64.0 2023-04-27 11:15:37,740 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=99824.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 11:16:08,396 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.453e+01 1.597e+02 1.870e+02 2.238e+02 4.750e+02, threshold=3.741e+02, percent-clipped=1.0 2023-04-27 11:16:19,143 INFO [finetune.py:976] (4/7) Epoch 18, batch 2500, loss[loss=0.1893, simple_loss=0.2776, pruned_loss=0.05052, over 4800.00 frames. ], tot_loss[loss=0.1747, simple_loss=0.2446, pruned_loss=0.05242, over 955864.30 frames. ], batch size: 45, lr: 3.34e-03, grad_scale: 64.0 2023-04-27 11:16:28,105 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6926, 1.9717, 1.7125, 1.9592, 1.4999, 1.6925, 1.7214, 1.3272], device='cuda:4'), covar=tensor([0.1763, 0.1334, 0.0851, 0.1061, 0.3546, 0.1227, 0.1779, 0.2568], device='cuda:4'), in_proj_covar=tensor([0.0286, 0.0306, 0.0219, 0.0281, 0.0312, 0.0260, 0.0249, 0.0266], device='cuda:4'), out_proj_covar=tensor([1.1500e-04, 1.2169e-04, 8.7153e-05, 1.1144e-04, 1.2687e-04, 1.0304e-04, 1.0079e-04, 1.0563e-04], device='cuda:4') 2023-04-27 11:16:28,688 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=99885.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:16:31,602 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8645, 1.3514, 1.6711, 1.7234, 1.6205, 1.3476, 0.7884, 1.3224], device='cuda:4'), covar=tensor([0.2814, 0.2926, 0.1527, 0.2101, 0.2181, 0.2395, 0.4108, 0.2028], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0245, 0.0226, 0.0315, 0.0218, 0.0231, 0.0228, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 11:16:52,961 INFO [finetune.py:976] (4/7) Epoch 18, batch 2550, loss[loss=0.2197, simple_loss=0.3023, pruned_loss=0.06855, over 4844.00 frames. ], tot_loss[loss=0.1779, simple_loss=0.2488, pruned_loss=0.05349, over 957738.78 frames. ], batch size: 49, lr: 3.34e-03, grad_scale: 64.0 2023-04-27 11:16:53,711 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5575, 1.9067, 1.7365, 2.5180, 2.6547, 2.1274, 2.0627, 1.8900], device='cuda:4'), covar=tensor([0.1726, 0.1593, 0.1887, 0.1214, 0.1037, 0.1805, 0.2011, 0.2333], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0312, 0.0350, 0.0288, 0.0329, 0.0308, 0.0300, 0.0367], device='cuda:4'), out_proj_covar=tensor([6.3782e-05, 6.4968e-05, 7.4474e-05, 5.8424e-05, 6.8331e-05, 6.4853e-05, 6.3185e-05, 7.8155e-05], device='cuda:4') 2023-04-27 11:16:54,890 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=99924.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:16:54,907 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=99924.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:17:00,810 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=99933.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:17:10,140 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=99945.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:17:15,411 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.47 vs. limit=2.0 2023-04-27 11:17:15,648 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8567, 1.5128, 1.9766, 2.3170, 1.9418, 1.8040, 1.9200, 1.8369], device='cuda:4'), covar=tensor([0.4681, 0.6689, 0.6201, 0.5817, 0.5753, 0.8127, 0.8014, 0.8616], device='cuda:4'), in_proj_covar=tensor([0.0425, 0.0409, 0.0500, 0.0504, 0.0452, 0.0479, 0.0484, 0.0490], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 11:17:16,096 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.051e+02 1.590e+02 1.888e+02 2.384e+02 3.870e+02, threshold=3.776e+02, percent-clipped=2.0 2023-04-27 11:17:26,893 INFO [finetune.py:976] (4/7) Epoch 18, batch 2600, loss[loss=0.201, simple_loss=0.2709, pruned_loss=0.06559, over 4865.00 frames. ], tot_loss[loss=0.1796, simple_loss=0.2509, pruned_loss=0.0542, over 957470.49 frames. ], batch size: 34, lr: 3.34e-03, grad_scale: 64.0 2023-04-27 11:17:27,554 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=99972.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:17:32,462 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=99980.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:17:52,667 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=100007.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 11:18:01,563 INFO [finetune.py:976] (4/7) Epoch 18, batch 2650, loss[loss=0.1664, simple_loss=0.2391, pruned_loss=0.04687, over 4896.00 frames. ], tot_loss[loss=0.1806, simple_loss=0.252, pruned_loss=0.05457, over 957883.24 frames. ], batch size: 32, lr: 3.34e-03, grad_scale: 64.0 2023-04-27 11:18:03,447 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.3573, 3.3351, 2.4680, 3.9182, 3.3546, 3.4143, 1.4864, 3.3134], device='cuda:4'), covar=tensor([0.1993, 0.1313, 0.3077, 0.2160, 0.3002, 0.1990, 0.5883, 0.2687], device='cuda:4'), in_proj_covar=tensor([0.0247, 0.0216, 0.0252, 0.0308, 0.0300, 0.0250, 0.0274, 0.0273], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 11:18:05,842 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=100028.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:18:10,673 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=100036.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:18:24,400 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.122e+02 1.666e+02 1.842e+02 2.203e+02 3.197e+02, threshold=3.685e+02, percent-clipped=0.0 2023-04-27 11:18:26,957 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2559, 1.6537, 1.5537, 1.8238, 1.7467, 1.9724, 1.5058, 3.5354], device='cuda:4'), covar=tensor([0.0592, 0.0759, 0.0727, 0.1147, 0.0636, 0.0485, 0.0764, 0.0142], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0040, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 11:18:30,634 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9024, 1.6500, 2.1076, 2.2322, 1.6272, 1.5047, 1.7894, 1.2306], device='cuda:4'), covar=tensor([0.0437, 0.0837, 0.0446, 0.0576, 0.0673, 0.1061, 0.0649, 0.0643], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0069, 0.0068, 0.0067, 0.0075, 0.0096, 0.0074, 0.0067], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 11:18:33,072 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=100068.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 11:18:35,203 INFO [finetune.py:976] (4/7) Epoch 18, batch 2700, loss[loss=0.1608, simple_loss=0.2292, pruned_loss=0.04619, over 4787.00 frames. ], tot_loss[loss=0.1791, simple_loss=0.2509, pruned_loss=0.05361, over 958118.39 frames. ], batch size: 51, lr: 3.34e-03, grad_scale: 64.0 2023-04-27 11:18:38,370 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=100076.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:18:42,670 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1118, 2.4794, 0.9891, 1.3777, 1.9015, 1.2450, 3.3312, 1.7763], device='cuda:4'), covar=tensor([0.0665, 0.0764, 0.0886, 0.1276, 0.0517, 0.1002, 0.0219, 0.0583], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0065, 0.0048, 0.0046, 0.0050, 0.0052, 0.0074, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 11:19:37,243 INFO [finetune.py:976] (4/7) Epoch 18, batch 2750, loss[loss=0.1653, simple_loss=0.2336, pruned_loss=0.04846, over 4936.00 frames. ], tot_loss[loss=0.1777, simple_loss=0.2485, pruned_loss=0.05343, over 956921.21 frames. ], batch size: 33, lr: 3.34e-03, grad_scale: 64.0 2023-04-27 11:19:37,682 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-27 11:19:39,079 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=100124.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:19:39,106 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=100124.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 11:20:02,150 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1389, 2.6394, 0.9794, 1.3722, 1.9513, 1.2270, 3.4967, 1.8537], device='cuda:4'), covar=tensor([0.0700, 0.0623, 0.0871, 0.1405, 0.0571, 0.1093, 0.0260, 0.0637], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0065, 0.0048, 0.0046, 0.0050, 0.0052, 0.0073, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 11:20:11,922 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.977e+01 1.603e+02 1.951e+02 2.448e+02 3.827e+02, threshold=3.902e+02, percent-clipped=2.0 2023-04-27 11:20:12,928 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.30 vs. limit=5.0 2023-04-27 11:20:28,164 INFO [finetune.py:976] (4/7) Epoch 18, batch 2800, loss[loss=0.171, simple_loss=0.2399, pruned_loss=0.05105, over 4826.00 frames. ], tot_loss[loss=0.1748, simple_loss=0.2449, pruned_loss=0.05231, over 957655.23 frames. ], batch size: 30, lr: 3.34e-03, grad_scale: 32.0 2023-04-27 11:20:34,238 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=100172.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 11:21:07,287 INFO [finetune.py:976] (4/7) Epoch 18, batch 2850, loss[loss=0.1701, simple_loss=0.2385, pruned_loss=0.0509, over 4800.00 frames. ], tot_loss[loss=0.1744, simple_loss=0.2442, pruned_loss=0.0523, over 957853.59 frames. ], batch size: 45, lr: 3.34e-03, grad_scale: 32.0 2023-04-27 11:21:15,066 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=100224.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:21:18,783 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1550, 1.4384, 1.2239, 1.3776, 1.1677, 1.2083, 1.2448, 1.0505], device='cuda:4'), covar=tensor([0.1956, 0.1581, 0.1141, 0.1454, 0.3572, 0.1386, 0.1823, 0.2300], device='cuda:4'), in_proj_covar=tensor([0.0285, 0.0304, 0.0217, 0.0279, 0.0311, 0.0258, 0.0248, 0.0265], device='cuda:4'), out_proj_covar=tensor([1.1451e-04, 1.2068e-04, 8.6222e-05, 1.1081e-04, 1.2632e-04, 1.0251e-04, 1.0027e-04, 1.0510e-04], device='cuda:4') 2023-04-27 11:21:29,442 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=100238.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:21:39,951 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=100245.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:21:52,186 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.041e+02 1.601e+02 1.939e+02 2.362e+02 3.802e+02, threshold=3.878e+02, percent-clipped=0.0 2023-04-27 11:22:14,790 INFO [finetune.py:976] (4/7) Epoch 18, batch 2900, loss[loss=0.1308, simple_loss=0.2, pruned_loss=0.03076, over 4762.00 frames. ], tot_loss[loss=0.177, simple_loss=0.247, pruned_loss=0.05348, over 957108.20 frames. ], batch size: 26, lr: 3.34e-03, grad_scale: 32.0 2023-04-27 11:22:15,467 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=100272.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:22:28,744 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=100293.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:22:32,444 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=100299.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:22:40,642 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8155, 2.3060, 1.8982, 2.2147, 1.5869, 1.9508, 1.9710, 1.4927], device='cuda:4'), covar=tensor([0.2133, 0.1093, 0.0817, 0.1138, 0.3390, 0.1160, 0.1828, 0.2622], device='cuda:4'), in_proj_covar=tensor([0.0287, 0.0305, 0.0218, 0.0280, 0.0312, 0.0260, 0.0250, 0.0266], device='cuda:4'), out_proj_covar=tensor([1.1515e-04, 1.2141e-04, 8.6799e-05, 1.1132e-04, 1.2682e-04, 1.0301e-04, 1.0106e-04, 1.0570e-04], device='cuda:4') 2023-04-27 11:22:48,679 INFO [finetune.py:976] (4/7) Epoch 18, batch 2950, loss[loss=0.172, simple_loss=0.2445, pruned_loss=0.04973, over 4197.00 frames. ], tot_loss[loss=0.1785, simple_loss=0.2499, pruned_loss=0.05354, over 955386.46 frames. ], batch size: 65, lr: 3.34e-03, grad_scale: 32.0 2023-04-27 11:22:57,898 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=100336.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:23:09,851 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.143e+02 1.535e+02 1.848e+02 2.418e+02 4.913e+02, threshold=3.697e+02, percent-clipped=1.0 2023-04-27 11:23:15,824 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=100363.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 11:23:22,041 INFO [finetune.py:976] (4/7) Epoch 18, batch 3000, loss[loss=0.2047, simple_loss=0.2693, pruned_loss=0.07007, over 4892.00 frames. ], tot_loss[loss=0.1807, simple_loss=0.2521, pruned_loss=0.05467, over 956151.16 frames. ], batch size: 37, lr: 3.34e-03, grad_scale: 32.0 2023-04-27 11:23:22,041 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 11:23:24,137 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.4252, 3.5499, 2.4916, 3.9030, 3.5797, 3.4920, 1.5459, 3.4486], device='cuda:4'), covar=tensor([0.1597, 0.1175, 0.3027, 0.1932, 0.2339, 0.1513, 0.5254, 0.2285], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0214, 0.0249, 0.0305, 0.0296, 0.0247, 0.0271, 0.0271], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 11:23:32,628 INFO [finetune.py:1010] (4/7) Epoch 18, validation: loss=0.1524, simple_loss=0.2231, pruned_loss=0.04086, over 2265189.00 frames. 2023-04-27 11:23:32,628 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-27 11:23:41,190 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=100384.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:23:52,514 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=100402.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:24:03,119 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-27 11:24:04,563 INFO [finetune.py:976] (4/7) Epoch 18, batch 3050, loss[loss=0.17, simple_loss=0.2458, pruned_loss=0.0471, over 4925.00 frames. ], tot_loss[loss=0.1802, simple_loss=0.2519, pruned_loss=0.05423, over 954553.98 frames. ], batch size: 33, lr: 3.34e-03, grad_scale: 32.0 2023-04-27 11:24:11,418 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9980, 2.3954, 0.9916, 1.3125, 1.7823, 1.2553, 3.0493, 1.6565], device='cuda:4'), covar=tensor([0.0668, 0.0565, 0.0741, 0.1263, 0.0507, 0.1014, 0.0237, 0.0626], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0065, 0.0048, 0.0046, 0.0049, 0.0052, 0.0073, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0007, 0.0008, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-27 11:24:43,546 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.094e+02 1.525e+02 1.858e+02 2.098e+02 3.467e+02, threshold=3.716e+02, percent-clipped=0.0 2023-04-27 11:24:55,015 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=100463.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 11:24:59,804 INFO [finetune.py:976] (4/7) Epoch 18, batch 3100, loss[loss=0.1602, simple_loss=0.2425, pruned_loss=0.039, over 4914.00 frames. ], tot_loss[loss=0.1782, simple_loss=0.2498, pruned_loss=0.05334, over 952805.42 frames. ], batch size: 35, lr: 3.34e-03, grad_scale: 32.0 2023-04-27 11:25:59,155 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5840, 1.4560, 0.4722, 1.3013, 1.4971, 1.4781, 1.3544, 1.4076], device='cuda:4'), covar=tensor([0.0465, 0.0329, 0.0399, 0.0508, 0.0282, 0.0477, 0.0452, 0.0534], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0029, 0.0019, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0045, 0.0038, 0.0051, 0.0038, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 11:26:02,764 INFO [finetune.py:976] (4/7) Epoch 18, batch 3150, loss[loss=0.1626, simple_loss=0.2308, pruned_loss=0.04718, over 4741.00 frames. ], tot_loss[loss=0.1759, simple_loss=0.2467, pruned_loss=0.05253, over 954793.59 frames. ], batch size: 54, lr: 3.34e-03, grad_scale: 32.0 2023-04-27 11:26:16,712 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([5.0061, 4.8936, 3.0791, 5.6311, 4.9815, 4.9955, 2.5364, 4.7799], device='cuda:4'), covar=tensor([0.1329, 0.1018, 0.2852, 0.0870, 0.2616, 0.1494, 0.5201, 0.2296], device='cuda:4'), in_proj_covar=tensor([0.0245, 0.0216, 0.0250, 0.0306, 0.0298, 0.0248, 0.0272, 0.0272], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 11:26:37,705 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.912e+01 1.520e+02 1.805e+02 2.268e+02 8.838e+02, threshold=3.610e+02, percent-clipped=4.0 2023-04-27 11:26:59,505 INFO [finetune.py:976] (4/7) Epoch 18, batch 3200, loss[loss=0.2188, simple_loss=0.2797, pruned_loss=0.07896, over 4752.00 frames. ], tot_loss[loss=0.1726, simple_loss=0.2428, pruned_loss=0.05123, over 954665.92 frames. ], batch size: 54, lr: 3.33e-03, grad_scale: 32.0 2023-04-27 11:27:33,066 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=100594.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:27:44,632 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.89 vs. limit=5.0 2023-04-27 11:28:06,116 INFO [finetune.py:976] (4/7) Epoch 18, batch 3250, loss[loss=0.1972, simple_loss=0.272, pruned_loss=0.06124, over 4872.00 frames. ], tot_loss[loss=0.1756, simple_loss=0.2453, pruned_loss=0.053, over 954222.08 frames. ], batch size: 31, lr: 3.33e-03, grad_scale: 32.0 2023-04-27 11:28:30,302 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.424e+01 1.529e+02 1.873e+02 2.219e+02 4.824e+02, threshold=3.746e+02, percent-clipped=4.0 2023-04-27 11:28:34,777 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7824, 1.6384, 1.9684, 2.1796, 1.6067, 1.4027, 1.7700, 1.0673], device='cuda:4'), covar=tensor([0.0507, 0.0691, 0.0386, 0.0481, 0.0694, 0.1183, 0.0577, 0.0662], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0068, 0.0067, 0.0067, 0.0074, 0.0095, 0.0073, 0.0066], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 11:28:35,350 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2600, 1.2915, 3.8573, 3.5754, 3.3937, 3.7297, 3.6603, 3.3907], device='cuda:4'), covar=tensor([0.7589, 0.5900, 0.1160, 0.1945, 0.1339, 0.2125, 0.1633, 0.1281], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0306, 0.0403, 0.0405, 0.0348, 0.0402, 0.0311, 0.0365], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 11:28:35,370 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=100663.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 11:28:40,637 INFO [finetune.py:976] (4/7) Epoch 18, batch 3300, loss[loss=0.1913, simple_loss=0.2515, pruned_loss=0.06555, over 4827.00 frames. ], tot_loss[loss=0.177, simple_loss=0.2474, pruned_loss=0.05331, over 953375.46 frames. ], batch size: 25, lr: 3.33e-03, grad_scale: 32.0 2023-04-27 11:29:02,300 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4342, 0.9559, 0.2761, 1.1261, 1.0640, 1.3271, 1.2018, 1.2139], device='cuda:4'), covar=tensor([0.0517, 0.0414, 0.0436, 0.0565, 0.0308, 0.0493, 0.0498, 0.0577], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0029, 0.0020, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0045, 0.0038, 0.0051, 0.0038, 0.0049, 0.0049, 0.0050], device='cuda:4') 2023-04-27 11:29:07,778 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=100711.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 11:29:13,775 INFO [finetune.py:976] (4/7) Epoch 18, batch 3350, loss[loss=0.1956, simple_loss=0.263, pruned_loss=0.06414, over 4895.00 frames. ], tot_loss[loss=0.1768, simple_loss=0.2478, pruned_loss=0.05288, over 951791.98 frames. ], batch size: 37, lr: 3.33e-03, grad_scale: 32.0 2023-04-27 11:29:19,926 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=100730.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 11:29:20,523 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6692, 1.5337, 1.6745, 1.9616, 2.0448, 1.5790, 1.2709, 1.8046], device='cuda:4'), covar=tensor([0.0818, 0.1105, 0.0702, 0.0611, 0.0551, 0.0872, 0.0799, 0.0554], device='cuda:4'), in_proj_covar=tensor([0.0189, 0.0202, 0.0182, 0.0173, 0.0178, 0.0182, 0.0151, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 11:29:26,518 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2714, 1.9568, 2.2204, 2.5999, 2.1813, 1.8176, 1.4686, 1.9806], device='cuda:4'), covar=tensor([0.3217, 0.2899, 0.1635, 0.1906, 0.2519, 0.2468, 0.3738, 0.1794], device='cuda:4'), in_proj_covar=tensor([0.0289, 0.0245, 0.0225, 0.0313, 0.0218, 0.0230, 0.0228, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 11:29:37,235 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.219e+01 1.703e+02 2.119e+02 2.646e+02 1.102e+03, threshold=4.237e+02, percent-clipped=5.0 2023-04-27 11:29:39,127 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=100758.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 11:29:47,009 INFO [finetune.py:976] (4/7) Epoch 18, batch 3400, loss[loss=0.1795, simple_loss=0.2442, pruned_loss=0.05737, over 4834.00 frames. ], tot_loss[loss=0.1773, simple_loss=0.2486, pruned_loss=0.05297, over 953007.84 frames. ], batch size: 49, lr: 3.33e-03, grad_scale: 32.0 2023-04-27 11:30:00,832 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=100791.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 11:30:20,270 INFO [finetune.py:976] (4/7) Epoch 18, batch 3450, loss[loss=0.1741, simple_loss=0.2329, pruned_loss=0.05765, over 4723.00 frames. ], tot_loss[loss=0.1779, simple_loss=0.249, pruned_loss=0.05345, over 952687.15 frames. ], batch size: 23, lr: 3.33e-03, grad_scale: 32.0 2023-04-27 11:30:54,165 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.079e+02 1.615e+02 1.948e+02 2.378e+02 4.172e+02, threshold=3.896e+02, percent-clipped=0.0 2023-04-27 11:31:09,718 INFO [finetune.py:976] (4/7) Epoch 18, batch 3500, loss[loss=0.1197, simple_loss=0.1867, pruned_loss=0.02631, over 4783.00 frames. ], tot_loss[loss=0.1765, simple_loss=0.2466, pruned_loss=0.05317, over 953049.96 frames. ], batch size: 26, lr: 3.33e-03, grad_scale: 32.0 2023-04-27 11:31:23,476 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2969, 1.7547, 1.5713, 2.0602, 1.8063, 2.1056, 1.5856, 4.1275], device='cuda:4'), covar=tensor([0.0522, 0.0773, 0.0752, 0.1058, 0.0581, 0.0522, 0.0714, 0.0094], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0039, 0.0037, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0012, 0.0013, 0.0016], device='cuda:4') 2023-04-27 11:31:28,316 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5063, 1.3420, 1.6039, 1.6550, 1.3925, 1.3051, 1.3156, 0.8402], device='cuda:4'), covar=tensor([0.0491, 0.0642, 0.0447, 0.0566, 0.0668, 0.1107, 0.0555, 0.0628], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0068, 0.0067, 0.0067, 0.0074, 0.0095, 0.0073, 0.0066], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 11:31:30,613 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=100894.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:31:49,441 INFO [finetune.py:976] (4/7) Epoch 18, batch 3550, loss[loss=0.1489, simple_loss=0.2283, pruned_loss=0.03476, over 4748.00 frames. ], tot_loss[loss=0.1732, simple_loss=0.2432, pruned_loss=0.05158, over 952147.91 frames. ], batch size: 28, lr: 3.33e-03, grad_scale: 32.0 2023-04-27 11:31:49,511 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.8892, 3.8963, 2.7865, 4.5521, 3.9308, 3.9394, 1.6948, 3.8088], device='cuda:4'), covar=tensor([0.1505, 0.1179, 0.2885, 0.1404, 0.2595, 0.1693, 0.5911, 0.2295], device='cuda:4'), in_proj_covar=tensor([0.0245, 0.0215, 0.0250, 0.0305, 0.0298, 0.0250, 0.0272, 0.0273], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 11:32:02,731 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=100942.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:32:04,709 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-27 11:32:23,485 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.154e+02 1.524e+02 1.758e+02 2.158e+02 5.191e+02, threshold=3.516e+02, percent-clipped=1.0 2023-04-27 11:32:24,118 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3641, 2.0021, 2.1619, 2.7833, 2.6485, 2.2936, 1.8844, 2.5246], device='cuda:4'), covar=tensor([0.0717, 0.1065, 0.0692, 0.0446, 0.0489, 0.0711, 0.0671, 0.0449], device='cuda:4'), in_proj_covar=tensor([0.0189, 0.0202, 0.0183, 0.0172, 0.0178, 0.0182, 0.0151, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 11:32:39,052 INFO [finetune.py:976] (4/7) Epoch 18, batch 3600, loss[loss=0.1298, simple_loss=0.2042, pruned_loss=0.02776, over 4828.00 frames. ], tot_loss[loss=0.1715, simple_loss=0.2413, pruned_loss=0.05085, over 952841.15 frames. ], batch size: 30, lr: 3.33e-03, grad_scale: 32.0 2023-04-27 11:32:46,121 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.63 vs. limit=5.0 2023-04-27 11:33:50,171 INFO [finetune.py:976] (4/7) Epoch 18, batch 3650, loss[loss=0.2002, simple_loss=0.2758, pruned_loss=0.06228, over 4868.00 frames. ], tot_loss[loss=0.1751, simple_loss=0.2454, pruned_loss=0.05237, over 952456.79 frames. ], batch size: 34, lr: 3.33e-03, grad_scale: 32.0 2023-04-27 11:34:14,153 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7980, 1.4522, 1.8594, 2.2809, 1.8809, 1.7684, 1.8105, 1.7986], device='cuda:4'), covar=tensor([0.4519, 0.6861, 0.6937, 0.5801, 0.6201, 0.8044, 0.7974, 0.8880], device='cuda:4'), in_proj_covar=tensor([0.0422, 0.0406, 0.0499, 0.0502, 0.0453, 0.0477, 0.0483, 0.0488], device='cuda:4'), out_proj_covar=tensor([1.0156e-04, 9.9916e-05, 1.1205e-04, 1.1981e-04, 1.0844e-04, 1.1444e-04, 1.1463e-04, 1.1507e-04], device='cuda:4') 2023-04-27 11:34:30,938 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4156, 1.5979, 1.7588, 1.8970, 1.7593, 1.8014, 1.8709, 1.8059], device='cuda:4'), covar=tensor([0.3795, 0.5618, 0.4543, 0.4458, 0.5536, 0.7506, 0.5589, 0.5024], device='cuda:4'), in_proj_covar=tensor([0.0331, 0.0374, 0.0320, 0.0332, 0.0344, 0.0394, 0.0356, 0.0326], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 11:34:31,992 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.015e+02 1.646e+02 1.981e+02 2.447e+02 4.488e+02, threshold=3.961e+02, percent-clipped=2.0 2023-04-27 11:34:34,903 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=101058.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:34:48,426 INFO [finetune.py:976] (4/7) Epoch 18, batch 3700, loss[loss=0.2201, simple_loss=0.293, pruned_loss=0.0736, over 4809.00 frames. ], tot_loss[loss=0.1775, simple_loss=0.249, pruned_loss=0.05303, over 953028.18 frames. ], batch size: 38, lr: 3.33e-03, grad_scale: 32.0 2023-04-27 11:34:57,630 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=101086.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 11:35:09,642 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7057, 1.7150, 1.6273, 1.1681, 1.3169, 1.3216, 1.6503, 1.2431], device='cuda:4'), covar=tensor([0.1603, 0.1465, 0.1282, 0.1487, 0.2032, 0.1738, 0.0884, 0.1792], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0213, 0.0169, 0.0205, 0.0200, 0.0184, 0.0156, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 11:35:10,810 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=101106.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:35:11,934 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=101107.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:35:14,938 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3663, 1.7488, 2.1826, 2.6529, 2.1258, 1.7270, 1.4487, 2.0239], device='cuda:4'), covar=tensor([0.3237, 0.3239, 0.1695, 0.2398, 0.2513, 0.2693, 0.4276, 0.2019], device='cuda:4'), in_proj_covar=tensor([0.0287, 0.0244, 0.0224, 0.0311, 0.0217, 0.0229, 0.0226, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 11:35:22,248 INFO [finetune.py:976] (4/7) Epoch 18, batch 3750, loss[loss=0.1825, simple_loss=0.2684, pruned_loss=0.04836, over 4912.00 frames. ], tot_loss[loss=0.1794, simple_loss=0.2507, pruned_loss=0.05403, over 952118.31 frames. ], batch size: 38, lr: 3.33e-03, grad_scale: 32.0 2023-04-27 11:35:43,439 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.081e+02 1.572e+02 1.895e+02 2.251e+02 5.219e+02, threshold=3.790e+02, percent-clipped=2.0 2023-04-27 11:35:54,464 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=101168.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:35:56,187 INFO [finetune.py:976] (4/7) Epoch 18, batch 3800, loss[loss=0.1407, simple_loss=0.2202, pruned_loss=0.03057, over 4191.00 frames. ], tot_loss[loss=0.1804, simple_loss=0.2521, pruned_loss=0.05433, over 952717.83 frames. ], batch size: 66, lr: 3.33e-03, grad_scale: 32.0 2023-04-27 11:36:14,750 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.7671, 3.4830, 2.6677, 3.0713, 2.2706, 2.1028, 2.7092, 2.1916], device='cuda:4'), covar=tensor([0.1534, 0.1256, 0.1389, 0.1265, 0.2086, 0.1863, 0.0886, 0.1870], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0211, 0.0167, 0.0204, 0.0199, 0.0183, 0.0155, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 11:36:30,056 INFO [finetune.py:976] (4/7) Epoch 18, batch 3850, loss[loss=0.1976, simple_loss=0.2703, pruned_loss=0.06241, over 4811.00 frames. ], tot_loss[loss=0.1777, simple_loss=0.2488, pruned_loss=0.05333, over 952230.56 frames. ], batch size: 51, lr: 3.33e-03, grad_scale: 32.0 2023-04-27 11:36:50,614 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.073e+02 1.513e+02 1.818e+02 2.217e+02 6.339e+02, threshold=3.636e+02, percent-clipped=4.0 2023-04-27 11:37:02,693 INFO [finetune.py:976] (4/7) Epoch 18, batch 3900, loss[loss=0.1531, simple_loss=0.2247, pruned_loss=0.04075, over 4823.00 frames. ], tot_loss[loss=0.1763, simple_loss=0.2467, pruned_loss=0.05298, over 954162.01 frames. ], batch size: 38, lr: 3.33e-03, grad_scale: 32.0 2023-04-27 11:37:35,479 INFO [finetune.py:976] (4/7) Epoch 18, batch 3950, loss[loss=0.1544, simple_loss=0.2327, pruned_loss=0.03807, over 4809.00 frames. ], tot_loss[loss=0.1729, simple_loss=0.2432, pruned_loss=0.05135, over 955910.33 frames. ], batch size: 25, lr: 3.33e-03, grad_scale: 32.0 2023-04-27 11:38:06,735 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6146, 1.1827, 1.3743, 1.2476, 1.7765, 1.4277, 1.1753, 1.3308], device='cuda:4'), covar=tensor([0.1539, 0.1426, 0.1954, 0.1452, 0.0805, 0.1330, 0.1978, 0.2378], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0311, 0.0349, 0.0286, 0.0327, 0.0307, 0.0300, 0.0368], device='cuda:4'), out_proj_covar=tensor([6.3296e-05, 6.4652e-05, 7.4193e-05, 5.8060e-05, 6.7922e-05, 6.4617e-05, 6.3009e-05, 7.8352e-05], device='cuda:4') 2023-04-27 11:38:08,565 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5456, 2.7935, 1.1940, 1.9161, 1.8349, 2.2298, 1.8781, 1.1875], device='cuda:4'), covar=tensor([0.1107, 0.0879, 0.1555, 0.0993, 0.0928, 0.0800, 0.1239, 0.1569], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0243, 0.0137, 0.0121, 0.0132, 0.0152, 0.0117, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 11:38:09,084 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.783e+01 1.521e+02 1.787e+02 2.150e+02 4.001e+02, threshold=3.574e+02, percent-clipped=1.0 2023-04-27 11:38:30,022 INFO [finetune.py:976] (4/7) Epoch 18, batch 4000, loss[loss=0.1789, simple_loss=0.265, pruned_loss=0.04638, over 4906.00 frames. ], tot_loss[loss=0.1727, simple_loss=0.2427, pruned_loss=0.05137, over 953934.88 frames. ], batch size: 37, lr: 3.33e-03, grad_scale: 32.0 2023-04-27 11:38:52,321 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=101386.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 11:39:35,328 INFO [finetune.py:976] (4/7) Epoch 18, batch 4050, loss[loss=0.1558, simple_loss=0.2314, pruned_loss=0.04006, over 4898.00 frames. ], tot_loss[loss=0.175, simple_loss=0.2448, pruned_loss=0.05262, over 952392.70 frames. ], batch size: 35, lr: 3.33e-03, grad_scale: 32.0 2023-04-27 11:39:45,784 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6352, 1.8536, 2.0115, 2.0947, 2.0109, 2.0655, 2.0585, 2.0705], device='cuda:4'), covar=tensor([0.3790, 0.5591, 0.4711, 0.4732, 0.5572, 0.7597, 0.5553, 0.4864], device='cuda:4'), in_proj_covar=tensor([0.0332, 0.0374, 0.0321, 0.0333, 0.0345, 0.0394, 0.0357, 0.0327], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 11:39:55,130 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=101434.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 11:40:07,832 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.179e+02 1.675e+02 1.988e+02 2.422e+02 4.320e+02, threshold=3.976e+02, percent-clipped=3.0 2023-04-27 11:40:12,816 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=101463.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:40:18,126 INFO [finetune.py:976] (4/7) Epoch 18, batch 4100, loss[loss=0.15, simple_loss=0.2211, pruned_loss=0.03943, over 4793.00 frames. ], tot_loss[loss=0.1772, simple_loss=0.2479, pruned_loss=0.05326, over 953892.62 frames. ], batch size: 25, lr: 3.33e-03, grad_scale: 32.0 2023-04-27 11:40:37,450 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.7263, 2.0733, 2.4958, 3.1826, 2.4653, 2.0216, 1.8229, 2.4517], device='cuda:4'), covar=tensor([0.3285, 0.3229, 0.1586, 0.2257, 0.2715, 0.2698, 0.4005, 0.1899], device='cuda:4'), in_proj_covar=tensor([0.0288, 0.0244, 0.0225, 0.0313, 0.0217, 0.0230, 0.0226, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 11:40:51,380 INFO [finetune.py:976] (4/7) Epoch 18, batch 4150, loss[loss=0.1442, simple_loss=0.2251, pruned_loss=0.0316, over 4772.00 frames. ], tot_loss[loss=0.1776, simple_loss=0.2488, pruned_loss=0.05314, over 953188.82 frames. ], batch size: 26, lr: 3.33e-03, grad_scale: 32.0 2023-04-27 11:40:52,164 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.58 vs. limit=2.0 2023-04-27 11:41:12,059 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9630, 2.6270, 2.3516, 2.1984, 1.5297, 1.5301, 2.3578, 1.4447], device='cuda:4'), covar=tensor([0.1563, 0.1392, 0.1137, 0.1377, 0.2073, 0.1669, 0.0783, 0.1873], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0214, 0.0169, 0.0207, 0.0202, 0.0186, 0.0157, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 11:41:13,340 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4568, 1.6617, 1.8196, 1.9709, 1.8888, 1.9236, 1.9073, 1.8989], device='cuda:4'), covar=tensor([0.3650, 0.6140, 0.4852, 0.4498, 0.5446, 0.7056, 0.5453, 0.5116], device='cuda:4'), in_proj_covar=tensor([0.0331, 0.0373, 0.0320, 0.0332, 0.0343, 0.0393, 0.0356, 0.0326], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 11:41:14,002 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.46 vs. limit=5.0 2023-04-27 11:41:14,402 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.437e+01 1.627e+02 2.025e+02 2.401e+02 3.721e+02, threshold=4.051e+02, percent-clipped=0.0 2023-04-27 11:41:24,217 INFO [finetune.py:976] (4/7) Epoch 18, batch 4200, loss[loss=0.1315, simple_loss=0.2, pruned_loss=0.03149, over 4174.00 frames. ], tot_loss[loss=0.1773, simple_loss=0.2488, pruned_loss=0.0529, over 954167.91 frames. ], batch size: 18, lr: 3.33e-03, grad_scale: 32.0 2023-04-27 11:41:55,739 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5459, 0.7212, 1.4727, 1.9604, 1.6530, 1.4722, 1.4536, 1.4734], device='cuda:4'), covar=tensor([0.3983, 0.5932, 0.5325, 0.5511, 0.5105, 0.6437, 0.6594, 0.7128], device='cuda:4'), in_proj_covar=tensor([0.0426, 0.0409, 0.0501, 0.0505, 0.0455, 0.0480, 0.0487, 0.0492], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 11:41:58,014 INFO [finetune.py:976] (4/7) Epoch 18, batch 4250, loss[loss=0.1934, simple_loss=0.2589, pruned_loss=0.06399, over 4863.00 frames. ], tot_loss[loss=0.1752, simple_loss=0.2466, pruned_loss=0.05188, over 954683.03 frames. ], batch size: 44, lr: 3.33e-03, grad_scale: 32.0 2023-04-27 11:42:21,985 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.254e+01 1.472e+02 1.774e+02 2.178e+02 3.033e+02, threshold=3.548e+02, percent-clipped=0.0 2023-04-27 11:42:31,625 INFO [finetune.py:976] (4/7) Epoch 18, batch 4300, loss[loss=0.1898, simple_loss=0.2542, pruned_loss=0.06276, over 4861.00 frames. ], tot_loss[loss=0.173, simple_loss=0.2439, pruned_loss=0.05105, over 956161.17 frames. ], batch size: 31, lr: 3.33e-03, grad_scale: 32.0 2023-04-27 11:42:40,120 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=101684.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:43:04,507 INFO [finetune.py:976] (4/7) Epoch 18, batch 4350, loss[loss=0.1781, simple_loss=0.2584, pruned_loss=0.04891, over 4842.00 frames. ], tot_loss[loss=0.1713, simple_loss=0.2418, pruned_loss=0.0504, over 957827.27 frames. ], batch size: 47, lr: 3.33e-03, grad_scale: 32.0 2023-04-27 11:43:20,689 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=101745.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:43:32,959 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.731e+01 1.464e+02 1.839e+02 2.155e+02 7.335e+02, threshold=3.679e+02, percent-clipped=2.0 2023-04-27 11:43:43,334 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=101763.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:43:48,147 INFO [finetune.py:976] (4/7) Epoch 18, batch 4400, loss[loss=0.1418, simple_loss=0.2111, pruned_loss=0.03628, over 4807.00 frames. ], tot_loss[loss=0.173, simple_loss=0.2432, pruned_loss=0.05137, over 956739.30 frames. ], batch size: 25, lr: 3.32e-03, grad_scale: 32.0 2023-04-27 11:43:56,408 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.90 vs. limit=2.0 2023-04-27 11:44:37,288 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=101811.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:44:48,685 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4399, 1.9728, 2.1877, 2.9671, 2.3774, 1.9062, 2.0937, 2.2339], device='cuda:4'), covar=tensor([0.2801, 0.3013, 0.1543, 0.2064, 0.2480, 0.2356, 0.3508, 0.2197], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0247, 0.0227, 0.0315, 0.0219, 0.0231, 0.0228, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 11:44:49,150 INFO [finetune.py:976] (4/7) Epoch 18, batch 4450, loss[loss=0.1455, simple_loss=0.232, pruned_loss=0.02949, over 4109.00 frames. ], tot_loss[loss=0.1767, simple_loss=0.2476, pruned_loss=0.0529, over 954444.55 frames. ], batch size: 65, lr: 3.32e-03, grad_scale: 32.0 2023-04-27 11:45:32,214 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.150e+02 1.703e+02 1.979e+02 2.478e+02 5.839e+02, threshold=3.957e+02, percent-clipped=3.0 2023-04-27 11:45:42,464 INFO [finetune.py:976] (4/7) Epoch 18, batch 4500, loss[loss=0.1822, simple_loss=0.2586, pruned_loss=0.05293, over 4898.00 frames. ], tot_loss[loss=0.1776, simple_loss=0.2492, pruned_loss=0.05298, over 954844.83 frames. ], batch size: 35, lr: 3.32e-03, grad_scale: 32.0 2023-04-27 11:46:15,946 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=101920.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 11:46:16,443 INFO [finetune.py:976] (4/7) Epoch 18, batch 4550, loss[loss=0.207, simple_loss=0.2699, pruned_loss=0.07208, over 4805.00 frames. ], tot_loss[loss=0.178, simple_loss=0.2497, pruned_loss=0.05316, over 954301.00 frames. ], batch size: 40, lr: 3.32e-03, grad_scale: 32.0 2023-04-27 11:46:25,176 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2456, 1.6374, 1.5718, 2.0887, 2.3147, 1.9221, 1.8246, 1.6927], device='cuda:4'), covar=tensor([0.1639, 0.1730, 0.1643, 0.1359, 0.1228, 0.1625, 0.2262, 0.2277], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0314, 0.0353, 0.0291, 0.0331, 0.0309, 0.0303, 0.0372], device='cuda:4'), out_proj_covar=tensor([6.3955e-05, 6.5322e-05, 7.4994e-05, 5.9067e-05, 6.8829e-05, 6.5012e-05, 6.3770e-05, 7.9291e-05], device='cuda:4') 2023-04-27 11:46:30,429 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-04-27 11:46:31,173 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1704, 1.4577, 1.4555, 1.7571, 1.6132, 1.9281, 1.3916, 3.7305], device='cuda:4'), covar=tensor([0.0629, 0.0849, 0.0825, 0.1226, 0.0672, 0.0520, 0.0808, 0.0143], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0039, 0.0037, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0012, 0.0013, 0.0016], device='cuda:4') 2023-04-27 11:46:38,530 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.069e+02 1.683e+02 1.915e+02 2.320e+02 4.613e+02, threshold=3.831e+02, percent-clipped=2.0 2023-04-27 11:46:49,881 INFO [finetune.py:976] (4/7) Epoch 18, batch 4600, loss[loss=0.1993, simple_loss=0.2816, pruned_loss=0.05848, over 4905.00 frames. ], tot_loss[loss=0.1762, simple_loss=0.2477, pruned_loss=0.05235, over 952865.69 frames. ], batch size: 37, lr: 3.32e-03, grad_scale: 32.0 2023-04-27 11:46:56,091 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=101981.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 11:46:59,715 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=101987.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:47:14,100 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2516, 1.4082, 1.3081, 1.6662, 1.5282, 1.7235, 1.3179, 3.0281], device='cuda:4'), covar=tensor([0.0612, 0.0818, 0.0829, 0.1183, 0.0663, 0.0513, 0.0776, 0.0164], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 11:47:24,661 INFO [finetune.py:976] (4/7) Epoch 18, batch 4650, loss[loss=0.1405, simple_loss=0.2122, pruned_loss=0.03437, over 4749.00 frames. ], tot_loss[loss=0.1753, simple_loss=0.2457, pruned_loss=0.05239, over 953032.92 frames. ], batch size: 27, lr: 3.32e-03, grad_scale: 32.0 2023-04-27 11:47:26,014 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2612, 2.8706, 2.3823, 2.6710, 2.1078, 2.4643, 2.6965, 1.8121], device='cuda:4'), covar=tensor([0.2014, 0.1300, 0.0667, 0.1042, 0.3007, 0.1058, 0.1806, 0.2546], device='cuda:4'), in_proj_covar=tensor([0.0285, 0.0304, 0.0217, 0.0279, 0.0310, 0.0258, 0.0248, 0.0265], device='cuda:4'), out_proj_covar=tensor([1.1439e-04, 1.2082e-04, 8.6081e-05, 1.1062e-04, 1.2582e-04, 1.0220e-04, 1.0027e-04, 1.0516e-04], device='cuda:4') 2023-04-27 11:47:36,314 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=102040.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:47:41,248 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=102048.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:47:45,411 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.010e+02 1.596e+02 1.902e+02 2.192e+02 3.719e+02, threshold=3.804e+02, percent-clipped=0.0 2023-04-27 11:47:58,103 INFO [finetune.py:976] (4/7) Epoch 18, batch 4700, loss[loss=0.1991, simple_loss=0.254, pruned_loss=0.07211, over 4898.00 frames. ], tot_loss[loss=0.1736, simple_loss=0.2434, pruned_loss=0.05189, over 953693.58 frames. ], batch size: 32, lr: 3.32e-03, grad_scale: 32.0 2023-04-27 11:48:18,213 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.7545, 3.5081, 2.7470, 4.3127, 3.5725, 3.7124, 1.5492, 3.7442], device='cuda:4'), covar=tensor([0.1769, 0.1409, 0.3994, 0.1390, 0.3265, 0.1874, 0.5799, 0.2265], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0215, 0.0251, 0.0304, 0.0299, 0.0250, 0.0273, 0.0272], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 11:48:30,320 INFO [finetune.py:976] (4/7) Epoch 18, batch 4750, loss[loss=0.1568, simple_loss=0.2261, pruned_loss=0.04379, over 4741.00 frames. ], tot_loss[loss=0.1723, simple_loss=0.2416, pruned_loss=0.05149, over 954906.15 frames. ], batch size: 23, lr: 3.32e-03, grad_scale: 32.0 2023-04-27 11:48:50,266 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8833, 1.4759, 1.4411, 1.5508, 2.0623, 1.6192, 1.3806, 1.3963], device='cuda:4'), covar=tensor([0.1423, 0.1270, 0.1872, 0.1265, 0.0785, 0.1437, 0.1890, 0.2044], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0312, 0.0352, 0.0290, 0.0330, 0.0308, 0.0302, 0.0370], device='cuda:4'), out_proj_covar=tensor([6.3364e-05, 6.4759e-05, 7.4857e-05, 5.8864e-05, 6.8523e-05, 6.4773e-05, 6.3523e-05, 7.8886e-05], device='cuda:4') 2023-04-27 11:48:51,944 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.089e+02 1.496e+02 1.838e+02 2.058e+02 4.267e+02, threshold=3.676e+02, percent-clipped=2.0 2023-04-27 11:49:09,304 INFO [finetune.py:976] (4/7) Epoch 18, batch 4800, loss[loss=0.2074, simple_loss=0.2715, pruned_loss=0.07162, over 4815.00 frames. ], tot_loss[loss=0.1755, simple_loss=0.2451, pruned_loss=0.05296, over 955560.97 frames. ], batch size: 51, lr: 3.32e-03, grad_scale: 64.0 2023-04-27 11:49:18,722 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6858, 0.9653, 1.6720, 2.1110, 1.7068, 1.5937, 1.6724, 1.6238], device='cuda:4'), covar=tensor([0.4701, 0.6686, 0.6515, 0.6094, 0.6208, 0.7596, 0.7791, 0.8473], device='cuda:4'), in_proj_covar=tensor([0.0422, 0.0407, 0.0499, 0.0501, 0.0451, 0.0476, 0.0484, 0.0488], device='cuda:4'), out_proj_covar=tensor([1.0171e-04, 9.9959e-05, 1.1213e-04, 1.1951e-04, 1.0805e-04, 1.1427e-04, 1.1464e-04, 1.1523e-04], device='cuda:4') 2023-04-27 11:49:54,663 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.4583, 1.3597, 1.4191, 0.9822, 1.3936, 1.2043, 1.7687, 1.2699], device='cuda:4'), covar=tensor([0.3475, 0.1897, 0.4844, 0.3006, 0.1699, 0.2360, 0.1576, 0.4786], device='cuda:4'), in_proj_covar=tensor([0.0342, 0.0349, 0.0430, 0.0359, 0.0385, 0.0384, 0.0373, 0.0424], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 11:50:15,445 INFO [finetune.py:976] (4/7) Epoch 18, batch 4850, loss[loss=0.1566, simple_loss=0.2327, pruned_loss=0.0403, over 4789.00 frames. ], tot_loss[loss=0.1758, simple_loss=0.2465, pruned_loss=0.05257, over 954964.13 frames. ], batch size: 25, lr: 3.32e-03, grad_scale: 64.0 2023-04-27 11:50:39,406 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-27 11:50:58,315 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.240e+02 1.715e+02 2.079e+02 2.399e+02 4.707e+02, threshold=4.157e+02, percent-clipped=1.0 2023-04-27 11:51:19,609 INFO [finetune.py:976] (4/7) Epoch 18, batch 4900, loss[loss=0.1949, simple_loss=0.2706, pruned_loss=0.05959, over 4844.00 frames. ], tot_loss[loss=0.177, simple_loss=0.248, pruned_loss=0.053, over 953535.26 frames. ], batch size: 44, lr: 3.32e-03, grad_scale: 64.0 2023-04-27 11:51:23,226 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=102276.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 11:51:32,354 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-04-27 11:52:04,393 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1575, 2.4927, 0.8522, 1.5223, 1.4736, 1.9151, 1.5848, 0.8364], device='cuda:4'), covar=tensor([0.1436, 0.1096, 0.1724, 0.1269, 0.1139, 0.0867, 0.1547, 0.1830], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0245, 0.0138, 0.0121, 0.0133, 0.0154, 0.0119, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 11:52:25,540 INFO [finetune.py:976] (4/7) Epoch 18, batch 4950, loss[loss=0.1715, simple_loss=0.2496, pruned_loss=0.04669, over 4893.00 frames. ], tot_loss[loss=0.1771, simple_loss=0.2485, pruned_loss=0.05283, over 954237.81 frames. ], batch size: 43, lr: 3.32e-03, grad_scale: 64.0 2023-04-27 11:52:50,227 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.90 vs. limit=5.0 2023-04-27 11:52:50,467 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=102340.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:52:51,183 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.60 vs. limit=2.0 2023-04-27 11:52:57,587 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=102343.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:53:04,801 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.066e+02 1.629e+02 1.929e+02 2.326e+02 4.533e+02, threshold=3.857e+02, percent-clipped=1.0 2023-04-27 11:53:14,474 INFO [finetune.py:976] (4/7) Epoch 18, batch 5000, loss[loss=0.1373, simple_loss=0.202, pruned_loss=0.03635, over 4932.00 frames. ], tot_loss[loss=0.1763, simple_loss=0.2473, pruned_loss=0.05263, over 954076.40 frames. ], batch size: 38, lr: 3.32e-03, grad_scale: 64.0 2023-04-27 11:53:28,307 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=102388.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:53:48,402 INFO [finetune.py:976] (4/7) Epoch 18, batch 5050, loss[loss=0.1904, simple_loss=0.245, pruned_loss=0.06785, over 4907.00 frames. ], tot_loss[loss=0.1743, simple_loss=0.2449, pruned_loss=0.05189, over 954519.06 frames. ], batch size: 36, lr: 3.32e-03, grad_scale: 64.0 2023-04-27 11:54:23,901 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.725e+01 1.630e+02 1.921e+02 2.242e+02 3.810e+02, threshold=3.842e+02, percent-clipped=0.0 2023-04-27 11:54:41,724 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.87 vs. limit=5.0 2023-04-27 11:54:42,289 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7861, 1.2964, 1.8637, 2.2933, 1.8518, 1.7617, 1.8596, 1.7689], device='cuda:4'), covar=tensor([0.4716, 0.6346, 0.6539, 0.5801, 0.5787, 0.7536, 0.7708, 0.8755], device='cuda:4'), in_proj_covar=tensor([0.0424, 0.0407, 0.0499, 0.0503, 0.0451, 0.0477, 0.0485, 0.0490], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 11:54:45,116 INFO [finetune.py:976] (4/7) Epoch 18, batch 5100, loss[loss=0.1603, simple_loss=0.2281, pruned_loss=0.04623, over 4902.00 frames. ], tot_loss[loss=0.171, simple_loss=0.2412, pruned_loss=0.05043, over 955101.84 frames. ], batch size: 43, lr: 3.32e-03, grad_scale: 64.0 2023-04-27 11:54:50,148 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9865, 1.7967, 2.3682, 2.4697, 1.8397, 1.6200, 1.8888, 1.1022], device='cuda:4'), covar=tensor([0.0635, 0.0815, 0.0392, 0.0779, 0.0857, 0.1062, 0.0883, 0.0793], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0069, 0.0068, 0.0067, 0.0075, 0.0095, 0.0073, 0.0066], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 11:54:53,022 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-27 11:55:18,948 INFO [finetune.py:976] (4/7) Epoch 18, batch 5150, loss[loss=0.1592, simple_loss=0.23, pruned_loss=0.04421, over 4848.00 frames. ], tot_loss[loss=0.1727, simple_loss=0.2424, pruned_loss=0.05151, over 952193.48 frames. ], batch size: 49, lr: 3.32e-03, grad_scale: 64.0 2023-04-27 11:55:42,255 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4750, 1.8507, 1.7191, 2.3520, 2.5758, 1.9912, 1.8892, 1.7295], device='cuda:4'), covar=tensor([0.1723, 0.1721, 0.1853, 0.1947, 0.0973, 0.2068, 0.2097, 0.2422], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0312, 0.0353, 0.0291, 0.0331, 0.0310, 0.0303, 0.0371], device='cuda:4'), out_proj_covar=tensor([6.3885e-05, 6.4890e-05, 7.4814e-05, 5.9048e-05, 6.8666e-05, 6.5177e-05, 6.3788e-05, 7.9124e-05], device='cuda:4') 2023-04-27 11:55:52,830 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.038e+02 1.715e+02 2.103e+02 2.437e+02 3.988e+02, threshold=4.205e+02, percent-clipped=0.0 2023-04-27 11:56:08,073 INFO [finetune.py:976] (4/7) Epoch 18, batch 5200, loss[loss=0.1746, simple_loss=0.2507, pruned_loss=0.04919, over 4855.00 frames. ], tot_loss[loss=0.175, simple_loss=0.2453, pruned_loss=0.05229, over 952338.30 frames. ], batch size: 44, lr: 3.32e-03, grad_scale: 64.0 2023-04-27 11:56:11,161 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=102576.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 11:56:25,967 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.84 vs. limit=5.0 2023-04-27 11:56:27,495 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0835, 2.6915, 2.0966, 2.0947, 1.5208, 1.4798, 2.2919, 1.4646], device='cuda:4'), covar=tensor([0.1494, 0.1455, 0.1343, 0.1696, 0.2194, 0.1785, 0.0937, 0.1955], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0214, 0.0169, 0.0206, 0.0201, 0.0186, 0.0156, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 11:56:42,076 INFO [finetune.py:976] (4/7) Epoch 18, batch 5250, loss[loss=0.2521, simple_loss=0.3148, pruned_loss=0.0947, over 4921.00 frames. ], tot_loss[loss=0.1758, simple_loss=0.2468, pruned_loss=0.05245, over 949928.78 frames. ], batch size: 42, lr: 3.32e-03, grad_scale: 64.0 2023-04-27 11:56:42,804 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1921, 2.5021, 1.0743, 1.3865, 2.0218, 1.2638, 3.3993, 1.8113], device='cuda:4'), covar=tensor([0.0593, 0.0539, 0.0769, 0.1257, 0.0501, 0.1003, 0.0250, 0.0600], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0065, 0.0048, 0.0046, 0.0050, 0.0052, 0.0073, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 11:56:43,956 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=102624.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 11:56:45,170 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7412, 1.5255, 1.7188, 2.0005, 2.1324, 1.6193, 1.4140, 1.8010], device='cuda:4'), covar=tensor([0.0771, 0.1115, 0.0724, 0.0558, 0.0544, 0.0766, 0.0747, 0.0584], device='cuda:4'), in_proj_covar=tensor([0.0189, 0.0201, 0.0182, 0.0171, 0.0177, 0.0181, 0.0150, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 11:56:57,432 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=102643.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:57:11,272 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.104e+02 1.529e+02 1.797e+02 2.306e+02 3.318e+02, threshold=3.594e+02, percent-clipped=0.0 2023-04-27 11:57:26,878 INFO [finetune.py:976] (4/7) Epoch 18, batch 5300, loss[loss=0.1558, simple_loss=0.2357, pruned_loss=0.03798, over 4771.00 frames. ], tot_loss[loss=0.1774, simple_loss=0.249, pruned_loss=0.05291, over 949715.33 frames. ], batch size: 29, lr: 3.32e-03, grad_scale: 64.0 2023-04-27 11:57:55,087 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=102691.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 11:58:31,736 INFO [finetune.py:976] (4/7) Epoch 18, batch 5350, loss[loss=0.1609, simple_loss=0.2387, pruned_loss=0.04154, over 4820.00 frames. ], tot_loss[loss=0.1768, simple_loss=0.2489, pruned_loss=0.0524, over 950497.41 frames. ], batch size: 30, lr: 3.32e-03, grad_scale: 64.0 2023-04-27 11:58:42,061 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.31 vs. limit=5.0 2023-04-27 11:58:50,054 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1631, 1.3867, 1.7386, 1.8388, 1.7568, 1.8069, 1.7437, 1.7644], device='cuda:4'), covar=tensor([0.4232, 0.5223, 0.4184, 0.4254, 0.5280, 0.6964, 0.4975, 0.4582], device='cuda:4'), in_proj_covar=tensor([0.0331, 0.0372, 0.0319, 0.0331, 0.0343, 0.0392, 0.0355, 0.0326], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 11:59:00,913 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0569, 2.7372, 2.1014, 2.0680, 1.5552, 1.4525, 2.2256, 1.4633], device='cuda:4'), covar=tensor([0.1747, 0.1473, 0.1463, 0.1733, 0.2366, 0.2019, 0.1012, 0.2078], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0213, 0.0168, 0.0206, 0.0201, 0.0185, 0.0156, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 11:59:22,080 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.383e+01 1.595e+02 1.837e+02 2.151e+02 4.228e+02, threshold=3.674e+02, percent-clipped=3.0 2023-04-27 11:59:37,387 INFO [finetune.py:976] (4/7) Epoch 18, batch 5400, loss[loss=0.1778, simple_loss=0.2481, pruned_loss=0.05376, over 4823.00 frames. ], tot_loss[loss=0.1735, simple_loss=0.2451, pruned_loss=0.051, over 950908.55 frames. ], batch size: 38, lr: 3.32e-03, grad_scale: 32.0 2023-04-27 12:00:39,502 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.8361, 1.2717, 3.2877, 3.0395, 2.9320, 3.2034, 3.1911, 2.8935], device='cuda:4'), covar=tensor([0.7458, 0.5352, 0.1480, 0.2181, 0.1415, 0.1977, 0.1715, 0.1727], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0311, 0.0411, 0.0413, 0.0355, 0.0412, 0.0317, 0.0370], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 12:00:40,209 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.81 vs. limit=5.0 2023-04-27 12:00:49,559 INFO [finetune.py:976] (4/7) Epoch 18, batch 5450, loss[loss=0.1644, simple_loss=0.2375, pruned_loss=0.04569, over 4864.00 frames. ], tot_loss[loss=0.1715, simple_loss=0.2424, pruned_loss=0.05035, over 949473.07 frames. ], batch size: 44, lr: 3.32e-03, grad_scale: 32.0 2023-04-27 12:01:02,813 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.9226, 4.6340, 3.2685, 5.5538, 4.8481, 4.8746, 2.1156, 4.8155], device='cuda:4'), covar=tensor([0.1523, 0.0944, 0.2740, 0.0797, 0.3093, 0.1493, 0.5630, 0.1975], device='cuda:4'), in_proj_covar=tensor([0.0243, 0.0214, 0.0249, 0.0305, 0.0299, 0.0249, 0.0270, 0.0273], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 12:01:23,560 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=102851.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:01:33,555 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.061e+02 1.567e+02 1.920e+02 2.316e+02 4.548e+02, threshold=3.840e+02, percent-clipped=3.0 2023-04-27 12:01:44,220 INFO [finetune.py:976] (4/7) Epoch 18, batch 5500, loss[loss=0.1688, simple_loss=0.2381, pruned_loss=0.0498, over 4839.00 frames. ], tot_loss[loss=0.1702, simple_loss=0.24, pruned_loss=0.05022, over 948631.48 frames. ], batch size: 47, lr: 3.32e-03, grad_scale: 32.0 2023-04-27 12:01:59,986 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=102887.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:02:10,151 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-04-27 12:02:17,499 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=102912.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:02:19,355 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1403, 1.5878, 1.9984, 2.5449, 2.0422, 1.5725, 1.4326, 1.9074], device='cuda:4'), covar=tensor([0.3663, 0.3695, 0.2155, 0.2749, 0.2776, 0.3167, 0.4322, 0.2202], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0248, 0.0228, 0.0315, 0.0220, 0.0232, 0.0229, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 12:02:23,854 INFO [finetune.py:976] (4/7) Epoch 18, batch 5550, loss[loss=0.193, simple_loss=0.2641, pruned_loss=0.061, over 4801.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2417, pruned_loss=0.0512, over 949056.17 frames. ], batch size: 45, lr: 3.32e-03, grad_scale: 32.0 2023-04-27 12:02:40,497 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=102948.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:02:45,149 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.004e+02 1.541e+02 1.915e+02 2.352e+02 3.794e+02, threshold=3.830e+02, percent-clipped=0.0 2023-04-27 12:02:54,418 INFO [finetune.py:976] (4/7) Epoch 18, batch 5600, loss[loss=0.1992, simple_loss=0.277, pruned_loss=0.06068, over 4899.00 frames. ], tot_loss[loss=0.1747, simple_loss=0.2456, pruned_loss=0.05194, over 950214.74 frames. ], batch size: 36, lr: 3.32e-03, grad_scale: 32.0 2023-04-27 12:03:24,720 INFO [finetune.py:976] (4/7) Epoch 18, batch 5650, loss[loss=0.1813, simple_loss=0.2421, pruned_loss=0.06026, over 4154.00 frames. ], tot_loss[loss=0.1771, simple_loss=0.2487, pruned_loss=0.05272, over 951548.66 frames. ], batch size: 18, lr: 3.31e-03, grad_scale: 32.0 2023-04-27 12:03:36,980 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7496, 1.3503, 1.6089, 1.5867, 1.6278, 1.3186, 0.7994, 1.2919], device='cuda:4'), covar=tensor([0.3033, 0.2800, 0.1515, 0.1900, 0.2263, 0.2400, 0.3749, 0.1842], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0246, 0.0226, 0.0313, 0.0219, 0.0230, 0.0227, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 12:03:46,251 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.096e+02 1.540e+02 1.863e+02 2.247e+02 6.465e+02, threshold=3.725e+02, percent-clipped=2.0 2023-04-27 12:03:55,213 INFO [finetune.py:976] (4/7) Epoch 18, batch 5700, loss[loss=0.1304, simple_loss=0.1888, pruned_loss=0.03599, over 4003.00 frames. ], tot_loss[loss=0.1736, simple_loss=0.244, pruned_loss=0.05157, over 929849.15 frames. ], batch size: 17, lr: 3.31e-03, grad_scale: 32.0 2023-04-27 12:04:24,054 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=103098.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:04:24,531 INFO [finetune.py:976] (4/7) Epoch 19, batch 0, loss[loss=0.1922, simple_loss=0.2559, pruned_loss=0.06427, over 4930.00 frames. ], tot_loss[loss=0.1922, simple_loss=0.2559, pruned_loss=0.06427, over 4930.00 frames. ], batch size: 41, lr: 3.31e-03, grad_scale: 32.0 2023-04-27 12:04:24,532 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 12:04:35,101 INFO [finetune.py:1010] (4/7) Epoch 19, validation: loss=0.1545, simple_loss=0.2248, pruned_loss=0.04209, over 2265189.00 frames. 2023-04-27 12:04:35,101 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-27 12:04:56,256 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=103132.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:05:02,846 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-27 12:05:07,369 INFO [finetune.py:976] (4/7) Epoch 19, batch 50, loss[loss=0.1845, simple_loss=0.2601, pruned_loss=0.0544, over 4903.00 frames. ], tot_loss[loss=0.1707, simple_loss=0.2431, pruned_loss=0.04917, over 217179.23 frames. ], batch size: 46, lr: 3.31e-03, grad_scale: 32.0 2023-04-27 12:05:13,146 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-27 12:05:13,472 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.041e+02 1.545e+02 1.855e+02 2.320e+02 4.324e+02, threshold=3.710e+02, percent-clipped=2.0 2023-04-27 12:05:15,442 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=103159.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:05:26,792 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9204, 0.6615, 0.8007, 0.6866, 1.0784, 0.8755, 0.8262, 0.8070], device='cuda:4'), covar=tensor([0.1319, 0.1383, 0.1651, 0.1626, 0.0959, 0.1250, 0.1356, 0.1701], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0311, 0.0350, 0.0289, 0.0329, 0.0307, 0.0301, 0.0369], device='cuda:4'), out_proj_covar=tensor([6.3235e-05, 6.4583e-05, 7.4181e-05, 5.8513e-05, 6.8346e-05, 6.4503e-05, 6.3268e-05, 7.8537e-05], device='cuda:4') 2023-04-27 12:05:57,694 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=103193.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:06:01,236 INFO [finetune.py:976] (4/7) Epoch 19, batch 100, loss[loss=0.1386, simple_loss=0.1997, pruned_loss=0.03878, over 4161.00 frames. ], tot_loss[loss=0.1733, simple_loss=0.2429, pruned_loss=0.05187, over 378799.59 frames. ], batch size: 65, lr: 3.31e-03, grad_scale: 32.0 2023-04-27 12:06:13,379 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=103207.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:07:01,895 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=103243.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:07:06,073 INFO [finetune.py:976] (4/7) Epoch 19, batch 150, loss[loss=0.1741, simple_loss=0.2353, pruned_loss=0.05645, over 4704.00 frames. ], tot_loss[loss=0.1698, simple_loss=0.2388, pruned_loss=0.05042, over 505850.53 frames. ], batch size: 23, lr: 3.31e-03, grad_scale: 32.0 2023-04-27 12:07:16,418 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.139e+02 1.567e+02 1.850e+02 2.251e+02 4.056e+02, threshold=3.701e+02, percent-clipped=1.0 2023-04-27 12:07:45,053 INFO [finetune.py:976] (4/7) Epoch 19, batch 200, loss[loss=0.1495, simple_loss=0.2219, pruned_loss=0.03855, over 4747.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.2359, pruned_loss=0.04887, over 604155.57 frames. ], batch size: 23, lr: 3.31e-03, grad_scale: 32.0 2023-04-27 12:07:48,058 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2209, 1.6012, 2.1176, 2.6756, 2.1178, 1.6398, 1.5269, 1.9432], device='cuda:4'), covar=tensor([0.3082, 0.3267, 0.1635, 0.2031, 0.2483, 0.2664, 0.4123, 0.1947], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0246, 0.0226, 0.0313, 0.0219, 0.0231, 0.0227, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 12:08:14,336 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0055, 2.5120, 1.0625, 1.4042, 1.8705, 1.2681, 3.0737, 1.6756], device='cuda:4'), covar=tensor([0.0639, 0.0660, 0.0754, 0.1092, 0.0464, 0.0912, 0.0239, 0.0552], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0052, 0.0073, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0009, 0.0007, 0.0007, 0.0008, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-27 12:08:33,912 INFO [finetune.py:976] (4/7) Epoch 19, batch 250, loss[loss=0.1464, simple_loss=0.2348, pruned_loss=0.02897, over 4862.00 frames. ], tot_loss[loss=0.1707, simple_loss=0.2408, pruned_loss=0.05028, over 680392.98 frames. ], batch size: 44, lr: 3.31e-03, grad_scale: 32.0 2023-04-27 12:08:44,148 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.956e+01 1.630e+02 1.978e+02 2.388e+02 4.596e+02, threshold=3.957e+02, percent-clipped=1.0 2023-04-27 12:08:56,543 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=103366.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 12:09:04,227 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3304, 1.4626, 1.3891, 1.9609, 1.6232, 1.7202, 1.4086, 3.1270], device='cuda:4'), covar=tensor([0.0713, 0.1013, 0.1034, 0.1203, 0.0828, 0.0567, 0.0945, 0.0268], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0039, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0014, 0.0013, 0.0012, 0.0013, 0.0016], device='cuda:4') 2023-04-27 12:09:14,458 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.68 vs. limit=5.0 2023-04-27 12:09:22,447 INFO [finetune.py:976] (4/7) Epoch 19, batch 300, loss[loss=0.1671, simple_loss=0.2364, pruned_loss=0.04895, over 4845.00 frames. ], tot_loss[loss=0.1729, simple_loss=0.2438, pruned_loss=0.05098, over 739958.39 frames. ], batch size: 30, lr: 3.31e-03, grad_scale: 32.0 2023-04-27 12:09:42,606 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=103427.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 12:09:55,321 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3060, 1.2211, 1.6299, 1.5585, 1.2182, 1.1190, 1.2723, 0.7895], device='cuda:4'), covar=tensor([0.0486, 0.0667, 0.0353, 0.0447, 0.0776, 0.1039, 0.0547, 0.0577], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0068, 0.0067, 0.0067, 0.0075, 0.0094, 0.0073, 0.0065], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 12:09:55,812 INFO [finetune.py:976] (4/7) Epoch 19, batch 350, loss[loss=0.1993, simple_loss=0.2686, pruned_loss=0.06497, over 4895.00 frames. ], tot_loss[loss=0.1759, simple_loss=0.2474, pruned_loss=0.05224, over 787595.05 frames. ], batch size: 43, lr: 3.31e-03, grad_scale: 32.0 2023-04-27 12:09:59,435 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=103454.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:10:00,583 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.010e+02 1.607e+02 1.938e+02 2.366e+02 5.284e+02, threshold=3.875e+02, percent-clipped=4.0 2023-04-27 12:10:22,421 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=103488.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:10:23,690 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7152, 1.2874, 1.3863, 1.4067, 1.8766, 1.5384, 1.2516, 1.3127], device='cuda:4'), covar=tensor([0.1396, 0.1473, 0.1822, 0.1312, 0.0861, 0.1511, 0.2028, 0.2325], device='cuda:4'), in_proj_covar=tensor([0.0306, 0.0310, 0.0349, 0.0288, 0.0328, 0.0306, 0.0301, 0.0367], device='cuda:4'), out_proj_covar=tensor([6.3112e-05, 6.4461e-05, 7.3942e-05, 5.8483e-05, 6.8205e-05, 6.4228e-05, 6.3217e-05, 7.8292e-05], device='cuda:4') 2023-04-27 12:10:24,898 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0050, 1.9602, 2.4104, 2.6002, 1.9569, 1.8134, 2.0552, 1.0585], device='cuda:4'), covar=tensor([0.0542, 0.0909, 0.0414, 0.0854, 0.0707, 0.1132, 0.0754, 0.0778], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0068, 0.0068, 0.0067, 0.0075, 0.0095, 0.0073, 0.0066], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 12:10:29,035 INFO [finetune.py:976] (4/7) Epoch 19, batch 400, loss[loss=0.1787, simple_loss=0.2484, pruned_loss=0.05454, over 4854.00 frames. ], tot_loss[loss=0.1781, simple_loss=0.2496, pruned_loss=0.05333, over 824048.53 frames. ], batch size: 31, lr: 3.31e-03, grad_scale: 32.0 2023-04-27 12:10:35,035 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=103507.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:10:58,906 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=103543.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:10:58,937 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8290, 1.3152, 1.5164, 1.4841, 1.9666, 1.6139, 1.3115, 1.4895], device='cuda:4'), covar=tensor([0.1473, 0.1350, 0.1804, 0.1420, 0.0827, 0.1395, 0.1807, 0.1971], device='cuda:4'), in_proj_covar=tensor([0.0306, 0.0311, 0.0349, 0.0289, 0.0329, 0.0306, 0.0301, 0.0368], device='cuda:4'), out_proj_covar=tensor([6.3190e-05, 6.4551e-05, 7.4044e-05, 5.8551e-05, 6.8299e-05, 6.4365e-05, 6.3264e-05, 7.8413e-05], device='cuda:4') 2023-04-27 12:11:02,472 INFO [finetune.py:976] (4/7) Epoch 19, batch 450, loss[loss=0.1832, simple_loss=0.2595, pruned_loss=0.0534, over 4883.00 frames. ], tot_loss[loss=0.1774, simple_loss=0.2487, pruned_loss=0.05306, over 853877.01 frames. ], batch size: 32, lr: 3.31e-03, grad_scale: 32.0 2023-04-27 12:11:06,185 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=103555.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:11:06,724 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.103e+02 1.516e+02 1.781e+02 2.073e+02 5.569e+02, threshold=3.563e+02, percent-clipped=1.0 2023-04-27 12:11:42,147 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=103591.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:11:42,297 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-27 12:11:46,989 INFO [finetune.py:976] (4/7) Epoch 19, batch 500, loss[loss=0.1752, simple_loss=0.2407, pruned_loss=0.05485, over 4906.00 frames. ], tot_loss[loss=0.1745, simple_loss=0.2454, pruned_loss=0.05184, over 875598.83 frames. ], batch size: 35, lr: 3.31e-03, grad_scale: 32.0 2023-04-27 12:12:14,025 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.90 vs. limit=2.0 2023-04-27 12:12:16,663 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1001, 3.2101, 0.8728, 1.6297, 1.5294, 2.3562, 1.8007, 1.1682], device='cuda:4'), covar=tensor([0.2048, 0.1685, 0.2504, 0.1848, 0.1476, 0.1287, 0.1670, 0.2053], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0244, 0.0137, 0.0121, 0.0132, 0.0153, 0.0118, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 12:12:28,872 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1778, 2.1994, 1.8334, 1.9689, 2.2765, 1.8116, 2.6857, 1.5113], device='cuda:4'), covar=tensor([0.3613, 0.2068, 0.4404, 0.2798, 0.1659, 0.2418, 0.1246, 0.4645], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0346, 0.0428, 0.0356, 0.0380, 0.0380, 0.0371, 0.0419], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 12:12:30,578 INFO [finetune.py:976] (4/7) Epoch 19, batch 550, loss[loss=0.1694, simple_loss=0.2279, pruned_loss=0.05546, over 4820.00 frames. ], tot_loss[loss=0.1723, simple_loss=0.2423, pruned_loss=0.05117, over 891492.23 frames. ], batch size: 30, lr: 3.31e-03, grad_scale: 32.0 2023-04-27 12:12:34,852 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.077e+02 1.551e+02 1.854e+02 2.248e+02 4.068e+02, threshold=3.707e+02, percent-clipped=1.0 2023-04-27 12:13:04,203 INFO [finetune.py:976] (4/7) Epoch 19, batch 600, loss[loss=0.1917, simple_loss=0.254, pruned_loss=0.06474, over 4831.00 frames. ], tot_loss[loss=0.1713, simple_loss=0.2417, pruned_loss=0.0505, over 907315.65 frames. ], batch size: 30, lr: 3.31e-03, grad_scale: 32.0 2023-04-27 12:13:19,608 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=103722.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 12:13:40,053 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.3288, 3.3925, 2.5620, 3.9344, 3.3211, 3.2770, 1.5811, 3.3936], device='cuda:4'), covar=tensor([0.1913, 0.1323, 0.4037, 0.1908, 0.3234, 0.1901, 0.5403, 0.2642], device='cuda:4'), in_proj_covar=tensor([0.0241, 0.0213, 0.0248, 0.0303, 0.0297, 0.0245, 0.0270, 0.0271], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 12:13:43,214 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=103740.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:13:43,781 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3200, 3.1648, 1.0655, 1.8160, 1.6305, 2.4150, 1.7268, 0.9072], device='cuda:4'), covar=tensor([0.1537, 0.1134, 0.1794, 0.1254, 0.1229, 0.0933, 0.1676, 0.2201], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0244, 0.0138, 0.0121, 0.0133, 0.0154, 0.0119, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 12:13:53,494 INFO [finetune.py:976] (4/7) Epoch 19, batch 650, loss[loss=0.1726, simple_loss=0.247, pruned_loss=0.04915, over 4928.00 frames. ], tot_loss[loss=0.176, simple_loss=0.2472, pruned_loss=0.0524, over 917431.66 frames. ], batch size: 33, lr: 3.31e-03, grad_scale: 32.0 2023-04-27 12:13:54,281 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.59 vs. limit=5.0 2023-04-27 12:14:01,784 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=103754.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:14:02,876 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.117e+02 1.593e+02 1.866e+02 2.212e+02 4.115e+02, threshold=3.733e+02, percent-clipped=2.0 2023-04-27 12:14:13,606 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7265, 1.2710, 1.4134, 1.4667, 1.8709, 1.4933, 1.2186, 1.3717], device='cuda:4'), covar=tensor([0.1536, 0.1427, 0.1911, 0.1238, 0.0877, 0.1640, 0.1962, 0.2309], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0313, 0.0351, 0.0290, 0.0330, 0.0308, 0.0302, 0.0369], device='cuda:4'), out_proj_covar=tensor([6.3449e-05, 6.5049e-05, 7.4312e-05, 5.8751e-05, 6.8537e-05, 6.4742e-05, 6.3495e-05, 7.8602e-05], device='cuda:4') 2023-04-27 12:14:14,716 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=103765.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:14:17,151 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-27 12:14:30,663 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=103788.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:14:37,247 INFO [finetune.py:976] (4/7) Epoch 19, batch 700, loss[loss=0.1641, simple_loss=0.2375, pruned_loss=0.04533, over 4756.00 frames. ], tot_loss[loss=0.1772, simple_loss=0.2486, pruned_loss=0.0529, over 921877.53 frames. ], batch size: 27, lr: 3.31e-03, grad_scale: 32.0 2023-04-27 12:14:38,723 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=103801.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:14:39,255 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=103802.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:14:47,042 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-27 12:14:55,329 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=103826.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:15:02,439 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=103836.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:15:10,879 INFO [finetune.py:976] (4/7) Epoch 19, batch 750, loss[loss=0.1547, simple_loss=0.2436, pruned_loss=0.03289, over 4918.00 frames. ], tot_loss[loss=0.1777, simple_loss=0.249, pruned_loss=0.05322, over 930165.98 frames. ], batch size: 38, lr: 3.30e-03, grad_scale: 32.0 2023-04-27 12:15:15,075 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.139e+02 1.609e+02 1.947e+02 2.389e+02 3.942e+02, threshold=3.894e+02, percent-clipped=2.0 2023-04-27 12:15:28,407 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=103876.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:15:44,008 INFO [finetune.py:976] (4/7) Epoch 19, batch 800, loss[loss=0.1793, simple_loss=0.2545, pruned_loss=0.05203, over 4892.00 frames. ], tot_loss[loss=0.1771, simple_loss=0.2489, pruned_loss=0.0527, over 936521.25 frames. ], batch size: 43, lr: 3.30e-03, grad_scale: 32.0 2023-04-27 12:16:09,187 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=103937.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 12:16:17,416 INFO [finetune.py:976] (4/7) Epoch 19, batch 850, loss[loss=0.1553, simple_loss=0.2238, pruned_loss=0.04341, over 4906.00 frames. ], tot_loss[loss=0.1765, simple_loss=0.2474, pruned_loss=0.05282, over 939130.22 frames. ], batch size: 35, lr: 3.30e-03, grad_scale: 32.0 2023-04-27 12:16:21,648 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.943e+01 1.500e+02 1.730e+02 2.145e+02 3.862e+02, threshold=3.461e+02, percent-clipped=0.0 2023-04-27 12:16:27,887 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.82 vs. limit=5.0 2023-04-27 12:16:31,215 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-04-27 12:16:53,881 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.85 vs. limit=5.0 2023-04-27 12:16:59,789 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=103990.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:17:05,186 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=103997.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:17:10,833 INFO [finetune.py:976] (4/7) Epoch 19, batch 900, loss[loss=0.1869, simple_loss=0.253, pruned_loss=0.06045, over 4825.00 frames. ], tot_loss[loss=0.175, simple_loss=0.2454, pruned_loss=0.05235, over 943744.85 frames. ], batch size: 38, lr: 3.30e-03, grad_scale: 32.0 2023-04-27 12:17:43,376 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=104022.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 12:17:57,249 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-04-27 12:18:05,358 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6267, 2.1522, 2.5242, 3.1319, 2.5036, 2.0092, 2.1163, 2.4389], device='cuda:4'), covar=tensor([0.3365, 0.3232, 0.1674, 0.2513, 0.2818, 0.2698, 0.3673, 0.2189], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0247, 0.0227, 0.0315, 0.0219, 0.0232, 0.0228, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 12:18:18,256 INFO [finetune.py:976] (4/7) Epoch 19, batch 950, loss[loss=0.1473, simple_loss=0.2296, pruned_loss=0.03251, over 4791.00 frames. ], tot_loss[loss=0.1721, simple_loss=0.2422, pruned_loss=0.05102, over 945881.61 frames. ], batch size: 29, lr: 3.30e-03, grad_scale: 32.0 2023-04-27 12:18:18,371 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=104049.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:18:19,613 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=104051.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:18:27,839 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.269e+02 1.627e+02 1.900e+02 2.232e+02 4.587e+02, threshold=3.799e+02, percent-clipped=1.0 2023-04-27 12:18:29,208 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=104058.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:18:30,526 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.80 vs. limit=2.0 2023-04-27 12:18:41,643 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=104070.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 12:19:02,244 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2058, 1.5067, 1.3685, 1.6837, 1.5306, 1.8656, 1.4093, 3.3166], device='cuda:4'), covar=tensor([0.0632, 0.0850, 0.0816, 0.1263, 0.0667, 0.0525, 0.0802, 0.0158], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 12:19:14,758 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=104096.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:19:21,307 INFO [finetune.py:976] (4/7) Epoch 19, batch 1000, loss[loss=0.1768, simple_loss=0.2499, pruned_loss=0.05183, over 4740.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.2446, pruned_loss=0.05161, over 948480.84 frames. ], batch size: 54, lr: 3.30e-03, grad_scale: 32.0 2023-04-27 12:19:33,859 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=104110.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:19:51,328 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=104121.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:19:55,065 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5408, 1.0817, 1.3899, 1.3123, 1.7319, 1.4498, 1.1468, 1.3338], device='cuda:4'), covar=tensor([0.2078, 0.1342, 0.2059, 0.1432, 0.0908, 0.1446, 0.2042, 0.2195], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0312, 0.0350, 0.0291, 0.0330, 0.0308, 0.0301, 0.0368], device='cuda:4'), out_proj_covar=tensor([6.3320e-05, 6.4798e-05, 7.4273e-05, 5.9030e-05, 6.8645e-05, 6.4693e-05, 6.3245e-05, 7.8461e-05], device='cuda:4') 2023-04-27 12:20:05,669 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0934, 2.4794, 1.1010, 1.4329, 1.9778, 1.3948, 3.2722, 1.8034], device='cuda:4'), covar=tensor([0.0602, 0.0589, 0.0703, 0.1239, 0.0458, 0.0902, 0.0320, 0.0626], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0065, 0.0048, 0.0046, 0.0050, 0.0052, 0.0074, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-27 12:20:16,092 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.68 vs. limit=5.0 2023-04-27 12:20:27,132 INFO [finetune.py:976] (4/7) Epoch 19, batch 1050, loss[loss=0.1963, simple_loss=0.2768, pruned_loss=0.05788, over 4899.00 frames. ], tot_loss[loss=0.1749, simple_loss=0.2463, pruned_loss=0.05175, over 949656.45 frames. ], batch size: 37, lr: 3.30e-03, grad_scale: 32.0 2023-04-27 12:20:37,628 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.418e+01 1.570e+02 1.795e+02 2.236e+02 4.469e+02, threshold=3.589e+02, percent-clipped=1.0 2023-04-27 12:20:38,390 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3610, 1.5906, 1.4586, 1.7288, 1.5968, 1.9240, 1.4381, 3.5866], device='cuda:4'), covar=tensor([0.0581, 0.0782, 0.0781, 0.1226, 0.0637, 0.0514, 0.0773, 0.0142], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 12:21:05,398 INFO [finetune.py:976] (4/7) Epoch 19, batch 1100, loss[loss=0.1626, simple_loss=0.2368, pruned_loss=0.04422, over 4839.00 frames. ], tot_loss[loss=0.1768, simple_loss=0.2483, pruned_loss=0.05266, over 951386.28 frames. ], batch size: 49, lr: 3.30e-03, grad_scale: 32.0 2023-04-27 12:21:27,608 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=104232.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 12:21:34,654 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5658, 1.3542, 4.3034, 4.0077, 3.7507, 4.1383, 4.0998, 3.8900], device='cuda:4'), covar=tensor([0.6851, 0.6009, 0.1097, 0.1848, 0.1162, 0.1600, 0.1514, 0.1338], device='cuda:4'), in_proj_covar=tensor([0.0306, 0.0307, 0.0403, 0.0407, 0.0350, 0.0407, 0.0312, 0.0366], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 12:21:39,371 INFO [finetune.py:976] (4/7) Epoch 19, batch 1150, loss[loss=0.1693, simple_loss=0.2271, pruned_loss=0.05577, over 4472.00 frames. ], tot_loss[loss=0.1788, simple_loss=0.2499, pruned_loss=0.0539, over 950651.02 frames. ], batch size: 19, lr: 3.30e-03, grad_scale: 32.0 2023-04-27 12:21:39,476 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8296, 2.3742, 0.9288, 1.2307, 1.6358, 1.1468, 2.8327, 1.4918], device='cuda:4'), covar=tensor([0.0897, 0.0864, 0.0989, 0.1628, 0.0662, 0.1308, 0.0360, 0.0888], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0065, 0.0048, 0.0046, 0.0050, 0.0052, 0.0074, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 12:21:44,616 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.137e+01 1.582e+02 1.911e+02 2.328e+02 3.877e+02, threshold=3.822e+02, percent-clipped=1.0 2023-04-27 12:22:12,706 INFO [finetune.py:976] (4/7) Epoch 19, batch 1200, loss[loss=0.1559, simple_loss=0.2232, pruned_loss=0.04435, over 4885.00 frames. ], tot_loss[loss=0.1759, simple_loss=0.2473, pruned_loss=0.05225, over 951159.55 frames. ], batch size: 32, lr: 3.30e-03, grad_scale: 32.0 2023-04-27 12:22:15,771 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1878, 2.8113, 2.1723, 2.1113, 1.6267, 1.5978, 2.2323, 1.4816], device='cuda:4'), covar=tensor([0.1676, 0.1418, 0.1341, 0.1755, 0.2372, 0.1906, 0.0959, 0.2011], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0212, 0.0168, 0.0206, 0.0201, 0.0185, 0.0156, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 12:22:29,830 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2122, 1.9699, 2.5205, 2.6719, 1.9744, 1.7753, 2.0054, 1.0969], device='cuda:4'), covar=tensor([0.0469, 0.0778, 0.0424, 0.0625, 0.0817, 0.1101, 0.0754, 0.0829], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0068, 0.0067, 0.0067, 0.0074, 0.0095, 0.0073, 0.0066], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 12:22:38,336 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8674, 1.7574, 2.0638, 2.2898, 1.7411, 1.5584, 1.8073, 0.9612], device='cuda:4'), covar=tensor([0.0554, 0.0592, 0.0536, 0.0694, 0.0776, 0.1054, 0.0700, 0.0787], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0068, 0.0067, 0.0067, 0.0075, 0.0095, 0.0074, 0.0066], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 12:22:44,720 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=104346.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:22:46,460 INFO [finetune.py:976] (4/7) Epoch 19, batch 1250, loss[loss=0.1663, simple_loss=0.239, pruned_loss=0.04679, over 4917.00 frames. ], tot_loss[loss=0.1741, simple_loss=0.2452, pruned_loss=0.05146, over 951353.95 frames. ], batch size: 37, lr: 3.30e-03, grad_scale: 32.0 2023-04-27 12:22:49,474 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=104353.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:22:51,236 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.445e+01 1.483e+02 1.801e+02 2.223e+02 4.756e+02, threshold=3.603e+02, percent-clipped=1.0 2023-04-27 12:23:00,949 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.2959, 3.2312, 2.5108, 3.8145, 3.2881, 3.2931, 1.4097, 3.1657], device='cuda:4'), covar=tensor([0.2024, 0.1346, 0.3498, 0.2553, 0.2701, 0.1979, 0.5608, 0.2683], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0215, 0.0249, 0.0308, 0.0299, 0.0248, 0.0272, 0.0274], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 12:23:45,757 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=104396.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:23:47,992 INFO [finetune.py:976] (4/7) Epoch 19, batch 1300, loss[loss=0.1969, simple_loss=0.2528, pruned_loss=0.07048, over 4893.00 frames. ], tot_loss[loss=0.1726, simple_loss=0.243, pruned_loss=0.05109, over 953965.67 frames. ], batch size: 35, lr: 3.30e-03, grad_scale: 32.0 2023-04-27 12:23:56,155 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=104405.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:24:18,874 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=104421.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:24:44,365 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.76 vs. limit=2.0 2023-04-27 12:24:44,514 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=104444.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:24:53,500 INFO [finetune.py:976] (4/7) Epoch 19, batch 1350, loss[loss=0.1912, simple_loss=0.2818, pruned_loss=0.05034, over 4916.00 frames. ], tot_loss[loss=0.1728, simple_loss=0.2433, pruned_loss=0.05113, over 954285.51 frames. ], batch size: 37, lr: 3.30e-03, grad_scale: 32.0 2023-04-27 12:25:03,934 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.093e+02 1.562e+02 1.851e+02 2.195e+02 3.087e+02, threshold=3.702e+02, percent-clipped=0.0 2023-04-27 12:25:23,558 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=104469.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:25:57,941 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2833, 1.7230, 2.1773, 2.6947, 2.2373, 1.7142, 1.5213, 1.9671], device='cuda:4'), covar=tensor([0.3283, 0.3348, 0.1686, 0.2233, 0.2722, 0.2675, 0.4197, 0.2217], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0247, 0.0228, 0.0316, 0.0220, 0.0232, 0.0228, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 12:25:58,404 INFO [finetune.py:976] (4/7) Epoch 19, batch 1400, loss[loss=0.1638, simple_loss=0.2393, pruned_loss=0.04418, over 4897.00 frames. ], tot_loss[loss=0.1753, simple_loss=0.2466, pruned_loss=0.05202, over 954581.96 frames. ], batch size: 35, lr: 3.30e-03, grad_scale: 32.0 2023-04-27 12:26:30,549 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=104532.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 12:26:37,385 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.54 vs. limit=5.0 2023-04-27 12:26:41,376 INFO [finetune.py:976] (4/7) Epoch 19, batch 1450, loss[loss=0.1884, simple_loss=0.2555, pruned_loss=0.06065, over 4825.00 frames. ], tot_loss[loss=0.1775, simple_loss=0.2489, pruned_loss=0.05299, over 954941.85 frames. ], batch size: 30, lr: 3.30e-03, grad_scale: 32.0 2023-04-27 12:26:46,134 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.128e+02 1.618e+02 1.883e+02 2.290e+02 5.063e+02, threshold=3.766e+02, percent-clipped=2.0 2023-04-27 12:27:03,309 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=104580.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:27:14,738 INFO [finetune.py:976] (4/7) Epoch 19, batch 1500, loss[loss=0.1477, simple_loss=0.2233, pruned_loss=0.03605, over 4917.00 frames. ], tot_loss[loss=0.1789, simple_loss=0.2506, pruned_loss=0.05358, over 956543.31 frames. ], batch size: 42, lr: 3.30e-03, grad_scale: 32.0 2023-04-27 12:27:46,361 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-27 12:27:46,866 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=104646.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:27:48,605 INFO [finetune.py:976] (4/7) Epoch 19, batch 1550, loss[loss=0.1647, simple_loss=0.2372, pruned_loss=0.04611, over 4789.00 frames. ], tot_loss[loss=0.1788, simple_loss=0.2504, pruned_loss=0.05363, over 957219.19 frames. ], batch size: 25, lr: 3.30e-03, grad_scale: 32.0 2023-04-27 12:27:51,638 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=104653.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:27:53,361 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.152e+02 1.681e+02 1.907e+02 2.272e+02 3.950e+02, threshold=3.814e+02, percent-clipped=3.0 2023-04-27 12:28:35,129 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=104694.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:28:43,382 INFO [finetune.py:976] (4/7) Epoch 19, batch 1600, loss[loss=0.1551, simple_loss=0.2319, pruned_loss=0.03912, over 4227.00 frames. ], tot_loss[loss=0.1759, simple_loss=0.2471, pruned_loss=0.0524, over 956094.63 frames. ], batch size: 18, lr: 3.30e-03, grad_scale: 32.0 2023-04-27 12:28:44,681 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=104701.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:28:53,300 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=104705.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:29:18,825 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3862, 1.7789, 1.6802, 2.0406, 1.9748, 1.9755, 1.6915, 3.9599], device='cuda:4'), covar=tensor([0.0552, 0.0732, 0.0716, 0.1032, 0.0556, 0.0608, 0.0650, 0.0123], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 12:29:25,256 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-04-27 12:29:27,150 INFO [finetune.py:976] (4/7) Epoch 19, batch 1650, loss[loss=0.154, simple_loss=0.2236, pruned_loss=0.0422, over 4827.00 frames. ], tot_loss[loss=0.174, simple_loss=0.2448, pruned_loss=0.05162, over 956680.89 frames. ], batch size: 38, lr: 3.30e-03, grad_scale: 64.0 2023-04-27 12:29:29,686 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=104753.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:29:31,431 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.063e+01 1.479e+02 1.732e+02 2.090e+02 3.270e+02, threshold=3.465e+02, percent-clipped=0.0 2023-04-27 12:30:01,076 INFO [finetune.py:976] (4/7) Epoch 19, batch 1700, loss[loss=0.2134, simple_loss=0.2702, pruned_loss=0.07834, over 4901.00 frames. ], tot_loss[loss=0.1717, simple_loss=0.2423, pruned_loss=0.0505, over 958319.15 frames. ], batch size: 37, lr: 3.30e-03, grad_scale: 64.0 2023-04-27 12:30:15,451 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9998, 2.1100, 2.0827, 1.7303, 2.0806, 1.8986, 2.6925, 1.7248], device='cuda:4'), covar=tensor([0.3311, 0.1597, 0.3564, 0.2478, 0.1554, 0.2016, 0.1419, 0.4027], device='cuda:4'), in_proj_covar=tensor([0.0341, 0.0348, 0.0428, 0.0354, 0.0383, 0.0381, 0.0373, 0.0420], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 12:30:41,125 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4514, 1.3617, 1.7050, 1.7057, 1.3986, 1.1640, 1.3396, 0.8946], device='cuda:4'), covar=tensor([0.0552, 0.0567, 0.0412, 0.0544, 0.0666, 0.1328, 0.0567, 0.0697], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0068, 0.0067, 0.0067, 0.0074, 0.0095, 0.0073, 0.0066], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 12:30:48,747 INFO [finetune.py:976] (4/7) Epoch 19, batch 1750, loss[loss=0.2197, simple_loss=0.2895, pruned_loss=0.07501, over 4905.00 frames. ], tot_loss[loss=0.1749, simple_loss=0.2452, pruned_loss=0.05227, over 957700.00 frames. ], batch size: 43, lr: 3.30e-03, grad_scale: 64.0 2023-04-27 12:30:52,561 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1892, 1.7171, 2.0370, 2.4056, 2.0395, 1.6160, 1.3364, 1.8467], device='cuda:4'), covar=tensor([0.3001, 0.2955, 0.1511, 0.2044, 0.2313, 0.2491, 0.3988, 0.1909], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0246, 0.0226, 0.0315, 0.0219, 0.0231, 0.0228, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 12:30:53,010 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.077e+02 1.623e+02 1.966e+02 2.378e+02 4.217e+02, threshold=3.932e+02, percent-clipped=2.0 2023-04-27 12:31:37,539 INFO [finetune.py:976] (4/7) Epoch 19, batch 1800, loss[loss=0.1823, simple_loss=0.2621, pruned_loss=0.05128, over 4902.00 frames. ], tot_loss[loss=0.1743, simple_loss=0.2453, pruned_loss=0.05168, over 956792.24 frames. ], batch size: 36, lr: 3.30e-03, grad_scale: 64.0 2023-04-27 12:32:10,889 INFO [finetune.py:976] (4/7) Epoch 19, batch 1850, loss[loss=0.1717, simple_loss=0.2364, pruned_loss=0.05355, over 4734.00 frames. ], tot_loss[loss=0.1768, simple_loss=0.2477, pruned_loss=0.05291, over 955235.09 frames. ], batch size: 54, lr: 3.30e-03, grad_scale: 64.0 2023-04-27 12:32:15,623 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.138e+02 1.634e+02 1.902e+02 2.257e+02 6.149e+02, threshold=3.804e+02, percent-clipped=1.0 2023-04-27 12:32:44,645 INFO [finetune.py:976] (4/7) Epoch 19, batch 1900, loss[loss=0.1697, simple_loss=0.242, pruned_loss=0.04867, over 4784.00 frames. ], tot_loss[loss=0.1784, simple_loss=0.2498, pruned_loss=0.05352, over 957296.42 frames. ], batch size: 25, lr: 3.30e-03, grad_scale: 64.0 2023-04-27 12:33:18,498 INFO [finetune.py:976] (4/7) Epoch 19, batch 1950, loss[loss=0.1611, simple_loss=0.2343, pruned_loss=0.04391, over 4848.00 frames. ], tot_loss[loss=0.176, simple_loss=0.2476, pruned_loss=0.05217, over 958576.60 frames. ], batch size: 49, lr: 3.30e-03, grad_scale: 64.0 2023-04-27 12:33:22,765 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.216e+02 1.554e+02 1.843e+02 2.166e+02 4.298e+02, threshold=3.686e+02, percent-clipped=1.0 2023-04-27 12:34:13,140 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4559, 1.7495, 1.8505, 2.1193, 1.9507, 2.1093, 1.7349, 4.6402], device='cuda:4'), covar=tensor([0.0558, 0.0761, 0.0756, 0.1180, 0.0647, 0.0517, 0.0711, 0.0117], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 12:34:18,468 INFO [finetune.py:976] (4/7) Epoch 19, batch 2000, loss[loss=0.1849, simple_loss=0.2585, pruned_loss=0.05567, over 4760.00 frames. ], tot_loss[loss=0.1732, simple_loss=0.2445, pruned_loss=0.05101, over 958172.70 frames. ], batch size: 54, lr: 3.29e-03, grad_scale: 64.0 2023-04-27 12:35:12,836 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7273, 2.1416, 1.9264, 2.0454, 1.5823, 1.8302, 1.7713, 1.4376], device='cuda:4'), covar=tensor([0.1784, 0.1124, 0.0704, 0.1087, 0.3321, 0.1051, 0.1722, 0.2247], device='cuda:4'), in_proj_covar=tensor([0.0286, 0.0302, 0.0215, 0.0278, 0.0309, 0.0257, 0.0249, 0.0264], device='cuda:4'), out_proj_covar=tensor([1.1499e-04, 1.1981e-04, 8.5392e-05, 1.1048e-04, 1.2546e-04, 1.0190e-04, 1.0060e-04, 1.0465e-04], device='cuda:4') 2023-04-27 12:35:14,515 INFO [finetune.py:976] (4/7) Epoch 19, batch 2050, loss[loss=0.1808, simple_loss=0.2416, pruned_loss=0.06, over 4045.00 frames. ], tot_loss[loss=0.1722, simple_loss=0.2426, pruned_loss=0.05094, over 956009.51 frames. ], batch size: 65, lr: 3.29e-03, grad_scale: 64.0 2023-04-27 12:35:18,783 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.655e+01 1.540e+02 1.837e+02 2.232e+02 4.472e+02, threshold=3.673e+02, percent-clipped=5.0 2023-04-27 12:35:52,487 INFO [finetune.py:976] (4/7) Epoch 19, batch 2100, loss[loss=0.2057, simple_loss=0.278, pruned_loss=0.06671, over 4923.00 frames. ], tot_loss[loss=0.1733, simple_loss=0.2433, pruned_loss=0.05163, over 955022.60 frames. ], batch size: 42, lr: 3.29e-03, grad_scale: 32.0 2023-04-27 12:36:09,485 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3531, 1.7682, 2.1761, 2.7078, 2.1783, 1.7349, 1.5587, 1.9491], device='cuda:4'), covar=tensor([0.3089, 0.3260, 0.1666, 0.2116, 0.2595, 0.2659, 0.3968, 0.2050], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0245, 0.0226, 0.0313, 0.0217, 0.0230, 0.0226, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 12:36:37,776 INFO [finetune.py:976] (4/7) Epoch 19, batch 2150, loss[loss=0.2961, simple_loss=0.3568, pruned_loss=0.1177, over 4859.00 frames. ], tot_loss[loss=0.1768, simple_loss=0.247, pruned_loss=0.05332, over 951705.37 frames. ], batch size: 44, lr: 3.29e-03, grad_scale: 32.0 2023-04-27 12:36:48,986 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.234e+01 1.652e+02 1.972e+02 2.377e+02 5.881e+02, threshold=3.945e+02, percent-clipped=2.0 2023-04-27 12:36:53,471 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1796, 2.1372, 1.7193, 1.8265, 2.2352, 1.7668, 2.7929, 1.4840], device='cuda:4'), covar=tensor([0.3876, 0.2055, 0.5341, 0.3085, 0.1727, 0.2618, 0.1351, 0.4888], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0345, 0.0425, 0.0352, 0.0380, 0.0377, 0.0369, 0.0416], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 12:37:03,655 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0368, 2.5587, 1.1191, 1.3230, 2.0102, 1.2611, 3.4639, 1.7796], device='cuda:4'), covar=tensor([0.0738, 0.0618, 0.0831, 0.1383, 0.0539, 0.1051, 0.0253, 0.0622], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0065, 0.0048, 0.0047, 0.0050, 0.0052, 0.0074, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 12:37:36,680 INFO [finetune.py:976] (4/7) Epoch 19, batch 2200, loss[loss=0.1845, simple_loss=0.2635, pruned_loss=0.05276, over 4787.00 frames. ], tot_loss[loss=0.1788, simple_loss=0.2494, pruned_loss=0.05408, over 952631.11 frames. ], batch size: 51, lr: 3.29e-03, grad_scale: 32.0 2023-04-27 12:37:54,575 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-04-27 12:38:15,462 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.63 vs. limit=5.0 2023-04-27 12:38:49,117 INFO [finetune.py:976] (4/7) Epoch 19, batch 2250, loss[loss=0.1452, simple_loss=0.2334, pruned_loss=0.02854, over 4766.00 frames. ], tot_loss[loss=0.179, simple_loss=0.2499, pruned_loss=0.05405, over 953039.63 frames. ], batch size: 28, lr: 3.29e-03, grad_scale: 32.0 2023-04-27 12:38:59,488 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.067e+02 1.635e+02 1.938e+02 2.374e+02 3.739e+02, threshold=3.876e+02, percent-clipped=0.0 2023-04-27 12:39:34,807 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=105388.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:39:53,432 INFO [finetune.py:976] (4/7) Epoch 19, batch 2300, loss[loss=0.2093, simple_loss=0.2734, pruned_loss=0.07263, over 4809.00 frames. ], tot_loss[loss=0.1791, simple_loss=0.2503, pruned_loss=0.05393, over 954454.17 frames. ], batch size: 40, lr: 3.29e-03, grad_scale: 32.0 2023-04-27 12:40:04,191 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2687, 3.1896, 1.0460, 1.7526, 1.8460, 2.2599, 1.9281, 1.0526], device='cuda:4'), covar=tensor([0.1377, 0.0877, 0.1709, 0.1177, 0.0957, 0.0929, 0.1363, 0.1811], device='cuda:4'), in_proj_covar=tensor([0.0115, 0.0240, 0.0136, 0.0119, 0.0130, 0.0150, 0.0114, 0.0117], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 12:40:28,523 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=5.25 vs. limit=5.0 2023-04-27 12:40:49,860 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-27 12:40:58,677 INFO [finetune.py:976] (4/7) Epoch 19, batch 2350, loss[loss=0.1706, simple_loss=0.2465, pruned_loss=0.04739, over 4804.00 frames. ], tot_loss[loss=0.1777, simple_loss=0.2484, pruned_loss=0.05353, over 953345.41 frames. ], batch size: 41, lr: 3.29e-03, grad_scale: 32.0 2023-04-27 12:40:59,296 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=105449.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 12:41:09,881 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.005e+02 1.591e+02 1.880e+02 2.234e+02 4.098e+02, threshold=3.760e+02, percent-clipped=1.0 2023-04-27 12:41:52,215 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=105486.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:42:06,556 INFO [finetune.py:976] (4/7) Epoch 19, batch 2400, loss[loss=0.1666, simple_loss=0.2253, pruned_loss=0.05393, over 4766.00 frames. ], tot_loss[loss=0.1744, simple_loss=0.2448, pruned_loss=0.05201, over 954763.51 frames. ], batch size: 26, lr: 3.29e-03, grad_scale: 32.0 2023-04-27 12:42:09,006 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=105502.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:42:25,152 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.67 vs. limit=2.0 2023-04-27 12:42:39,427 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=105547.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:42:40,534 INFO [finetune.py:976] (4/7) Epoch 19, batch 2450, loss[loss=0.1857, simple_loss=0.2441, pruned_loss=0.06359, over 4199.00 frames. ], tot_loss[loss=0.1728, simple_loss=0.2427, pruned_loss=0.05149, over 955548.22 frames. ], batch size: 65, lr: 3.29e-03, grad_scale: 32.0 2023-04-27 12:42:45,853 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.021e+02 1.599e+02 1.963e+02 2.373e+02 4.323e+02, threshold=3.926e+02, percent-clipped=3.0 2023-04-27 12:42:50,132 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=105563.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:42:56,460 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.8427, 1.1335, 3.3213, 3.0828, 2.9731, 3.2228, 3.2543, 2.9376], device='cuda:4'), covar=tensor([0.7538, 0.5604, 0.1411, 0.2106, 0.1361, 0.2294, 0.1537, 0.1656], device='cuda:4'), in_proj_covar=tensor([0.0305, 0.0303, 0.0403, 0.0404, 0.0348, 0.0405, 0.0313, 0.0364], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 12:42:58,935 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5859, 3.1780, 0.9987, 1.7496, 2.3854, 1.6520, 4.4925, 2.4722], device='cuda:4'), covar=tensor([0.0608, 0.0626, 0.0872, 0.1289, 0.0533, 0.0998, 0.0296, 0.0553], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0073, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0009, 0.0007, 0.0007, 0.0008, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-27 12:43:14,477 INFO [finetune.py:976] (4/7) Epoch 19, batch 2500, loss[loss=0.2505, simple_loss=0.3183, pruned_loss=0.09139, over 4741.00 frames. ], tot_loss[loss=0.1732, simple_loss=0.243, pruned_loss=0.05173, over 954795.62 frames. ], batch size: 59, lr: 3.29e-03, grad_scale: 32.0 2023-04-27 12:43:46,226 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8158, 1.5505, 4.3776, 4.0612, 3.9319, 4.1459, 4.2238, 3.9925], device='cuda:4'), covar=tensor([0.5885, 0.5790, 0.1175, 0.1907, 0.1178, 0.1795, 0.1019, 0.1637], device='cuda:4'), in_proj_covar=tensor([0.0305, 0.0304, 0.0404, 0.0405, 0.0349, 0.0406, 0.0313, 0.0365], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 12:43:47,976 INFO [finetune.py:976] (4/7) Epoch 19, batch 2550, loss[loss=0.157, simple_loss=0.2446, pruned_loss=0.03473, over 4909.00 frames. ], tot_loss[loss=0.177, simple_loss=0.2478, pruned_loss=0.05313, over 956113.31 frames. ], batch size: 36, lr: 3.29e-03, grad_scale: 32.0 2023-04-27 12:43:53,308 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.125e+02 1.536e+02 1.856e+02 2.246e+02 3.753e+02, threshold=3.711e+02, percent-clipped=0.0 2023-04-27 12:44:27,713 INFO [finetune.py:976] (4/7) Epoch 19, batch 2600, loss[loss=0.2022, simple_loss=0.2783, pruned_loss=0.06301, over 4837.00 frames. ], tot_loss[loss=0.177, simple_loss=0.2484, pruned_loss=0.05281, over 953471.54 frames. ], batch size: 47, lr: 3.29e-03, grad_scale: 32.0 2023-04-27 12:44:58,142 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=105744.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 12:44:58,822 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5016, 2.0266, 2.3967, 3.0270, 2.3561, 1.8572, 1.9412, 2.3298], device='cuda:4'), covar=tensor([0.3257, 0.3126, 0.1550, 0.2202, 0.2587, 0.2582, 0.3634, 0.1906], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0245, 0.0226, 0.0314, 0.0217, 0.0231, 0.0226, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 12:45:01,139 INFO [finetune.py:976] (4/7) Epoch 19, batch 2650, loss[loss=0.1778, simple_loss=0.256, pruned_loss=0.04977, over 4864.00 frames. ], tot_loss[loss=0.1777, simple_loss=0.2498, pruned_loss=0.05285, over 954555.61 frames. ], batch size: 31, lr: 3.29e-03, grad_scale: 32.0 2023-04-27 12:45:06,397 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.206e+02 1.583e+02 1.907e+02 2.231e+02 5.599e+02, threshold=3.814e+02, percent-clipped=2.0 2023-04-27 12:45:26,480 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1613, 1.7756, 1.9816, 2.4198, 2.0718, 1.6847, 1.5245, 1.9316], device='cuda:4'), covar=tensor([0.2247, 0.2543, 0.1367, 0.1632, 0.1984, 0.2132, 0.3902, 0.1770], device='cuda:4'), in_proj_covar=tensor([0.0291, 0.0245, 0.0227, 0.0315, 0.0218, 0.0231, 0.0227, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 12:45:30,667 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=105792.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:45:34,814 INFO [finetune.py:976] (4/7) Epoch 19, batch 2700, loss[loss=0.1563, simple_loss=0.2311, pruned_loss=0.04076, over 4754.00 frames. ], tot_loss[loss=0.1753, simple_loss=0.2475, pruned_loss=0.05154, over 953732.99 frames. ], batch size: 27, lr: 3.29e-03, grad_scale: 32.0 2023-04-27 12:45:40,204 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=105799.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:45:44,467 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0524, 1.3305, 1.2572, 1.5740, 1.4163, 1.5437, 1.2803, 2.4271], device='cuda:4'), covar=tensor([0.0617, 0.0772, 0.0797, 0.1115, 0.0639, 0.0520, 0.0744, 0.0244], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 12:46:05,287 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.33 vs. limit=2.0 2023-04-27 12:46:05,645 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.69 vs. limit=2.0 2023-04-27 12:46:36,275 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=105842.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:46:40,483 INFO [finetune.py:976] (4/7) Epoch 19, batch 2750, loss[loss=0.2072, simple_loss=0.2664, pruned_loss=0.07406, over 4759.00 frames. ], tot_loss[loss=0.1728, simple_loss=0.2443, pruned_loss=0.05071, over 951875.58 frames. ], batch size: 28, lr: 3.29e-03, grad_scale: 32.0 2023-04-27 12:46:43,049 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=105853.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:46:45,897 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.998e+01 1.495e+02 1.843e+02 2.455e+02 4.470e+02, threshold=3.686e+02, percent-clipped=1.0 2023-04-27 12:46:46,580 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=105858.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:46:47,886 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=105860.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:47:27,510 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5536, 3.4514, 1.1969, 1.8537, 1.8461, 2.4870, 2.0654, 1.0235], device='cuda:4'), covar=tensor([0.1319, 0.0961, 0.1744, 0.1246, 0.1111, 0.0938, 0.1429, 0.1987], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0241, 0.0137, 0.0120, 0.0131, 0.0152, 0.0115, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 12:47:35,996 INFO [finetune.py:976] (4/7) Epoch 19, batch 2800, loss[loss=0.1647, simple_loss=0.237, pruned_loss=0.04616, over 4836.00 frames. ], tot_loss[loss=0.1713, simple_loss=0.2421, pruned_loss=0.05024, over 953154.63 frames. ], batch size: 33, lr: 3.29e-03, grad_scale: 32.0 2023-04-27 12:48:29,438 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0593, 2.2487, 0.7611, 1.2481, 1.2815, 1.6008, 1.3947, 0.8431], device='cuda:4'), covar=tensor([0.0890, 0.1110, 0.1281, 0.0883, 0.0731, 0.0756, 0.0999, 0.1166], device='cuda:4'), in_proj_covar=tensor([0.0115, 0.0240, 0.0136, 0.0119, 0.0131, 0.0151, 0.0115, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 12:48:43,334 INFO [finetune.py:976] (4/7) Epoch 19, batch 2850, loss[loss=0.1652, simple_loss=0.2392, pruned_loss=0.04558, over 4706.00 frames. ], tot_loss[loss=0.1717, simple_loss=0.2419, pruned_loss=0.05078, over 952071.32 frames. ], batch size: 23, lr: 3.29e-03, grad_scale: 32.0 2023-04-27 12:48:51,837 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.90 vs. limit=5.0 2023-04-27 12:48:53,316 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.7204, 1.9965, 2.0679, 2.1216, 2.0084, 2.1384, 2.1771, 2.1028], device='cuda:4'), covar=tensor([0.3846, 0.5530, 0.4929, 0.4605, 0.5570, 0.6784, 0.5393, 0.5107], device='cuda:4'), in_proj_covar=tensor([0.0334, 0.0370, 0.0318, 0.0332, 0.0343, 0.0393, 0.0355, 0.0326], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 12:48:53,760 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.059e+02 1.431e+02 1.804e+02 2.150e+02 4.441e+02, threshold=3.609e+02, percent-clipped=4.0 2023-04-27 12:49:49,878 INFO [finetune.py:976] (4/7) Epoch 19, batch 2900, loss[loss=0.187, simple_loss=0.2774, pruned_loss=0.04833, over 4906.00 frames. ], tot_loss[loss=0.1734, simple_loss=0.2443, pruned_loss=0.05128, over 952685.92 frames. ], batch size: 43, lr: 3.29e-03, grad_scale: 32.0 2023-04-27 12:50:40,773 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=106044.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:50:43,675 INFO [finetune.py:976] (4/7) Epoch 19, batch 2950, loss[loss=0.1851, simple_loss=0.2549, pruned_loss=0.05763, over 4883.00 frames. ], tot_loss[loss=0.1756, simple_loss=0.2477, pruned_loss=0.05175, over 953675.74 frames. ], batch size: 32, lr: 3.29e-03, grad_scale: 32.0 2023-04-27 12:50:44,120 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.34 vs. limit=2.0 2023-04-27 12:50:48,559 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.113e+02 1.617e+02 1.806e+02 2.204e+02 5.278e+02, threshold=3.611e+02, percent-clipped=1.0 2023-04-27 12:50:58,102 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2937, 3.0332, 0.9012, 1.5667, 2.0782, 1.4948, 4.0920, 1.9742], device='cuda:4'), covar=tensor([0.0710, 0.0725, 0.0883, 0.1221, 0.0538, 0.0977, 0.0233, 0.0608], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0048, 0.0047, 0.0051, 0.0052, 0.0074, 0.0052], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 12:51:01,777 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=106077.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:51:07,083 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1284, 2.7677, 2.1732, 2.1651, 1.5673, 1.5229, 2.2672, 1.5050], device='cuda:4'), covar=tensor([0.1584, 0.1389, 0.1289, 0.1583, 0.2068, 0.1818, 0.0909, 0.1918], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0210, 0.0168, 0.0203, 0.0198, 0.0183, 0.0155, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 12:51:12,301 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=106092.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:51:17,512 INFO [finetune.py:976] (4/7) Epoch 19, batch 3000, loss[loss=0.1716, simple_loss=0.2478, pruned_loss=0.04774, over 4891.00 frames. ], tot_loss[loss=0.1755, simple_loss=0.2483, pruned_loss=0.05134, over 953565.94 frames. ], batch size: 35, lr: 3.29e-03, grad_scale: 32.0 2023-04-27 12:51:17,512 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 12:51:21,422 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1748, 1.6155, 1.9925, 2.3014, 1.9772, 1.6110, 1.1850, 1.6869], device='cuda:4'), covar=tensor([0.3351, 0.3204, 0.1735, 0.2138, 0.2561, 0.2758, 0.4277, 0.2210], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0245, 0.0227, 0.0315, 0.0218, 0.0231, 0.0227, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 12:51:33,615 INFO [finetune.py:1010] (4/7) Epoch 19, validation: loss=0.1523, simple_loss=0.2226, pruned_loss=0.04099, over 2265189.00 frames. 2023-04-27 12:51:33,616 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-27 12:51:34,949 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6025, 1.3398, 4.4798, 4.1861, 3.9611, 4.3125, 4.2460, 3.9748], device='cuda:4'), covar=tensor([0.7014, 0.6303, 0.0994, 0.1811, 0.1151, 0.1456, 0.1045, 0.1518], device='cuda:4'), in_proj_covar=tensor([0.0303, 0.0301, 0.0400, 0.0402, 0.0346, 0.0401, 0.0309, 0.0362], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 12:51:43,208 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-27 12:52:18,486 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=106138.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:52:26,617 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=106142.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:52:36,646 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=106148.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:52:37,147 INFO [finetune.py:976] (4/7) Epoch 19, batch 3050, loss[loss=0.1765, simple_loss=0.2466, pruned_loss=0.05325, over 4856.00 frames. ], tot_loss[loss=0.1753, simple_loss=0.2485, pruned_loss=0.05102, over 954966.21 frames. ], batch size: 44, lr: 3.29e-03, grad_scale: 32.0 2023-04-27 12:52:37,467 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-27 12:52:46,796 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=106155.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:52:48,481 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.007e+02 1.467e+02 1.745e+02 2.097e+02 4.231e+02, threshold=3.490e+02, percent-clipped=3.0 2023-04-27 12:52:49,215 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=106158.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:52:58,198 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7571, 2.1534, 1.7268, 1.4490, 1.3497, 1.3157, 1.7732, 1.2745], device='cuda:4'), covar=tensor([0.1695, 0.1260, 0.1480, 0.1825, 0.2261, 0.1871, 0.1025, 0.2059], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0211, 0.0167, 0.0203, 0.0198, 0.0183, 0.0154, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 12:53:31,155 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=106190.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:53:42,130 INFO [finetune.py:976] (4/7) Epoch 19, batch 3100, loss[loss=0.1713, simple_loss=0.2418, pruned_loss=0.05038, over 4927.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.2465, pruned_loss=0.05061, over 957594.61 frames. ], batch size: 33, lr: 3.29e-03, grad_scale: 32.0 2023-04-27 12:53:51,785 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=106206.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:54:06,161 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.59 vs. limit=5.0 2023-04-27 12:54:35,398 INFO [finetune.py:976] (4/7) Epoch 19, batch 3150, loss[loss=0.1582, simple_loss=0.2268, pruned_loss=0.0448, over 4831.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2437, pruned_loss=0.05015, over 955889.58 frames. ], batch size: 30, lr: 3.29e-03, grad_scale: 32.0 2023-04-27 12:54:36,070 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4554, 1.6285, 1.8686, 1.9958, 1.9106, 1.9613, 1.9533, 1.9855], device='cuda:4'), covar=tensor([0.3950, 0.5452, 0.4725, 0.4873, 0.5601, 0.7376, 0.5363, 0.4910], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0374, 0.0321, 0.0335, 0.0345, 0.0396, 0.0359, 0.0329], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 12:54:40,720 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.012e+02 1.609e+02 1.934e+02 2.307e+02 5.875e+02, threshold=3.868e+02, percent-clipped=2.0 2023-04-27 12:54:49,432 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=106268.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:55:23,648 INFO [finetune.py:976] (4/7) Epoch 19, batch 3200, loss[loss=0.178, simple_loss=0.2516, pruned_loss=0.05218, over 4862.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2403, pruned_loss=0.04921, over 953898.75 frames. ], batch size: 44, lr: 3.28e-03, grad_scale: 32.0 2023-04-27 12:55:26,269 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-27 12:55:50,417 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=106329.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:55:54,734 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=106336.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:56:03,601 INFO [finetune.py:976] (4/7) Epoch 19, batch 3250, loss[loss=0.1649, simple_loss=0.2423, pruned_loss=0.0438, over 4897.00 frames. ], tot_loss[loss=0.1705, simple_loss=0.2414, pruned_loss=0.04981, over 955649.05 frames. ], batch size: 35, lr: 3.28e-03, grad_scale: 32.0 2023-04-27 12:56:06,822 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9036, 2.5025, 2.0283, 1.9020, 1.4031, 1.4210, 2.1202, 1.3663], device='cuda:4'), covar=tensor([0.1533, 0.1206, 0.1295, 0.1656, 0.2246, 0.1794, 0.0895, 0.1919], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0212, 0.0168, 0.0203, 0.0198, 0.0184, 0.0155, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 12:56:08,528 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.061e+02 1.638e+02 1.854e+02 2.216e+02 4.860e+02, threshold=3.708e+02, percent-clipped=2.0 2023-04-27 12:56:25,634 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.20 vs. limit=5.0 2023-04-27 12:56:35,777 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=106397.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:56:37,430 INFO [finetune.py:976] (4/7) Epoch 19, batch 3300, loss[loss=0.1535, simple_loss=0.2369, pruned_loss=0.03507, over 4792.00 frames. ], tot_loss[loss=0.174, simple_loss=0.2459, pruned_loss=0.05103, over 955622.54 frames. ], batch size: 29, lr: 3.28e-03, grad_scale: 32.0 2023-04-27 12:56:47,330 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5813, 3.7500, 0.8710, 1.9006, 1.9267, 2.5056, 2.0769, 0.9752], device='cuda:4'), covar=tensor([0.1409, 0.0919, 0.2132, 0.1314, 0.1057, 0.1039, 0.1641, 0.1994], device='cuda:4'), in_proj_covar=tensor([0.0115, 0.0239, 0.0136, 0.0118, 0.0130, 0.0150, 0.0114, 0.0117], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 12:57:00,443 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=106433.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:57:02,923 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4771, 1.6467, 1.4501, 1.0644, 1.1550, 1.1221, 1.4554, 1.0744], device='cuda:4'), covar=tensor([0.1902, 0.1389, 0.1490, 0.1903, 0.2282, 0.2061, 0.1043, 0.2158], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0212, 0.0168, 0.0204, 0.0199, 0.0185, 0.0155, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 12:57:09,678 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=106448.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:57:10,219 INFO [finetune.py:976] (4/7) Epoch 19, batch 3350, loss[loss=0.1452, simple_loss=0.2173, pruned_loss=0.03651, over 4619.00 frames. ], tot_loss[loss=0.1752, simple_loss=0.2468, pruned_loss=0.05181, over 952761.09 frames. ], batch size: 20, lr: 3.28e-03, grad_scale: 32.0 2023-04-27 12:57:14,361 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=106454.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:57:14,940 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=106455.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:57:16,073 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.216e+02 1.680e+02 1.880e+02 2.311e+02 4.481e+02, threshold=3.759e+02, percent-clipped=2.0 2023-04-27 12:57:42,143 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=106496.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:57:43,907 INFO [finetune.py:976] (4/7) Epoch 19, batch 3400, loss[loss=0.1749, simple_loss=0.2508, pruned_loss=0.04945, over 4791.00 frames. ], tot_loss[loss=0.1773, simple_loss=0.249, pruned_loss=0.05278, over 954316.59 frames. ], batch size: 29, lr: 3.28e-03, grad_scale: 32.0 2023-04-27 12:57:45,984 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-27 12:57:46,913 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=106503.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:57:55,249 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=106515.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:58:03,197 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.47 vs. limit=5.0 2023-04-27 12:58:17,778 INFO [finetune.py:976] (4/7) Epoch 19, batch 3450, loss[loss=0.1449, simple_loss=0.22, pruned_loss=0.03487, over 4755.00 frames. ], tot_loss[loss=0.1772, simple_loss=0.2484, pruned_loss=0.05295, over 953489.01 frames. ], batch size: 27, lr: 3.28e-03, grad_scale: 32.0 2023-04-27 12:58:23,138 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.135e+02 1.592e+02 1.859e+02 2.190e+02 3.296e+02, threshold=3.717e+02, percent-clipped=0.0 2023-04-27 12:58:57,660 INFO [finetune.py:976] (4/7) Epoch 19, batch 3500, loss[loss=0.2027, simple_loss=0.2589, pruned_loss=0.07325, over 4908.00 frames. ], tot_loss[loss=0.1765, simple_loss=0.2471, pruned_loss=0.05301, over 956629.98 frames. ], batch size: 43, lr: 3.28e-03, grad_scale: 32.0 2023-04-27 12:59:11,522 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.3875, 3.2849, 2.4220, 3.8871, 3.3682, 3.4066, 1.4146, 3.2878], device='cuda:4'), covar=tensor([0.1868, 0.1565, 0.3422, 0.2449, 0.3514, 0.2004, 0.6035, 0.2774], device='cuda:4'), in_proj_covar=tensor([0.0245, 0.0216, 0.0248, 0.0308, 0.0298, 0.0247, 0.0270, 0.0270], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 12:59:14,552 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=106624.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 12:59:36,352 INFO [finetune.py:976] (4/7) Epoch 19, batch 3550, loss[loss=0.147, simple_loss=0.2131, pruned_loss=0.04047, over 4832.00 frames. ], tot_loss[loss=0.1741, simple_loss=0.2442, pruned_loss=0.05203, over 958488.21 frames. ], batch size: 30, lr: 3.28e-03, grad_scale: 32.0 2023-04-27 12:59:41,159 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.386e+01 1.482e+02 1.819e+02 2.170e+02 5.506e+02, threshold=3.638e+02, percent-clipped=1.0 2023-04-27 12:59:45,406 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9801, 1.2745, 4.8491, 4.5511, 4.2488, 4.5579, 4.2750, 4.2972], device='cuda:4'), covar=tensor([0.6582, 0.6112, 0.0935, 0.1686, 0.1082, 0.1406, 0.2019, 0.1550], device='cuda:4'), in_proj_covar=tensor([0.0302, 0.0300, 0.0402, 0.0400, 0.0344, 0.0399, 0.0309, 0.0361], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 12:59:49,568 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.7171, 1.8212, 1.8751, 1.4561, 1.7887, 1.5990, 2.3301, 1.6419], device='cuda:4'), covar=tensor([0.3531, 0.1646, 0.4714, 0.2527, 0.1620, 0.2160, 0.1474, 0.4325], device='cuda:4'), in_proj_covar=tensor([0.0339, 0.0345, 0.0429, 0.0354, 0.0383, 0.0378, 0.0374, 0.0419], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 12:59:57,896 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7204, 1.3347, 1.8459, 2.2515, 1.8436, 1.7089, 1.7855, 1.7566], device='cuda:4'), covar=tensor([0.3984, 0.5412, 0.5111, 0.4702, 0.4807, 0.6566, 0.6598, 0.7527], device='cuda:4'), in_proj_covar=tensor([0.0425, 0.0407, 0.0500, 0.0503, 0.0452, 0.0481, 0.0486, 0.0492], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 13:00:02,388 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.7393, 4.6933, 3.0941, 5.4600, 4.8203, 4.6958, 2.1129, 4.7931], device='cuda:4'), covar=tensor([0.1446, 0.0940, 0.2807, 0.0831, 0.2656, 0.1605, 0.5258, 0.1745], device='cuda:4'), in_proj_covar=tensor([0.0243, 0.0215, 0.0247, 0.0306, 0.0296, 0.0246, 0.0269, 0.0268], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 13:00:05,460 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=106692.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:00:14,734 INFO [finetune.py:976] (4/7) Epoch 19, batch 3600, loss[loss=0.1506, simple_loss=0.2179, pruned_loss=0.04163, over 4836.00 frames. ], tot_loss[loss=0.1725, simple_loss=0.2417, pruned_loss=0.05161, over 957583.21 frames. ], batch size: 33, lr: 3.28e-03, grad_scale: 32.0 2023-04-27 13:00:58,235 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=106733.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:01:14,592 INFO [finetune.py:976] (4/7) Epoch 19, batch 3650, loss[loss=0.1661, simple_loss=0.2255, pruned_loss=0.05338, over 4245.00 frames. ], tot_loss[loss=0.1744, simple_loss=0.2443, pruned_loss=0.05223, over 955450.62 frames. ], batch size: 18, lr: 3.28e-03, grad_scale: 32.0 2023-04-27 13:01:19,419 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.799e+01 1.618e+02 2.033e+02 2.562e+02 4.341e+02, threshold=4.066e+02, percent-clipped=3.0 2023-04-27 13:01:35,806 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=106781.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:01:41,588 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8421, 2.3784, 1.0358, 1.3927, 1.8542, 1.2220, 2.9940, 1.7209], device='cuda:4'), covar=tensor([0.0737, 0.0543, 0.0676, 0.1186, 0.0466, 0.1023, 0.0225, 0.0615], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0065, 0.0048, 0.0047, 0.0050, 0.0052, 0.0074, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 13:01:48,135 INFO [finetune.py:976] (4/7) Epoch 19, batch 3700, loss[loss=0.1758, simple_loss=0.2528, pruned_loss=0.04936, over 4772.00 frames. ], tot_loss[loss=0.1755, simple_loss=0.2467, pruned_loss=0.05213, over 954955.12 frames. ], batch size: 27, lr: 3.28e-03, grad_scale: 32.0 2023-04-27 13:01:55,080 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=106810.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:02:14,806 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4733, 1.8172, 1.7815, 1.9330, 1.8513, 1.9082, 1.8596, 1.8533], device='cuda:4'), covar=tensor([0.3995, 0.5336, 0.4575, 0.4286, 0.5462, 0.7331, 0.5367, 0.5269], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0374, 0.0321, 0.0335, 0.0345, 0.0395, 0.0359, 0.0329], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 13:02:22,253 INFO [finetune.py:976] (4/7) Epoch 19, batch 3750, loss[loss=0.1441, simple_loss=0.2042, pruned_loss=0.04203, over 4729.00 frames. ], tot_loss[loss=0.1762, simple_loss=0.2473, pruned_loss=0.05259, over 956080.34 frames. ], batch size: 23, lr: 3.28e-03, grad_scale: 32.0 2023-04-27 13:02:25,129 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-04-27 13:02:27,101 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.601e+01 1.567e+02 1.830e+02 2.266e+02 4.856e+02, threshold=3.659e+02, percent-clipped=1.0 2023-04-27 13:02:40,829 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-27 13:02:55,383 INFO [finetune.py:976] (4/7) Epoch 19, batch 3800, loss[loss=0.1554, simple_loss=0.2189, pruned_loss=0.04596, over 4721.00 frames. ], tot_loss[loss=0.1778, simple_loss=0.249, pruned_loss=0.05329, over 957416.87 frames. ], batch size: 23, lr: 3.28e-03, grad_scale: 32.0 2023-04-27 13:03:03,477 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.14 vs. limit=2.0 2023-04-27 13:03:10,688 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=106924.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:03:13,273 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-04-27 13:03:15,313 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5947, 1.5766, 0.6379, 1.3169, 1.5693, 1.5003, 1.4191, 1.4646], device='cuda:4'), covar=tensor([0.0502, 0.0367, 0.0383, 0.0548, 0.0286, 0.0503, 0.0494, 0.0544], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0029, 0.0020, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0045, 0.0038, 0.0052, 0.0038, 0.0050, 0.0050, 0.0051], device='cuda:4') 2023-04-27 13:03:27,662 INFO [finetune.py:976] (4/7) Epoch 19, batch 3850, loss[loss=0.1991, simple_loss=0.2625, pruned_loss=0.06785, over 4899.00 frames. ], tot_loss[loss=0.1766, simple_loss=0.2478, pruned_loss=0.05271, over 958873.85 frames. ], batch size: 32, lr: 3.28e-03, grad_scale: 32.0 2023-04-27 13:03:33,088 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.149e+02 1.551e+02 1.789e+02 2.154e+02 4.176e+02, threshold=3.578e+02, percent-clipped=2.0 2023-04-27 13:03:42,316 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=106972.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:03:56,049 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=106992.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:04:00,615 INFO [finetune.py:976] (4/7) Epoch 19, batch 3900, loss[loss=0.1519, simple_loss=0.2329, pruned_loss=0.03545, over 4899.00 frames. ], tot_loss[loss=0.1756, simple_loss=0.2461, pruned_loss=0.05257, over 958369.51 frames. ], batch size: 35, lr: 3.28e-03, grad_scale: 32.0 2023-04-27 13:04:15,059 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3508, 2.0403, 2.4031, 2.7145, 2.7490, 2.2617, 1.7447, 2.5260], device='cuda:4'), covar=tensor([0.0803, 0.1060, 0.0577, 0.0552, 0.0585, 0.0823, 0.0788, 0.0532], device='cuda:4'), in_proj_covar=tensor([0.0186, 0.0200, 0.0182, 0.0172, 0.0175, 0.0181, 0.0150, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 13:04:33,284 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=107040.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:04:41,525 INFO [finetune.py:976] (4/7) Epoch 19, batch 3950, loss[loss=0.1685, simple_loss=0.2415, pruned_loss=0.04781, over 4895.00 frames. ], tot_loss[loss=0.173, simple_loss=0.2431, pruned_loss=0.05145, over 959752.34 frames. ], batch size: 32, lr: 3.28e-03, grad_scale: 32.0 2023-04-27 13:04:54,081 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.559e+01 1.455e+02 1.767e+02 2.125e+02 3.704e+02, threshold=3.534e+02, percent-clipped=1.0 2023-04-27 13:05:26,565 INFO [finetune.py:976] (4/7) Epoch 19, batch 4000, loss[loss=0.1669, simple_loss=0.2532, pruned_loss=0.04027, over 4731.00 frames. ], tot_loss[loss=0.1729, simple_loss=0.243, pruned_loss=0.05142, over 960248.57 frames. ], batch size: 59, lr: 3.28e-03, grad_scale: 32.0 2023-04-27 13:05:35,380 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=107110.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:06:15,751 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=107140.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 13:06:28,916 INFO [finetune.py:976] (4/7) Epoch 19, batch 4050, loss[loss=0.1682, simple_loss=0.2453, pruned_loss=0.04561, over 4817.00 frames. ], tot_loss[loss=0.1771, simple_loss=0.2478, pruned_loss=0.05326, over 960070.10 frames. ], batch size: 51, lr: 3.28e-03, grad_scale: 32.0 2023-04-27 13:06:30,823 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=107152.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:06:35,808 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.479e+01 1.702e+02 1.942e+02 2.364e+02 4.795e+02, threshold=3.885e+02, percent-clipped=4.0 2023-04-27 13:06:36,474 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=107158.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:07:15,574 INFO [finetune.py:976] (4/7) Epoch 19, batch 4100, loss[loss=0.1802, simple_loss=0.2538, pruned_loss=0.05327, over 4830.00 frames. ], tot_loss[loss=0.1784, simple_loss=0.2496, pruned_loss=0.05362, over 957963.20 frames. ], batch size: 49, lr: 3.28e-03, grad_scale: 64.0 2023-04-27 13:07:17,387 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=107201.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 13:07:26,146 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=107213.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 13:07:44,168 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0951, 2.6665, 1.0341, 1.4256, 1.9843, 1.2657, 3.2567, 1.7574], device='cuda:4'), covar=tensor([0.0645, 0.0689, 0.0770, 0.1083, 0.0469, 0.0936, 0.0242, 0.0603], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0065, 0.0048, 0.0047, 0.0050, 0.0052, 0.0074, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 13:07:48,307 INFO [finetune.py:976] (4/7) Epoch 19, batch 4150, loss[loss=0.2081, simple_loss=0.2822, pruned_loss=0.06699, over 4756.00 frames. ], tot_loss[loss=0.1793, simple_loss=0.2509, pruned_loss=0.05388, over 957967.56 frames. ], batch size: 26, lr: 3.28e-03, grad_scale: 64.0 2023-04-27 13:07:54,579 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.951e+01 1.595e+02 1.845e+02 2.201e+02 3.820e+02, threshold=3.690e+02, percent-clipped=0.0 2023-04-27 13:08:03,692 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.45 vs. limit=2.0 2023-04-27 13:08:17,797 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5708, 1.5162, 0.6870, 1.2845, 1.4660, 1.4471, 1.3455, 1.4391], device='cuda:4'), covar=tensor([0.0521, 0.0393, 0.0380, 0.0577, 0.0296, 0.0516, 0.0533, 0.0572], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0029, 0.0019, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0045, 0.0038, 0.0052, 0.0038, 0.0049, 0.0049, 0.0051], device='cuda:4') 2023-04-27 13:08:21,575 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9885, 1.6312, 2.1139, 2.4044, 2.0657, 1.9046, 1.9729, 2.0222], device='cuda:4'), covar=tensor([0.5009, 0.7483, 0.7947, 0.5934, 0.6122, 0.9503, 0.9484, 1.0226], device='cuda:4'), in_proj_covar=tensor([0.0428, 0.0410, 0.0505, 0.0507, 0.0456, 0.0483, 0.0489, 0.0496], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 13:08:22,019 INFO [finetune.py:976] (4/7) Epoch 19, batch 4200, loss[loss=0.1612, simple_loss=0.2366, pruned_loss=0.04288, over 4802.00 frames. ], tot_loss[loss=0.1782, simple_loss=0.2501, pruned_loss=0.05316, over 956098.43 frames. ], batch size: 40, lr: 3.28e-03, grad_scale: 64.0 2023-04-27 13:08:55,959 INFO [finetune.py:976] (4/7) Epoch 19, batch 4250, loss[loss=0.1419, simple_loss=0.2081, pruned_loss=0.03782, over 4696.00 frames. ], tot_loss[loss=0.1752, simple_loss=0.2466, pruned_loss=0.05195, over 957052.58 frames. ], batch size: 23, lr: 3.28e-03, grad_scale: 64.0 2023-04-27 13:09:01,364 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.574e+01 1.612e+02 1.786e+02 2.196e+02 3.792e+02, threshold=3.571e+02, percent-clipped=1.0 2023-04-27 13:09:05,576 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7985, 1.7586, 2.0741, 2.1984, 1.6953, 1.4393, 1.7485, 1.0074], device='cuda:4'), covar=tensor([0.0676, 0.0781, 0.0474, 0.0934, 0.0871, 0.1151, 0.0788, 0.0762], device='cuda:4'), in_proj_covar=tensor([0.0067, 0.0068, 0.0066, 0.0066, 0.0074, 0.0095, 0.0073, 0.0065], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 13:09:29,659 INFO [finetune.py:976] (4/7) Epoch 19, batch 4300, loss[loss=0.2109, simple_loss=0.2835, pruned_loss=0.06917, over 4834.00 frames. ], tot_loss[loss=0.1737, simple_loss=0.2441, pruned_loss=0.05168, over 959186.35 frames. ], batch size: 39, lr: 3.28e-03, grad_scale: 32.0 2023-04-27 13:09:40,817 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-27 13:09:50,816 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2065, 1.6544, 2.0287, 2.4616, 1.9533, 1.6642, 1.2361, 1.7941], device='cuda:4'), covar=tensor([0.3002, 0.3260, 0.1664, 0.2152, 0.2636, 0.2645, 0.4422, 0.2163], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0247, 0.0228, 0.0317, 0.0220, 0.0232, 0.0229, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 13:10:15,946 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9718, 1.4384, 1.7497, 1.7543, 1.6873, 1.4558, 0.8821, 1.4478], device='cuda:4'), covar=tensor([0.2981, 0.3153, 0.1638, 0.1969, 0.2368, 0.2399, 0.3942, 0.1935], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0246, 0.0227, 0.0316, 0.0219, 0.0232, 0.0228, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 13:10:16,563 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7111, 1.9724, 1.8291, 1.4638, 1.2604, 1.3298, 1.9155, 1.2158], device='cuda:4'), covar=tensor([0.1715, 0.1392, 0.1379, 0.1748, 0.2347, 0.2008, 0.0930, 0.2082], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0211, 0.0168, 0.0204, 0.0200, 0.0185, 0.0156, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 13:10:31,834 INFO [finetune.py:976] (4/7) Epoch 19, batch 4350, loss[loss=0.1571, simple_loss=0.2332, pruned_loss=0.04053, over 4763.00 frames. ], tot_loss[loss=0.1724, simple_loss=0.2422, pruned_loss=0.05127, over 958023.77 frames. ], batch size: 28, lr: 3.28e-03, grad_scale: 32.0 2023-04-27 13:10:37,327 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.020e+02 1.596e+02 1.862e+02 2.295e+02 4.192e+02, threshold=3.723e+02, percent-clipped=1.0 2023-04-27 13:10:59,410 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5202, 1.1462, 4.0985, 3.8218, 3.5619, 3.8512, 3.8023, 3.6272], device='cuda:4'), covar=tensor([0.7429, 0.6285, 0.1221, 0.1979, 0.1274, 0.1708, 0.1912, 0.1691], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0304, 0.0408, 0.0406, 0.0350, 0.0408, 0.0314, 0.0367], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 13:11:03,013 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=107496.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 13:11:04,780 INFO [finetune.py:976] (4/7) Epoch 19, batch 4400, loss[loss=0.1725, simple_loss=0.241, pruned_loss=0.052, over 4754.00 frames. ], tot_loss[loss=0.1735, simple_loss=0.2435, pruned_loss=0.05178, over 958390.37 frames. ], batch size: 27, lr: 3.27e-03, grad_scale: 32.0 2023-04-27 13:11:10,415 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=107508.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 13:11:11,649 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=107510.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:11:49,529 INFO [finetune.py:976] (4/7) Epoch 19, batch 4450, loss[loss=0.2187, simple_loss=0.2905, pruned_loss=0.0734, over 4811.00 frames. ], tot_loss[loss=0.1761, simple_loss=0.2473, pruned_loss=0.05247, over 955774.54 frames. ], batch size: 38, lr: 3.27e-03, grad_scale: 32.0 2023-04-27 13:12:00,087 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.014e+02 1.632e+02 1.928e+02 2.326e+02 3.887e+02, threshold=3.856e+02, percent-clipped=1.0 2023-04-27 13:12:19,104 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=107571.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:12:43,863 INFO [finetune.py:976] (4/7) Epoch 19, batch 4500, loss[loss=0.1721, simple_loss=0.2508, pruned_loss=0.04671, over 4859.00 frames. ], tot_loss[loss=0.1765, simple_loss=0.2482, pruned_loss=0.05245, over 955026.66 frames. ], batch size: 44, lr: 3.27e-03, grad_scale: 32.0 2023-04-27 13:13:14,107 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7758, 1.7282, 2.1515, 2.3313, 1.6619, 1.5221, 1.8436, 0.9890], device='cuda:4'), covar=tensor([0.0655, 0.0769, 0.0442, 0.0636, 0.0884, 0.1154, 0.0724, 0.0753], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0068, 0.0067, 0.0067, 0.0075, 0.0096, 0.0073, 0.0065], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 13:13:16,556 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4367, 1.6817, 1.8111, 1.9092, 1.7803, 1.8369, 1.9128, 1.8905], device='cuda:4'), covar=tensor([0.3802, 0.5675, 0.4487, 0.4450, 0.5700, 0.7740, 0.5155, 0.4832], device='cuda:4'), in_proj_covar=tensor([0.0336, 0.0372, 0.0320, 0.0334, 0.0345, 0.0395, 0.0357, 0.0327], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 13:13:17,594 INFO [finetune.py:976] (4/7) Epoch 19, batch 4550, loss[loss=0.1716, simple_loss=0.2525, pruned_loss=0.0453, over 4884.00 frames. ], tot_loss[loss=0.1783, simple_loss=0.25, pruned_loss=0.05334, over 955763.11 frames. ], batch size: 43, lr: 3.27e-03, grad_scale: 32.0 2023-04-27 13:13:23,030 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.130e+02 1.628e+02 1.814e+02 2.190e+02 5.311e+02, threshold=3.629e+02, percent-clipped=2.0 2023-04-27 13:13:23,749 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=107659.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:13:51,123 INFO [finetune.py:976] (4/7) Epoch 19, batch 4600, loss[loss=0.1586, simple_loss=0.2341, pruned_loss=0.04151, over 4839.00 frames. ], tot_loss[loss=0.1781, simple_loss=0.2497, pruned_loss=0.05324, over 955993.76 frames. ], batch size: 49, lr: 3.27e-03, grad_scale: 32.0 2023-04-27 13:14:04,034 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=107720.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:14:07,543 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4892, 3.5594, 0.9077, 1.9818, 1.9279, 2.3949, 2.0075, 1.0741], device='cuda:4'), covar=tensor([0.1361, 0.0747, 0.1954, 0.1171, 0.1030, 0.0981, 0.1426, 0.1980], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0242, 0.0138, 0.0119, 0.0132, 0.0152, 0.0117, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 13:14:16,204 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=107737.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:14:24,299 INFO [finetune.py:976] (4/7) Epoch 19, batch 4650, loss[loss=0.1635, simple_loss=0.2221, pruned_loss=0.05249, over 4829.00 frames. ], tot_loss[loss=0.1753, simple_loss=0.2463, pruned_loss=0.05219, over 955910.51 frames. ], batch size: 25, lr: 3.27e-03, grad_scale: 32.0 2023-04-27 13:14:29,768 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.154e+02 1.588e+02 1.876e+02 2.187e+02 4.173e+02, threshold=3.752e+02, percent-clipped=2.0 2023-04-27 13:14:36,599 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2406, 2.8368, 2.3081, 2.6708, 2.1787, 2.2904, 2.5117, 1.7243], device='cuda:4'), covar=tensor([0.2086, 0.1272, 0.0763, 0.1239, 0.2663, 0.1215, 0.1962, 0.2930], device='cuda:4'), in_proj_covar=tensor([0.0288, 0.0303, 0.0217, 0.0279, 0.0312, 0.0260, 0.0251, 0.0266], device='cuda:4'), out_proj_covar=tensor([1.1543e-04, 1.2027e-04, 8.5949e-05, 1.1051e-04, 1.2678e-04, 1.0300e-04, 1.0142e-04, 1.0557e-04], device='cuda:4') 2023-04-27 13:15:06,656 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=107796.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 13:15:07,932 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=107798.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 13:15:08,402 INFO [finetune.py:976] (4/7) Epoch 19, batch 4700, loss[loss=0.178, simple_loss=0.242, pruned_loss=0.05703, over 4754.00 frames. ], tot_loss[loss=0.1715, simple_loss=0.2421, pruned_loss=0.05044, over 956666.01 frames. ], batch size: 27, lr: 3.27e-03, grad_scale: 32.0 2023-04-27 13:15:19,602 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=107808.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:15:44,930 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1962, 2.7123, 1.2612, 1.5497, 2.2048, 1.4510, 3.2286, 1.7639], device='cuda:4'), covar=tensor([0.0591, 0.0663, 0.0748, 0.0928, 0.0372, 0.0835, 0.0173, 0.0554], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0065, 0.0048, 0.0046, 0.0050, 0.0052, 0.0073, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 13:15:48,352 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=107844.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 13:15:52,874 INFO [finetune.py:976] (4/7) Epoch 19, batch 4750, loss[loss=0.1645, simple_loss=0.2364, pruned_loss=0.04628, over 4900.00 frames. ], tot_loss[loss=0.1701, simple_loss=0.2403, pruned_loss=0.04991, over 955970.74 frames. ], batch size: 35, lr: 3.27e-03, grad_scale: 32.0 2023-04-27 13:15:54,811 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.7929, 2.2329, 2.0629, 2.1610, 2.0310, 2.0796, 2.1645, 2.0893], device='cuda:4'), covar=tensor([0.3755, 0.5556, 0.5281, 0.4453, 0.5737, 0.7005, 0.6189, 0.5710], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0372, 0.0321, 0.0335, 0.0345, 0.0393, 0.0358, 0.0328], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 13:15:57,097 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=107856.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:15:58,278 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.851e+01 1.548e+02 1.806e+02 2.304e+02 4.398e+02, threshold=3.612e+02, percent-clipped=1.0 2023-04-27 13:16:09,256 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=107866.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:16:23,232 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.21 vs. limit=5.0 2023-04-27 13:16:53,224 INFO [finetune.py:976] (4/7) Epoch 19, batch 4800, loss[loss=0.1841, simple_loss=0.2358, pruned_loss=0.06617, over 4212.00 frames. ], tot_loss[loss=0.1712, simple_loss=0.2418, pruned_loss=0.0503, over 955310.07 frames. ], batch size: 18, lr: 3.27e-03, grad_scale: 32.0 2023-04-27 13:17:30,282 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.4389, 1.3547, 1.4505, 1.0624, 1.4326, 1.2978, 1.7585, 1.3907], device='cuda:4'), covar=tensor([0.3894, 0.1979, 0.4932, 0.2713, 0.1639, 0.2355, 0.1821, 0.4649], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0345, 0.0424, 0.0352, 0.0380, 0.0376, 0.0370, 0.0414], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 13:17:31,834 INFO [finetune.py:976] (4/7) Epoch 19, batch 4850, loss[loss=0.1886, simple_loss=0.2403, pruned_loss=0.0685, over 4272.00 frames. ], tot_loss[loss=0.176, simple_loss=0.2468, pruned_loss=0.05258, over 955042.05 frames. ], batch size: 18, lr: 3.27e-03, grad_scale: 32.0 2023-04-27 13:17:37,758 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.098e+02 1.607e+02 1.917e+02 2.218e+02 6.571e+02, threshold=3.834e+02, percent-clipped=2.0 2023-04-27 13:17:39,066 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3699, 1.6037, 1.4425, 2.0295, 1.8114, 2.0408, 1.4693, 4.4429], device='cuda:4'), covar=tensor([0.0697, 0.1089, 0.1103, 0.1416, 0.0809, 0.0686, 0.1054, 0.0178], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 13:18:04,092 INFO [finetune.py:976] (4/7) Epoch 19, batch 4900, loss[loss=0.1977, simple_loss=0.2781, pruned_loss=0.05862, over 4777.00 frames. ], tot_loss[loss=0.1777, simple_loss=0.2492, pruned_loss=0.05314, over 957011.74 frames. ], batch size: 51, lr: 3.27e-03, grad_scale: 32.0 2023-04-27 13:18:16,726 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=108015.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:18:16,740 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5384, 3.0528, 1.4195, 1.8128, 2.5045, 1.7663, 3.7423, 2.2719], device='cuda:4'), covar=tensor([0.0563, 0.0543, 0.0669, 0.1033, 0.0401, 0.0806, 0.0262, 0.0517], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0065, 0.0048, 0.0046, 0.0050, 0.0052, 0.0073, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 13:18:38,322 INFO [finetune.py:976] (4/7) Epoch 19, batch 4950, loss[loss=0.1947, simple_loss=0.2646, pruned_loss=0.06243, over 4819.00 frames. ], tot_loss[loss=0.1792, simple_loss=0.2509, pruned_loss=0.05376, over 955647.28 frames. ], batch size: 40, lr: 3.27e-03, grad_scale: 32.0 2023-04-27 13:18:45,329 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.105e+02 1.581e+02 1.882e+02 2.406e+02 4.521e+02, threshold=3.765e+02, percent-clipped=4.0 2023-04-27 13:19:01,894 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.11 vs. limit=5.0 2023-04-27 13:19:07,764 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=108093.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 13:19:11,336 INFO [finetune.py:976] (4/7) Epoch 19, batch 5000, loss[loss=0.1628, simple_loss=0.2342, pruned_loss=0.04566, over 4892.00 frames. ], tot_loss[loss=0.1782, simple_loss=0.2494, pruned_loss=0.05355, over 955207.91 frames. ], batch size: 35, lr: 3.27e-03, grad_scale: 32.0 2023-04-27 13:19:37,027 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.9667, 2.1946, 2.1334, 2.1875, 2.0278, 2.1045, 2.3277, 2.1986], device='cuda:4'), covar=tensor([0.3425, 0.5425, 0.4637, 0.4277, 0.5276, 0.6789, 0.4994, 0.5142], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0373, 0.0322, 0.0335, 0.0345, 0.0394, 0.0358, 0.0328], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 13:19:44,759 INFO [finetune.py:976] (4/7) Epoch 19, batch 5050, loss[loss=0.1689, simple_loss=0.2249, pruned_loss=0.05641, over 3985.00 frames. ], tot_loss[loss=0.1758, simple_loss=0.2465, pruned_loss=0.05259, over 954616.24 frames. ], batch size: 17, lr: 3.27e-03, grad_scale: 32.0 2023-04-27 13:19:46,103 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0387, 1.4957, 1.6017, 1.8835, 2.1769, 1.8051, 1.5435, 1.4956], device='cuda:4'), covar=tensor([0.1468, 0.1864, 0.2106, 0.1476, 0.1206, 0.1983, 0.2105, 0.2354], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0313, 0.0352, 0.0291, 0.0329, 0.0309, 0.0300, 0.0372], device='cuda:4'), out_proj_covar=tensor([6.3797e-05, 6.4925e-05, 7.4542e-05, 5.8903e-05, 6.8234e-05, 6.4828e-05, 6.2934e-05, 7.9251e-05], device='cuda:4') 2023-04-27 13:19:51,164 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.142e+01 1.502e+02 1.806e+02 2.114e+02 4.959e+02, threshold=3.611e+02, percent-clipped=2.0 2023-04-27 13:19:57,093 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=108166.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:20:29,632 INFO [finetune.py:976] (4/7) Epoch 19, batch 5100, loss[loss=0.1627, simple_loss=0.235, pruned_loss=0.04516, over 4830.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2424, pruned_loss=0.05085, over 955773.44 frames. ], batch size: 39, lr: 3.27e-03, grad_scale: 32.0 2023-04-27 13:20:39,735 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5517, 2.0636, 2.3285, 2.7460, 2.6846, 2.3160, 1.9581, 2.5579], device='cuda:4'), covar=tensor([0.0696, 0.1142, 0.0675, 0.0540, 0.0593, 0.0823, 0.0734, 0.0514], device='cuda:4'), in_proj_covar=tensor([0.0189, 0.0204, 0.0184, 0.0174, 0.0179, 0.0183, 0.0153, 0.0182], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 13:20:50,603 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=108214.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:20:58,548 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.67 vs. limit=5.0 2023-04-27 13:21:13,453 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.16 vs. limit=5.0 2023-04-27 13:21:34,995 INFO [finetune.py:976] (4/7) Epoch 19, batch 5150, loss[loss=0.2072, simple_loss=0.2779, pruned_loss=0.06824, over 4818.00 frames. ], tot_loss[loss=0.1704, simple_loss=0.2405, pruned_loss=0.05017, over 953518.85 frames. ], batch size: 40, lr: 3.27e-03, grad_scale: 32.0 2023-04-27 13:21:44,948 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.855e+01 1.529e+02 1.881e+02 2.347e+02 3.720e+02, threshold=3.762e+02, percent-clipped=1.0 2023-04-27 13:22:28,072 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2340, 1.6008, 2.0294, 2.6181, 2.0482, 1.5694, 1.3553, 1.9533], device='cuda:4'), covar=tensor([0.3115, 0.3333, 0.1755, 0.2150, 0.2810, 0.2789, 0.4308, 0.1954], device='cuda:4'), in_proj_covar=tensor([0.0291, 0.0246, 0.0227, 0.0316, 0.0219, 0.0232, 0.0227, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 13:22:29,164 INFO [finetune.py:976] (4/7) Epoch 19, batch 5200, loss[loss=0.2217, simple_loss=0.2968, pruned_loss=0.07328, over 4811.00 frames. ], tot_loss[loss=0.1736, simple_loss=0.2449, pruned_loss=0.05116, over 951754.08 frames. ], batch size: 51, lr: 3.27e-03, grad_scale: 32.0 2023-04-27 13:22:32,973 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.7940, 2.0025, 1.9632, 2.1321, 1.9255, 2.0077, 2.0817, 2.0170], device='cuda:4'), covar=tensor([0.4086, 0.6425, 0.5270, 0.4598, 0.6411, 0.7751, 0.6606, 0.6205], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0373, 0.0322, 0.0335, 0.0345, 0.0395, 0.0358, 0.0329], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 13:22:51,813 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=108315.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:23:05,671 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=108325.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:23:13,351 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.89 vs. limit=5.0 2023-04-27 13:23:36,576 INFO [finetune.py:976] (4/7) Epoch 19, batch 5250, loss[loss=0.1563, simple_loss=0.2421, pruned_loss=0.03525, over 4913.00 frames. ], tot_loss[loss=0.1768, simple_loss=0.2486, pruned_loss=0.05246, over 954462.89 frames. ], batch size: 36, lr: 3.27e-03, grad_scale: 32.0 2023-04-27 13:23:44,340 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.109e+02 1.684e+02 2.011e+02 2.395e+02 3.683e+02, threshold=4.022e+02, percent-clipped=0.0 2023-04-27 13:23:53,519 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=108363.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:24:19,481 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=108386.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:24:23,706 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=108393.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:24:27,188 INFO [finetune.py:976] (4/7) Epoch 19, batch 5300, loss[loss=0.1509, simple_loss=0.2372, pruned_loss=0.03225, over 4757.00 frames. ], tot_loss[loss=0.1788, simple_loss=0.2508, pruned_loss=0.05346, over 955631.02 frames. ], batch size: 27, lr: 3.27e-03, grad_scale: 32.0 2023-04-27 13:24:50,797 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=108432.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:24:56,167 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=108441.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:25:01,039 INFO [finetune.py:976] (4/7) Epoch 19, batch 5350, loss[loss=0.1795, simple_loss=0.2514, pruned_loss=0.05383, over 4863.00 frames. ], tot_loss[loss=0.1768, simple_loss=0.2488, pruned_loss=0.05241, over 953679.14 frames. ], batch size: 31, lr: 3.27e-03, grad_scale: 32.0 2023-04-27 13:25:06,497 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.103e+02 1.659e+02 1.988e+02 2.329e+02 5.055e+02, threshold=3.977e+02, percent-clipped=1.0 2023-04-27 13:25:27,051 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9581, 1.4235, 1.4707, 1.7684, 2.1419, 1.7421, 1.4959, 1.4179], device='cuda:4'), covar=tensor([0.1625, 0.1572, 0.2050, 0.1429, 0.0867, 0.1661, 0.1995, 0.2305], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0313, 0.0350, 0.0290, 0.0328, 0.0308, 0.0300, 0.0371], device='cuda:4'), out_proj_covar=tensor([6.3765e-05, 6.4894e-05, 7.4143e-05, 5.8609e-05, 6.7968e-05, 6.4730e-05, 6.2859e-05, 7.9086e-05], device='cuda:4') 2023-04-27 13:25:31,255 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=108493.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 13:25:31,846 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7314, 1.7569, 2.1288, 2.2914, 1.6816, 1.4114, 1.7994, 0.9850], device='cuda:4'), covar=tensor([0.0612, 0.0838, 0.0497, 0.0821, 0.0893, 0.1241, 0.0983, 0.0816], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0069, 0.0067, 0.0067, 0.0075, 0.0097, 0.0073, 0.0066], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 13:25:34,726 INFO [finetune.py:976] (4/7) Epoch 19, batch 5400, loss[loss=0.1844, simple_loss=0.2446, pruned_loss=0.06211, over 4829.00 frames. ], tot_loss[loss=0.1748, simple_loss=0.2463, pruned_loss=0.05162, over 953685.71 frames. ], batch size: 25, lr: 3.27e-03, grad_scale: 32.0 2023-04-27 13:25:36,128 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9911, 1.5569, 1.5593, 1.7321, 2.1374, 1.7430, 1.5120, 1.4411], device='cuda:4'), covar=tensor([0.1421, 0.1421, 0.1958, 0.1274, 0.0833, 0.1574, 0.1839, 0.2349], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0312, 0.0350, 0.0289, 0.0327, 0.0308, 0.0299, 0.0371], device='cuda:4'), out_proj_covar=tensor([6.3657e-05, 6.4780e-05, 7.4033e-05, 5.8519e-05, 6.7895e-05, 6.4649e-05, 6.2732e-05, 7.8940e-05], device='cuda:4') 2023-04-27 13:26:05,906 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-27 13:26:09,393 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3548, 1.2669, 1.5747, 1.5684, 1.2348, 1.1555, 1.2659, 0.8450], device='cuda:4'), covar=tensor([0.0521, 0.0659, 0.0343, 0.0618, 0.0781, 0.0978, 0.0538, 0.0515], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0068, 0.0067, 0.0067, 0.0074, 0.0096, 0.0073, 0.0065], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 13:26:19,346 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0933, 2.0287, 1.7890, 1.7574, 2.1343, 1.6791, 2.4848, 1.5190], device='cuda:4'), covar=tensor([0.3511, 0.1875, 0.4288, 0.2795, 0.1550, 0.2465, 0.1445, 0.4182], device='cuda:4'), in_proj_covar=tensor([0.0338, 0.0345, 0.0426, 0.0354, 0.0382, 0.0377, 0.0372, 0.0417], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 13:26:31,591 INFO [finetune.py:976] (4/7) Epoch 19, batch 5450, loss[loss=0.1643, simple_loss=0.2339, pruned_loss=0.0473, over 4939.00 frames. ], tot_loss[loss=0.1724, simple_loss=0.2432, pruned_loss=0.05081, over 954737.89 frames. ], batch size: 33, lr: 3.27e-03, grad_scale: 32.0 2023-04-27 13:26:42,381 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.852e+01 1.533e+02 1.903e+02 2.214e+02 4.592e+02, threshold=3.806e+02, percent-clipped=2.0 2023-04-27 13:27:00,283 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.6975, 1.5908, 1.7211, 1.3136, 1.7271, 1.3668, 2.0819, 1.4775], device='cuda:4'), covar=tensor([0.3418, 0.1796, 0.3961, 0.2772, 0.1411, 0.2346, 0.1837, 0.4039], device='cuda:4'), in_proj_covar=tensor([0.0339, 0.0345, 0.0427, 0.0354, 0.0383, 0.0378, 0.0373, 0.0418], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 13:27:04,689 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.20 vs. limit=5.0 2023-04-27 13:27:10,224 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=108577.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:27:25,457 INFO [finetune.py:976] (4/7) Epoch 19, batch 5500, loss[loss=0.1667, simple_loss=0.2487, pruned_loss=0.04232, over 4760.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2411, pruned_loss=0.05055, over 954904.86 frames. ], batch size: 54, lr: 3.27e-03, grad_scale: 32.0 2023-04-27 13:27:35,788 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.59 vs. limit=5.0 2023-04-27 13:27:51,049 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=108638.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:27:58,828 INFO [finetune.py:976] (4/7) Epoch 19, batch 5550, loss[loss=0.1723, simple_loss=0.2489, pruned_loss=0.04783, over 4841.00 frames. ], tot_loss[loss=0.1716, simple_loss=0.2421, pruned_loss=0.05055, over 955953.10 frames. ], batch size: 49, lr: 3.27e-03, grad_scale: 32.0 2023-04-27 13:28:09,998 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.131e+01 1.592e+02 1.937e+02 2.305e+02 3.019e+02, threshold=3.873e+02, percent-clipped=0.0 2023-04-27 13:28:41,417 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=108681.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:29:03,284 INFO [finetune.py:976] (4/7) Epoch 19, batch 5600, loss[loss=0.1908, simple_loss=0.2588, pruned_loss=0.06134, over 4916.00 frames. ], tot_loss[loss=0.1745, simple_loss=0.2456, pruned_loss=0.0517, over 953468.13 frames. ], batch size: 42, lr: 3.27e-03, grad_scale: 32.0 2023-04-27 13:29:12,630 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0289, 1.6995, 1.9942, 2.4126, 2.3066, 1.9790, 1.6452, 2.1986], device='cuda:4'), covar=tensor([0.0964, 0.1196, 0.0719, 0.0619, 0.0668, 0.0878, 0.0829, 0.0570], device='cuda:4'), in_proj_covar=tensor([0.0189, 0.0203, 0.0184, 0.0173, 0.0180, 0.0183, 0.0153, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 13:29:27,148 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-27 13:30:00,557 INFO [finetune.py:976] (4/7) Epoch 19, batch 5650, loss[loss=0.2061, simple_loss=0.2802, pruned_loss=0.06598, over 4798.00 frames. ], tot_loss[loss=0.1768, simple_loss=0.2488, pruned_loss=0.05242, over 951518.98 frames. ], batch size: 45, lr: 3.26e-03, grad_scale: 32.0 2023-04-27 13:30:17,476 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.047e+02 1.527e+02 1.868e+02 2.111e+02 3.831e+02, threshold=3.735e+02, percent-clipped=0.0 2023-04-27 13:30:39,808 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4086, 2.9103, 2.4344, 2.7914, 2.2487, 2.5908, 2.5085, 2.1076], device='cuda:4'), covar=tensor([0.1835, 0.1116, 0.0707, 0.1053, 0.2925, 0.1057, 0.2027, 0.2768], device='cuda:4'), in_proj_covar=tensor([0.0288, 0.0303, 0.0217, 0.0280, 0.0314, 0.0260, 0.0251, 0.0265], device='cuda:4'), out_proj_covar=tensor([1.1547e-04, 1.2046e-04, 8.6164e-05, 1.1102e-04, 1.2733e-04, 1.0304e-04, 1.0145e-04, 1.0509e-04], device='cuda:4') 2023-04-27 13:30:52,599 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=108788.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 13:30:54,416 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=108791.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:31:04,844 INFO [finetune.py:976] (4/7) Epoch 19, batch 5700, loss[loss=0.1719, simple_loss=0.2327, pruned_loss=0.0556, over 4057.00 frames. ], tot_loss[loss=0.174, simple_loss=0.2446, pruned_loss=0.0517, over 931342.62 frames. ], batch size: 17, lr: 3.26e-03, grad_scale: 32.0 2023-04-27 13:31:40,719 INFO [finetune.py:976] (4/7) Epoch 20, batch 0, loss[loss=0.1986, simple_loss=0.2572, pruned_loss=0.07001, over 4889.00 frames. ], tot_loss[loss=0.1986, simple_loss=0.2572, pruned_loss=0.07001, over 4889.00 frames. ], batch size: 35, lr: 3.26e-03, grad_scale: 32.0 2023-04-27 13:31:40,719 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 13:31:48,451 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1514, 2.6233, 1.0477, 1.4257, 1.8123, 1.3606, 3.0447, 1.7777], device='cuda:4'), covar=tensor([0.0679, 0.0597, 0.0736, 0.1204, 0.0491, 0.0910, 0.0271, 0.0577], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0065, 0.0048, 0.0047, 0.0050, 0.0052, 0.0074, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 13:31:49,850 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3119, 1.4919, 1.7282, 1.8998, 1.8547, 1.9526, 1.8026, 1.8546], device='cuda:4'), covar=tensor([0.4299, 0.5630, 0.5036, 0.4640, 0.5787, 0.7457, 0.5733, 0.5118], device='cuda:4'), in_proj_covar=tensor([0.0338, 0.0373, 0.0321, 0.0334, 0.0345, 0.0394, 0.0357, 0.0328], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 13:31:57,318 INFO [finetune.py:1010] (4/7) Epoch 20, validation: loss=0.1536, simple_loss=0.2249, pruned_loss=0.04109, over 2265189.00 frames. 2023-04-27 13:31:57,318 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-27 13:32:13,986 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=108852.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:32:17,454 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.016e+02 1.429e+02 1.789e+02 2.182e+02 4.169e+02, threshold=3.578e+02, percent-clipped=1.0 2023-04-27 13:32:21,702 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1942, 1.5738, 1.4452, 1.7683, 1.6539, 2.0551, 1.4648, 3.4495], device='cuda:4'), covar=tensor([0.0587, 0.0813, 0.0758, 0.1168, 0.0634, 0.0434, 0.0726, 0.0151], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 13:32:30,439 INFO [finetune.py:976] (4/7) Epoch 20, batch 50, loss[loss=0.1795, simple_loss=0.2485, pruned_loss=0.0552, over 4894.00 frames. ], tot_loss[loss=0.1768, simple_loss=0.2479, pruned_loss=0.05282, over 217567.06 frames. ], batch size: 36, lr: 3.26e-03, grad_scale: 32.0 2023-04-27 13:33:03,608 INFO [finetune.py:976] (4/7) Epoch 20, batch 100, loss[loss=0.1628, simple_loss=0.2344, pruned_loss=0.04562, over 4864.00 frames. ], tot_loss[loss=0.1735, simple_loss=0.244, pruned_loss=0.05146, over 383348.04 frames. ], batch size: 31, lr: 3.26e-03, grad_scale: 32.0 2023-04-27 13:33:08,741 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=108933.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:33:13,040 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1762, 2.7662, 2.1553, 2.2449, 1.5839, 1.5389, 2.2489, 1.5039], device='cuda:4'), covar=tensor([0.1582, 0.1417, 0.1307, 0.1587, 0.2189, 0.1903, 0.0976, 0.1990], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0213, 0.0170, 0.0204, 0.0201, 0.0186, 0.0156, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 13:33:23,935 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.275e+01 1.453e+02 1.763e+02 2.128e+02 3.737e+02, threshold=3.527e+02, percent-clipped=2.0 2023-04-27 13:33:36,947 INFO [finetune.py:976] (4/7) Epoch 20, batch 150, loss[loss=0.1678, simple_loss=0.226, pruned_loss=0.05473, over 4740.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2383, pruned_loss=0.04976, over 509662.00 frames. ], batch size: 54, lr: 3.26e-03, grad_scale: 32.0 2023-04-27 13:33:39,881 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=108981.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:34:02,754 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3920, 1.4543, 1.8346, 1.8228, 1.3533, 1.2481, 1.4976, 1.1168], device='cuda:4'), covar=tensor([0.0640, 0.0586, 0.0362, 0.0627, 0.0870, 0.1205, 0.0614, 0.0557], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0068, 0.0067, 0.0067, 0.0074, 0.0096, 0.0073, 0.0065], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 13:34:09,747 INFO [finetune.py:976] (4/7) Epoch 20, batch 200, loss[loss=0.1728, simple_loss=0.2469, pruned_loss=0.04935, over 4860.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.238, pruned_loss=0.04988, over 608209.63 frames. ], batch size: 44, lr: 3.26e-03, grad_scale: 32.0 2023-04-27 13:34:09,840 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2414, 3.0399, 0.9525, 1.6635, 1.6778, 2.1824, 1.7727, 0.9760], device='cuda:4'), covar=tensor([0.1459, 0.1004, 0.1814, 0.1255, 0.1120, 0.0977, 0.1571, 0.1899], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0242, 0.0139, 0.0119, 0.0132, 0.0152, 0.0117, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 13:34:11,042 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=109029.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:34:16,938 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=109037.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 13:34:29,595 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.062e+02 1.602e+02 1.908e+02 2.213e+02 3.365e+02, threshold=3.817e+02, percent-clipped=0.0 2023-04-27 13:34:30,320 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3737, 2.9786, 1.3076, 1.6016, 2.1854, 1.5426, 3.4709, 1.9590], device='cuda:4'), covar=tensor([0.0601, 0.0741, 0.0805, 0.1039, 0.0461, 0.0818, 0.0212, 0.0548], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0066, 0.0048, 0.0047, 0.0050, 0.0052, 0.0075, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 13:34:40,888 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-27 13:34:42,777 INFO [finetune.py:976] (4/7) Epoch 20, batch 250, loss[loss=0.2023, simple_loss=0.2785, pruned_loss=0.06306, over 4817.00 frames. ], tot_loss[loss=0.1717, simple_loss=0.2415, pruned_loss=0.05094, over 685541.66 frames. ], batch size: 51, lr: 3.26e-03, grad_scale: 32.0 2023-04-27 13:35:02,387 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=109088.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:35:13,158 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=109098.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 13:35:31,288 INFO [finetune.py:976] (4/7) Epoch 20, batch 300, loss[loss=0.1743, simple_loss=0.2501, pruned_loss=0.04927, over 4864.00 frames. ], tot_loss[loss=0.1734, simple_loss=0.2431, pruned_loss=0.05183, over 741997.94 frames. ], batch size: 34, lr: 3.26e-03, grad_scale: 32.0 2023-04-27 13:35:49,171 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=109136.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:35:56,827 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=109147.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:36:03,480 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.124e+02 1.663e+02 2.046e+02 2.623e+02 4.738e+02, threshold=4.093e+02, percent-clipped=3.0 2023-04-27 13:36:25,682 INFO [finetune.py:976] (4/7) Epoch 20, batch 350, loss[loss=0.1967, simple_loss=0.2698, pruned_loss=0.06178, over 4744.00 frames. ], tot_loss[loss=0.1747, simple_loss=0.2453, pruned_loss=0.05203, over 790870.92 frames. ], batch size: 59, lr: 3.26e-03, grad_scale: 32.0 2023-04-27 13:36:40,557 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-04-27 13:36:47,068 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=109194.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:36:47,483 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.58 vs. limit=5.0 2023-04-27 13:36:59,306 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5782, 1.2761, 4.3383, 4.1202, 3.8372, 4.1894, 4.0489, 3.7879], device='cuda:4'), covar=tensor([0.6939, 0.5984, 0.1195, 0.1774, 0.1154, 0.1444, 0.1548, 0.1790], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0302, 0.0405, 0.0404, 0.0349, 0.0407, 0.0312, 0.0366], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 13:37:01,163 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1382, 2.6397, 1.0343, 1.4190, 2.1328, 1.3126, 3.5306, 1.8874], device='cuda:4'), covar=tensor([0.0676, 0.0590, 0.0755, 0.1317, 0.0489, 0.1037, 0.0276, 0.0652], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0065, 0.0048, 0.0047, 0.0050, 0.0052, 0.0074, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 13:37:21,409 INFO [finetune.py:976] (4/7) Epoch 20, batch 400, loss[loss=0.1587, simple_loss=0.2432, pruned_loss=0.03707, over 4900.00 frames. ], tot_loss[loss=0.1747, simple_loss=0.2461, pruned_loss=0.05169, over 828804.20 frames. ], batch size: 36, lr: 3.26e-03, grad_scale: 32.0 2023-04-27 13:37:31,619 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=109233.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:38:03,281 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8878, 2.5301, 1.9316, 1.9005, 1.3472, 1.3742, 2.0121, 1.3456], device='cuda:4'), covar=tensor([0.1611, 0.1332, 0.1439, 0.1646, 0.2270, 0.1858, 0.0975, 0.1983], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0211, 0.0168, 0.0202, 0.0199, 0.0183, 0.0154, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 13:38:03,877 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=109255.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:38:05,569 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.152e+02 1.658e+02 1.948e+02 2.358e+02 4.073e+02, threshold=3.896e+02, percent-clipped=0.0 2023-04-27 13:38:22,933 INFO [finetune.py:976] (4/7) Epoch 20, batch 450, loss[loss=0.2083, simple_loss=0.2702, pruned_loss=0.07322, over 4848.00 frames. ], tot_loss[loss=0.1743, simple_loss=0.2457, pruned_loss=0.05147, over 858759.16 frames. ], batch size: 44, lr: 3.26e-03, grad_scale: 32.0 2023-04-27 13:38:25,939 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=109281.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:38:34,969 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=109294.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:38:50,559 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=109318.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:38:54,450 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.51 vs. limit=2.0 2023-04-27 13:38:55,901 INFO [finetune.py:976] (4/7) Epoch 20, batch 500, loss[loss=0.1768, simple_loss=0.251, pruned_loss=0.05128, over 4872.00 frames. ], tot_loss[loss=0.1726, simple_loss=0.2434, pruned_loss=0.05096, over 878722.65 frames. ], batch size: 31, lr: 3.26e-03, grad_scale: 32.0 2023-04-27 13:39:16,051 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=109355.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:39:17,760 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.900e+01 1.534e+02 1.939e+02 2.475e+02 3.939e+02, threshold=3.877e+02, percent-clipped=2.0 2023-04-27 13:39:24,025 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8459, 1.6441, 1.4726, 1.7189, 2.0875, 1.7456, 1.4676, 1.4075], device='cuda:4'), covar=tensor([0.1400, 0.1251, 0.1698, 0.1283, 0.0714, 0.1387, 0.1744, 0.1979], device='cuda:4'), in_proj_covar=tensor([0.0303, 0.0306, 0.0344, 0.0284, 0.0320, 0.0302, 0.0295, 0.0364], device='cuda:4'), out_proj_covar=tensor([6.2256e-05, 6.3329e-05, 7.2701e-05, 5.7506e-05, 6.6126e-05, 6.3451e-05, 6.1823e-05, 7.7397e-05], device='cuda:4') 2023-04-27 13:39:29,356 INFO [finetune.py:976] (4/7) Epoch 20, batch 550, loss[loss=0.186, simple_loss=0.2449, pruned_loss=0.06351, over 4873.00 frames. ], tot_loss[loss=0.1718, simple_loss=0.2416, pruned_loss=0.05096, over 894681.35 frames. ], batch size: 34, lr: 3.26e-03, grad_scale: 32.0 2023-04-27 13:39:30,087 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4600, 1.3635, 1.7022, 1.7323, 1.3234, 1.2191, 1.4111, 0.9830], device='cuda:4'), covar=tensor([0.0544, 0.0657, 0.0351, 0.0555, 0.0785, 0.1217, 0.0634, 0.0600], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0068, 0.0067, 0.0067, 0.0074, 0.0096, 0.0073, 0.0065], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 13:39:30,711 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=109379.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 13:39:40,649 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=109393.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 13:40:01,835 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.7996, 2.0364, 1.9898, 2.1489, 1.9413, 1.9993, 2.0420, 2.0448], device='cuda:4'), covar=tensor([0.4138, 0.6330, 0.5000, 0.4563, 0.5887, 0.7445, 0.6268, 0.5432], device='cuda:4'), in_proj_covar=tensor([0.0336, 0.0373, 0.0322, 0.0335, 0.0345, 0.0394, 0.0358, 0.0328], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 13:40:02,926 INFO [finetune.py:976] (4/7) Epoch 20, batch 600, loss[loss=0.1943, simple_loss=0.2639, pruned_loss=0.06234, over 4866.00 frames. ], tot_loss[loss=0.1723, simple_loss=0.2424, pruned_loss=0.05107, over 908999.58 frames. ], batch size: 44, lr: 3.26e-03, grad_scale: 64.0 2023-04-27 13:40:10,670 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7638, 1.4756, 1.9749, 2.1326, 1.8359, 1.7748, 1.8083, 1.8410], device='cuda:4'), covar=tensor([0.5870, 0.7682, 0.7633, 0.8032, 0.7378, 0.9580, 0.9985, 1.1535], device='cuda:4'), in_proj_covar=tensor([0.0430, 0.0411, 0.0506, 0.0506, 0.0457, 0.0485, 0.0492, 0.0499], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 13:40:16,606 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=109447.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:40:24,270 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.064e+02 1.599e+02 2.024e+02 2.481e+02 5.741e+02, threshold=4.048e+02, percent-clipped=1.0 2023-04-27 13:40:33,462 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.59 vs. limit=2.0 2023-04-27 13:40:36,277 INFO [finetune.py:976] (4/7) Epoch 20, batch 650, loss[loss=0.1676, simple_loss=0.2442, pruned_loss=0.04547, over 4831.00 frames. ], tot_loss[loss=0.1738, simple_loss=0.2447, pruned_loss=0.05147, over 919318.58 frames. ], batch size: 49, lr: 3.26e-03, grad_scale: 64.0 2023-04-27 13:40:48,775 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=109495.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:40:56,592 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6445, 2.1261, 1.6075, 1.4506, 1.2458, 1.2596, 1.6417, 1.1688], device='cuda:4'), covar=tensor([0.1637, 0.1239, 0.1392, 0.1701, 0.2240, 0.1940, 0.0952, 0.2062], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0212, 0.0169, 0.0203, 0.0200, 0.0184, 0.0155, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 13:41:10,003 INFO [finetune.py:976] (4/7) Epoch 20, batch 700, loss[loss=0.1868, simple_loss=0.2565, pruned_loss=0.05855, over 4863.00 frames. ], tot_loss[loss=0.1751, simple_loss=0.2463, pruned_loss=0.05197, over 925698.11 frames. ], batch size: 31, lr: 3.26e-03, grad_scale: 64.0 2023-04-27 13:41:25,581 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=109550.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:41:30,317 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.167e+02 1.590e+02 1.898e+02 2.337e+02 3.520e+02, threshold=3.796e+02, percent-clipped=0.0 2023-04-27 13:41:36,185 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6116, 1.4410, 1.3290, 1.5049, 1.8975, 1.5535, 1.3191, 1.2343], device='cuda:4'), covar=tensor([0.1443, 0.1118, 0.1582, 0.1259, 0.0581, 0.1333, 0.1760, 0.1918], device='cuda:4'), in_proj_covar=tensor([0.0304, 0.0307, 0.0345, 0.0285, 0.0321, 0.0303, 0.0296, 0.0366], device='cuda:4'), out_proj_covar=tensor([6.2583e-05, 6.3569e-05, 7.3073e-05, 5.7642e-05, 6.6440e-05, 6.3682e-05, 6.2068e-05, 7.7813e-05], device='cuda:4') 2023-04-27 13:41:43,811 INFO [finetune.py:976] (4/7) Epoch 20, batch 750, loss[loss=0.1563, simple_loss=0.2384, pruned_loss=0.03709, over 4891.00 frames. ], tot_loss[loss=0.1753, simple_loss=0.2467, pruned_loss=0.05195, over 931413.07 frames. ], batch size: 37, lr: 3.26e-03, grad_scale: 64.0 2023-04-27 13:41:45,086 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=109579.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:42:49,901 INFO [finetune.py:976] (4/7) Epoch 20, batch 800, loss[loss=0.1989, simple_loss=0.257, pruned_loss=0.07037, over 4797.00 frames. ], tot_loss[loss=0.1742, simple_loss=0.2461, pruned_loss=0.05117, over 936961.76 frames. ], batch size: 51, lr: 3.25e-03, grad_scale: 64.0 2023-04-27 13:42:56,880 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-27 13:43:07,617 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=109640.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:43:19,839 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=109650.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:43:27,910 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.098e+02 1.469e+02 1.785e+02 2.276e+02 4.254e+02, threshold=3.571e+02, percent-clipped=2.0 2023-04-27 13:43:28,676 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=109659.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:43:50,500 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=109674.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 13:43:52,232 INFO [finetune.py:976] (4/7) Epoch 20, batch 850, loss[loss=0.2153, simple_loss=0.2781, pruned_loss=0.07623, over 4879.00 frames. ], tot_loss[loss=0.1717, simple_loss=0.2433, pruned_loss=0.0501, over 941470.48 frames. ], batch size: 32, lr: 3.25e-03, grad_scale: 64.0 2023-04-27 13:44:08,049 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=109693.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 13:44:26,932 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=109720.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:44:31,595 INFO [finetune.py:976] (4/7) Epoch 20, batch 900, loss[loss=0.1877, simple_loss=0.2386, pruned_loss=0.06836, over 4833.00 frames. ], tot_loss[loss=0.1693, simple_loss=0.2403, pruned_loss=0.04912, over 943749.23 frames. ], batch size: 33, lr: 3.25e-03, grad_scale: 64.0 2023-04-27 13:44:40,145 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=109741.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 13:44:50,922 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.301e+02 1.595e+02 1.882e+02 2.269e+02 5.455e+02, threshold=3.764e+02, percent-clipped=0.0 2023-04-27 13:45:03,938 INFO [finetune.py:976] (4/7) Epoch 20, batch 950, loss[loss=0.2178, simple_loss=0.2787, pruned_loss=0.07841, over 4219.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.2397, pruned_loss=0.04967, over 945670.70 frames. ], batch size: 18, lr: 3.25e-03, grad_scale: 32.0 2023-04-27 13:45:16,155 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5244, 1.7893, 1.8556, 1.9758, 1.8340, 1.9197, 1.9327, 1.9087], device='cuda:4'), covar=tensor([0.3698, 0.5819, 0.4364, 0.4298, 0.5412, 0.6887, 0.5110, 0.5078], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0373, 0.0321, 0.0335, 0.0345, 0.0394, 0.0358, 0.0328], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 13:45:38,014 INFO [finetune.py:976] (4/7) Epoch 20, batch 1000, loss[loss=0.1854, simple_loss=0.2511, pruned_loss=0.05987, over 4135.00 frames. ], tot_loss[loss=0.1718, simple_loss=0.2427, pruned_loss=0.05046, over 946566.43 frames. ], batch size: 65, lr: 3.25e-03, grad_scale: 32.0 2023-04-27 13:45:39,656 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.85 vs. limit=2.0 2023-04-27 13:45:52,581 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=109850.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:45:58,962 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.126e+02 1.623e+02 1.911e+02 2.262e+02 3.883e+02, threshold=3.822e+02, percent-clipped=2.0 2023-04-27 13:46:01,511 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.6489, 3.7324, 2.7340, 4.2408, 3.7163, 3.6489, 1.5799, 3.5859], device='cuda:4'), covar=tensor([0.1878, 0.1196, 0.3215, 0.1697, 0.3194, 0.1731, 0.5766, 0.2414], device='cuda:4'), in_proj_covar=tensor([0.0245, 0.0215, 0.0249, 0.0304, 0.0294, 0.0246, 0.0272, 0.0271], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 13:46:08,740 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6009, 1.6892, 0.7556, 1.3083, 1.7545, 1.4306, 1.3442, 1.4890], device='cuda:4'), covar=tensor([0.0493, 0.0354, 0.0361, 0.0533, 0.0276, 0.0503, 0.0497, 0.0536], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0029, 0.0019, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0045, 0.0038, 0.0051, 0.0038, 0.0049, 0.0049, 0.0051], device='cuda:4') 2023-04-27 13:46:11,418 INFO [finetune.py:976] (4/7) Epoch 20, batch 1050, loss[loss=0.179, simple_loss=0.2429, pruned_loss=0.05752, over 4734.00 frames. ], tot_loss[loss=0.1752, simple_loss=0.2471, pruned_loss=0.05165, over 949980.87 frames. ], batch size: 59, lr: 3.25e-03, grad_scale: 32.0 2023-04-27 13:46:11,709 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-27 13:46:25,237 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=109898.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:46:43,735 INFO [finetune.py:976] (4/7) Epoch 20, batch 1100, loss[loss=0.1742, simple_loss=0.261, pruned_loss=0.04364, over 4789.00 frames. ], tot_loss[loss=0.1758, simple_loss=0.2479, pruned_loss=0.05183, over 951742.40 frames. ], batch size: 51, lr: 3.25e-03, grad_scale: 32.0 2023-04-27 13:46:50,147 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=109935.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:46:59,848 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=109950.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:47:05,249 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.994e+01 1.686e+02 1.942e+02 2.391e+02 4.510e+02, threshold=3.883e+02, percent-clipped=3.0 2023-04-27 13:47:16,127 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=109974.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:47:17,879 INFO [finetune.py:976] (4/7) Epoch 20, batch 1150, loss[loss=0.1494, simple_loss=0.2275, pruned_loss=0.03562, over 4885.00 frames. ], tot_loss[loss=0.1763, simple_loss=0.2487, pruned_loss=0.0519, over 952341.43 frames. ], batch size: 43, lr: 3.25e-03, grad_scale: 32.0 2023-04-27 13:47:34,816 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.89 vs. limit=5.0 2023-04-27 13:47:40,684 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=109998.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:47:52,882 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=110015.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:47:54,547 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=110017.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:47:57,537 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=110022.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:48:00,997 INFO [finetune.py:976] (4/7) Epoch 20, batch 1200, loss[loss=0.1441, simple_loss=0.2235, pruned_loss=0.03231, over 4808.00 frames. ], tot_loss[loss=0.1748, simple_loss=0.2471, pruned_loss=0.05132, over 954271.21 frames. ], batch size: 51, lr: 3.25e-03, grad_scale: 32.0 2023-04-27 13:48:17,794 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=110044.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:48:37,464 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.027e+02 1.685e+02 1.958e+02 2.263e+02 7.161e+02, threshold=3.916e+02, percent-clipped=1.0 2023-04-27 13:49:01,340 INFO [finetune.py:976] (4/7) Epoch 20, batch 1250, loss[loss=0.2533, simple_loss=0.301, pruned_loss=0.1028, over 4936.00 frames. ], tot_loss[loss=0.1735, simple_loss=0.2448, pruned_loss=0.05113, over 953483.66 frames. ], batch size: 38, lr: 3.25e-03, grad_scale: 32.0 2023-04-27 13:49:02,093 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=110078.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:49:13,730 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2072, 3.0307, 2.3982, 2.7544, 2.1864, 2.5278, 2.6612, 1.8886], device='cuda:4'), covar=tensor([0.2359, 0.1217, 0.0822, 0.1341, 0.2929, 0.1206, 0.1821, 0.3025], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0304, 0.0219, 0.0280, 0.0314, 0.0260, 0.0252, 0.0265], device='cuda:4'), out_proj_covar=tensor([1.1607e-04, 1.2077e-04, 8.6746e-05, 1.1080e-04, 1.2744e-04, 1.0294e-04, 1.0187e-04, 1.0498e-04], device='cuda:4') 2023-04-27 13:49:14,845 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7351, 2.0323, 1.7336, 1.4268, 1.2690, 1.2858, 1.7559, 1.2749], device='cuda:4'), covar=tensor([0.1728, 0.1358, 0.1443, 0.1775, 0.2537, 0.2023, 0.1055, 0.2082], device='cuda:4'), in_proj_covar=tensor([0.0194, 0.0209, 0.0167, 0.0202, 0.0199, 0.0183, 0.0154, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 13:49:25,649 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=110105.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:49:28,681 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=110110.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:49:31,772 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=110115.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:49:37,605 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8141, 1.4443, 1.4107, 1.7038, 2.0247, 1.6871, 1.3854, 1.3104], device='cuda:4'), covar=tensor([0.1435, 0.1503, 0.1841, 0.1310, 0.0851, 0.1543, 0.2022, 0.2239], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0309, 0.0349, 0.0287, 0.0324, 0.0307, 0.0298, 0.0367], device='cuda:4'), out_proj_covar=tensor([6.3176e-05, 6.4090e-05, 7.3883e-05, 5.7986e-05, 6.7081e-05, 6.4382e-05, 6.2503e-05, 7.8096e-05], device='cuda:4') 2023-04-27 13:49:40,408 INFO [finetune.py:976] (4/7) Epoch 20, batch 1300, loss[loss=0.1301, simple_loss=0.1956, pruned_loss=0.03234, over 4812.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2419, pruned_loss=0.05013, over 956086.04 frames. ], batch size: 25, lr: 3.25e-03, grad_scale: 32.0 2023-04-27 13:50:01,335 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0513, 2.6922, 2.0813, 2.0805, 1.4484, 1.4630, 2.1443, 1.4423], device='cuda:4'), covar=tensor([0.1605, 0.1415, 0.1396, 0.1734, 0.2302, 0.1927, 0.0977, 0.2066], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0211, 0.0168, 0.0204, 0.0201, 0.0185, 0.0155, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 13:50:01,794 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.082e+02 1.513e+02 1.801e+02 2.249e+02 8.027e+02, threshold=3.601e+02, percent-clipped=2.0 2023-04-27 13:50:09,737 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=110171.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:50:13,297 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=110176.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:50:13,763 INFO [finetune.py:976] (4/7) Epoch 20, batch 1350, loss[loss=0.1781, simple_loss=0.2467, pruned_loss=0.05474, over 4756.00 frames. ], tot_loss[loss=0.1726, simple_loss=0.2431, pruned_loss=0.05112, over 954342.60 frames. ], batch size: 27, lr: 3.25e-03, grad_scale: 32.0 2023-04-27 13:50:27,459 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0773, 0.7900, 0.8822, 0.7658, 1.1985, 0.9414, 0.8183, 0.8919], device='cuda:4'), covar=tensor([0.2015, 0.1480, 0.2160, 0.1745, 0.0973, 0.1470, 0.1965, 0.2305], device='cuda:4'), in_proj_covar=tensor([0.0306, 0.0308, 0.0348, 0.0286, 0.0323, 0.0305, 0.0297, 0.0366], device='cuda:4'), out_proj_covar=tensor([6.2980e-05, 6.3761e-05, 7.3614e-05, 5.7735e-05, 6.6837e-05, 6.4069e-05, 6.2261e-05, 7.7874e-05], device='cuda:4') 2023-04-27 13:50:29,998 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3674, 1.7749, 1.6189, 1.9295, 1.9698, 1.9381, 1.6866, 4.2021], device='cuda:4'), covar=tensor([0.0546, 0.0812, 0.0757, 0.1186, 0.0582, 0.0539, 0.0726, 0.0112], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 13:50:47,125 INFO [finetune.py:976] (4/7) Epoch 20, batch 1400, loss[loss=0.194, simple_loss=0.2745, pruned_loss=0.05677, over 4797.00 frames. ], tot_loss[loss=0.174, simple_loss=0.2456, pruned_loss=0.05125, over 954845.29 frames. ], batch size: 45, lr: 3.25e-03, grad_scale: 32.0 2023-04-27 13:50:52,566 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=110235.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:51:09,029 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.151e+02 1.619e+02 1.811e+02 2.225e+02 6.587e+02, threshold=3.621e+02, percent-clipped=5.0 2023-04-27 13:51:19,999 INFO [finetune.py:976] (4/7) Epoch 20, batch 1450, loss[loss=0.1883, simple_loss=0.2727, pruned_loss=0.05194, over 4799.00 frames. ], tot_loss[loss=0.176, simple_loss=0.2482, pruned_loss=0.05191, over 956200.27 frames. ], batch size: 41, lr: 3.25e-03, grad_scale: 32.0 2023-04-27 13:51:25,120 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=110283.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:51:25,189 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=110283.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:51:31,843 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.82 vs. limit=2.0 2023-04-27 13:51:38,788 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=110303.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:51:41,445 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.78 vs. limit=2.0 2023-04-27 13:51:45,989 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7439, 2.1979, 1.7432, 1.4853, 1.2849, 1.3103, 1.7147, 1.2643], device='cuda:4'), covar=tensor([0.1697, 0.1236, 0.1429, 0.1774, 0.2437, 0.2044, 0.1021, 0.2090], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0210, 0.0168, 0.0202, 0.0199, 0.0184, 0.0155, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 13:51:46,557 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=110315.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:51:48,390 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=110318.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:51:53,751 INFO [finetune.py:976] (4/7) Epoch 20, batch 1500, loss[loss=0.1479, simple_loss=0.2138, pruned_loss=0.04104, over 3985.00 frames. ], tot_loss[loss=0.1762, simple_loss=0.2484, pruned_loss=0.05197, over 953479.32 frames. ], batch size: 17, lr: 3.25e-03, grad_scale: 32.0 2023-04-27 13:52:05,649 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=110344.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:52:12,963 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.93 vs. limit=2.0 2023-04-27 13:52:16,150 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.004e+02 1.560e+02 1.895e+02 2.185e+02 4.822e+02, threshold=3.791e+02, percent-clipped=2.0 2023-04-27 13:52:19,125 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=110363.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:52:19,781 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=110364.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:52:25,265 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=110373.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:52:27,675 INFO [finetune.py:976] (4/7) Epoch 20, batch 1550, loss[loss=0.127, simple_loss=0.2064, pruned_loss=0.02384, over 4711.00 frames. ], tot_loss[loss=0.1759, simple_loss=0.2483, pruned_loss=0.05174, over 955002.28 frames. ], batch size: 23, lr: 3.25e-03, grad_scale: 32.0 2023-04-27 13:52:29,081 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=110379.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:53:00,007 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=110400.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:53:08,395 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2918, 2.1505, 2.3066, 2.6784, 2.6889, 2.2877, 1.8387, 2.4092], device='cuda:4'), covar=tensor([0.0811, 0.0968, 0.0702, 0.0566, 0.0573, 0.0748, 0.0704, 0.0520], device='cuda:4'), in_proj_covar=tensor([0.0186, 0.0199, 0.0182, 0.0172, 0.0177, 0.0180, 0.0151, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 13:53:22,803 INFO [finetune.py:976] (4/7) Epoch 20, batch 1600, loss[loss=0.1867, simple_loss=0.2665, pruned_loss=0.05343, over 4903.00 frames. ], tot_loss[loss=0.1755, simple_loss=0.247, pruned_loss=0.05197, over 953869.15 frames. ], batch size: 35, lr: 3.25e-03, grad_scale: 32.0 2023-04-27 13:53:25,367 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4877, 2.9958, 1.1040, 1.5395, 2.3022, 1.4972, 4.0373, 2.0482], device='cuda:4'), covar=tensor([0.0643, 0.0887, 0.0929, 0.1243, 0.0498, 0.0976, 0.0211, 0.0594], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0065, 0.0048, 0.0047, 0.0050, 0.0052, 0.0074, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 13:54:08,137 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.013e+02 1.546e+02 1.968e+02 2.341e+02 5.862e+02, threshold=3.936e+02, percent-clipped=2.0 2023-04-27 13:54:17,901 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=110466.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:54:21,383 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=110471.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:54:22,030 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0665, 2.3418, 0.8850, 1.3728, 1.3805, 1.7191, 1.5457, 0.8761], device='cuda:4'), covar=tensor([0.1789, 0.1823, 0.1999, 0.1753, 0.1361, 0.1291, 0.1728, 0.2106], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0245, 0.0139, 0.0120, 0.0134, 0.0154, 0.0119, 0.0122], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 13:54:30,175 INFO [finetune.py:976] (4/7) Epoch 20, batch 1650, loss[loss=0.1798, simple_loss=0.2353, pruned_loss=0.06216, over 4854.00 frames. ], tot_loss[loss=0.1736, simple_loss=0.2442, pruned_loss=0.05148, over 954390.85 frames. ], batch size: 44, lr: 3.25e-03, grad_scale: 32.0 2023-04-27 13:54:57,106 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.7478, 1.7139, 1.7644, 1.3509, 1.7891, 1.4252, 2.3110, 1.5589], device='cuda:4'), covar=tensor([0.3979, 0.2054, 0.5039, 0.2988, 0.1683, 0.2602, 0.1621, 0.4869], device='cuda:4'), in_proj_covar=tensor([0.0338, 0.0346, 0.0424, 0.0350, 0.0381, 0.0374, 0.0371, 0.0417], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 13:55:01,854 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0242, 1.5725, 2.1407, 2.4892, 2.1133, 2.0495, 2.1263, 2.0860], device='cuda:4'), covar=tensor([0.4598, 0.6322, 0.6654, 0.5469, 0.5821, 0.8240, 0.7783, 0.8068], device='cuda:4'), in_proj_covar=tensor([0.0426, 0.0409, 0.0502, 0.0506, 0.0454, 0.0482, 0.0488, 0.0494], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 13:55:09,359 INFO [finetune.py:976] (4/7) Epoch 20, batch 1700, loss[loss=0.2212, simple_loss=0.2762, pruned_loss=0.0831, over 4836.00 frames. ], tot_loss[loss=0.1725, simple_loss=0.2424, pruned_loss=0.05125, over 956025.59 frames. ], batch size: 33, lr: 3.25e-03, grad_scale: 32.0 2023-04-27 13:55:13,416 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.90 vs. limit=2.0 2023-04-27 13:55:29,554 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5929, 1.0929, 1.7006, 2.1491, 1.7022, 1.6107, 1.6722, 1.6202], device='cuda:4'), covar=tensor([0.4372, 0.6337, 0.6004, 0.5133, 0.5539, 0.7442, 0.7142, 0.8046], device='cuda:4'), in_proj_covar=tensor([0.0427, 0.0410, 0.0504, 0.0507, 0.0456, 0.0483, 0.0489, 0.0495], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 13:55:30,313 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-04-27 13:55:31,158 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.338e+01 1.550e+02 1.876e+02 2.294e+02 4.137e+02, threshold=3.752e+02, percent-clipped=2.0 2023-04-27 13:55:43,125 INFO [finetune.py:976] (4/7) Epoch 20, batch 1750, loss[loss=0.1608, simple_loss=0.2392, pruned_loss=0.04117, over 4824.00 frames. ], tot_loss[loss=0.1748, simple_loss=0.2448, pruned_loss=0.05237, over 954653.31 frames. ], batch size: 30, lr: 3.25e-03, grad_scale: 32.0 2023-04-27 13:55:43,623 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.70 vs. limit=2.0 2023-04-27 13:55:56,401 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1942, 1.5027, 1.3914, 1.6374, 1.5534, 1.8287, 1.3675, 3.2658], device='cuda:4'), covar=tensor([0.0621, 0.0843, 0.0750, 0.1151, 0.0594, 0.0605, 0.0750, 0.0139], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 13:56:06,134 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=110611.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:56:16,747 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9958, 2.0034, 1.7935, 1.6526, 2.1218, 1.6362, 2.6955, 1.5783], device='cuda:4'), covar=tensor([0.3555, 0.1788, 0.4528, 0.2949, 0.1676, 0.2463, 0.1310, 0.4222], device='cuda:4'), in_proj_covar=tensor([0.0339, 0.0347, 0.0426, 0.0352, 0.0382, 0.0375, 0.0373, 0.0418], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 13:56:17,234 INFO [finetune.py:976] (4/7) Epoch 20, batch 1800, loss[loss=0.206, simple_loss=0.2785, pruned_loss=0.06672, over 4927.00 frames. ], tot_loss[loss=0.1776, simple_loss=0.2491, pruned_loss=0.05305, over 956893.89 frames. ], batch size: 38, lr: 3.25e-03, grad_scale: 32.0 2023-04-27 13:56:24,701 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=110639.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:56:32,883 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.8259, 2.0509, 2.0817, 2.2312, 1.9388, 2.0638, 2.1971, 2.1402], device='cuda:4'), covar=tensor([0.3788, 0.5833, 0.4956, 0.4144, 0.5430, 0.6781, 0.5551, 0.5132], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0373, 0.0323, 0.0335, 0.0344, 0.0394, 0.0357, 0.0328], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 13:56:32,892 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9796, 1.7081, 2.1570, 2.4298, 2.0483, 1.9651, 2.0805, 2.0267], device='cuda:4'), covar=tensor([0.4530, 0.6653, 0.6743, 0.5412, 0.5592, 0.7979, 0.8433, 0.9461], device='cuda:4'), in_proj_covar=tensor([0.0428, 0.0410, 0.0504, 0.0508, 0.0456, 0.0484, 0.0490, 0.0496], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 13:56:38,212 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.583e+02 1.893e+02 2.198e+02 3.811e+02, threshold=3.786e+02, percent-clipped=1.0 2023-04-27 13:56:38,804 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=110659.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:56:47,741 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=110672.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:56:48,306 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=110673.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:56:48,909 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=110674.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:56:50,675 INFO [finetune.py:976] (4/7) Epoch 20, batch 1850, loss[loss=0.1896, simple_loss=0.268, pruned_loss=0.05559, over 4805.00 frames. ], tot_loss[loss=0.178, simple_loss=0.2498, pruned_loss=0.05311, over 956897.68 frames. ], batch size: 40, lr: 3.25e-03, grad_scale: 32.0 2023-04-27 13:56:56,834 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1629, 2.4777, 2.0405, 2.3382, 1.7281, 2.1629, 2.2231, 1.6740], device='cuda:4'), covar=tensor([0.1956, 0.1164, 0.0859, 0.1327, 0.3421, 0.1264, 0.2196, 0.3064], device='cuda:4'), in_proj_covar=tensor([0.0285, 0.0299, 0.0216, 0.0277, 0.0310, 0.0257, 0.0249, 0.0262], device='cuda:4'), out_proj_covar=tensor([1.1439e-04, 1.1870e-04, 8.5637e-05, 1.0979e-04, 1.2578e-04, 1.0189e-04, 1.0046e-04, 1.0380e-04], device='cuda:4') 2023-04-27 13:57:06,442 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=110700.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:57:21,095 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=110721.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:57:24,743 INFO [finetune.py:976] (4/7) Epoch 20, batch 1900, loss[loss=0.1633, simple_loss=0.2517, pruned_loss=0.0375, over 4899.00 frames. ], tot_loss[loss=0.1785, simple_loss=0.2504, pruned_loss=0.05327, over 955408.01 frames. ], batch size: 46, lr: 3.25e-03, grad_scale: 32.0 2023-04-27 13:57:30,868 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=110736.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:57:32,735 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.63 vs. limit=2.0 2023-04-27 13:57:38,120 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=110748.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:57:46,353 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.050e+01 1.575e+02 1.913e+02 2.383e+02 8.631e+02, threshold=3.825e+02, percent-clipped=5.0 2023-04-27 13:57:50,213 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2291, 2.7610, 2.1966, 2.2296, 1.5828, 1.5611, 2.3421, 1.5324], device='cuda:4'), covar=tensor([0.1633, 0.1398, 0.1407, 0.1594, 0.2343, 0.1800, 0.0950, 0.1963], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0212, 0.0169, 0.0203, 0.0200, 0.0184, 0.0155, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 13:57:51,430 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=110766.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:57:55,036 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=110771.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:57:59,104 INFO [finetune.py:976] (4/7) Epoch 20, batch 1950, loss[loss=0.1631, simple_loss=0.2305, pruned_loss=0.0479, over 4835.00 frames. ], tot_loss[loss=0.1749, simple_loss=0.247, pruned_loss=0.05136, over 955220.99 frames. ], batch size: 30, lr: 3.25e-03, grad_scale: 32.0 2023-04-27 13:58:29,587 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=110797.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 13:58:52,671 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=110814.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:59:01,953 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=110819.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:59:11,705 INFO [finetune.py:976] (4/7) Epoch 20, batch 2000, loss[loss=0.1844, simple_loss=0.2468, pruned_loss=0.06106, over 4931.00 frames. ], tot_loss[loss=0.1732, simple_loss=0.2448, pruned_loss=0.05083, over 955145.44 frames. ], batch size: 38, lr: 3.25e-03, grad_scale: 32.0 2023-04-27 13:59:33,917 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.034e+02 1.549e+02 1.813e+02 2.155e+02 3.715e+02, threshold=3.626e+02, percent-clipped=0.0 2023-04-27 13:59:45,784 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=110868.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 13:59:48,928 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-27 13:59:56,924 INFO [finetune.py:976] (4/7) Epoch 20, batch 2050, loss[loss=0.1443, simple_loss=0.1995, pruned_loss=0.04453, over 4303.00 frames. ], tot_loss[loss=0.1718, simple_loss=0.2425, pruned_loss=0.05057, over 955195.89 frames. ], batch size: 18, lr: 3.24e-03, grad_scale: 32.0 2023-04-27 14:00:53,302 INFO [finetune.py:976] (4/7) Epoch 20, batch 2100, loss[loss=0.2131, simple_loss=0.2645, pruned_loss=0.08084, over 4917.00 frames. ], tot_loss[loss=0.171, simple_loss=0.2416, pruned_loss=0.05015, over 955317.93 frames. ], batch size: 36, lr: 3.24e-03, grad_scale: 32.0 2023-04-27 14:00:54,675 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=110929.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:00:55,779 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.8226, 1.9164, 1.7085, 1.3716, 1.7452, 1.5705, 2.2699, 1.3780], device='cuda:4'), covar=tensor([0.3261, 0.1483, 0.5011, 0.2520, 0.1653, 0.2131, 0.1374, 0.4808], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0349, 0.0428, 0.0354, 0.0384, 0.0376, 0.0374, 0.0418], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 14:01:01,683 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=110939.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:01:13,869 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=110958.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:01:14,355 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.165e+02 1.601e+02 1.866e+02 2.314e+02 3.983e+02, threshold=3.732e+02, percent-clipped=4.0 2023-04-27 14:01:14,965 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=110959.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:01:20,706 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=110967.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:01:25,026 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=110974.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:01:27,215 INFO [finetune.py:976] (4/7) Epoch 20, batch 2150, loss[loss=0.2277, simple_loss=0.2984, pruned_loss=0.07848, over 4807.00 frames. ], tot_loss[loss=0.1733, simple_loss=0.2442, pruned_loss=0.0512, over 953485.74 frames. ], batch size: 39, lr: 3.24e-03, grad_scale: 32.0 2023-04-27 14:01:34,418 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=110987.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:01:47,106 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=111007.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:01:55,863 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=111019.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:01:57,638 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=111022.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:02:00,613 INFO [finetune.py:976] (4/7) Epoch 20, batch 2200, loss[loss=0.2045, simple_loss=0.2694, pruned_loss=0.0698, over 4925.00 frames. ], tot_loss[loss=0.1738, simple_loss=0.2458, pruned_loss=0.05092, over 953826.54 frames. ], batch size: 33, lr: 3.24e-03, grad_scale: 32.0 2023-04-27 14:02:03,084 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0521, 2.5579, 2.1586, 1.9610, 1.4605, 1.5346, 2.1922, 1.4226], device='cuda:4'), covar=tensor([0.1602, 0.1429, 0.1357, 0.1681, 0.2295, 0.1910, 0.0925, 0.2052], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0210, 0.0168, 0.0202, 0.0199, 0.0184, 0.0154, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 14:02:22,146 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.079e+02 1.590e+02 1.963e+02 2.328e+02 5.181e+02, threshold=3.927e+02, percent-clipped=3.0 2023-04-27 14:02:34,619 INFO [finetune.py:976] (4/7) Epoch 20, batch 2250, loss[loss=0.15, simple_loss=0.2189, pruned_loss=0.04053, over 4795.00 frames. ], tot_loss[loss=0.1749, simple_loss=0.2468, pruned_loss=0.05153, over 955438.72 frames. ], batch size: 26, lr: 3.24e-03, grad_scale: 32.0 2023-04-27 14:02:44,878 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=111092.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 14:02:49,914 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-27 14:03:08,139 INFO [finetune.py:976] (4/7) Epoch 20, batch 2300, loss[loss=0.1712, simple_loss=0.2369, pruned_loss=0.05276, over 4778.00 frames. ], tot_loss[loss=0.1746, simple_loss=0.2473, pruned_loss=0.05097, over 957352.16 frames. ], batch size: 26, lr: 3.24e-03, grad_scale: 32.0 2023-04-27 14:03:29,169 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.334e+01 1.540e+02 1.823e+02 2.154e+02 3.608e+02, threshold=3.645e+02, percent-clipped=0.0 2023-04-27 14:03:41,035 INFO [finetune.py:976] (4/7) Epoch 20, batch 2350, loss[loss=0.1783, simple_loss=0.2537, pruned_loss=0.05143, over 4924.00 frames. ], tot_loss[loss=0.1748, simple_loss=0.2467, pruned_loss=0.05144, over 956832.07 frames. ], batch size: 38, lr: 3.24e-03, grad_scale: 32.0 2023-04-27 14:04:27,663 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=111224.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:04:28,325 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=111225.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:04:29,450 INFO [finetune.py:976] (4/7) Epoch 20, batch 2400, loss[loss=0.174, simple_loss=0.2471, pruned_loss=0.05042, over 4853.00 frames. ], tot_loss[loss=0.1729, simple_loss=0.2438, pruned_loss=0.05094, over 956878.97 frames. ], batch size: 44, lr: 3.24e-03, grad_scale: 32.0 2023-04-27 14:04:52,749 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5353, 1.4913, 1.8909, 1.9541, 1.4587, 1.3927, 1.5268, 1.0071], device='cuda:4'), covar=tensor([0.0547, 0.0681, 0.0367, 0.0548, 0.0726, 0.1047, 0.0634, 0.0615], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0069, 0.0067, 0.0068, 0.0075, 0.0096, 0.0073, 0.0065], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 14:05:02,140 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.830e+01 1.579e+02 1.923e+02 2.296e+02 7.513e+02, threshold=3.847e+02, percent-clipped=3.0 2023-04-27 14:05:07,622 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=111267.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:05:19,546 INFO [finetune.py:976] (4/7) Epoch 20, batch 2450, loss[loss=0.2231, simple_loss=0.2776, pruned_loss=0.08436, over 4902.00 frames. ], tot_loss[loss=0.1721, simple_loss=0.2421, pruned_loss=0.05107, over 958249.96 frames. ], batch size: 35, lr: 3.24e-03, grad_scale: 32.0 2023-04-27 14:05:36,546 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3321, 3.4536, 0.7741, 1.6578, 1.7985, 2.2824, 1.8536, 0.8630], device='cuda:4'), covar=tensor([0.1520, 0.0897, 0.2131, 0.1412, 0.1171, 0.1158, 0.1546, 0.2135], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0242, 0.0138, 0.0120, 0.0133, 0.0153, 0.0118, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 14:05:36,589 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=111286.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:05:48,444 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.53 vs. limit=2.0 2023-04-27 14:06:01,576 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-04-27 14:06:05,496 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=111314.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:06:06,093 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=111315.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:06:13,838 INFO [finetune.py:976] (4/7) Epoch 20, batch 2500, loss[loss=0.1642, simple_loss=0.2322, pruned_loss=0.0481, over 4778.00 frames. ], tot_loss[loss=0.1738, simple_loss=0.2437, pruned_loss=0.05193, over 954868.64 frames. ], batch size: 26, lr: 3.24e-03, grad_scale: 32.0 2023-04-27 14:06:35,391 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.107e+02 1.671e+02 1.878e+02 2.269e+02 4.413e+02, threshold=3.755e+02, percent-clipped=2.0 2023-04-27 14:06:42,016 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9899, 1.7576, 1.9127, 2.3756, 2.4111, 1.9455, 1.6562, 2.2095], device='cuda:4'), covar=tensor([0.0861, 0.1161, 0.0782, 0.0603, 0.0652, 0.0880, 0.0799, 0.0522], device='cuda:4'), in_proj_covar=tensor([0.0188, 0.0201, 0.0184, 0.0173, 0.0179, 0.0181, 0.0154, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 14:06:46,746 INFO [finetune.py:976] (4/7) Epoch 20, batch 2550, loss[loss=0.1443, simple_loss=0.2144, pruned_loss=0.03708, over 4890.00 frames. ], tot_loss[loss=0.175, simple_loss=0.2461, pruned_loss=0.05192, over 955341.84 frames. ], batch size: 32, lr: 3.24e-03, grad_scale: 32.0 2023-04-27 14:06:53,132 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5381, 1.4876, 0.7195, 1.2419, 1.6251, 1.3945, 1.3253, 1.3849], device='cuda:4'), covar=tensor([0.0503, 0.0386, 0.0364, 0.0568, 0.0276, 0.0509, 0.0510, 0.0558], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0029, 0.0020, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0045, 0.0038, 0.0052, 0.0038, 0.0049, 0.0050, 0.0051], device='cuda:4') 2023-04-27 14:06:57,283 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=111392.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 14:07:15,379 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.95 vs. limit=5.0 2023-04-27 14:07:18,288 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2588, 1.7547, 2.1811, 2.5676, 2.1708, 1.7458, 1.4641, 2.0152], device='cuda:4'), covar=tensor([0.3236, 0.3033, 0.1597, 0.2441, 0.2403, 0.2703, 0.4174, 0.1972], device='cuda:4'), in_proj_covar=tensor([0.0295, 0.0248, 0.0228, 0.0318, 0.0219, 0.0235, 0.0231, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 14:07:20,616 INFO [finetune.py:976] (4/7) Epoch 20, batch 2600, loss[loss=0.1262, simple_loss=0.1922, pruned_loss=0.03006, over 3992.00 frames. ], tot_loss[loss=0.1745, simple_loss=0.2463, pruned_loss=0.05132, over 954247.49 frames. ], batch size: 17, lr: 3.24e-03, grad_scale: 32.0 2023-04-27 14:07:23,164 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3908, 1.6829, 1.6200, 1.9255, 1.8555, 1.8376, 1.5967, 3.5726], device='cuda:4'), covar=tensor([0.0539, 0.0749, 0.0685, 0.1078, 0.0550, 0.0491, 0.0674, 0.0156], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0040, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 14:07:29,613 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=111440.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:07:42,989 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.116e+02 1.630e+02 1.917e+02 2.393e+02 4.047e+02, threshold=3.834e+02, percent-clipped=2.0 2023-04-27 14:07:48,663 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=111468.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:07:54,484 INFO [finetune.py:976] (4/7) Epoch 20, batch 2650, loss[loss=0.1416, simple_loss=0.2138, pruned_loss=0.03469, over 4750.00 frames. ], tot_loss[loss=0.1745, simple_loss=0.247, pruned_loss=0.05102, over 952945.05 frames. ], batch size: 27, lr: 3.24e-03, grad_scale: 32.0 2023-04-27 14:08:01,724 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9232, 2.4112, 1.9117, 2.1825, 1.5858, 2.0549, 2.0567, 1.6132], device='cuda:4'), covar=tensor([0.2080, 0.1244, 0.0884, 0.1208, 0.3550, 0.1101, 0.1903, 0.2622], device='cuda:4'), in_proj_covar=tensor([0.0286, 0.0300, 0.0216, 0.0277, 0.0311, 0.0258, 0.0249, 0.0261], device='cuda:4'), out_proj_covar=tensor([1.1461e-04, 1.1889e-04, 8.5762e-05, 1.0987e-04, 1.2603e-04, 1.0208e-04, 1.0042e-04, 1.0327e-04], device='cuda:4') 2023-04-27 14:08:26,366 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=111524.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:08:28,082 INFO [finetune.py:976] (4/7) Epoch 20, batch 2700, loss[loss=0.2057, simple_loss=0.2734, pruned_loss=0.06903, over 4890.00 frames. ], tot_loss[loss=0.1754, simple_loss=0.2476, pruned_loss=0.05159, over 955171.95 frames. ], batch size: 35, lr: 3.24e-03, grad_scale: 32.0 2023-04-27 14:08:29,415 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=111529.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:08:49,485 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.017e+02 1.491e+02 1.860e+02 2.271e+02 3.945e+02, threshold=3.719e+02, percent-clipped=1.0 2023-04-27 14:08:52,584 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.69 vs. limit=5.0 2023-04-27 14:08:58,401 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=111572.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:09:01,904 INFO [finetune.py:976] (4/7) Epoch 20, batch 2750, loss[loss=0.1608, simple_loss=0.2371, pruned_loss=0.0422, over 4869.00 frames. ], tot_loss[loss=0.1738, simple_loss=0.2452, pruned_loss=0.05118, over 954604.95 frames. ], batch size: 31, lr: 3.24e-03, grad_scale: 32.0 2023-04-27 14:09:04,323 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=111581.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:09:08,626 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4449, 1.5705, 1.3738, 1.0446, 1.1237, 1.1027, 1.3944, 1.0594], device='cuda:4'), covar=tensor([0.1555, 0.1161, 0.1304, 0.1447, 0.2075, 0.1812, 0.0921, 0.1797], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0211, 0.0169, 0.0204, 0.0200, 0.0186, 0.0156, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 14:09:37,578 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=111614.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:09:38,739 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3969, 1.3067, 1.5940, 1.6388, 1.2739, 1.1988, 1.2749, 0.8962], device='cuda:4'), covar=tensor([0.0488, 0.0690, 0.0341, 0.0527, 0.0769, 0.1275, 0.0556, 0.0600], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0069, 0.0067, 0.0067, 0.0075, 0.0097, 0.0073, 0.0065], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 14:09:57,465 INFO [finetune.py:976] (4/7) Epoch 20, batch 2800, loss[loss=0.144, simple_loss=0.2105, pruned_loss=0.03872, over 4398.00 frames. ], tot_loss[loss=0.1721, simple_loss=0.2427, pruned_loss=0.05079, over 953638.38 frames. ], batch size: 19, lr: 3.24e-03, grad_scale: 32.0 2023-04-27 14:10:24,460 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.707e+01 1.587e+02 1.829e+02 2.177e+02 6.182e+02, threshold=3.659e+02, percent-clipped=2.0 2023-04-27 14:10:26,840 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=111662.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:10:32,246 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3851, 2.7993, 1.0656, 1.7533, 1.6797, 2.1831, 1.7413, 1.1379], device='cuda:4'), covar=tensor([0.1212, 0.1030, 0.1643, 0.1155, 0.0994, 0.0858, 0.1350, 0.2036], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0243, 0.0139, 0.0121, 0.0134, 0.0153, 0.0118, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 14:10:36,928 INFO [finetune.py:976] (4/7) Epoch 20, batch 2850, loss[loss=0.1758, simple_loss=0.2523, pruned_loss=0.04967, over 4737.00 frames. ], tot_loss[loss=0.1708, simple_loss=0.2411, pruned_loss=0.05028, over 954846.52 frames. ], batch size: 54, lr: 3.24e-03, grad_scale: 32.0 2023-04-27 14:10:39,988 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.8915, 2.8956, 2.1829, 3.2915, 2.9270, 2.8902, 1.1979, 2.7958], device='cuda:4'), covar=tensor([0.2220, 0.1615, 0.3652, 0.2965, 0.4379, 0.2182, 0.6007, 0.2964], device='cuda:4'), in_proj_covar=tensor([0.0245, 0.0214, 0.0250, 0.0303, 0.0295, 0.0247, 0.0273, 0.0271], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 14:10:58,239 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0760, 1.4561, 1.8908, 2.0945, 1.8390, 1.4828, 1.1133, 1.7037], device='cuda:4'), covar=tensor([0.2698, 0.2950, 0.1518, 0.2018, 0.2278, 0.2527, 0.4275, 0.1810], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0246, 0.0227, 0.0316, 0.0219, 0.0233, 0.0229, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 14:11:31,378 INFO [finetune.py:976] (4/7) Epoch 20, batch 2900, loss[loss=0.1806, simple_loss=0.2559, pruned_loss=0.05259, over 4857.00 frames. ], tot_loss[loss=0.1749, simple_loss=0.2458, pruned_loss=0.05203, over 955621.25 frames. ], batch size: 44, lr: 3.24e-03, grad_scale: 32.0 2023-04-27 14:12:12,495 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.123e+02 1.619e+02 1.984e+02 2.443e+02 6.162e+02, threshold=3.968e+02, percent-clipped=7.0 2023-04-27 14:12:33,213 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0237, 2.6333, 1.8886, 2.0548, 1.4562, 1.4395, 2.0491, 1.3416], device='cuda:4'), covar=tensor([0.1565, 0.1468, 0.1346, 0.1661, 0.2087, 0.1937, 0.0946, 0.1991], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0212, 0.0169, 0.0204, 0.0200, 0.0186, 0.0156, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 14:12:36,595 INFO [finetune.py:976] (4/7) Epoch 20, batch 2950, loss[loss=0.1686, simple_loss=0.2414, pruned_loss=0.04792, over 4783.00 frames. ], tot_loss[loss=0.1774, simple_loss=0.2495, pruned_loss=0.05262, over 956920.44 frames. ], batch size: 29, lr: 3.24e-03, grad_scale: 64.0 2023-04-27 14:12:54,526 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.5552, 1.3792, 1.3490, 1.0580, 1.3211, 1.1094, 1.6177, 1.1537], device='cuda:4'), covar=tensor([0.3533, 0.1973, 0.5239, 0.2843, 0.1884, 0.2462, 0.1843, 0.5111], device='cuda:4'), in_proj_covar=tensor([0.0339, 0.0351, 0.0427, 0.0355, 0.0383, 0.0377, 0.0375, 0.0418], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 14:12:56,815 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=111791.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 14:13:29,964 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=111824.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:13:32,109 INFO [finetune.py:976] (4/7) Epoch 20, batch 3000, loss[loss=0.1868, simple_loss=0.2563, pruned_loss=0.05865, over 4837.00 frames. ], tot_loss[loss=0.1793, simple_loss=0.2515, pruned_loss=0.05356, over 957299.50 frames. ], batch size: 47, lr: 3.24e-03, grad_scale: 64.0 2023-04-27 14:13:32,110 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 14:13:49,409 INFO [finetune.py:1010] (4/7) Epoch 20, validation: loss=0.1527, simple_loss=0.2229, pruned_loss=0.04123, over 2265189.00 frames. 2023-04-27 14:13:49,410 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-27 14:14:20,689 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=111852.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 14:14:29,830 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.150e+02 1.660e+02 1.972e+02 2.446e+02 4.506e+02, threshold=3.944e+02, percent-clipped=2.0 2023-04-27 14:14:58,617 INFO [finetune.py:976] (4/7) Epoch 20, batch 3050, loss[loss=0.1865, simple_loss=0.2573, pruned_loss=0.05786, over 4886.00 frames. ], tot_loss[loss=0.1773, simple_loss=0.2499, pruned_loss=0.05237, over 957202.26 frames. ], batch size: 43, lr: 3.24e-03, grad_scale: 64.0 2023-04-27 14:15:01,100 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=111881.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:15:35,932 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5387, 1.3545, 4.0047, 3.7630, 3.5053, 3.7735, 3.7404, 3.5077], device='cuda:4'), covar=tensor([0.6796, 0.5403, 0.1059, 0.1574, 0.1077, 0.1793, 0.2150, 0.1560], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0309, 0.0409, 0.0410, 0.0354, 0.0411, 0.0317, 0.0368], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 14:15:37,190 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2267, 1.6809, 2.0647, 2.4259, 2.0834, 1.6357, 1.3332, 1.8181], device='cuda:4'), covar=tensor([0.2976, 0.3000, 0.1587, 0.1977, 0.2268, 0.2616, 0.4220, 0.2029], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0244, 0.0225, 0.0313, 0.0217, 0.0232, 0.0227, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 14:15:48,561 INFO [finetune.py:976] (4/7) Epoch 20, batch 3100, loss[loss=0.2016, simple_loss=0.2583, pruned_loss=0.07243, over 4815.00 frames. ], tot_loss[loss=0.1754, simple_loss=0.2474, pruned_loss=0.05169, over 956354.19 frames. ], batch size: 30, lr: 3.24e-03, grad_scale: 32.0 2023-04-27 14:15:50,329 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=111929.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:16:10,764 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5685, 1.4040, 0.4779, 1.2340, 1.4076, 1.4165, 1.2941, 1.3702], device='cuda:4'), covar=tensor([0.0515, 0.0397, 0.0397, 0.0567, 0.0294, 0.0518, 0.0506, 0.0610], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0029, 0.0019, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0050, 0.0045, 0.0038, 0.0051, 0.0038, 0.0049, 0.0049, 0.0051], device='cuda:4') 2023-04-27 14:16:16,419 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.135e+02 1.458e+02 1.721e+02 2.109e+02 3.576e+02, threshold=3.441e+02, percent-clipped=0.0 2023-04-27 14:16:19,296 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-04-27 14:16:27,779 INFO [finetune.py:976] (4/7) Epoch 20, batch 3150, loss[loss=0.1607, simple_loss=0.23, pruned_loss=0.04573, over 4934.00 frames. ], tot_loss[loss=0.173, simple_loss=0.2446, pruned_loss=0.05069, over 955938.76 frames. ], batch size: 33, lr: 3.24e-03, grad_scale: 32.0 2023-04-27 14:16:27,914 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3570, 1.8249, 2.2408, 2.6691, 2.2919, 1.7747, 1.5040, 2.1065], device='cuda:4'), covar=tensor([0.3162, 0.2923, 0.1661, 0.2247, 0.2344, 0.2572, 0.4066, 0.1850], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0245, 0.0226, 0.0313, 0.0218, 0.0232, 0.0228, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 14:16:33,783 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.9219, 2.3412, 5.6346, 5.0328, 5.0259, 5.3471, 4.8167, 4.6761], device='cuda:4'), covar=tensor([0.6891, 0.6843, 0.1060, 0.2387, 0.1430, 0.2908, 0.1489, 0.2484], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0307, 0.0406, 0.0406, 0.0351, 0.0407, 0.0314, 0.0366], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 14:16:53,762 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.0143, 3.9653, 2.8019, 4.6475, 4.0580, 4.0270, 1.6588, 4.0099], device='cuda:4'), covar=tensor([0.1620, 0.1208, 0.2869, 0.1380, 0.2496, 0.1703, 0.5604, 0.2116], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0213, 0.0248, 0.0301, 0.0294, 0.0246, 0.0272, 0.0269], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 14:17:02,002 INFO [finetune.py:976] (4/7) Epoch 20, batch 3200, loss[loss=0.2108, simple_loss=0.2662, pruned_loss=0.0777, over 4849.00 frames. ], tot_loss[loss=0.1704, simple_loss=0.2413, pruned_loss=0.04976, over 954521.58 frames. ], batch size: 47, lr: 3.24e-03, grad_scale: 32.0 2023-04-27 14:17:35,518 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.097e+02 1.584e+02 1.831e+02 2.274e+02 4.743e+02, threshold=3.663e+02, percent-clipped=4.0 2023-04-27 14:17:36,850 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7471, 1.4011, 1.4894, 1.4692, 1.9367, 1.5470, 1.2686, 1.3844], device='cuda:4'), covar=tensor([0.1671, 0.1264, 0.1809, 0.1380, 0.0783, 0.1561, 0.1993, 0.2399], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0311, 0.0352, 0.0288, 0.0326, 0.0307, 0.0301, 0.0371], device='cuda:4'), out_proj_covar=tensor([6.3806e-05, 6.4304e-05, 7.4443e-05, 5.8304e-05, 6.7443e-05, 6.4467e-05, 6.3147e-05, 7.8904e-05], device='cuda:4') 2023-04-27 14:17:39,745 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=112066.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:17:46,879 INFO [finetune.py:976] (4/7) Epoch 20, batch 3250, loss[loss=0.1845, simple_loss=0.2522, pruned_loss=0.05843, over 4936.00 frames. ], tot_loss[loss=0.1724, simple_loss=0.2429, pruned_loss=0.05099, over 954824.09 frames. ], batch size: 33, lr: 3.23e-03, grad_scale: 32.0 2023-04-27 14:17:57,668 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9380, 1.5724, 1.5155, 1.8520, 2.1807, 1.7392, 1.5267, 1.4612], device='cuda:4'), covar=tensor([0.1791, 0.1658, 0.2138, 0.1454, 0.1038, 0.1788, 0.2129, 0.2358], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0312, 0.0353, 0.0290, 0.0327, 0.0309, 0.0302, 0.0373], device='cuda:4'), out_proj_covar=tensor([6.4094e-05, 6.4573e-05, 7.4697e-05, 5.8521e-05, 6.7629e-05, 6.4730e-05, 6.3386e-05, 7.9221e-05], device='cuda:4') 2023-04-27 14:18:18,443 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=112124.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:18:20,136 INFO [finetune.py:976] (4/7) Epoch 20, batch 3300, loss[loss=0.1469, simple_loss=0.2223, pruned_loss=0.03579, over 4773.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.2453, pruned_loss=0.05119, over 955199.98 frames. ], batch size: 26, lr: 3.23e-03, grad_scale: 32.0 2023-04-27 14:18:20,256 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=112127.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:18:44,825 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=112147.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 14:18:52,672 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.675e+01 1.599e+02 1.882e+02 2.331e+02 3.289e+02, threshold=3.764e+02, percent-clipped=0.0 2023-04-27 14:19:00,555 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=112172.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:19:04,166 INFO [finetune.py:976] (4/7) Epoch 20, batch 3350, loss[loss=0.1892, simple_loss=0.2657, pruned_loss=0.05634, over 4910.00 frames. ], tot_loss[loss=0.1742, simple_loss=0.2464, pruned_loss=0.05102, over 953751.47 frames. ], batch size: 37, lr: 3.23e-03, grad_scale: 32.0 2023-04-27 14:19:37,509 INFO [finetune.py:976] (4/7) Epoch 20, batch 3400, loss[loss=0.1504, simple_loss=0.2247, pruned_loss=0.03801, over 4151.00 frames. ], tot_loss[loss=0.1741, simple_loss=0.2468, pruned_loss=0.05073, over 953919.05 frames. ], batch size: 65, lr: 3.23e-03, grad_scale: 32.0 2023-04-27 14:20:15,437 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.071e+02 1.600e+02 1.895e+02 2.268e+02 5.639e+02, threshold=3.789e+02, percent-clipped=3.0 2023-04-27 14:20:25,422 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.7974, 3.5640, 2.8668, 4.3701, 3.6897, 3.7397, 1.6314, 3.7167], device='cuda:4'), covar=tensor([0.1660, 0.1242, 0.3642, 0.1391, 0.3290, 0.1880, 0.5666, 0.2402], device='cuda:4'), in_proj_covar=tensor([0.0243, 0.0214, 0.0249, 0.0301, 0.0294, 0.0246, 0.0272, 0.0269], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 14:20:37,465 INFO [finetune.py:976] (4/7) Epoch 20, batch 3450, loss[loss=0.1679, simple_loss=0.2508, pruned_loss=0.04253, over 4799.00 frames. ], tot_loss[loss=0.1731, simple_loss=0.2461, pruned_loss=0.05007, over 955487.21 frames. ], batch size: 29, lr: 3.23e-03, grad_scale: 32.0 2023-04-27 14:21:10,402 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1416, 1.3866, 1.2592, 1.6515, 1.4850, 1.6566, 1.2789, 3.0404], device='cuda:4'), covar=tensor([0.0632, 0.0806, 0.0839, 0.1238, 0.0642, 0.0499, 0.0798, 0.0179], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0040, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 14:21:22,277 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6175, 1.6626, 0.8217, 1.3954, 1.8509, 1.4948, 1.4341, 1.5013], device='cuda:4'), covar=tensor([0.0501, 0.0367, 0.0336, 0.0548, 0.0248, 0.0503, 0.0471, 0.0560], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0029, 0.0019, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0045, 0.0038, 0.0052, 0.0038, 0.0049, 0.0050, 0.0051], device='cuda:4') 2023-04-27 14:21:35,778 INFO [finetune.py:976] (4/7) Epoch 20, batch 3500, loss[loss=0.2372, simple_loss=0.2963, pruned_loss=0.08909, over 4253.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2444, pruned_loss=0.04976, over 954355.43 frames. ], batch size: 65, lr: 3.23e-03, grad_scale: 32.0 2023-04-27 14:21:49,531 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5441, 1.3542, 4.2900, 4.0036, 3.7381, 4.0585, 3.9189, 3.7683], device='cuda:4'), covar=tensor([0.6855, 0.5423, 0.0980, 0.1634, 0.1035, 0.1349, 0.1715, 0.1578], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0307, 0.0407, 0.0406, 0.0351, 0.0409, 0.0315, 0.0366], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 14:21:59,357 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.140e+02 1.563e+02 1.868e+02 2.164e+02 4.202e+02, threshold=3.735e+02, percent-clipped=1.0 2023-04-27 14:22:09,770 INFO [finetune.py:976] (4/7) Epoch 20, batch 3550, loss[loss=0.1683, simple_loss=0.2387, pruned_loss=0.04892, over 4760.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2428, pruned_loss=0.04972, over 954569.50 frames. ], batch size: 26, lr: 3.23e-03, grad_scale: 32.0 2023-04-27 14:22:18,187 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=112389.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:22:40,998 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=112422.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:22:44,030 INFO [finetune.py:976] (4/7) Epoch 20, batch 3600, loss[loss=0.1625, simple_loss=0.2198, pruned_loss=0.05264, over 4823.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2398, pruned_loss=0.04886, over 951770.25 frames. ], batch size: 39, lr: 3.23e-03, grad_scale: 32.0 2023-04-27 14:22:57,488 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=112447.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 14:22:59,365 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=112450.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:23:06,297 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.099e+02 1.532e+02 1.802e+02 2.141e+02 3.664e+02, threshold=3.603e+02, percent-clipped=0.0 2023-04-27 14:23:13,550 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6245, 1.6503, 0.7245, 1.2768, 1.8530, 1.4534, 1.3788, 1.4120], device='cuda:4'), covar=tensor([0.0519, 0.0394, 0.0371, 0.0581, 0.0277, 0.0549, 0.0533, 0.0595], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0029, 0.0019, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0045, 0.0038, 0.0052, 0.0038, 0.0050, 0.0050, 0.0051], device='cuda:4') 2023-04-27 14:23:18,451 INFO [finetune.py:976] (4/7) Epoch 20, batch 3650, loss[loss=0.1833, simple_loss=0.2601, pruned_loss=0.05329, over 4835.00 frames. ], tot_loss[loss=0.1708, simple_loss=0.2422, pruned_loss=0.04968, over 954166.88 frames. ], batch size: 49, lr: 3.23e-03, grad_scale: 32.0 2023-04-27 14:23:30,564 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=112495.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 14:23:58,092 INFO [finetune.py:976] (4/7) Epoch 20, batch 3700, loss[loss=0.1841, simple_loss=0.2595, pruned_loss=0.05436, over 4742.00 frames. ], tot_loss[loss=0.174, simple_loss=0.2463, pruned_loss=0.0509, over 952847.13 frames. ], batch size: 27, lr: 3.23e-03, grad_scale: 32.0 2023-04-27 14:24:41,781 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.085e+02 1.566e+02 1.857e+02 2.073e+02 3.044e+02, threshold=3.713e+02, percent-clipped=0.0 2023-04-27 14:24:48,659 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=112562.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:25:05,135 INFO [finetune.py:976] (4/7) Epoch 20, batch 3750, loss[loss=0.1731, simple_loss=0.2449, pruned_loss=0.05068, over 4892.00 frames. ], tot_loss[loss=0.1752, simple_loss=0.2474, pruned_loss=0.05143, over 954081.19 frames. ], batch size: 35, lr: 3.23e-03, grad_scale: 32.0 2023-04-27 14:25:05,267 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=112577.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:25:26,547 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7544, 1.8981, 1.0898, 1.4560, 2.1656, 1.5321, 1.5086, 1.5869], device='cuda:4'), covar=tensor([0.0490, 0.0349, 0.0296, 0.0544, 0.0225, 0.0510, 0.0493, 0.0550], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0029, 0.0019, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0045, 0.0038, 0.0052, 0.0038, 0.0049, 0.0050, 0.0051], device='cuda:4') 2023-04-27 14:26:00,859 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=112617.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:26:11,160 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=112623.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:26:19,269 INFO [finetune.py:976] (4/7) Epoch 20, batch 3800, loss[loss=0.141, simple_loss=0.2211, pruned_loss=0.03045, over 4829.00 frames. ], tot_loss[loss=0.1762, simple_loss=0.2486, pruned_loss=0.05186, over 954523.16 frames. ], batch size: 30, lr: 3.23e-03, grad_scale: 32.0 2023-04-27 14:26:31,100 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=112638.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 14:26:42,519 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7373, 2.1099, 1.8713, 2.0061, 1.6596, 1.7911, 1.8446, 1.4655], device='cuda:4'), covar=tensor([0.1849, 0.1150, 0.0772, 0.1104, 0.3128, 0.1061, 0.1782, 0.2335], device='cuda:4'), in_proj_covar=tensor([0.0284, 0.0298, 0.0213, 0.0275, 0.0308, 0.0255, 0.0245, 0.0258], device='cuda:4'), out_proj_covar=tensor([1.1367e-04, 1.1812e-04, 8.4402e-05, 1.0878e-04, 1.2481e-04, 1.0125e-04, 9.9227e-05, 1.0235e-04], device='cuda:4') 2023-04-27 14:27:02,945 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.113e+02 1.520e+02 1.863e+02 2.243e+02 5.060e+02, threshold=3.726e+02, percent-clipped=6.0 2023-04-27 14:27:06,895 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.78 vs. limit=2.0 2023-04-27 14:27:25,226 INFO [finetune.py:976] (4/7) Epoch 20, batch 3850, loss[loss=0.1598, simple_loss=0.2356, pruned_loss=0.04198, over 4840.00 frames. ], tot_loss[loss=0.174, simple_loss=0.2464, pruned_loss=0.0508, over 953656.55 frames. ], batch size: 47, lr: 3.23e-03, grad_scale: 32.0 2023-04-27 14:27:25,947 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=112678.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 14:28:09,555 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=112722.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:28:13,013 INFO [finetune.py:976] (4/7) Epoch 20, batch 3900, loss[loss=0.169, simple_loss=0.2405, pruned_loss=0.04874, over 4764.00 frames. ], tot_loss[loss=0.1721, simple_loss=0.2439, pruned_loss=0.05013, over 952897.47 frames. ], batch size: 59, lr: 3.23e-03, grad_scale: 32.0 2023-04-27 14:28:25,004 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=112745.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:28:34,457 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.220e+01 1.530e+02 1.887e+02 2.253e+02 4.348e+02, threshold=3.774e+02, percent-clipped=2.0 2023-04-27 14:28:39,451 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0713, 1.4437, 1.3492, 1.6283, 1.5972, 1.5264, 1.4075, 2.3941], device='cuda:4'), covar=tensor([0.0639, 0.0857, 0.0797, 0.1242, 0.0618, 0.0471, 0.0712, 0.0240], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0040, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 14:28:41,692 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=112770.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:28:45,926 INFO [finetune.py:976] (4/7) Epoch 20, batch 3950, loss[loss=0.1725, simple_loss=0.2357, pruned_loss=0.05463, over 4813.00 frames. ], tot_loss[loss=0.169, simple_loss=0.2402, pruned_loss=0.04893, over 953940.73 frames. ], batch size: 51, lr: 3.23e-03, grad_scale: 32.0 2023-04-27 14:28:50,193 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.22 vs. limit=5.0 2023-04-27 14:28:53,818 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-04-27 14:29:16,287 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.81 vs. limit=5.0 2023-04-27 14:29:20,148 INFO [finetune.py:976] (4/7) Epoch 20, batch 4000, loss[loss=0.1763, simple_loss=0.2418, pruned_loss=0.05543, over 4040.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.2401, pruned_loss=0.0494, over 953472.59 frames. ], batch size: 17, lr: 3.23e-03, grad_scale: 32.0 2023-04-27 14:29:21,494 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2770, 1.6271, 2.0445, 2.4500, 2.0460, 1.6410, 1.3782, 1.8012], device='cuda:4'), covar=tensor([0.2949, 0.3076, 0.1611, 0.2048, 0.2532, 0.2537, 0.3986, 0.2031], device='cuda:4'), in_proj_covar=tensor([0.0291, 0.0246, 0.0227, 0.0313, 0.0219, 0.0233, 0.0228, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 14:29:40,401 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7706, 0.9961, 1.6985, 2.2042, 1.8203, 1.6604, 1.6957, 1.7110], device='cuda:4'), covar=tensor([0.4803, 0.7340, 0.6390, 0.6363, 0.6677, 0.8401, 0.8707, 0.8114], device='cuda:4'), in_proj_covar=tensor([0.0431, 0.0411, 0.0506, 0.0508, 0.0457, 0.0486, 0.0494, 0.0498], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 14:29:42,039 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.029e+02 1.556e+02 1.866e+02 2.319e+02 5.001e+02, threshold=3.732e+02, percent-clipped=1.0 2023-04-27 14:29:46,123 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6788, 1.7461, 0.8278, 1.4182, 2.0692, 1.5172, 1.4778, 1.5362], device='cuda:4'), covar=tensor([0.0480, 0.0378, 0.0334, 0.0556, 0.0258, 0.0510, 0.0474, 0.0560], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0029, 0.0019, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0045, 0.0038, 0.0052, 0.0038, 0.0050, 0.0050, 0.0051], device='cuda:4') 2023-04-27 14:29:53,756 INFO [finetune.py:976] (4/7) Epoch 20, batch 4050, loss[loss=0.2037, simple_loss=0.2821, pruned_loss=0.06263, over 4744.00 frames. ], tot_loss[loss=0.1735, simple_loss=0.2446, pruned_loss=0.05117, over 955095.37 frames. ], batch size: 54, lr: 3.23e-03, grad_scale: 32.0 2023-04-27 14:30:17,049 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1220, 2.7285, 2.3726, 2.4233, 1.8895, 2.3989, 2.5018, 1.8778], device='cuda:4'), covar=tensor([0.2289, 0.1116, 0.0678, 0.1466, 0.3583, 0.1102, 0.2030, 0.2892], device='cuda:4'), in_proj_covar=tensor([0.0281, 0.0294, 0.0211, 0.0272, 0.0304, 0.0252, 0.0243, 0.0255], device='cuda:4'), out_proj_covar=tensor([1.1257e-04, 1.1670e-04, 8.3520e-05, 1.0780e-04, 1.2345e-04, 9.9883e-05, 9.8311e-05, 1.0114e-04], device='cuda:4') 2023-04-27 14:30:21,648 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=112918.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:30:23,578 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=112921.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:30:27,669 INFO [finetune.py:976] (4/7) Epoch 20, batch 4100, loss[loss=0.1286, simple_loss=0.2035, pruned_loss=0.02687, over 4763.00 frames. ], tot_loss[loss=0.1741, simple_loss=0.2459, pruned_loss=0.05117, over 952551.90 frames. ], batch size: 26, lr: 3.23e-03, grad_scale: 32.0 2023-04-27 14:30:31,347 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=112933.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 14:31:11,352 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.236e+02 1.611e+02 1.943e+02 2.315e+02 4.443e+02, threshold=3.885e+02, percent-clipped=3.0 2023-04-27 14:31:12,531 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.71 vs. limit=2.0 2023-04-27 14:31:22,763 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9789, 2.4108, 2.1905, 2.1916, 1.6447, 2.0920, 2.0424, 1.7404], device='cuda:4'), covar=tensor([0.1761, 0.0929, 0.0602, 0.1210, 0.3076, 0.0848, 0.1559, 0.2281], device='cuda:4'), in_proj_covar=tensor([0.0282, 0.0294, 0.0211, 0.0273, 0.0305, 0.0252, 0.0244, 0.0255], device='cuda:4'), out_proj_covar=tensor([1.1266e-04, 1.1678e-04, 8.3480e-05, 1.0801e-04, 1.2393e-04, 9.9970e-05, 9.8465e-05, 1.0109e-04], device='cuda:4') 2023-04-27 14:31:25,617 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=112973.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 14:31:33,460 INFO [finetune.py:976] (4/7) Epoch 20, batch 4150, loss[loss=0.1648, simple_loss=0.2148, pruned_loss=0.05738, over 4264.00 frames. ], tot_loss[loss=0.1771, simple_loss=0.2491, pruned_loss=0.05257, over 953716.79 frames. ], batch size: 18, lr: 3.23e-03, grad_scale: 32.0 2023-04-27 14:31:42,877 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=112982.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:32:06,228 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5619, 1.7473, 1.4588, 1.1870, 1.1798, 1.1778, 1.4812, 1.1387], device='cuda:4'), covar=tensor([0.1721, 0.1313, 0.1527, 0.1629, 0.2259, 0.2012, 0.1041, 0.2030], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0212, 0.0169, 0.0204, 0.0200, 0.0186, 0.0156, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 14:32:29,692 INFO [finetune.py:976] (4/7) Epoch 20, batch 4200, loss[loss=0.2034, simple_loss=0.2828, pruned_loss=0.06197, over 4307.00 frames. ], tot_loss[loss=0.1778, simple_loss=0.2502, pruned_loss=0.05276, over 955456.21 frames. ], batch size: 66, lr: 3.23e-03, grad_scale: 32.0 2023-04-27 14:32:41,736 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=113045.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:32:42,849 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3024, 1.2943, 1.3930, 1.6068, 1.6583, 1.2916, 0.8828, 1.4773], device='cuda:4'), covar=tensor([0.0802, 0.1277, 0.0854, 0.0615, 0.0633, 0.0822, 0.0882, 0.0576], device='cuda:4'), in_proj_covar=tensor([0.0187, 0.0202, 0.0184, 0.0172, 0.0177, 0.0182, 0.0153, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 14:32:57,028 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.082e+02 1.553e+02 1.872e+02 2.220e+02 4.301e+02, threshold=3.745e+02, percent-clipped=1.0 2023-04-27 14:33:18,845 INFO [finetune.py:976] (4/7) Epoch 20, batch 4250, loss[loss=0.1644, simple_loss=0.232, pruned_loss=0.04846, over 4715.00 frames. ], tot_loss[loss=0.1766, simple_loss=0.2485, pruned_loss=0.0524, over 957197.85 frames. ], batch size: 54, lr: 3.23e-03, grad_scale: 32.0 2023-04-27 14:33:40,239 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=113093.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:33:42,105 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3731, 1.5655, 1.8108, 1.9190, 1.7354, 1.8794, 1.8874, 1.8965], device='cuda:4'), covar=tensor([0.3805, 0.5327, 0.4251, 0.4242, 0.5670, 0.7368, 0.4807, 0.4816], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0374, 0.0322, 0.0336, 0.0346, 0.0395, 0.0357, 0.0328], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 14:34:04,771 INFO [finetune.py:976] (4/7) Epoch 20, batch 4300, loss[loss=0.1498, simple_loss=0.2235, pruned_loss=0.03806, over 4908.00 frames. ], tot_loss[loss=0.1742, simple_loss=0.2457, pruned_loss=0.05137, over 957500.67 frames. ], batch size: 36, lr: 3.23e-03, grad_scale: 32.0 2023-04-27 14:34:13,191 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.6371, 4.4518, 3.1909, 5.2878, 4.6049, 4.5654, 1.9384, 4.5437], device='cuda:4'), covar=tensor([0.1497, 0.1048, 0.3040, 0.0937, 0.3102, 0.1617, 0.5438, 0.2123], device='cuda:4'), in_proj_covar=tensor([0.0245, 0.0215, 0.0250, 0.0306, 0.0295, 0.0247, 0.0273, 0.0272], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 14:34:14,470 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0917, 1.9912, 1.9647, 1.7069, 2.0904, 1.6799, 2.5511, 1.7298], device='cuda:4'), covar=tensor([0.2753, 0.1368, 0.3016, 0.2136, 0.1233, 0.1986, 0.1135, 0.3186], device='cuda:4'), in_proj_covar=tensor([0.0338, 0.0345, 0.0421, 0.0351, 0.0377, 0.0371, 0.0368, 0.0413], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 14:34:27,841 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.100e+02 1.602e+02 1.982e+02 2.267e+02 5.350e+02, threshold=3.963e+02, percent-clipped=3.0 2023-04-27 14:34:30,031 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.70 vs. limit=2.0 2023-04-27 14:34:38,150 INFO [finetune.py:976] (4/7) Epoch 20, batch 4350, loss[loss=0.1516, simple_loss=0.2249, pruned_loss=0.03913, over 4747.00 frames. ], tot_loss[loss=0.1726, simple_loss=0.2428, pruned_loss=0.05121, over 954750.07 frames. ], batch size: 28, lr: 3.23e-03, grad_scale: 32.0 2023-04-27 14:35:06,246 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=113218.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:35:11,650 INFO [finetune.py:976] (4/7) Epoch 20, batch 4400, loss[loss=0.2452, simple_loss=0.3025, pruned_loss=0.09394, over 4906.00 frames. ], tot_loss[loss=0.173, simple_loss=0.2432, pruned_loss=0.05139, over 955483.40 frames. ], batch size: 36, lr: 3.23e-03, grad_scale: 32.0 2023-04-27 14:35:15,442 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=113233.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:35:16,046 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=113234.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:35:34,719 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.136e+02 1.671e+02 1.942e+02 2.362e+02 3.791e+02, threshold=3.884e+02, percent-clipped=0.0 2023-04-27 14:35:38,475 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=113266.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:35:42,695 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=113273.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 14:35:43,952 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.2179, 2.9917, 3.2281, 3.6702, 3.5269, 3.2867, 2.6665, 3.5008], device='cuda:4'), covar=tensor([0.0677, 0.0808, 0.0495, 0.0443, 0.0427, 0.0555, 0.0627, 0.0390], device='cuda:4'), in_proj_covar=tensor([0.0187, 0.0202, 0.0184, 0.0172, 0.0178, 0.0182, 0.0153, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 14:35:45,063 INFO [finetune.py:976] (4/7) Epoch 20, batch 4450, loss[loss=0.2138, simple_loss=0.2903, pruned_loss=0.06863, over 4831.00 frames. ], tot_loss[loss=0.1754, simple_loss=0.2469, pruned_loss=0.05196, over 956258.35 frames. ], batch size: 33, lr: 3.23e-03, grad_scale: 32.0 2023-04-27 14:35:45,128 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=113277.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:35:47,574 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=113281.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:35:57,005 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=113295.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:36:15,056 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=113321.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:36:18,629 INFO [finetune.py:976] (4/7) Epoch 20, batch 4500, loss[loss=0.2183, simple_loss=0.2763, pruned_loss=0.08017, over 4259.00 frames. ], tot_loss[loss=0.1763, simple_loss=0.2479, pruned_loss=0.05235, over 954373.32 frames. ], batch size: 65, lr: 3.22e-03, grad_scale: 32.0 2023-04-27 14:37:03,577 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.082e+01 1.625e+02 1.957e+02 2.322e+02 4.205e+02, threshold=3.914e+02, percent-clipped=1.0 2023-04-27 14:37:25,452 INFO [finetune.py:976] (4/7) Epoch 20, batch 4550, loss[loss=0.1951, simple_loss=0.27, pruned_loss=0.06012, over 4918.00 frames. ], tot_loss[loss=0.177, simple_loss=0.2494, pruned_loss=0.05229, over 956026.85 frames. ], batch size: 42, lr: 3.22e-03, grad_scale: 32.0 2023-04-27 14:37:53,074 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4111, 3.1490, 0.9173, 1.6935, 1.8595, 2.2827, 1.8171, 0.9611], device='cuda:4'), covar=tensor([0.1499, 0.1099, 0.1968, 0.1366, 0.1177, 0.1036, 0.1630, 0.1862], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0241, 0.0137, 0.0120, 0.0134, 0.0152, 0.0117, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 14:37:54,862 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2048, 1.3535, 1.6982, 1.7773, 1.6674, 1.7774, 1.7156, 1.7345], device='cuda:4'), covar=tensor([0.3385, 0.5282, 0.4198, 0.4365, 0.5427, 0.7385, 0.5039, 0.4666], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0373, 0.0320, 0.0335, 0.0345, 0.0394, 0.0356, 0.0327], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 14:38:02,717 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0094, 1.6132, 1.4573, 1.8781, 2.1278, 1.7927, 1.5502, 1.3708], device='cuda:4'), covar=tensor([0.1394, 0.1525, 0.2059, 0.1319, 0.0920, 0.1497, 0.1949, 0.2525], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0314, 0.0352, 0.0290, 0.0328, 0.0308, 0.0302, 0.0375], device='cuda:4'), out_proj_covar=tensor([6.4325e-05, 6.5051e-05, 7.4380e-05, 5.8552e-05, 6.7711e-05, 6.4590e-05, 6.3312e-05, 7.9683e-05], device='cuda:4') 2023-04-27 14:38:05,040 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.4710, 4.4937, 3.1459, 5.1452, 4.5281, 4.4422, 1.7634, 4.4256], device='cuda:4'), covar=tensor([0.1606, 0.0855, 0.3206, 0.1094, 0.2748, 0.1687, 0.6174, 0.2397], device='cuda:4'), in_proj_covar=tensor([0.0248, 0.0217, 0.0253, 0.0308, 0.0298, 0.0250, 0.0276, 0.0273], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 14:38:22,955 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6168, 1.9154, 2.1034, 2.1658, 2.0026, 2.0097, 2.1268, 2.0933], device='cuda:4'), covar=tensor([0.3937, 0.5210, 0.4366, 0.4005, 0.5434, 0.7233, 0.5178, 0.4622], device='cuda:4'), in_proj_covar=tensor([0.0335, 0.0372, 0.0319, 0.0334, 0.0344, 0.0392, 0.0354, 0.0326], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 14:38:24,624 INFO [finetune.py:976] (4/7) Epoch 20, batch 4600, loss[loss=0.1957, simple_loss=0.2527, pruned_loss=0.0694, over 4906.00 frames. ], tot_loss[loss=0.1755, simple_loss=0.2481, pruned_loss=0.05148, over 955664.34 frames. ], batch size: 32, lr: 3.22e-03, grad_scale: 32.0 2023-04-27 14:38:48,544 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=2.01 vs. limit=2.0 2023-04-27 14:39:00,869 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.917e+01 1.466e+02 1.708e+02 2.051e+02 4.040e+02, threshold=3.416e+02, percent-clipped=1.0 2023-04-27 14:39:02,948 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.67 vs. limit=5.0 2023-04-27 14:39:11,483 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7341, 1.1334, 1.7698, 2.1996, 1.7945, 1.6510, 1.7250, 1.6950], device='cuda:4'), covar=tensor([0.4559, 0.7030, 0.6148, 0.5529, 0.6175, 0.7725, 0.7605, 0.8505], device='cuda:4'), in_proj_covar=tensor([0.0430, 0.0411, 0.0505, 0.0505, 0.0458, 0.0485, 0.0493, 0.0499], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 14:39:13,176 INFO [finetune.py:976] (4/7) Epoch 20, batch 4650, loss[loss=0.1897, simple_loss=0.2593, pruned_loss=0.05999, over 4904.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.2458, pruned_loss=0.05106, over 956003.72 frames. ], batch size: 35, lr: 3.22e-03, grad_scale: 32.0 2023-04-27 14:39:40,665 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2628, 1.9926, 2.1695, 2.6496, 2.5918, 2.0983, 1.7978, 2.3248], device='cuda:4'), covar=tensor([0.0761, 0.0947, 0.0685, 0.0513, 0.0538, 0.0759, 0.0710, 0.0537], device='cuda:4'), in_proj_covar=tensor([0.0188, 0.0203, 0.0184, 0.0172, 0.0178, 0.0182, 0.0153, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 14:39:47,028 INFO [finetune.py:976] (4/7) Epoch 20, batch 4700, loss[loss=0.1392, simple_loss=0.2219, pruned_loss=0.02825, over 4766.00 frames. ], tot_loss[loss=0.1708, simple_loss=0.2421, pruned_loss=0.04975, over 954906.95 frames. ], batch size: 28, lr: 3.22e-03, grad_scale: 32.0 2023-04-27 14:40:02,398 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5909, 2.9647, 0.8225, 1.6665, 2.2123, 1.5305, 4.3158, 2.2559], device='cuda:4'), covar=tensor([0.0590, 0.0778, 0.0906, 0.1222, 0.0517, 0.0988, 0.0189, 0.0574], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0073, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0007, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 14:40:08,114 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.331e+01 1.541e+02 1.809e+02 2.201e+02 3.607e+02, threshold=3.618e+02, percent-clipped=1.0 2023-04-27 14:40:19,994 INFO [finetune.py:976] (4/7) Epoch 20, batch 4750, loss[loss=0.1882, simple_loss=0.2622, pruned_loss=0.05714, over 4853.00 frames. ], tot_loss[loss=0.1683, simple_loss=0.2393, pruned_loss=0.04865, over 956616.84 frames. ], batch size: 47, lr: 3.22e-03, grad_scale: 32.0 2023-04-27 14:40:20,113 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=113577.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:40:29,023 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=113590.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:40:36,588 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.13 vs. limit=2.0 2023-04-27 14:40:46,847 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=113617.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:40:52,167 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=113625.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:40:53,787 INFO [finetune.py:976] (4/7) Epoch 20, batch 4800, loss[loss=0.2171, simple_loss=0.2887, pruned_loss=0.07276, over 4911.00 frames. ], tot_loss[loss=0.1709, simple_loss=0.242, pruned_loss=0.04992, over 955544.47 frames. ], batch size: 42, lr: 3.22e-03, grad_scale: 32.0 2023-04-27 14:41:05,320 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5157, 1.3738, 1.7641, 1.8443, 1.3433, 1.2294, 1.5169, 0.9677], device='cuda:4'), covar=tensor([0.0500, 0.0670, 0.0401, 0.0542, 0.0742, 0.1211, 0.0593, 0.0650], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0069, 0.0068, 0.0068, 0.0076, 0.0097, 0.0074, 0.0066], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 14:41:06,516 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=113646.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:41:15,406 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.167e+02 1.571e+02 1.898e+02 2.197e+02 3.707e+02, threshold=3.796e+02, percent-clipped=1.0 2023-04-27 14:41:27,083 INFO [finetune.py:976] (4/7) Epoch 20, batch 4850, loss[loss=0.1789, simple_loss=0.2545, pruned_loss=0.05165, over 4849.00 frames. ], tot_loss[loss=0.1743, simple_loss=0.2458, pruned_loss=0.05141, over 954492.89 frames. ], batch size: 47, lr: 3.22e-03, grad_scale: 32.0 2023-04-27 14:41:27,787 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=113678.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:41:46,455 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=113707.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:41:58,263 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.58 vs. limit=2.0 2023-04-27 14:42:00,311 INFO [finetune.py:976] (4/7) Epoch 20, batch 4900, loss[loss=0.1602, simple_loss=0.2351, pruned_loss=0.04264, over 4800.00 frames. ], tot_loss[loss=0.1751, simple_loss=0.2469, pruned_loss=0.05172, over 954726.70 frames. ], batch size: 29, lr: 3.22e-03, grad_scale: 32.0 2023-04-27 14:42:27,118 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.158e+02 1.699e+02 1.953e+02 2.350e+02 6.319e+02, threshold=3.906e+02, percent-clipped=4.0 2023-04-27 14:42:38,976 INFO [finetune.py:976] (4/7) Epoch 20, batch 4950, loss[loss=0.2006, simple_loss=0.2682, pruned_loss=0.06652, over 4825.00 frames. ], tot_loss[loss=0.1772, simple_loss=0.2492, pruned_loss=0.05259, over 954982.27 frames. ], batch size: 30, lr: 3.22e-03, grad_scale: 32.0 2023-04-27 14:43:30,855 INFO [finetune.py:976] (4/7) Epoch 20, batch 5000, loss[loss=0.1812, simple_loss=0.2384, pruned_loss=0.06195, over 4825.00 frames. ], tot_loss[loss=0.1755, simple_loss=0.2475, pruned_loss=0.05178, over 955817.09 frames. ], batch size: 30, lr: 3.22e-03, grad_scale: 32.0 2023-04-27 14:43:42,492 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.36 vs. limit=5.0 2023-04-27 14:44:05,777 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5211, 1.7227, 1.5251, 2.0244, 1.8332, 1.9800, 1.5877, 3.6793], device='cuda:4'), covar=tensor([0.0621, 0.0974, 0.1001, 0.1146, 0.0689, 0.0636, 0.0911, 0.0215], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 14:44:14,003 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.052e+02 1.541e+02 1.777e+02 2.201e+02 3.232e+02, threshold=3.554e+02, percent-clipped=0.0 2023-04-27 14:44:22,336 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2523, 1.5367, 1.3735, 1.7906, 1.7153, 1.6968, 1.4247, 3.0659], device='cuda:4'), covar=tensor([0.0618, 0.0850, 0.0838, 0.1183, 0.0580, 0.0460, 0.0727, 0.0192], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 14:44:32,033 INFO [finetune.py:976] (4/7) Epoch 20, batch 5050, loss[loss=0.1743, simple_loss=0.2357, pruned_loss=0.05645, over 4117.00 frames. ], tot_loss[loss=0.1742, simple_loss=0.2455, pruned_loss=0.0515, over 957297.84 frames. ], batch size: 65, lr: 3.22e-03, grad_scale: 32.0 2023-04-27 14:44:52,896 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=113890.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:44:52,984 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.09 vs. limit=5.0 2023-04-27 14:45:34,210 INFO [finetune.py:976] (4/7) Epoch 20, batch 5100, loss[loss=0.1875, simple_loss=0.2512, pruned_loss=0.06185, over 4893.00 frames. ], tot_loss[loss=0.1705, simple_loss=0.2413, pruned_loss=0.04984, over 956810.32 frames. ], batch size: 35, lr: 3.22e-03, grad_scale: 64.0 2023-04-27 14:45:42,070 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=113938.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:45:57,339 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.243e+01 1.576e+02 1.837e+02 2.238e+02 4.174e+02, threshold=3.673e+02, percent-clipped=2.0 2023-04-27 14:45:58,687 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8711, 1.8199, 2.1699, 2.3372, 1.7564, 1.5645, 1.8360, 1.0371], device='cuda:4'), covar=tensor([0.0551, 0.0571, 0.0408, 0.0687, 0.0706, 0.0967, 0.0635, 0.0726], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0068, 0.0067, 0.0068, 0.0076, 0.0096, 0.0073, 0.0066], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 14:46:05,377 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=113973.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:46:08,237 INFO [finetune.py:976] (4/7) Epoch 20, batch 5150, loss[loss=0.1804, simple_loss=0.2548, pruned_loss=0.053, over 4769.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2418, pruned_loss=0.05017, over 957408.49 frames. ], batch size: 28, lr: 3.22e-03, grad_scale: 64.0 2023-04-27 14:46:27,908 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=114002.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:46:32,282 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0350, 1.7742, 1.9753, 2.4826, 2.4686, 1.9894, 1.6002, 2.1457], device='cuda:4'), covar=tensor([0.0871, 0.1114, 0.0751, 0.0551, 0.0575, 0.0827, 0.0859, 0.0563], device='cuda:4'), in_proj_covar=tensor([0.0188, 0.0202, 0.0184, 0.0172, 0.0178, 0.0181, 0.0153, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 14:46:39,846 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-27 14:46:43,546 INFO [finetune.py:976] (4/7) Epoch 20, batch 5200, loss[loss=0.1779, simple_loss=0.2597, pruned_loss=0.04808, over 4724.00 frames. ], tot_loss[loss=0.175, simple_loss=0.246, pruned_loss=0.05198, over 957124.59 frames. ], batch size: 23, lr: 3.22e-03, grad_scale: 64.0 2023-04-27 14:47:06,607 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.860e+01 1.682e+02 1.942e+02 2.331e+02 4.447e+02, threshold=3.884e+02, percent-clipped=1.0 2023-04-27 14:47:16,855 INFO [finetune.py:976] (4/7) Epoch 20, batch 5250, loss[loss=0.1929, simple_loss=0.2672, pruned_loss=0.05927, over 4921.00 frames. ], tot_loss[loss=0.1768, simple_loss=0.2483, pruned_loss=0.05264, over 956301.62 frames. ], batch size: 38, lr: 3.22e-03, grad_scale: 64.0 2023-04-27 14:47:50,697 INFO [finetune.py:976] (4/7) Epoch 20, batch 5300, loss[loss=0.1769, simple_loss=0.2582, pruned_loss=0.04777, over 4886.00 frames. ], tot_loss[loss=0.1765, simple_loss=0.2485, pruned_loss=0.05228, over 956011.96 frames. ], batch size: 43, lr: 3.22e-03, grad_scale: 64.0 2023-04-27 14:47:50,806 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=114127.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:48:13,280 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.642e+01 1.650e+02 1.908e+02 2.204e+02 4.976e+02, threshold=3.816e+02, percent-clipped=2.0 2023-04-27 14:48:24,237 INFO [finetune.py:976] (4/7) Epoch 20, batch 5350, loss[loss=0.1736, simple_loss=0.243, pruned_loss=0.05209, over 4892.00 frames. ], tot_loss[loss=0.1761, simple_loss=0.2484, pruned_loss=0.05185, over 956486.05 frames. ], batch size: 43, lr: 3.22e-03, grad_scale: 64.0 2023-04-27 14:48:31,427 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=114188.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:48:56,951 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2005, 1.5920, 1.4466, 1.7204, 1.6537, 1.9587, 1.3905, 3.5676], device='cuda:4'), covar=tensor([0.0643, 0.0815, 0.0797, 0.1225, 0.0642, 0.0546, 0.0763, 0.0130], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 14:48:58,063 INFO [finetune.py:976] (4/7) Epoch 20, batch 5400, loss[loss=0.1967, simple_loss=0.2631, pruned_loss=0.06515, over 4905.00 frames. ], tot_loss[loss=0.1747, simple_loss=0.2465, pruned_loss=0.0514, over 954992.32 frames. ], batch size: 43, lr: 3.22e-03, grad_scale: 64.0 2023-04-27 14:49:05,311 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1560, 2.5055, 0.9406, 1.3451, 1.6372, 1.1511, 3.0070, 1.5453], device='cuda:4'), covar=tensor([0.0617, 0.0523, 0.0740, 0.1204, 0.0510, 0.0999, 0.0259, 0.0662], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0065, 0.0048, 0.0046, 0.0049, 0.0052, 0.0073, 0.0052], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 14:49:33,868 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.776e+01 1.553e+02 1.821e+02 2.286e+02 4.099e+02, threshold=3.642e+02, percent-clipped=1.0 2023-04-27 14:49:46,385 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0230, 1.7788, 1.9756, 2.2954, 2.3371, 1.8963, 1.4665, 2.0633], device='cuda:4'), covar=tensor([0.0722, 0.1025, 0.0674, 0.0570, 0.0512, 0.0826, 0.0797, 0.0519], device='cuda:4'), in_proj_covar=tensor([0.0187, 0.0201, 0.0183, 0.0171, 0.0177, 0.0181, 0.0152, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 14:49:53,932 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=114273.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:49:56,284 INFO [finetune.py:976] (4/7) Epoch 20, batch 5450, loss[loss=0.1641, simple_loss=0.2262, pruned_loss=0.051, over 4762.00 frames. ], tot_loss[loss=0.1731, simple_loss=0.2444, pruned_loss=0.05088, over 956175.24 frames. ], batch size: 27, lr: 3.22e-03, grad_scale: 64.0 2023-04-27 14:50:28,224 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2410, 1.7576, 1.6542, 1.8488, 1.8324, 1.9601, 1.4809, 3.8206], device='cuda:4'), covar=tensor([0.0641, 0.0799, 0.0755, 0.1190, 0.0606, 0.0488, 0.0762, 0.0120], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 14:50:28,231 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=114302.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:50:48,638 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4824, 1.3068, 1.7484, 1.7207, 1.3464, 1.2395, 1.3660, 0.8918], device='cuda:4'), covar=tensor([0.0520, 0.0664, 0.0354, 0.0661, 0.0804, 0.1068, 0.0587, 0.0584], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0068, 0.0067, 0.0068, 0.0075, 0.0096, 0.0073, 0.0065], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 14:50:58,437 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=114321.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:51:02,104 INFO [finetune.py:976] (4/7) Epoch 20, batch 5500, loss[loss=0.1709, simple_loss=0.2345, pruned_loss=0.05358, over 4820.00 frames. ], tot_loss[loss=0.171, simple_loss=0.2416, pruned_loss=0.05021, over 956973.41 frames. ], batch size: 39, lr: 3.22e-03, grad_scale: 64.0 2023-04-27 14:51:22,114 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=114350.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:51:28,689 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.001e+02 1.538e+02 1.938e+02 2.403e+02 5.552e+02, threshold=3.877e+02, percent-clipped=2.0 2023-04-27 14:51:33,235 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-27 14:51:41,082 INFO [finetune.py:976] (4/7) Epoch 20, batch 5550, loss[loss=0.1581, simple_loss=0.2379, pruned_loss=0.0392, over 4778.00 frames. ], tot_loss[loss=0.1731, simple_loss=0.2435, pruned_loss=0.05133, over 958119.27 frames. ], batch size: 29, lr: 3.22e-03, grad_scale: 64.0 2023-04-27 14:51:41,796 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4048, 2.5971, 1.1851, 1.5165, 2.0581, 1.4401, 3.3894, 1.8361], device='cuda:4'), covar=tensor([0.0560, 0.0576, 0.0718, 0.1163, 0.0476, 0.0912, 0.0394, 0.0623], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0065, 0.0048, 0.0046, 0.0050, 0.0052, 0.0074, 0.0052], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 14:51:50,167 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.73 vs. limit=2.0 2023-04-27 14:51:53,667 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.97 vs. limit=2.0 2023-04-27 14:52:13,279 INFO [finetune.py:976] (4/7) Epoch 20, batch 5600, loss[loss=0.1584, simple_loss=0.2373, pruned_loss=0.03975, over 4819.00 frames. ], tot_loss[loss=0.1763, simple_loss=0.2476, pruned_loss=0.0525, over 956264.03 frames. ], batch size: 39, lr: 3.22e-03, grad_scale: 64.0 2023-04-27 14:52:32,509 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.241e+02 1.607e+02 1.931e+02 2.416e+02 4.718e+02, threshold=3.861e+02, percent-clipped=3.0 2023-04-27 14:52:42,450 INFO [finetune.py:976] (4/7) Epoch 20, batch 5650, loss[loss=0.1308, simple_loss=0.2116, pruned_loss=0.02504, over 4760.00 frames. ], tot_loss[loss=0.1771, simple_loss=0.2494, pruned_loss=0.0524, over 957375.21 frames. ], batch size: 26, lr: 3.22e-03, grad_scale: 64.0 2023-04-27 14:52:46,425 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=114483.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:52:58,911 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3434, 1.9174, 2.2480, 2.5980, 2.2421, 1.8945, 1.4878, 1.9880], device='cuda:4'), covar=tensor([0.3438, 0.3107, 0.1620, 0.2025, 0.2503, 0.2639, 0.4053, 0.1925], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0245, 0.0226, 0.0313, 0.0220, 0.0233, 0.0227, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 14:53:12,738 INFO [finetune.py:976] (4/7) Epoch 20, batch 5700, loss[loss=0.1588, simple_loss=0.2001, pruned_loss=0.05872, over 4061.00 frames. ], tot_loss[loss=0.1736, simple_loss=0.2455, pruned_loss=0.05088, over 945084.07 frames. ], batch size: 17, lr: 3.22e-03, grad_scale: 64.0 2023-04-27 14:53:39,071 INFO [finetune.py:976] (4/7) Epoch 21, batch 0, loss[loss=0.1822, simple_loss=0.2544, pruned_loss=0.055, over 4772.00 frames. ], tot_loss[loss=0.1822, simple_loss=0.2544, pruned_loss=0.055, over 4772.00 frames. ], batch size: 28, lr: 3.21e-03, grad_scale: 64.0 2023-04-27 14:53:39,072 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 14:53:46,366 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6246, 1.6305, 3.6469, 3.4275, 3.3018, 3.3923, 3.5681, 3.2598], device='cuda:4'), covar=tensor([0.5857, 0.4449, 0.1114, 0.1617, 0.1066, 0.1567, 0.0820, 0.1388], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0306, 0.0403, 0.0403, 0.0347, 0.0407, 0.0311, 0.0365], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 14:53:58,292 INFO [finetune.py:1010] (4/7) Epoch 21, validation: loss=0.1544, simple_loss=0.2245, pruned_loss=0.04212, over 2265189.00 frames. 2023-04-27 14:53:58,293 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-27 14:54:02,564 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 7.383e+01 1.462e+02 1.753e+02 2.110e+02 4.375e+02, threshold=3.507e+02, percent-clipped=1.0 2023-04-27 14:54:37,746 INFO [finetune.py:976] (4/7) Epoch 21, batch 50, loss[loss=0.1793, simple_loss=0.2516, pruned_loss=0.0535, over 4723.00 frames. ], tot_loss[loss=0.1751, simple_loss=0.2474, pruned_loss=0.0514, over 217766.98 frames. ], batch size: 54, lr: 3.21e-03, grad_scale: 64.0 2023-04-27 14:55:37,894 INFO [finetune.py:976] (4/7) Epoch 21, batch 100, loss[loss=0.1456, simple_loss=0.2081, pruned_loss=0.04154, over 4080.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2397, pruned_loss=0.04956, over 378776.36 frames. ], batch size: 66, lr: 3.21e-03, grad_scale: 64.0 2023-04-27 14:55:48,249 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.033e+02 1.526e+02 1.765e+02 2.101e+02 5.147e+02, threshold=3.531e+02, percent-clipped=4.0 2023-04-27 14:56:23,858 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7104, 1.0014, 1.6811, 2.1874, 1.8132, 1.6551, 1.6740, 1.6383], device='cuda:4'), covar=tensor([0.3846, 0.5929, 0.5197, 0.4908, 0.4716, 0.6698, 0.6352, 0.7671], device='cuda:4'), in_proj_covar=tensor([0.0427, 0.0410, 0.0504, 0.0504, 0.0455, 0.0483, 0.0493, 0.0496], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 14:56:44,431 INFO [finetune.py:976] (4/7) Epoch 21, batch 150, loss[loss=0.1886, simple_loss=0.2536, pruned_loss=0.06184, over 4899.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2375, pruned_loss=0.05008, over 507555.11 frames. ], batch size: 36, lr: 3.21e-03, grad_scale: 64.0 2023-04-27 14:56:45,178 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6413, 1.9356, 1.9671, 2.0754, 1.9837, 2.0479, 2.0621, 2.0794], device='cuda:4'), covar=tensor([0.4059, 0.5556, 0.4644, 0.4625, 0.5698, 0.7360, 0.5671, 0.4838], device='cuda:4'), in_proj_covar=tensor([0.0339, 0.0375, 0.0321, 0.0336, 0.0347, 0.0394, 0.0357, 0.0329], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 14:56:57,800 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=114716.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:57:22,715 INFO [finetune.py:976] (4/7) Epoch 21, batch 200, loss[loss=0.1345, simple_loss=0.2188, pruned_loss=0.02512, over 4768.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.2361, pruned_loss=0.04907, over 607505.87 frames. ], batch size: 26, lr: 3.21e-03, grad_scale: 64.0 2023-04-27 14:57:26,739 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.046e+02 1.480e+02 1.757e+02 1.981e+02 3.579e+02, threshold=3.513e+02, percent-clipped=1.0 2023-04-27 14:57:28,027 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.8154, 2.4091, 2.9255, 3.4579, 2.8326, 2.6960, 2.8930, 2.1216], device='cuda:4'), covar=tensor([0.0372, 0.0659, 0.0266, 0.0324, 0.0457, 0.0594, 0.0442, 0.0522], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0069, 0.0067, 0.0067, 0.0075, 0.0096, 0.0073, 0.0065], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 14:57:38,723 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=114777.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:57:42,328 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=114783.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:57:56,199 INFO [finetune.py:976] (4/7) Epoch 21, batch 250, loss[loss=0.2154, simple_loss=0.2847, pruned_loss=0.07305, over 4796.00 frames. ], tot_loss[loss=0.1707, simple_loss=0.2405, pruned_loss=0.05049, over 684795.68 frames. ], batch size: 45, lr: 3.21e-03, grad_scale: 32.0 2023-04-27 14:58:12,782 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6225, 1.9628, 1.7620, 2.5290, 2.6270, 2.2225, 2.0879, 1.7892], device='cuda:4'), covar=tensor([0.1541, 0.1737, 0.1868, 0.1367, 0.1047, 0.1477, 0.2044, 0.2261], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0313, 0.0351, 0.0290, 0.0327, 0.0309, 0.0303, 0.0374], device='cuda:4'), out_proj_covar=tensor([6.4143e-05, 6.4905e-05, 7.4186e-05, 5.8575e-05, 6.7459e-05, 6.4820e-05, 6.3520e-05, 7.9509e-05], device='cuda:4') 2023-04-27 14:58:15,123 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=114831.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:58:30,017 INFO [finetune.py:976] (4/7) Epoch 21, batch 300, loss[loss=0.1488, simple_loss=0.2132, pruned_loss=0.04219, over 4698.00 frames. ], tot_loss[loss=0.1723, simple_loss=0.2433, pruned_loss=0.05065, over 744118.92 frames. ], batch size: 23, lr: 3.21e-03, grad_scale: 32.0 2023-04-27 14:58:34,665 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.745e+01 1.667e+02 1.887e+02 2.264e+02 4.948e+02, threshold=3.774e+02, percent-clipped=2.0 2023-04-27 14:58:42,288 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=114871.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:59:03,089 INFO [finetune.py:976] (4/7) Epoch 21, batch 350, loss[loss=0.1506, simple_loss=0.2266, pruned_loss=0.03727, over 4759.00 frames. ], tot_loss[loss=0.1761, simple_loss=0.2472, pruned_loss=0.05248, over 791906.02 frames. ], batch size: 27, lr: 3.21e-03, grad_scale: 32.0 2023-04-27 14:59:22,184 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=114932.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 14:59:32,265 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2737, 2.8307, 2.4030, 2.6696, 1.9471, 2.4690, 2.5692, 1.8470], device='cuda:4'), covar=tensor([0.1963, 0.0992, 0.0694, 0.1099, 0.3118, 0.1017, 0.2003, 0.2539], device='cuda:4'), in_proj_covar=tensor([0.0283, 0.0298, 0.0214, 0.0274, 0.0309, 0.0253, 0.0246, 0.0260], device='cuda:4'), out_proj_covar=tensor([1.1338e-04, 1.1853e-04, 8.4804e-05, 1.0851e-04, 1.2547e-04, 1.0004e-04, 9.9484e-05, 1.0321e-04], device='cuda:4') 2023-04-27 14:59:36,225 INFO [finetune.py:976] (4/7) Epoch 21, batch 400, loss[loss=0.1404, simple_loss=0.2178, pruned_loss=0.03153, over 4893.00 frames. ], tot_loss[loss=0.1737, simple_loss=0.2458, pruned_loss=0.05082, over 828775.99 frames. ], batch size: 43, lr: 3.21e-03, grad_scale: 32.0 2023-04-27 14:59:40,904 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.017e+02 1.660e+02 1.983e+02 2.165e+02 4.861e+02, threshold=3.966e+02, percent-clipped=1.0 2023-04-27 14:59:58,630 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1336, 1.8761, 2.1610, 2.4835, 2.4914, 1.8962, 1.4723, 2.0986], device='cuda:4'), covar=tensor([0.0853, 0.1051, 0.0653, 0.0559, 0.0574, 0.0882, 0.0853, 0.0624], device='cuda:4'), in_proj_covar=tensor([0.0190, 0.0204, 0.0185, 0.0173, 0.0179, 0.0183, 0.0154, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 15:00:10,133 INFO [finetune.py:976] (4/7) Epoch 21, batch 450, loss[loss=0.1538, simple_loss=0.2187, pruned_loss=0.04443, over 4745.00 frames. ], tot_loss[loss=0.1737, simple_loss=0.2457, pruned_loss=0.05081, over 856035.54 frames. ], batch size: 27, lr: 3.21e-03, grad_scale: 32.0 2023-04-27 15:00:44,848 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8587, 2.3921, 2.1061, 2.3180, 1.8023, 2.0660, 2.0200, 1.4715], device='cuda:4'), covar=tensor([0.1947, 0.1144, 0.0776, 0.1142, 0.2786, 0.1146, 0.1892, 0.2536], device='cuda:4'), in_proj_covar=tensor([0.0284, 0.0299, 0.0215, 0.0274, 0.0309, 0.0253, 0.0247, 0.0261], device='cuda:4'), out_proj_covar=tensor([1.1361e-04, 1.1869e-04, 8.4968e-05, 1.0851e-04, 1.2533e-04, 1.0010e-04, 9.9689e-05, 1.0340e-04], device='cuda:4') 2023-04-27 15:00:59,033 INFO [finetune.py:976] (4/7) Epoch 21, batch 500, loss[loss=0.1488, simple_loss=0.2293, pruned_loss=0.03414, over 4827.00 frames. ], tot_loss[loss=0.1713, simple_loss=0.2432, pruned_loss=0.04973, over 877304.90 frames. ], batch size: 41, lr: 3.21e-03, grad_scale: 32.0 2023-04-27 15:00:59,216 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.58 vs. limit=5.0 2023-04-27 15:01:09,616 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.974e+01 1.571e+02 1.812e+02 2.235e+02 3.288e+02, threshold=3.623e+02, percent-clipped=0.0 2023-04-27 15:01:16,898 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=115072.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:01:32,716 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=115086.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:01:54,522 INFO [finetune.py:976] (4/7) Epoch 21, batch 550, loss[loss=0.1836, simple_loss=0.2498, pruned_loss=0.05867, over 4830.00 frames. ], tot_loss[loss=0.171, simple_loss=0.2418, pruned_loss=0.05005, over 894838.70 frames. ], batch size: 30, lr: 3.21e-03, grad_scale: 32.0 2023-04-27 15:02:53,399 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=115147.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:03:02,662 INFO [finetune.py:976] (4/7) Epoch 21, batch 600, loss[loss=0.1309, simple_loss=0.2047, pruned_loss=0.02853, over 4767.00 frames. ], tot_loss[loss=0.1703, simple_loss=0.2413, pruned_loss=0.04967, over 910317.02 frames. ], batch size: 27, lr: 3.21e-03, grad_scale: 32.0 2023-04-27 15:03:12,534 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.066e+02 1.494e+02 1.729e+02 2.315e+02 4.392e+02, threshold=3.458e+02, percent-clipped=4.0 2023-04-27 15:04:02,676 INFO [finetune.py:976] (4/7) Epoch 21, batch 650, loss[loss=0.1695, simple_loss=0.2415, pruned_loss=0.04874, over 4927.00 frames. ], tot_loss[loss=0.1713, simple_loss=0.2431, pruned_loss=0.04979, over 919044.53 frames. ], batch size: 38, lr: 3.21e-03, grad_scale: 32.0 2023-04-27 15:04:17,233 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1548, 2.8816, 2.4181, 2.6097, 1.9456, 2.5439, 2.5631, 1.9671], device='cuda:4'), covar=tensor([0.2116, 0.1105, 0.0754, 0.1206, 0.3266, 0.1033, 0.1945, 0.2690], device='cuda:4'), in_proj_covar=tensor([0.0287, 0.0301, 0.0216, 0.0277, 0.0312, 0.0255, 0.0249, 0.0263], device='cuda:4'), out_proj_covar=tensor([1.1481e-04, 1.1966e-04, 8.5668e-05, 1.0961e-04, 1.2633e-04, 1.0100e-04, 1.0062e-04, 1.0422e-04], device='cuda:4') 2023-04-27 15:04:17,775 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=115227.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:04:36,594 INFO [finetune.py:976] (4/7) Epoch 21, batch 700, loss[loss=0.165, simple_loss=0.2405, pruned_loss=0.04473, over 4879.00 frames. ], tot_loss[loss=0.1717, simple_loss=0.2439, pruned_loss=0.04974, over 924751.49 frames. ], batch size: 32, lr: 3.21e-03, grad_scale: 32.0 2023-04-27 15:04:40,862 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.165e+02 1.680e+02 1.934e+02 2.254e+02 3.960e+02, threshold=3.868e+02, percent-clipped=2.0 2023-04-27 15:04:44,485 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0415, 1.7672, 2.0250, 2.3876, 2.3979, 1.9509, 1.5278, 2.2174], device='cuda:4'), covar=tensor([0.0879, 0.1123, 0.0734, 0.0579, 0.0624, 0.0890, 0.0845, 0.0552], device='cuda:4'), in_proj_covar=tensor([0.0189, 0.0203, 0.0185, 0.0172, 0.0178, 0.0181, 0.0153, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 15:05:08,182 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1291, 2.5515, 1.0832, 1.4434, 1.9702, 1.2101, 3.3069, 1.7737], device='cuda:4'), covar=tensor([0.0652, 0.0566, 0.0699, 0.1228, 0.0469, 0.1033, 0.0341, 0.0652], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0072, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0007, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 15:05:08,848 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1586, 1.6170, 1.9797, 2.1702, 1.9789, 1.6015, 1.0868, 1.7207], device='cuda:4'), covar=tensor([0.3121, 0.3186, 0.1733, 0.2130, 0.2462, 0.2646, 0.4180, 0.1913], device='cuda:4'), in_proj_covar=tensor([0.0291, 0.0244, 0.0225, 0.0312, 0.0219, 0.0232, 0.0225, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 15:05:10,553 INFO [finetune.py:976] (4/7) Epoch 21, batch 750, loss[loss=0.176, simple_loss=0.2423, pruned_loss=0.05487, over 4155.00 frames. ], tot_loss[loss=0.1726, simple_loss=0.2449, pruned_loss=0.05015, over 931838.21 frames. ], batch size: 65, lr: 3.21e-03, grad_scale: 32.0 2023-04-27 15:05:38,650 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3478, 2.0933, 2.4028, 2.7631, 2.8260, 2.1915, 1.8680, 2.4538], device='cuda:4'), covar=tensor([0.0855, 0.1105, 0.0673, 0.0609, 0.0595, 0.0911, 0.0822, 0.0603], device='cuda:4'), in_proj_covar=tensor([0.0188, 0.0202, 0.0184, 0.0172, 0.0177, 0.0180, 0.0152, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 15:05:44,407 INFO [finetune.py:976] (4/7) Epoch 21, batch 800, loss[loss=0.2176, simple_loss=0.2747, pruned_loss=0.08028, over 4906.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2444, pruned_loss=0.04977, over 937281.32 frames. ], batch size: 37, lr: 3.21e-03, grad_scale: 32.0 2023-04-27 15:05:48,603 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.653e+01 1.488e+02 1.739e+02 2.068e+02 3.121e+02, threshold=3.478e+02, percent-clipped=0.0 2023-04-27 15:05:54,639 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=115370.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:05:55,232 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=115371.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:05:55,832 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=115372.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:06:08,212 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8667, 1.1023, 1.7166, 1.8169, 1.7838, 1.8572, 1.6900, 1.7173], device='cuda:4'), covar=tensor([0.3598, 0.5023, 0.4051, 0.4288, 0.5339, 0.6762, 0.4670, 0.4361], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0374, 0.0321, 0.0335, 0.0345, 0.0394, 0.0356, 0.0327], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 15:06:18,145 INFO [finetune.py:976] (4/7) Epoch 21, batch 850, loss[loss=0.1742, simple_loss=0.2385, pruned_loss=0.05498, over 4860.00 frames. ], tot_loss[loss=0.1696, simple_loss=0.2417, pruned_loss=0.04877, over 940288.92 frames. ], batch size: 31, lr: 3.21e-03, grad_scale: 32.0 2023-04-27 15:06:26,351 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.96 vs. limit=2.0 2023-04-27 15:06:27,930 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=115420.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:06:35,535 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=115431.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:06:36,151 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=115432.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:06:43,131 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=115442.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:07:01,736 INFO [finetune.py:976] (4/7) Epoch 21, batch 900, loss[loss=0.1358, simple_loss=0.2171, pruned_loss=0.02724, over 4921.00 frames. ], tot_loss[loss=0.1673, simple_loss=0.2389, pruned_loss=0.04789, over 944085.17 frames. ], batch size: 43, lr: 3.20e-03, grad_scale: 32.0 2023-04-27 15:07:06,007 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.637e+01 1.508e+02 1.757e+02 2.120e+02 3.037e+02, threshold=3.515e+02, percent-clipped=0.0 2023-04-27 15:07:34,517 INFO [finetune.py:976] (4/7) Epoch 21, batch 950, loss[loss=0.1709, simple_loss=0.247, pruned_loss=0.04741, over 4817.00 frames. ], tot_loss[loss=0.1685, simple_loss=0.2397, pruned_loss=0.04863, over 947821.08 frames. ], batch size: 45, lr: 3.20e-03, grad_scale: 32.0 2023-04-27 15:07:55,181 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3953, 1.7322, 1.5967, 2.1698, 1.8418, 2.1363, 1.6232, 4.3903], device='cuda:4'), covar=tensor([0.0555, 0.0818, 0.0776, 0.1106, 0.0649, 0.0534, 0.0730, 0.0109], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 15:08:04,466 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0015, 2.4564, 2.0986, 2.3231, 1.8692, 2.1032, 2.0217, 1.7585], device='cuda:4'), covar=tensor([0.1566, 0.0974, 0.0711, 0.0964, 0.2739, 0.0998, 0.1614, 0.2131], device='cuda:4'), in_proj_covar=tensor([0.0285, 0.0300, 0.0215, 0.0276, 0.0311, 0.0254, 0.0248, 0.0262], device='cuda:4'), out_proj_covar=tensor([1.1431e-04, 1.1896e-04, 8.5228e-05, 1.0916e-04, 1.2588e-04, 1.0051e-04, 1.0011e-04, 1.0401e-04], device='cuda:4') 2023-04-27 15:08:06,313 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=115527.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:08:08,687 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6144, 2.0766, 2.5667, 3.0510, 2.4236, 2.0209, 1.8382, 2.4560], device='cuda:4'), covar=tensor([0.3215, 0.3310, 0.1732, 0.2528, 0.2814, 0.2821, 0.4090, 0.2056], device='cuda:4'), in_proj_covar=tensor([0.0291, 0.0244, 0.0225, 0.0313, 0.0219, 0.0232, 0.0226, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 15:08:39,544 INFO [finetune.py:976] (4/7) Epoch 21, batch 1000, loss[loss=0.2216, simple_loss=0.2904, pruned_loss=0.07641, over 4746.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2433, pruned_loss=0.05038, over 945938.28 frames. ], batch size: 59, lr: 3.20e-03, grad_scale: 32.0 2023-04-27 15:08:49,080 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.133e+02 1.678e+02 2.025e+02 2.586e+02 4.511e+02, threshold=4.050e+02, percent-clipped=4.0 2023-04-27 15:09:08,298 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=115575.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:09:45,484 INFO [finetune.py:976] (4/7) Epoch 21, batch 1050, loss[loss=0.1746, simple_loss=0.2545, pruned_loss=0.04736, over 4759.00 frames. ], tot_loss[loss=0.173, simple_loss=0.2453, pruned_loss=0.0504, over 949364.07 frames. ], batch size: 54, lr: 3.20e-03, grad_scale: 32.0 2023-04-27 15:10:18,956 INFO [finetune.py:976] (4/7) Epoch 21, batch 1100, loss[loss=0.232, simple_loss=0.298, pruned_loss=0.08299, over 4833.00 frames. ], tot_loss[loss=0.1733, simple_loss=0.2456, pruned_loss=0.05047, over 949436.56 frames. ], batch size: 47, lr: 3.20e-03, grad_scale: 32.0 2023-04-27 15:10:19,073 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0391, 2.3393, 1.3369, 1.7626, 2.3803, 1.9322, 1.8201, 1.9385], device='cuda:4'), covar=tensor([0.0466, 0.0322, 0.0275, 0.0527, 0.0226, 0.0476, 0.0471, 0.0514], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0028, 0.0019, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0045, 0.0038, 0.0052, 0.0038, 0.0050, 0.0050, 0.0051], device='cuda:4') 2023-04-27 15:10:19,690 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9598, 1.7337, 1.9255, 2.3389, 2.3754, 1.7454, 1.5243, 2.1146], device='cuda:4'), covar=tensor([0.0787, 0.1113, 0.0742, 0.0511, 0.0511, 0.0888, 0.0775, 0.0507], device='cuda:4'), in_proj_covar=tensor([0.0186, 0.0200, 0.0182, 0.0170, 0.0176, 0.0179, 0.0151, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 15:10:23,752 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.026e+02 1.551e+02 1.797e+02 2.277e+02 4.571e+02, threshold=3.594e+02, percent-clipped=1.0 2023-04-27 15:10:33,353 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4821, 3.3552, 0.9432, 1.8100, 1.9183, 2.5447, 1.9073, 0.9824], device='cuda:4'), covar=tensor([0.1342, 0.0898, 0.1951, 0.1198, 0.1066, 0.0873, 0.1455, 0.2033], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0239, 0.0135, 0.0119, 0.0132, 0.0151, 0.0116, 0.0117], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 15:10:52,443 INFO [finetune.py:976] (4/7) Epoch 21, batch 1150, loss[loss=0.1233, simple_loss=0.2005, pruned_loss=0.02309, over 4792.00 frames. ], tot_loss[loss=0.1732, simple_loss=0.2461, pruned_loss=0.05013, over 951570.34 frames. ], batch size: 26, lr: 3.20e-03, grad_scale: 32.0 2023-04-27 15:11:07,917 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=115726.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:11:08,529 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=115727.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:11:18,222 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=115742.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:11:18,267 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1971, 2.7297, 2.1666, 2.0933, 1.6026, 1.5552, 2.3415, 1.4832], device='cuda:4'), covar=tensor([0.1591, 0.1477, 0.1392, 0.1718, 0.2221, 0.1893, 0.0939, 0.2024], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0210, 0.0168, 0.0203, 0.0199, 0.0183, 0.0155, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 15:11:22,976 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6207, 2.1123, 2.5957, 3.0739, 2.5271, 2.0165, 2.0321, 2.4951], device='cuda:4'), covar=tensor([0.2933, 0.2832, 0.1432, 0.2211, 0.2556, 0.2436, 0.3690, 0.1874], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0245, 0.0226, 0.0314, 0.0219, 0.0232, 0.0227, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 15:11:25,942 INFO [finetune.py:976] (4/7) Epoch 21, batch 1200, loss[loss=0.1776, simple_loss=0.2474, pruned_loss=0.05392, over 4811.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2442, pruned_loss=0.04985, over 953293.19 frames. ], batch size: 45, lr: 3.20e-03, grad_scale: 32.0 2023-04-27 15:11:31,125 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.085e+02 1.519e+02 1.757e+02 2.035e+02 5.645e+02, threshold=3.514e+02, percent-clipped=1.0 2023-04-27 15:11:50,225 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=115790.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:11:59,754 INFO [finetune.py:976] (4/7) Epoch 21, batch 1250, loss[loss=0.1714, simple_loss=0.2557, pruned_loss=0.04352, over 4823.00 frames. ], tot_loss[loss=0.1706, simple_loss=0.2428, pruned_loss=0.04924, over 954363.70 frames. ], batch size: 39, lr: 3.20e-03, grad_scale: 32.0 2023-04-27 15:12:55,528 INFO [finetune.py:976] (4/7) Epoch 21, batch 1300, loss[loss=0.1392, simple_loss=0.2174, pruned_loss=0.03053, over 4755.00 frames. ], tot_loss[loss=0.1687, simple_loss=0.2403, pruned_loss=0.04854, over 952985.13 frames. ], batch size: 28, lr: 3.20e-03, grad_scale: 32.0 2023-04-27 15:12:56,894 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5522, 1.6632, 1.4631, 1.1122, 1.1726, 1.1645, 1.4234, 1.1063], device='cuda:4'), covar=tensor([0.1723, 0.1367, 0.1526, 0.1875, 0.2377, 0.2051, 0.1094, 0.2082], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0209, 0.0168, 0.0203, 0.0198, 0.0183, 0.0155, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 15:12:59,792 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.963e+01 1.535e+02 1.763e+02 2.248e+02 3.829e+02, threshold=3.527e+02, percent-clipped=1.0 2023-04-27 15:13:21,245 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=115892.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 15:13:25,515 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2289, 1.3906, 1.2689, 1.8086, 1.4860, 1.5961, 1.2802, 3.1068], device='cuda:4'), covar=tensor([0.0693, 0.0997, 0.1053, 0.1231, 0.0809, 0.0569, 0.0947, 0.0253], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 15:13:29,446 INFO [finetune.py:976] (4/7) Epoch 21, batch 1350, loss[loss=0.2034, simple_loss=0.2739, pruned_loss=0.06643, over 4802.00 frames. ], tot_loss[loss=0.169, simple_loss=0.2402, pruned_loss=0.04886, over 953047.82 frames. ], batch size: 45, lr: 3.20e-03, grad_scale: 32.0 2023-04-27 15:13:56,434 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-27 15:14:07,145 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4853, 2.5538, 2.0321, 2.2946, 2.6314, 2.0854, 3.5321, 1.8852], device='cuda:4'), covar=tensor([0.4377, 0.2600, 0.4567, 0.3502, 0.2309, 0.2952, 0.1770, 0.4686], device='cuda:4'), in_proj_covar=tensor([0.0345, 0.0350, 0.0429, 0.0356, 0.0382, 0.0377, 0.0373, 0.0422], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 15:14:34,416 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=115953.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 15:14:34,904 INFO [finetune.py:976] (4/7) Epoch 21, batch 1400, loss[loss=0.2632, simple_loss=0.3251, pruned_loss=0.1007, over 4270.00 frames. ], tot_loss[loss=0.1716, simple_loss=0.2431, pruned_loss=0.05006, over 949437.64 frames. ], batch size: 65, lr: 3.20e-03, grad_scale: 32.0 2023-04-27 15:14:44,906 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.084e+02 1.645e+02 1.882e+02 2.243e+02 4.994e+02, threshold=3.764e+02, percent-clipped=1.0 2023-04-27 15:15:22,121 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8615, 2.5541, 1.9878, 1.7473, 1.3779, 1.3897, 2.0939, 1.3114], device='cuda:4'), covar=tensor([0.1729, 0.1251, 0.1371, 0.1752, 0.2256, 0.1936, 0.0955, 0.2030], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0210, 0.0168, 0.0203, 0.0199, 0.0184, 0.0155, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 15:15:26,317 INFO [finetune.py:976] (4/7) Epoch 21, batch 1450, loss[loss=0.1692, simple_loss=0.2388, pruned_loss=0.04982, over 4809.00 frames. ], tot_loss[loss=0.1736, simple_loss=0.2456, pruned_loss=0.05075, over 951226.92 frames. ], batch size: 38, lr: 3.20e-03, grad_scale: 32.0 2023-04-27 15:15:33,426 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=116014.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:15:42,236 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=116026.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:15:42,816 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=116027.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:16:00,012 INFO [finetune.py:976] (4/7) Epoch 21, batch 1500, loss[loss=0.2003, simple_loss=0.2751, pruned_loss=0.06273, over 4812.00 frames. ], tot_loss[loss=0.1737, simple_loss=0.246, pruned_loss=0.05069, over 951659.75 frames. ], batch size: 40, lr: 3.20e-03, grad_scale: 32.0 2023-04-27 15:16:05,187 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.070e+02 1.610e+02 1.919e+02 2.360e+02 3.995e+02, threshold=3.837e+02, percent-clipped=2.0 2023-04-27 15:16:08,375 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8527, 2.2556, 1.1033, 1.4966, 2.1337, 1.6800, 1.5872, 1.6837], device='cuda:4'), covar=tensor([0.0455, 0.0330, 0.0288, 0.0528, 0.0235, 0.0489, 0.0476, 0.0519], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0028, 0.0019, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0045, 0.0038, 0.0051, 0.0038, 0.0050, 0.0050, 0.0051], device='cuda:4') 2023-04-27 15:16:13,680 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=116074.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:16:14,789 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=116075.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:16:14,856 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=116075.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 15:16:33,606 INFO [finetune.py:976] (4/7) Epoch 21, batch 1550, loss[loss=0.1665, simple_loss=0.2389, pruned_loss=0.04704, over 4827.00 frames. ], tot_loss[loss=0.1735, simple_loss=0.2459, pruned_loss=0.05056, over 953284.32 frames. ], batch size: 30, lr: 3.20e-03, grad_scale: 32.0 2023-04-27 15:16:34,348 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=116105.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:16:41,991 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=116116.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:16:46,522 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.73 vs. limit=2.0 2023-04-27 15:16:59,563 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0698, 2.8006, 2.1292, 2.2607, 1.5015, 1.4918, 2.2288, 1.5049], device='cuda:4'), covar=tensor([0.1632, 0.1457, 0.1448, 0.1555, 0.2204, 0.1823, 0.0939, 0.1959], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0210, 0.0169, 0.0204, 0.0199, 0.0184, 0.0156, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 15:17:06,680 INFO [finetune.py:976] (4/7) Epoch 21, batch 1600, loss[loss=0.1876, simple_loss=0.2565, pruned_loss=0.05937, over 4058.00 frames. ], tot_loss[loss=0.1732, simple_loss=0.2451, pruned_loss=0.05069, over 953986.73 frames. ], batch size: 65, lr: 3.20e-03, grad_scale: 32.0 2023-04-27 15:17:10,688 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.50 vs. limit=2.0 2023-04-27 15:17:10,953 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.072e+02 1.579e+02 1.851e+02 2.317e+02 5.378e+02, threshold=3.702e+02, percent-clipped=3.0 2023-04-27 15:17:14,620 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=116166.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:17:21,317 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=116177.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:17:39,788 INFO [finetune.py:976] (4/7) Epoch 21, batch 1650, loss[loss=0.1915, simple_loss=0.2585, pruned_loss=0.06228, over 4744.00 frames. ], tot_loss[loss=0.1701, simple_loss=0.2415, pruned_loss=0.04938, over 955641.13 frames. ], batch size: 54, lr: 3.20e-03, grad_scale: 32.0 2023-04-27 15:18:04,934 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0214, 2.7000, 2.0706, 2.0690, 1.4667, 1.4620, 2.1439, 1.4609], device='cuda:4'), covar=tensor([0.1574, 0.1172, 0.1424, 0.1609, 0.2209, 0.1899, 0.0939, 0.2031], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0212, 0.0170, 0.0205, 0.0200, 0.0184, 0.0156, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 15:18:14,845 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=116248.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 15:18:18,422 INFO [finetune.py:976] (4/7) Epoch 21, batch 1700, loss[loss=0.1963, simple_loss=0.2633, pruned_loss=0.06462, over 4864.00 frames. ], tot_loss[loss=0.1684, simple_loss=0.2394, pruned_loss=0.04866, over 958106.81 frames. ], batch size: 31, lr: 3.20e-03, grad_scale: 32.0 2023-04-27 15:18:28,118 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.038e+02 1.499e+02 1.782e+02 2.142e+02 3.522e+02, threshold=3.563e+02, percent-clipped=0.0 2023-04-27 15:19:10,214 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.83 vs. limit=5.0 2023-04-27 15:19:31,552 INFO [finetune.py:976] (4/7) Epoch 21, batch 1750, loss[loss=0.2264, simple_loss=0.2877, pruned_loss=0.08256, over 4908.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.2406, pruned_loss=0.04922, over 957084.14 frames. ], batch size: 36, lr: 3.20e-03, grad_scale: 32.0 2023-04-27 15:19:55,827 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=116323.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:20:33,794 INFO [finetune.py:976] (4/7) Epoch 21, batch 1800, loss[loss=0.1658, simple_loss=0.2452, pruned_loss=0.04314, over 4903.00 frames. ], tot_loss[loss=0.1723, simple_loss=0.2448, pruned_loss=0.04995, over 957974.03 frames. ], batch size: 36, lr: 3.20e-03, grad_scale: 32.0 2023-04-27 15:20:38,069 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.160e+02 1.681e+02 1.987e+02 2.405e+02 5.932e+02, threshold=3.974e+02, percent-clipped=5.0 2023-04-27 15:20:43,622 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=116370.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 15:20:53,269 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=116384.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:20:56,794 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2787, 1.3679, 1.3030, 1.6930, 1.4287, 1.7674, 1.2933, 3.4554], device='cuda:4'), covar=tensor([0.0680, 0.1140, 0.1078, 0.1397, 0.0876, 0.0659, 0.1022, 0.0250], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0044, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 15:21:07,566 INFO [finetune.py:976] (4/7) Epoch 21, batch 1850, loss[loss=0.1446, simple_loss=0.2152, pruned_loss=0.037, over 4761.00 frames. ], tot_loss[loss=0.1741, simple_loss=0.2464, pruned_loss=0.05086, over 958235.60 frames. ], batch size: 23, lr: 3.20e-03, grad_scale: 32.0 2023-04-27 15:21:35,317 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=116446.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:21:40,091 INFO [finetune.py:976] (4/7) Epoch 21, batch 1900, loss[loss=0.1283, simple_loss=0.1971, pruned_loss=0.02972, over 4722.00 frames. ], tot_loss[loss=0.1747, simple_loss=0.2471, pruned_loss=0.05112, over 954942.31 frames. ], batch size: 23, lr: 3.20e-03, grad_scale: 32.0 2023-04-27 15:21:42,981 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2016, 2.7888, 2.4839, 2.4128, 1.6009, 1.6448, 2.6862, 1.6748], device='cuda:4'), covar=tensor([0.1707, 0.1572, 0.1208, 0.1430, 0.2145, 0.1846, 0.0838, 0.1895], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0210, 0.0169, 0.0203, 0.0199, 0.0184, 0.0155, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 15:21:45,218 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.938e+01 1.604e+02 1.932e+02 2.429e+02 3.655e+02, threshold=3.864e+02, percent-clipped=0.0 2023-04-27 15:21:45,307 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=116461.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:21:52,092 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=116472.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:22:13,563 INFO [finetune.py:976] (4/7) Epoch 21, batch 1950, loss[loss=0.128, simple_loss=0.1967, pruned_loss=0.0296, over 4475.00 frames. ], tot_loss[loss=0.1731, simple_loss=0.2456, pruned_loss=0.05025, over 955434.20 frames. ], batch size: 19, lr: 3.20e-03, grad_scale: 32.0 2023-04-27 15:22:14,292 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=116505.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:22:15,998 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=116507.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 15:22:23,129 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2830, 1.5839, 5.4584, 5.1219, 4.7126, 5.1561, 4.7564, 4.9834], device='cuda:4'), covar=tensor([0.6590, 0.6109, 0.1039, 0.1739, 0.1001, 0.1844, 0.1112, 0.1390], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0306, 0.0402, 0.0403, 0.0347, 0.0406, 0.0310, 0.0366], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 15:22:43,309 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=116548.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 15:22:46,839 INFO [finetune.py:976] (4/7) Epoch 21, batch 2000, loss[loss=0.1659, simple_loss=0.2379, pruned_loss=0.04693, over 4834.00 frames. ], tot_loss[loss=0.1717, simple_loss=0.2438, pruned_loss=0.04978, over 956092.10 frames. ], batch size: 38, lr: 3.20e-03, grad_scale: 32.0 2023-04-27 15:22:48,778 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6662, 1.5628, 2.0338, 1.9954, 1.5371, 1.4692, 1.7128, 0.9339], device='cuda:4'), covar=tensor([0.0466, 0.0533, 0.0329, 0.0550, 0.0686, 0.0966, 0.0497, 0.0599], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0067, 0.0066, 0.0067, 0.0074, 0.0095, 0.0072, 0.0065], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 15:22:51,556 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.079e+02 1.496e+02 1.815e+02 2.157e+02 3.594e+02, threshold=3.630e+02, percent-clipped=0.0 2023-04-27 15:22:55,205 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=116566.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:23:06,567 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4335, 2.2360, 1.8378, 1.9498, 2.2735, 1.8262, 2.4954, 1.6171], device='cuda:4'), covar=tensor([0.3404, 0.1647, 0.3982, 0.2829, 0.1577, 0.2286, 0.1644, 0.4451], device='cuda:4'), in_proj_covar=tensor([0.0346, 0.0350, 0.0430, 0.0355, 0.0383, 0.0378, 0.0374, 0.0422], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 15:23:14,413 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=116596.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 15:23:20,569 INFO [finetune.py:976] (4/7) Epoch 21, batch 2050, loss[loss=0.1629, simple_loss=0.2262, pruned_loss=0.04974, over 4766.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.2411, pruned_loss=0.04895, over 956827.43 frames. ], batch size: 27, lr: 3.20e-03, grad_scale: 32.0 2023-04-27 15:23:30,127 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=116618.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:23:59,066 INFO [finetune.py:976] (4/7) Epoch 21, batch 2100, loss[loss=0.2309, simple_loss=0.2966, pruned_loss=0.08264, over 4154.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2407, pruned_loss=0.04905, over 955150.61 frames. ], batch size: 65, lr: 3.20e-03, grad_scale: 32.0 2023-04-27 15:24:03,921 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.802e+01 1.616e+02 1.834e+02 2.363e+02 4.673e+02, threshold=3.668e+02, percent-clipped=1.0 2023-04-27 15:24:21,184 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=116670.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 15:24:32,898 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=116679.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:24:32,961 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=116679.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:24:43,194 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9154, 1.1533, 3.2746, 3.0130, 2.9165, 3.1730, 3.1684, 2.8477], device='cuda:4'), covar=tensor([0.7635, 0.5795, 0.1526, 0.2429, 0.1563, 0.2160, 0.1584, 0.1838], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0308, 0.0405, 0.0405, 0.0348, 0.0408, 0.0311, 0.0367], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 15:25:00,268 INFO [finetune.py:976] (4/7) Epoch 21, batch 2150, loss[loss=0.1715, simple_loss=0.2491, pruned_loss=0.04697, over 4728.00 frames. ], tot_loss[loss=0.1716, simple_loss=0.2437, pruned_loss=0.04978, over 955677.46 frames. ], batch size: 59, lr: 3.20e-03, grad_scale: 32.0 2023-04-27 15:25:05,365 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=116704.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:25:21,300 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=116718.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:25:28,803 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5780, 1.1015, 1.3230, 1.1772, 1.6341, 1.3099, 1.0606, 1.2852], device='cuda:4'), covar=tensor([0.1739, 0.1575, 0.2023, 0.1540, 0.1024, 0.1668, 0.2105, 0.2424], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0314, 0.0355, 0.0292, 0.0330, 0.0313, 0.0306, 0.0376], device='cuda:4'), out_proj_covar=tensor([6.4174e-05, 6.4921e-05, 7.4975e-05, 5.9015e-05, 6.8233e-05, 6.5666e-05, 6.4099e-05, 7.9938e-05], device='cuda:4') 2023-04-27 15:26:05,621 INFO [finetune.py:976] (4/7) Epoch 21, batch 2200, loss[loss=0.1742, simple_loss=0.2507, pruned_loss=0.04884, over 4830.00 frames. ], tot_loss[loss=0.1727, simple_loss=0.2455, pruned_loss=0.05, over 957100.03 frames. ], batch size: 39, lr: 3.19e-03, grad_scale: 32.0 2023-04-27 15:26:10,852 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.212e+01 1.600e+02 1.964e+02 2.414e+02 3.602e+02, threshold=3.928e+02, percent-clipped=0.0 2023-04-27 15:26:10,950 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=116761.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:26:13,914 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=116765.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:26:18,741 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=116772.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:26:32,506 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8232, 2.4059, 1.8533, 1.7665, 1.3074, 1.3405, 1.9776, 1.2738], device='cuda:4'), covar=tensor([0.1729, 0.1294, 0.1507, 0.1651, 0.2350, 0.1938, 0.0952, 0.2053], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0211, 0.0169, 0.0203, 0.0200, 0.0185, 0.0156, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 15:26:38,521 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=116802.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 15:26:38,567 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=116802.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:26:39,696 INFO [finetune.py:976] (4/7) Epoch 21, batch 2250, loss[loss=0.1648, simple_loss=0.2447, pruned_loss=0.04242, over 4836.00 frames. ], tot_loss[loss=0.1722, simple_loss=0.2453, pruned_loss=0.04959, over 958021.16 frames. ], batch size: 30, lr: 3.19e-03, grad_scale: 64.0 2023-04-27 15:26:42,838 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=116809.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:26:49,388 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1471, 1.3611, 1.3236, 1.6300, 1.5427, 1.7031, 1.2443, 2.9740], device='cuda:4'), covar=tensor([0.0628, 0.0882, 0.0785, 0.1233, 0.0650, 0.0526, 0.0783, 0.0198], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 15:26:50,612 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=116820.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:27:06,035 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.83 vs. limit=5.0 2023-04-27 15:27:13,183 INFO [finetune.py:976] (4/7) Epoch 21, batch 2300, loss[loss=0.1646, simple_loss=0.2458, pruned_loss=0.04177, over 4819.00 frames. ], tot_loss[loss=0.1723, simple_loss=0.2457, pruned_loss=0.0495, over 958638.72 frames. ], batch size: 40, lr: 3.19e-03, grad_scale: 64.0 2023-04-27 15:27:17,446 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.071e+02 1.557e+02 1.911e+02 2.274e+02 3.749e+02, threshold=3.822e+02, percent-clipped=0.0 2023-04-27 15:27:17,552 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=116861.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:27:18,831 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=116863.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:27:38,932 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.7105, 2.1683, 2.6509, 3.2845, 2.5432, 2.0572, 2.1153, 2.4820], device='cuda:4'), covar=tensor([0.3490, 0.3417, 0.1696, 0.2387, 0.2607, 0.2633, 0.3818, 0.2234], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0245, 0.0226, 0.0313, 0.0218, 0.0231, 0.0228, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 15:27:46,962 INFO [finetune.py:976] (4/7) Epoch 21, batch 2350, loss[loss=0.2259, simple_loss=0.2799, pruned_loss=0.08597, over 4192.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2442, pruned_loss=0.04988, over 956452.06 frames. ], batch size: 65, lr: 3.19e-03, grad_scale: 64.0 2023-04-27 15:28:01,859 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.50 vs. limit=2.0 2023-04-27 15:28:12,162 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2882, 2.7806, 2.5221, 2.5177, 1.7003, 1.7462, 2.7760, 1.7498], device='cuda:4'), covar=tensor([0.1620, 0.1633, 0.1257, 0.1489, 0.2300, 0.1773, 0.0826, 0.1874], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0211, 0.0170, 0.0204, 0.0201, 0.0185, 0.0157, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 15:28:18,096 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1988, 2.6014, 0.8917, 1.3477, 2.0678, 1.2532, 3.6166, 1.8949], device='cuda:4'), covar=tensor([0.0689, 0.0738, 0.0921, 0.1309, 0.0562, 0.1051, 0.0249, 0.0643], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0065, 0.0047, 0.0046, 0.0049, 0.0052, 0.0073, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 15:28:20,894 INFO [finetune.py:976] (4/7) Epoch 21, batch 2400, loss[loss=0.1956, simple_loss=0.2633, pruned_loss=0.06393, over 4875.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2425, pruned_loss=0.0499, over 957118.70 frames. ], batch size: 31, lr: 3.19e-03, grad_scale: 64.0 2023-04-27 15:28:25,119 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.043e+02 1.594e+02 1.806e+02 2.173e+02 4.519e+02, threshold=3.612e+02, percent-clipped=1.0 2023-04-27 15:28:34,737 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=116974.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:28:37,805 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=116979.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:28:54,617 INFO [finetune.py:976] (4/7) Epoch 21, batch 2450, loss[loss=0.1311, simple_loss=0.2109, pruned_loss=0.02567, over 4906.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2395, pruned_loss=0.04913, over 955314.26 frames. ], batch size: 37, lr: 3.19e-03, grad_scale: 64.0 2023-04-27 15:29:03,242 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8188, 2.4544, 1.9051, 1.8304, 1.3237, 1.3581, 2.0244, 1.2726], device='cuda:4'), covar=tensor([0.1769, 0.1399, 0.1471, 0.1773, 0.2424, 0.2040, 0.0972, 0.2194], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0212, 0.0171, 0.0205, 0.0202, 0.0186, 0.0158, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 15:29:03,998 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-27 15:29:10,690 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=117027.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:29:13,243 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-27 15:29:28,082 INFO [finetune.py:976] (4/7) Epoch 21, batch 2500, loss[loss=0.2103, simple_loss=0.2944, pruned_loss=0.06311, over 4889.00 frames. ], tot_loss[loss=0.1709, simple_loss=0.2415, pruned_loss=0.05018, over 955049.58 frames. ], batch size: 32, lr: 3.19e-03, grad_scale: 64.0 2023-04-27 15:29:32,763 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=117060.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:29:33,282 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.097e+02 1.592e+02 1.814e+02 2.143e+02 3.626e+02, threshold=3.628e+02, percent-clipped=1.0 2023-04-27 15:30:29,012 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=117102.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:30:30,122 INFO [finetune.py:976] (4/7) Epoch 21, batch 2550, loss[loss=0.2031, simple_loss=0.2817, pruned_loss=0.06225, over 4811.00 frames. ], tot_loss[loss=0.1733, simple_loss=0.2446, pruned_loss=0.05104, over 954229.75 frames. ], batch size: 38, lr: 3.19e-03, grad_scale: 64.0 2023-04-27 15:31:33,713 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=117150.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:31:41,893 INFO [finetune.py:976] (4/7) Epoch 21, batch 2600, loss[loss=0.1559, simple_loss=0.2296, pruned_loss=0.04117, over 4819.00 frames. ], tot_loss[loss=0.174, simple_loss=0.2455, pruned_loss=0.05122, over 951591.44 frames. ], batch size: 39, lr: 3.19e-03, grad_scale: 64.0 2023-04-27 15:31:42,644 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0331, 1.5456, 1.6495, 1.7698, 2.2041, 1.7233, 1.4008, 1.4467], device='cuda:4'), covar=tensor([0.1622, 0.1461, 0.1904, 0.1412, 0.0914, 0.1693, 0.2149, 0.2522], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0312, 0.0351, 0.0289, 0.0326, 0.0310, 0.0303, 0.0372], device='cuda:4'), out_proj_covar=tensor([6.3925e-05, 6.4623e-05, 7.4032e-05, 5.8371e-05, 6.7344e-05, 6.4965e-05, 6.3604e-05, 7.9047e-05], device='cuda:4') 2023-04-27 15:31:44,415 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=117158.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:31:52,560 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.120e+02 1.651e+02 1.953e+02 2.341e+02 4.282e+02, threshold=3.906e+02, percent-clipped=1.0 2023-04-27 15:31:52,677 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=117161.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:32:20,898 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-27 15:32:27,548 INFO [finetune.py:976] (4/7) Epoch 21, batch 2650, loss[loss=0.1647, simple_loss=0.2408, pruned_loss=0.0443, over 4747.00 frames. ], tot_loss[loss=0.1743, simple_loss=0.2461, pruned_loss=0.05128, over 950350.85 frames. ], batch size: 54, lr: 3.19e-03, grad_scale: 64.0 2023-04-27 15:32:30,681 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=117209.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:32:39,172 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=117221.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:32:55,689 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-27 15:33:01,274 INFO [finetune.py:976] (4/7) Epoch 21, batch 2700, loss[loss=0.1504, simple_loss=0.2333, pruned_loss=0.03377, over 4785.00 frames. ], tot_loss[loss=0.1733, simple_loss=0.2456, pruned_loss=0.05047, over 952417.73 frames. ], batch size: 29, lr: 3.19e-03, grad_scale: 64.0 2023-04-27 15:33:06,004 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.896e+01 1.368e+02 1.680e+02 2.087e+02 5.894e+02, threshold=3.359e+02, percent-clipped=1.0 2023-04-27 15:33:09,145 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8699, 3.9690, 0.7852, 2.0967, 2.2627, 2.7825, 2.2832, 0.8970], device='cuda:4'), covar=tensor([0.1222, 0.0920, 0.2066, 0.1101, 0.0959, 0.0921, 0.1362, 0.2126], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0239, 0.0136, 0.0119, 0.0132, 0.0152, 0.0116, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 15:33:15,077 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=117274.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:33:18,754 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.5142, 1.3867, 1.4289, 1.1321, 1.3777, 1.2046, 1.7225, 1.3732], device='cuda:4'), covar=tensor([0.3567, 0.1822, 0.4890, 0.2267, 0.1569, 0.2119, 0.1469, 0.4651], device='cuda:4'), in_proj_covar=tensor([0.0342, 0.0349, 0.0424, 0.0353, 0.0382, 0.0375, 0.0370, 0.0419], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 15:33:19,970 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=117282.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:33:35,176 INFO [finetune.py:976] (4/7) Epoch 21, batch 2750, loss[loss=0.1691, simple_loss=0.2407, pruned_loss=0.04872, over 4786.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2437, pruned_loss=0.05018, over 950858.29 frames. ], batch size: 29, lr: 3.19e-03, grad_scale: 64.0 2023-04-27 15:33:47,711 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=117322.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:33:52,064 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=117329.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:33:52,108 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5751, 1.7933, 1.9433, 2.0239, 1.8713, 2.0524, 2.1158, 2.0564], device='cuda:4'), covar=tensor([0.3652, 0.5026, 0.4352, 0.4434, 0.5364, 0.6628, 0.4550, 0.4773], device='cuda:4'), in_proj_covar=tensor([0.0335, 0.0371, 0.0321, 0.0334, 0.0344, 0.0392, 0.0354, 0.0327], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 15:34:08,696 INFO [finetune.py:976] (4/7) Epoch 21, batch 2800, loss[loss=0.1186, simple_loss=0.1928, pruned_loss=0.02223, over 4880.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.24, pruned_loss=0.04878, over 949477.50 frames. ], batch size: 34, lr: 3.19e-03, grad_scale: 64.0 2023-04-27 15:34:12,941 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=117360.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:34:13,411 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.017e+02 1.550e+02 1.821e+02 2.376e+02 5.325e+02, threshold=3.642e+02, percent-clipped=3.0 2023-04-27 15:34:31,544 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=117390.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:34:38,545 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1406, 0.7707, 0.9780, 0.8815, 1.2526, 1.0253, 0.8892, 0.9815], device='cuda:4'), covar=tensor([0.2148, 0.1817, 0.2431, 0.1823, 0.1293, 0.1708, 0.1996, 0.2643], device='cuda:4'), in_proj_covar=tensor([0.0314, 0.0315, 0.0355, 0.0294, 0.0330, 0.0313, 0.0307, 0.0375], device='cuda:4'), out_proj_covar=tensor([6.4621e-05, 6.5288e-05, 7.4869e-05, 5.9269e-05, 6.8174e-05, 6.5601e-05, 6.4364e-05, 7.9631e-05], device='cuda:4') 2023-04-27 15:34:41,632 INFO [finetune.py:976] (4/7) Epoch 21, batch 2850, loss[loss=0.176, simple_loss=0.2496, pruned_loss=0.05115, over 4917.00 frames. ], tot_loss[loss=0.1682, simple_loss=0.2388, pruned_loss=0.04878, over 949633.69 frames. ], batch size: 37, lr: 3.19e-03, grad_scale: 64.0 2023-04-27 15:34:44,560 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=117408.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:35:14,767 INFO [finetune.py:976] (4/7) Epoch 21, batch 2900, loss[loss=0.1967, simple_loss=0.2716, pruned_loss=0.06086, over 4814.00 frames. ], tot_loss[loss=0.1713, simple_loss=0.2424, pruned_loss=0.05011, over 950707.68 frames. ], batch size: 40, lr: 3.19e-03, grad_scale: 64.0 2023-04-27 15:35:17,825 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=117458.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:35:19,565 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.085e+02 1.585e+02 1.858e+02 2.394e+02 5.975e+02, threshold=3.717e+02, percent-clipped=5.0 2023-04-27 15:36:16,933 INFO [finetune.py:976] (4/7) Epoch 21, batch 2950, loss[loss=0.1735, simple_loss=0.2484, pruned_loss=0.04932, over 4929.00 frames. ], tot_loss[loss=0.1727, simple_loss=0.2451, pruned_loss=0.05012, over 950876.86 frames. ], batch size: 33, lr: 3.19e-03, grad_scale: 64.0 2023-04-27 15:36:18,222 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=117506.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:36:47,986 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9046, 2.3222, 1.9096, 2.2624, 1.7250, 1.9576, 1.9221, 1.5636], device='cuda:4'), covar=tensor([0.1838, 0.1198, 0.0934, 0.1136, 0.3401, 0.1263, 0.1804, 0.2655], device='cuda:4'), in_proj_covar=tensor([0.0286, 0.0299, 0.0215, 0.0276, 0.0310, 0.0252, 0.0246, 0.0262], device='cuda:4'), out_proj_covar=tensor([1.1436e-04, 1.1839e-04, 8.5339e-05, 1.0938e-04, 1.2552e-04, 1.0005e-04, 9.9386e-05, 1.0391e-04], device='cuda:4') 2023-04-27 15:36:57,909 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.8555, 3.7564, 2.6914, 4.3784, 3.8100, 3.8303, 1.5613, 3.7714], device='cuda:4'), covar=tensor([0.1579, 0.1153, 0.3000, 0.1421, 0.2860, 0.1522, 0.5526, 0.2210], device='cuda:4'), in_proj_covar=tensor([0.0242, 0.0213, 0.0249, 0.0302, 0.0293, 0.0243, 0.0271, 0.0269], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 15:36:59,219 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5557, 1.8383, 2.0095, 2.0528, 1.9477, 2.0250, 2.0435, 2.0260], device='cuda:4'), covar=tensor([0.4132, 0.5439, 0.4439, 0.4425, 0.5297, 0.7173, 0.5057, 0.4725], device='cuda:4'), in_proj_covar=tensor([0.0336, 0.0373, 0.0322, 0.0335, 0.0345, 0.0393, 0.0355, 0.0327], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 15:37:19,504 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.7051, 3.7009, 2.6511, 4.2996, 3.7884, 3.7282, 1.5218, 3.7062], device='cuda:4'), covar=tensor([0.1599, 0.1137, 0.2955, 0.1593, 0.2643, 0.1593, 0.5726, 0.2260], device='cuda:4'), in_proj_covar=tensor([0.0242, 0.0213, 0.0248, 0.0302, 0.0293, 0.0243, 0.0271, 0.0269], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 15:37:23,421 INFO [finetune.py:976] (4/7) Epoch 21, batch 3000, loss[loss=0.1635, simple_loss=0.2458, pruned_loss=0.04058, over 4884.00 frames. ], tot_loss[loss=0.1744, simple_loss=0.2469, pruned_loss=0.05095, over 952684.62 frames. ], batch size: 32, lr: 3.19e-03, grad_scale: 64.0 2023-04-27 15:37:23,421 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 15:37:30,536 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8198, 2.2579, 1.8481, 1.5812, 1.3945, 1.3939, 1.8820, 1.3585], device='cuda:4'), covar=tensor([0.1575, 0.1165, 0.1376, 0.1607, 0.2274, 0.1936, 0.0983, 0.2025], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0210, 0.0168, 0.0203, 0.0199, 0.0184, 0.0156, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 15:37:43,671 INFO [finetune.py:1010] (4/7) Epoch 21, validation: loss=0.1531, simple_loss=0.2228, pruned_loss=0.04164, over 2265189.00 frames. 2023-04-27 15:37:43,671 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-27 15:37:54,340 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.170e+02 1.590e+02 1.926e+02 2.493e+02 6.945e+02, threshold=3.852e+02, percent-clipped=2.0 2023-04-27 15:37:54,466 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1237, 1.4797, 1.4378, 1.6245, 1.6160, 1.7819, 1.3185, 3.2790], device='cuda:4'), covar=tensor([0.0662, 0.0818, 0.0770, 0.1199, 0.0627, 0.0519, 0.0780, 0.0154], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 15:38:15,506 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=117577.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:38:49,964 INFO [finetune.py:976] (4/7) Epoch 21, batch 3050, loss[loss=0.1604, simple_loss=0.2408, pruned_loss=0.04001, over 4793.00 frames. ], tot_loss[loss=0.1742, simple_loss=0.2469, pruned_loss=0.05072, over 952868.18 frames. ], batch size: 29, lr: 3.19e-03, grad_scale: 64.0 2023-04-27 15:39:08,814 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.8804, 1.9907, 1.7303, 1.5780, 1.9279, 1.5541, 2.5169, 1.3760], device='cuda:4'), covar=tensor([0.3472, 0.1696, 0.4380, 0.2850, 0.1771, 0.2321, 0.1220, 0.4490], device='cuda:4'), in_proj_covar=tensor([0.0342, 0.0348, 0.0426, 0.0354, 0.0382, 0.0376, 0.0369, 0.0420], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 15:39:26,059 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3925, 1.8251, 1.6674, 2.2583, 2.4817, 1.9940, 1.9594, 1.8009], device='cuda:4'), covar=tensor([0.2271, 0.2006, 0.2072, 0.1964, 0.1315, 0.2161, 0.2327, 0.2627], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0312, 0.0350, 0.0289, 0.0326, 0.0309, 0.0303, 0.0370], device='cuda:4'), out_proj_covar=tensor([6.3720e-05, 6.4606e-05, 7.3804e-05, 5.8397e-05, 6.7376e-05, 6.4809e-05, 6.3467e-05, 7.8621e-05], device='cuda:4') 2023-04-27 15:39:28,405 INFO [finetune.py:976] (4/7) Epoch 21, batch 3100, loss[loss=0.1708, simple_loss=0.2416, pruned_loss=0.05003, over 4149.00 frames. ], tot_loss[loss=0.1722, simple_loss=0.2453, pruned_loss=0.0496, over 954099.70 frames. ], batch size: 65, lr: 3.19e-03, grad_scale: 64.0 2023-04-27 15:39:29,166 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5437, 1.8486, 1.9884, 2.0650, 1.9825, 2.0147, 2.0439, 2.0445], device='cuda:4'), covar=tensor([0.3586, 0.5200, 0.4423, 0.4177, 0.5066, 0.6774, 0.4992, 0.4268], device='cuda:4'), in_proj_covar=tensor([0.0338, 0.0374, 0.0323, 0.0337, 0.0347, 0.0394, 0.0356, 0.0329], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 15:39:33,614 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.782e+01 1.629e+02 1.831e+02 2.140e+02 4.594e+02, threshold=3.661e+02, percent-clipped=1.0 2023-04-27 15:39:40,388 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1041, 2.5919, 1.0417, 1.2426, 1.9306, 1.2692, 3.3042, 1.5383], device='cuda:4'), covar=tensor([0.0718, 0.0662, 0.0780, 0.1393, 0.0540, 0.1040, 0.0304, 0.0740], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0073, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0007, 0.0008, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-27 15:39:49,485 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=117685.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:40:01,891 INFO [finetune.py:976] (4/7) Epoch 21, batch 3150, loss[loss=0.1358, simple_loss=0.2033, pruned_loss=0.03415, over 3918.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2429, pruned_loss=0.04963, over 954938.82 frames. ], batch size: 17, lr: 3.19e-03, grad_scale: 64.0 2023-04-27 15:40:19,242 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5603, 2.0028, 2.4840, 3.0862, 2.4400, 1.9078, 1.9540, 2.4049], device='cuda:4'), covar=tensor([0.2784, 0.2834, 0.1356, 0.2183, 0.2493, 0.2491, 0.3387, 0.1768], device='cuda:4'), in_proj_covar=tensor([0.0291, 0.0243, 0.0226, 0.0313, 0.0218, 0.0231, 0.0227, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 15:40:34,883 INFO [finetune.py:976] (4/7) Epoch 21, batch 3200, loss[loss=0.143, simple_loss=0.2215, pruned_loss=0.03226, over 4827.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2401, pruned_loss=0.04885, over 956651.20 frames. ], batch size: 38, lr: 3.19e-03, grad_scale: 32.0 2023-04-27 15:40:40,627 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.053e+02 1.488e+02 1.740e+02 2.134e+02 4.816e+02, threshold=3.479e+02, percent-clipped=1.0 2023-04-27 15:41:31,203 INFO [finetune.py:976] (4/7) Epoch 21, batch 3250, loss[loss=0.1854, simple_loss=0.2578, pruned_loss=0.05649, over 4833.00 frames. ], tot_loss[loss=0.1701, simple_loss=0.2413, pruned_loss=0.04948, over 955100.68 frames. ], batch size: 33, lr: 3.19e-03, grad_scale: 32.0 2023-04-27 15:41:31,969 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8700, 1.1592, 1.6087, 1.7134, 1.6925, 1.7710, 1.6100, 1.6159], device='cuda:4'), covar=tensor([0.3974, 0.5147, 0.4303, 0.4665, 0.5352, 0.6983, 0.5044, 0.4527], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0373, 0.0322, 0.0336, 0.0346, 0.0393, 0.0356, 0.0328], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 15:42:33,314 INFO [finetune.py:976] (4/7) Epoch 21, batch 3300, loss[loss=0.2134, simple_loss=0.271, pruned_loss=0.07795, over 4827.00 frames. ], tot_loss[loss=0.1736, simple_loss=0.2448, pruned_loss=0.05119, over 955858.45 frames. ], batch size: 33, lr: 3.19e-03, grad_scale: 32.0 2023-04-27 15:42:45,022 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.168e+02 1.642e+02 1.977e+02 2.287e+02 4.163e+02, threshold=3.954e+02, percent-clipped=4.0 2023-04-27 15:43:01,372 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=117877.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:43:07,274 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1036, 1.3864, 1.3338, 1.6089, 1.5218, 1.6857, 1.3031, 3.0214], device='cuda:4'), covar=tensor([0.0678, 0.0858, 0.0822, 0.1254, 0.0674, 0.0522, 0.0796, 0.0175], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 15:43:15,732 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4064, 1.7733, 1.7927, 1.9209, 1.8166, 1.8962, 1.8355, 1.8765], device='cuda:4'), covar=tensor([0.3885, 0.5526, 0.4456, 0.4348, 0.5259, 0.6741, 0.5693, 0.4833], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0373, 0.0322, 0.0336, 0.0346, 0.0393, 0.0356, 0.0328], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 15:43:18,018 INFO [finetune.py:976] (4/7) Epoch 21, batch 3350, loss[loss=0.1375, simple_loss=0.218, pruned_loss=0.02847, over 4738.00 frames. ], tot_loss[loss=0.1753, simple_loss=0.2471, pruned_loss=0.05175, over 957301.55 frames. ], batch size: 27, lr: 3.19e-03, grad_scale: 32.0 2023-04-27 15:43:37,873 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=117925.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:43:52,728 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-04-27 15:43:53,839 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9509, 2.4298, 1.0422, 1.3222, 1.8435, 1.2628, 2.9295, 1.6008], device='cuda:4'), covar=tensor([0.0671, 0.0590, 0.0707, 0.1210, 0.0459, 0.0965, 0.0225, 0.0635], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0072, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0009, 0.0007, 0.0007, 0.0008, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-27 15:44:02,873 INFO [finetune.py:976] (4/7) Epoch 21, batch 3400, loss[loss=0.1858, simple_loss=0.272, pruned_loss=0.04985, over 4912.00 frames. ], tot_loss[loss=0.1757, simple_loss=0.248, pruned_loss=0.05167, over 956571.76 frames. ], batch size: 46, lr: 3.19e-03, grad_scale: 32.0 2023-04-27 15:44:13,574 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.012e+02 1.581e+02 1.978e+02 2.353e+02 4.308e+02, threshold=3.955e+02, percent-clipped=3.0 2023-04-27 15:44:44,673 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=117985.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:44:57,996 INFO [finetune.py:976] (4/7) Epoch 21, batch 3450, loss[loss=0.2182, simple_loss=0.2587, pruned_loss=0.08883, over 4768.00 frames. ], tot_loss[loss=0.1749, simple_loss=0.2473, pruned_loss=0.05127, over 954951.19 frames. ], batch size: 26, lr: 3.18e-03, grad_scale: 32.0 2023-04-27 15:44:59,424 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.82 vs. limit=5.0 2023-04-27 15:45:00,135 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.63 vs. limit=5.0 2023-04-27 15:45:11,150 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7344, 1.5443, 1.3785, 1.6632, 1.9697, 1.6409, 1.4227, 1.3357], device='cuda:4'), covar=tensor([0.1532, 0.1156, 0.1703, 0.1435, 0.0965, 0.1403, 0.1912, 0.1803], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0310, 0.0350, 0.0287, 0.0325, 0.0308, 0.0301, 0.0371], device='cuda:4'), out_proj_covar=tensor([6.3446e-05, 6.4224e-05, 7.3965e-05, 5.8027e-05, 6.7060e-05, 6.4540e-05, 6.3154e-05, 7.8698e-05], device='cuda:4') 2023-04-27 15:45:18,121 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=118033.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:45:18,153 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6427, 3.9813, 0.6710, 2.0464, 2.1155, 2.4625, 2.2819, 0.8785], device='cuda:4'), covar=tensor([0.1355, 0.0887, 0.2080, 0.1206, 0.1117, 0.1199, 0.1362, 0.2199], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0238, 0.0135, 0.0118, 0.0132, 0.0151, 0.0115, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 15:45:22,465 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=118040.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:45:26,011 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5603, 1.3578, 1.2736, 1.3849, 1.7958, 1.4224, 1.2045, 1.2027], device='cuda:4'), covar=tensor([0.1482, 0.1141, 0.1811, 0.1268, 0.0632, 0.1293, 0.1746, 0.1930], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0310, 0.0350, 0.0287, 0.0325, 0.0308, 0.0301, 0.0370], device='cuda:4'), out_proj_covar=tensor([6.3421e-05, 6.4188e-05, 7.3871e-05, 5.7941e-05, 6.7011e-05, 6.4446e-05, 6.3127e-05, 7.8610e-05], device='cuda:4') 2023-04-27 15:45:31,129 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.70 vs. limit=2.0 2023-04-27 15:45:31,326 INFO [finetune.py:976] (4/7) Epoch 21, batch 3500, loss[loss=0.205, simple_loss=0.2677, pruned_loss=0.0711, over 4877.00 frames. ], tot_loss[loss=0.1715, simple_loss=0.2435, pruned_loss=0.04976, over 954620.83 frames. ], batch size: 34, lr: 3.18e-03, grad_scale: 32.0 2023-04-27 15:45:36,183 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.106e+01 1.494e+02 1.798e+02 2.113e+02 5.768e+02, threshold=3.596e+02, percent-clipped=1.0 2023-04-27 15:46:03,422 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=118101.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:46:05,142 INFO [finetune.py:976] (4/7) Epoch 21, batch 3550, loss[loss=0.1476, simple_loss=0.219, pruned_loss=0.03812, over 4911.00 frames. ], tot_loss[loss=0.1705, simple_loss=0.242, pruned_loss=0.04951, over 955023.82 frames. ], batch size: 36, lr: 3.18e-03, grad_scale: 32.0 2023-04-27 15:46:39,047 INFO [finetune.py:976] (4/7) Epoch 21, batch 3600, loss[loss=0.2303, simple_loss=0.2862, pruned_loss=0.08717, over 4730.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2397, pruned_loss=0.04902, over 955369.15 frames. ], batch size: 59, lr: 3.18e-03, grad_scale: 32.0 2023-04-27 15:46:43,924 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.120e+02 1.483e+02 1.865e+02 2.338e+02 3.870e+02, threshold=3.730e+02, percent-clipped=2.0 2023-04-27 15:46:47,631 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=118168.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:47:29,758 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9550, 4.1367, 0.6698, 2.3720, 2.3592, 2.8230, 2.4808, 1.0666], device='cuda:4'), covar=tensor([0.1125, 0.0914, 0.2053, 0.1026, 0.0908, 0.0971, 0.1300, 0.1963], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0237, 0.0135, 0.0118, 0.0131, 0.0151, 0.0115, 0.0117], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 15:47:39,452 INFO [finetune.py:976] (4/7) Epoch 21, batch 3650, loss[loss=0.2001, simple_loss=0.2812, pruned_loss=0.05945, over 4901.00 frames. ], tot_loss[loss=0.1713, simple_loss=0.2422, pruned_loss=0.05023, over 955521.62 frames. ], batch size: 35, lr: 3.18e-03, grad_scale: 32.0 2023-04-27 15:47:59,083 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6357, 1.9916, 1.7539, 1.9311, 1.4328, 1.7576, 1.8398, 1.3186], device='cuda:4'), covar=tensor([0.1601, 0.1072, 0.0759, 0.1104, 0.3069, 0.0974, 0.1404, 0.2249], device='cuda:4'), in_proj_covar=tensor([0.0282, 0.0297, 0.0213, 0.0276, 0.0308, 0.0250, 0.0243, 0.0261], device='cuda:4'), out_proj_covar=tensor([1.1293e-04, 1.1766e-04, 8.4440e-05, 1.0924e-04, 1.2488e-04, 9.9156e-05, 9.8190e-05, 1.0346e-04], device='cuda:4') 2023-04-27 15:48:09,948 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=118229.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 15:48:43,768 INFO [finetune.py:976] (4/7) Epoch 21, batch 3700, loss[loss=0.1706, simple_loss=0.2476, pruned_loss=0.04675, over 4908.00 frames. ], tot_loss[loss=0.1728, simple_loss=0.245, pruned_loss=0.05028, over 956574.78 frames. ], batch size: 43, lr: 3.18e-03, grad_scale: 32.0 2023-04-27 15:48:47,617 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6288, 1.4910, 1.9598, 1.9982, 1.4812, 1.3281, 1.5791, 0.9811], device='cuda:4'), covar=tensor([0.0521, 0.0615, 0.0402, 0.0503, 0.0688, 0.1108, 0.0582, 0.0619], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0068, 0.0067, 0.0068, 0.0075, 0.0096, 0.0073, 0.0065], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 15:48:54,006 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.129e+02 1.556e+02 1.839e+02 2.205e+02 3.757e+02, threshold=3.678e+02, percent-clipped=1.0 2023-04-27 15:48:57,440 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.10 vs. limit=2.0 2023-04-27 15:48:58,380 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=118269.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:49:03,964 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2127, 1.7776, 1.4924, 2.1262, 2.2651, 1.8665, 1.8427, 1.6417], device='cuda:4'), covar=tensor([0.1567, 0.1672, 0.2209, 0.1513, 0.1399, 0.2008, 0.2445, 0.2458], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0312, 0.0351, 0.0289, 0.0326, 0.0309, 0.0302, 0.0371], device='cuda:4'), out_proj_covar=tensor([6.3791e-05, 6.4471e-05, 7.4048e-05, 5.8225e-05, 6.7334e-05, 6.4708e-05, 6.3356e-05, 7.8702e-05], device='cuda:4') 2023-04-27 15:49:16,483 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=118289.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:49:27,410 INFO [finetune.py:976] (4/7) Epoch 21, batch 3750, loss[loss=0.1864, simple_loss=0.2594, pruned_loss=0.05669, over 4815.00 frames. ], tot_loss[loss=0.1738, simple_loss=0.246, pruned_loss=0.05076, over 955218.60 frames. ], batch size: 40, lr: 3.18e-03, grad_scale: 32.0 2023-04-27 15:49:30,587 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8287, 1.9546, 1.2275, 1.5939, 2.0319, 1.6546, 1.6275, 1.6982], device='cuda:4'), covar=tensor([0.0395, 0.0280, 0.0289, 0.0419, 0.0249, 0.0394, 0.0393, 0.0437], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0029, 0.0019, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0045, 0.0038, 0.0052, 0.0038, 0.0050, 0.0050, 0.0051], device='cuda:4') 2023-04-27 15:49:43,206 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=118329.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:49:43,813 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=118330.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:49:45,709 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6776, 1.4592, 1.9574, 2.0878, 1.4938, 1.3831, 1.5887, 0.9737], device='cuda:4'), covar=tensor([0.0448, 0.0762, 0.0410, 0.0576, 0.0748, 0.1160, 0.0687, 0.0648], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0068, 0.0067, 0.0067, 0.0074, 0.0095, 0.0072, 0.0064], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 15:50:10,269 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=118350.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:50:12,621 INFO [finetune.py:976] (4/7) Epoch 21, batch 3800, loss[loss=0.1434, simple_loss=0.2286, pruned_loss=0.02906, over 4919.00 frames. ], tot_loss[loss=0.1738, simple_loss=0.2467, pruned_loss=0.05048, over 953965.39 frames. ], batch size: 42, lr: 3.18e-03, grad_scale: 32.0 2023-04-27 15:50:23,126 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.061e+02 1.645e+02 2.008e+02 2.325e+02 4.479e+02, threshold=4.015e+02, percent-clipped=4.0 2023-04-27 15:50:52,597 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=118390.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:50:56,194 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=118396.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:51:02,181 INFO [finetune.py:976] (4/7) Epoch 21, batch 3850, loss[loss=0.1999, simple_loss=0.2713, pruned_loss=0.06427, over 4863.00 frames. ], tot_loss[loss=0.1721, simple_loss=0.2452, pruned_loss=0.04951, over 954443.87 frames. ], batch size: 31, lr: 3.18e-03, grad_scale: 32.0 2023-04-27 15:51:35,149 INFO [finetune.py:976] (4/7) Epoch 21, batch 3900, loss[loss=0.1365, simple_loss=0.2148, pruned_loss=0.02908, over 4788.00 frames. ], tot_loss[loss=0.1696, simple_loss=0.2422, pruned_loss=0.04852, over 955729.15 frames. ], batch size: 29, lr: 3.18e-03, grad_scale: 32.0 2023-04-27 15:51:35,934 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.49 vs. limit=2.0 2023-04-27 15:51:40,428 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.017e+02 1.638e+02 1.891e+02 2.265e+02 7.777e+02, threshold=3.782e+02, percent-clipped=1.0 2023-04-27 15:51:50,917 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1500, 1.7996, 2.2589, 2.5233, 2.1298, 2.0690, 2.1445, 2.1534], device='cuda:4'), covar=tensor([0.5231, 0.7541, 0.8220, 0.6145, 0.6329, 0.9246, 0.9108, 1.0296], device='cuda:4'), in_proj_covar=tensor([0.0428, 0.0411, 0.0504, 0.0505, 0.0458, 0.0487, 0.0494, 0.0501], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 15:52:07,431 INFO [finetune.py:976] (4/7) Epoch 21, batch 3950, loss[loss=0.1668, simple_loss=0.2419, pruned_loss=0.0458, over 4904.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.2395, pruned_loss=0.04795, over 957635.04 frames. ], batch size: 32, lr: 3.18e-03, grad_scale: 32.0 2023-04-27 15:52:21,112 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=118524.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 15:52:27,363 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.7374, 1.6763, 1.7008, 2.6100, 2.5069, 2.1326, 2.1992, 1.6601], device='cuda:4'), covar=tensor([0.1250, 0.1720, 0.1584, 0.1246, 0.0916, 0.1671, 0.1712, 0.2190], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0309, 0.0349, 0.0287, 0.0324, 0.0306, 0.0299, 0.0368], device='cuda:4'), out_proj_covar=tensor([6.3284e-05, 6.3906e-05, 7.3610e-05, 5.7890e-05, 6.6878e-05, 6.4187e-05, 6.2759e-05, 7.8050e-05], device='cuda:4') 2023-04-27 15:52:40,874 INFO [finetune.py:976] (4/7) Epoch 21, batch 4000, loss[loss=0.1652, simple_loss=0.2371, pruned_loss=0.04663, over 4830.00 frames. ], tot_loss[loss=0.1676, simple_loss=0.2386, pruned_loss=0.04829, over 958489.96 frames. ], batch size: 33, lr: 3.18e-03, grad_scale: 32.0 2023-04-27 15:52:47,271 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.678e+01 1.545e+02 1.901e+02 2.367e+02 3.489e+02, threshold=3.803e+02, percent-clipped=0.0 2023-04-27 15:52:50,285 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5255, 1.0224, 0.4446, 1.1725, 1.0957, 1.3806, 1.2896, 1.2677], device='cuda:4'), covar=tensor([0.0498, 0.0450, 0.0425, 0.0595, 0.0301, 0.0531, 0.0516, 0.0610], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0029, 0.0019, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0045, 0.0038, 0.0052, 0.0038, 0.0050, 0.0050, 0.0051], device='cuda:4') 2023-04-27 15:52:58,171 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.8728, 2.8434, 2.1759, 3.2753, 2.8487, 2.8892, 1.3134, 2.7831], device='cuda:4'), covar=tensor([0.2422, 0.1616, 0.3723, 0.3062, 0.3021, 0.2003, 0.5383, 0.2818], device='cuda:4'), in_proj_covar=tensor([0.0245, 0.0215, 0.0252, 0.0306, 0.0296, 0.0246, 0.0275, 0.0271], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 15:53:30,869 INFO [finetune.py:976] (4/7) Epoch 21, batch 4050, loss[loss=0.1772, simple_loss=0.2467, pruned_loss=0.05383, over 4916.00 frames. ], tot_loss[loss=0.1706, simple_loss=0.2421, pruned_loss=0.04951, over 958058.66 frames. ], batch size: 37, lr: 3.18e-03, grad_scale: 32.0 2023-04-27 15:53:56,792 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=118625.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:54:06,622 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=118631.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:54:26,640 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=118645.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:54:32,680 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.34 vs. limit=5.0 2023-04-27 15:54:33,133 INFO [finetune.py:976] (4/7) Epoch 21, batch 4100, loss[loss=0.189, simple_loss=0.2624, pruned_loss=0.05781, over 4824.00 frames. ], tot_loss[loss=0.1705, simple_loss=0.2429, pruned_loss=0.04902, over 957967.87 frames. ], batch size: 33, lr: 3.18e-03, grad_scale: 32.0 2023-04-27 15:54:38,507 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.195e+02 1.657e+02 1.981e+02 2.296e+02 3.754e+02, threshold=3.963e+02, percent-clipped=0.0 2023-04-27 15:54:54,352 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=118685.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:54:58,636 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=118692.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:55:00,960 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=118696.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:55:06,701 INFO [finetune.py:976] (4/7) Epoch 21, batch 4150, loss[loss=0.1849, simple_loss=0.2507, pruned_loss=0.05951, over 4821.00 frames. ], tot_loss[loss=0.1724, simple_loss=0.2449, pruned_loss=0.04999, over 955634.71 frames. ], batch size: 33, lr: 3.18e-03, grad_scale: 32.0 2023-04-27 15:55:56,016 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=118744.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:55:56,731 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8663, 1.4703, 1.9540, 2.3570, 1.9612, 1.8082, 1.8738, 1.8726], device='cuda:4'), covar=tensor([0.4452, 0.6877, 0.6418, 0.5206, 0.5638, 0.8140, 0.7686, 0.9065], device='cuda:4'), in_proj_covar=tensor([0.0427, 0.0410, 0.0503, 0.0503, 0.0456, 0.0485, 0.0492, 0.0499], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 15:56:07,490 INFO [finetune.py:976] (4/7) Epoch 21, batch 4200, loss[loss=0.1378, simple_loss=0.2066, pruned_loss=0.03452, over 4280.00 frames. ], tot_loss[loss=0.1722, simple_loss=0.2451, pruned_loss=0.04959, over 952585.24 frames. ], batch size: 65, lr: 3.18e-03, grad_scale: 32.0 2023-04-27 15:56:19,456 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.047e+02 1.568e+02 1.859e+02 2.228e+02 3.642e+02, threshold=3.719e+02, percent-clipped=0.0 2023-04-27 15:56:22,490 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0076, 2.3896, 1.0298, 1.3903, 1.8776, 1.2680, 2.9171, 1.6248], device='cuda:4'), covar=tensor([0.0678, 0.0665, 0.0759, 0.1166, 0.0455, 0.0919, 0.0234, 0.0597], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0073, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0007, 0.0008, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-27 15:56:58,125 INFO [finetune.py:976] (4/7) Epoch 21, batch 4250, loss[loss=0.2069, simple_loss=0.2559, pruned_loss=0.07898, over 4932.00 frames. ], tot_loss[loss=0.1709, simple_loss=0.2434, pruned_loss=0.04917, over 954186.99 frames. ], batch size: 33, lr: 3.18e-03, grad_scale: 32.0 2023-04-27 15:57:03,377 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2852, 1.5074, 1.3571, 1.7707, 1.6830, 1.8957, 1.3799, 3.3117], device='cuda:4'), covar=tensor([0.0586, 0.0794, 0.0781, 0.1156, 0.0591, 0.0468, 0.0747, 0.0152], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 15:57:13,410 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=118824.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:57:32,158 INFO [finetune.py:976] (4/7) Epoch 21, batch 4300, loss[loss=0.1347, simple_loss=0.2051, pruned_loss=0.03215, over 4829.00 frames. ], tot_loss[loss=0.1687, simple_loss=0.2408, pruned_loss=0.04826, over 955803.43 frames. ], batch size: 39, lr: 3.18e-03, grad_scale: 32.0 2023-04-27 15:57:37,518 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.744e+01 1.518e+02 1.726e+02 2.140e+02 3.725e+02, threshold=3.451e+02, percent-clipped=1.0 2023-04-27 15:57:44,623 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=118872.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:58:06,089 INFO [finetune.py:976] (4/7) Epoch 21, batch 4350, loss[loss=0.179, simple_loss=0.2493, pruned_loss=0.05429, over 4829.00 frames. ], tot_loss[loss=0.1673, simple_loss=0.2388, pruned_loss=0.04793, over 956065.24 frames. ], batch size: 30, lr: 3.18e-03, grad_scale: 32.0 2023-04-27 15:58:20,439 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=118925.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:58:39,669 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=118945.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:58:51,396 INFO [finetune.py:976] (4/7) Epoch 21, batch 4400, loss[loss=0.2426, simple_loss=0.3068, pruned_loss=0.08926, over 4833.00 frames. ], tot_loss[loss=0.1693, simple_loss=0.2404, pruned_loss=0.04913, over 955478.65 frames. ], batch size: 40, lr: 3.18e-03, grad_scale: 32.0 2023-04-27 15:59:00,773 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.017e+02 1.550e+02 1.866e+02 2.301e+02 5.364e+02, threshold=3.732e+02, percent-clipped=5.0 2023-04-27 15:59:14,268 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=118973.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:59:35,145 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=118985.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:59:36,316 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=118987.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:59:45,376 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=118993.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 15:59:52,146 INFO [finetune.py:976] (4/7) Epoch 21, batch 4450, loss[loss=0.1951, simple_loss=0.2797, pruned_loss=0.05526, over 4853.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2432, pruned_loss=0.04954, over 956875.86 frames. ], batch size: 44, lr: 3.18e-03, grad_scale: 32.0 2023-04-27 16:00:12,475 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=119033.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:00:25,718 INFO [finetune.py:976] (4/7) Epoch 21, batch 4500, loss[loss=0.1842, simple_loss=0.2475, pruned_loss=0.06044, over 4896.00 frames. ], tot_loss[loss=0.1718, simple_loss=0.2439, pruned_loss=0.04981, over 954575.31 frames. ], batch size: 35, lr: 3.18e-03, grad_scale: 32.0 2023-04-27 16:00:30,592 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.010e+02 1.788e+02 2.087e+02 2.542e+02 6.471e+02, threshold=4.174e+02, percent-clipped=4.0 2023-04-27 16:00:39,548 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=119075.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:00:41,819 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1028, 1.4451, 1.3633, 1.7053, 1.5655, 1.8078, 1.3308, 3.3819], device='cuda:4'), covar=tensor([0.0730, 0.1090, 0.1019, 0.1358, 0.0811, 0.0737, 0.1030, 0.0243], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 16:00:58,652 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=119093.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:01:15,261 INFO [finetune.py:976] (4/7) Epoch 21, batch 4550, loss[loss=0.1757, simple_loss=0.2483, pruned_loss=0.05156, over 4804.00 frames. ], tot_loss[loss=0.1746, simple_loss=0.247, pruned_loss=0.05105, over 954938.61 frames. ], batch size: 45, lr: 3.18e-03, grad_scale: 32.0 2023-04-27 16:01:59,491 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=119136.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:01:59,927 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-04-27 16:02:03,042 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-27 16:02:16,876 INFO [finetune.py:976] (4/7) Epoch 21, batch 4600, loss[loss=0.1702, simple_loss=0.2329, pruned_loss=0.05377, over 4745.00 frames. ], tot_loss[loss=0.1735, simple_loss=0.2461, pruned_loss=0.05044, over 955693.71 frames. ], batch size: 54, lr: 3.18e-03, grad_scale: 32.0 2023-04-27 16:02:16,986 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=119154.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 16:02:21,814 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.037e+02 1.477e+02 1.832e+02 2.213e+02 3.294e+02, threshold=3.663e+02, percent-clipped=0.0 2023-04-27 16:02:50,905 INFO [finetune.py:976] (4/7) Epoch 21, batch 4650, loss[loss=0.1687, simple_loss=0.2243, pruned_loss=0.0566, over 4740.00 frames. ], tot_loss[loss=0.1715, simple_loss=0.2436, pruned_loss=0.04974, over 954225.11 frames. ], batch size: 23, lr: 3.18e-03, grad_scale: 32.0 2023-04-27 16:03:09,481 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.2918, 3.1943, 2.4479, 3.8393, 3.2851, 3.2939, 1.3503, 3.2949], device='cuda:4'), covar=tensor([0.2148, 0.1642, 0.3611, 0.2552, 0.3192, 0.2130, 0.6253, 0.2764], device='cuda:4'), in_proj_covar=tensor([0.0247, 0.0217, 0.0254, 0.0307, 0.0298, 0.0248, 0.0278, 0.0274], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 16:03:21,164 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7614, 1.9728, 0.7857, 1.4824, 2.0157, 1.6379, 1.5177, 1.6666], device='cuda:4'), covar=tensor([0.0517, 0.0351, 0.0339, 0.0555, 0.0257, 0.0499, 0.0496, 0.0573], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0029, 0.0019, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0045, 0.0038, 0.0052, 0.0038, 0.0050, 0.0050, 0.0051], device='cuda:4') 2023-04-27 16:03:24,688 INFO [finetune.py:976] (4/7) Epoch 21, batch 4700, loss[loss=0.1546, simple_loss=0.2235, pruned_loss=0.04288, over 4868.00 frames. ], tot_loss[loss=0.1687, simple_loss=0.24, pruned_loss=0.04872, over 952231.72 frames. ], batch size: 31, lr: 3.18e-03, grad_scale: 32.0 2023-04-27 16:03:29,613 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.111e+02 1.611e+02 1.939e+02 2.351e+02 3.791e+02, threshold=3.879e+02, percent-clipped=1.0 2023-04-27 16:03:45,980 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=119287.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:03:58,706 INFO [finetune.py:976] (4/7) Epoch 21, batch 4750, loss[loss=0.182, simple_loss=0.2518, pruned_loss=0.05604, over 4893.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.2377, pruned_loss=0.04798, over 949825.94 frames. ], batch size: 35, lr: 3.17e-03, grad_scale: 32.0 2023-04-27 16:04:33,042 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3579, 1.5303, 3.6926, 3.4489, 3.2487, 3.4962, 3.4775, 3.2838], device='cuda:4'), covar=tensor([0.7139, 0.4925, 0.1225, 0.1891, 0.1388, 0.1898, 0.2769, 0.1632], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0306, 0.0404, 0.0403, 0.0346, 0.0408, 0.0310, 0.0365], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 16:04:39,496 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=119335.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:04:45,208 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.71 vs. limit=2.0 2023-04-27 16:04:54,475 INFO [finetune.py:976] (4/7) Epoch 21, batch 4800, loss[loss=0.1834, simple_loss=0.2536, pruned_loss=0.05663, over 4901.00 frames. ], tot_loss[loss=0.1697, simple_loss=0.2408, pruned_loss=0.04932, over 951597.47 frames. ], batch size: 35, lr: 3.17e-03, grad_scale: 32.0 2023-04-27 16:04:55,824 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0922, 1.8782, 1.6462, 1.5187, 2.1128, 1.5808, 2.4141, 1.4703], device='cuda:4'), covar=tensor([0.2958, 0.1684, 0.4073, 0.2791, 0.1236, 0.2284, 0.1317, 0.4067], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0348, 0.0423, 0.0352, 0.0379, 0.0374, 0.0366, 0.0416], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 16:04:59,371 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.005e+02 1.524e+02 1.812e+02 2.153e+02 4.144e+02, threshold=3.624e+02, percent-clipped=1.0 2023-04-27 16:05:03,752 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9383, 2.3577, 1.9874, 2.2558, 1.6175, 2.0116, 2.0950, 1.4263], device='cuda:4'), covar=tensor([0.1837, 0.0965, 0.0747, 0.1095, 0.3090, 0.1130, 0.1754, 0.2609], device='cuda:4'), in_proj_covar=tensor([0.0284, 0.0297, 0.0214, 0.0276, 0.0311, 0.0254, 0.0247, 0.0263], device='cuda:4'), out_proj_covar=tensor([1.1384e-04, 1.1778e-04, 8.4660e-05, 1.0943e-04, 1.2607e-04, 1.0051e-04, 9.9766e-05, 1.0429e-04], device='cuda:4') 2023-04-27 16:05:26,790 INFO [finetune.py:976] (4/7) Epoch 21, batch 4850, loss[loss=0.1541, simple_loss=0.2264, pruned_loss=0.04085, over 4708.00 frames. ], tot_loss[loss=0.1718, simple_loss=0.2438, pruned_loss=0.04995, over 950842.05 frames. ], batch size: 23, lr: 3.17e-03, grad_scale: 32.0 2023-04-27 16:05:40,587 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.6727, 4.6151, 3.0507, 5.3614, 4.6573, 4.6583, 2.0941, 4.6606], device='cuda:4'), covar=tensor([0.1640, 0.0962, 0.3375, 0.0895, 0.3681, 0.1535, 0.5412, 0.2116], device='cuda:4'), in_proj_covar=tensor([0.0246, 0.0216, 0.0253, 0.0305, 0.0297, 0.0247, 0.0276, 0.0272], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 16:05:43,553 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=119431.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:05:55,526 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=119449.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 16:05:58,988 INFO [finetune.py:976] (4/7) Epoch 21, batch 4900, loss[loss=0.2107, simple_loss=0.272, pruned_loss=0.07466, over 4811.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.2459, pruned_loss=0.05097, over 951848.05 frames. ], batch size: 40, lr: 3.17e-03, grad_scale: 32.0 2023-04-27 16:06:04,799 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.059e+02 1.637e+02 1.876e+02 2.288e+02 4.482e+02, threshold=3.752e+02, percent-clipped=1.0 2023-04-27 16:06:20,071 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5783, 2.6084, 1.8761, 2.2057, 2.6594, 2.0449, 3.3221, 1.8334], device='cuda:4'), covar=tensor([0.3649, 0.2186, 0.5224, 0.3214, 0.1899, 0.2808, 0.1696, 0.4701], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0347, 0.0423, 0.0352, 0.0379, 0.0373, 0.0367, 0.0416], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 16:06:35,841 INFO [finetune.py:976] (4/7) Epoch 21, batch 4950, loss[loss=0.1385, simple_loss=0.2085, pruned_loss=0.03431, over 4223.00 frames. ], tot_loss[loss=0.1737, simple_loss=0.2463, pruned_loss=0.05052, over 953255.88 frames. ], batch size: 66, lr: 3.17e-03, grad_scale: 16.0 2023-04-27 16:06:49,602 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=119514.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:07:32,623 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.38 vs. limit=5.0 2023-04-27 16:07:42,889 INFO [finetune.py:976] (4/7) Epoch 21, batch 5000, loss[loss=0.1739, simple_loss=0.2495, pruned_loss=0.04914, over 4789.00 frames. ], tot_loss[loss=0.1721, simple_loss=0.2448, pruned_loss=0.04969, over 953410.65 frames. ], batch size: 26, lr: 3.17e-03, grad_scale: 16.0 2023-04-27 16:07:56,236 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.721e+01 1.533e+02 1.932e+02 2.312e+02 5.244e+02, threshold=3.864e+02, percent-clipped=4.0 2023-04-27 16:08:15,222 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=119575.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:08:44,470 INFO [finetune.py:976] (4/7) Epoch 21, batch 5050, loss[loss=0.1455, simple_loss=0.2271, pruned_loss=0.03189, over 4813.00 frames. ], tot_loss[loss=0.1691, simple_loss=0.2414, pruned_loss=0.04837, over 953546.78 frames. ], batch size: 25, lr: 3.17e-03, grad_scale: 16.0 2023-04-27 16:09:03,488 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.8389, 2.1989, 2.5129, 3.2149, 3.0757, 2.5894, 2.2274, 2.7997], device='cuda:4'), covar=tensor([0.0732, 0.1122, 0.0738, 0.0485, 0.0501, 0.0827, 0.0701, 0.0558], device='cuda:4'), in_proj_covar=tensor([0.0188, 0.0203, 0.0185, 0.0175, 0.0178, 0.0181, 0.0153, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 16:09:28,918 INFO [finetune.py:976] (4/7) Epoch 21, batch 5100, loss[loss=0.143, simple_loss=0.2142, pruned_loss=0.03594, over 4855.00 frames. ], tot_loss[loss=0.167, simple_loss=0.2389, pruned_loss=0.0476, over 955413.55 frames. ], batch size: 44, lr: 3.17e-03, grad_scale: 16.0 2023-04-27 16:09:41,674 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.943e+01 1.499e+02 1.766e+02 2.161e+02 3.760e+02, threshold=3.532e+02, percent-clipped=0.0 2023-04-27 16:09:43,049 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=119665.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:10:31,238 INFO [finetune.py:976] (4/7) Epoch 21, batch 5150, loss[loss=0.2625, simple_loss=0.3283, pruned_loss=0.09836, over 4842.00 frames. ], tot_loss[loss=0.167, simple_loss=0.2389, pruned_loss=0.04756, over 957255.48 frames. ], batch size: 47, lr: 3.17e-03, grad_scale: 16.0 2023-04-27 16:11:05,041 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=119726.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:11:08,033 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=119731.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:11:18,317 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0898, 2.8143, 2.1266, 2.2255, 1.4742, 1.4968, 2.2636, 1.4382], device='cuda:4'), covar=tensor([0.1690, 0.1504, 0.1321, 0.1707, 0.2270, 0.1883, 0.0944, 0.2023], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0211, 0.0168, 0.0203, 0.0199, 0.0184, 0.0155, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 16:11:30,485 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=119749.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 16:11:33,427 INFO [finetune.py:976] (4/7) Epoch 21, batch 5200, loss[loss=0.227, simple_loss=0.3064, pruned_loss=0.07378, over 4919.00 frames. ], tot_loss[loss=0.1714, simple_loss=0.2439, pruned_loss=0.04944, over 958499.68 frames. ], batch size: 36, lr: 3.17e-03, grad_scale: 16.0 2023-04-27 16:11:39,368 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.908e+01 1.576e+02 2.065e+02 2.326e+02 4.790e+02, threshold=4.130e+02, percent-clipped=1.0 2023-04-27 16:11:51,104 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=119779.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:12:02,213 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=119797.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:12:06,570 INFO [finetune.py:976] (4/7) Epoch 21, batch 5250, loss[loss=0.1954, simple_loss=0.2719, pruned_loss=0.05948, over 4795.00 frames. ], tot_loss[loss=0.1734, simple_loss=0.2461, pruned_loss=0.05037, over 955889.54 frames. ], batch size: 41, lr: 3.17e-03, grad_scale: 16.0 2023-04-27 16:12:42,164 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-27 16:12:51,201 INFO [finetune.py:976] (4/7) Epoch 21, batch 5300, loss[loss=0.1608, simple_loss=0.2342, pruned_loss=0.04367, over 4760.00 frames. ], tot_loss[loss=0.174, simple_loss=0.2471, pruned_loss=0.05046, over 956205.17 frames. ], batch size: 54, lr: 3.17e-03, grad_scale: 16.0 2023-04-27 16:13:02,329 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.110e+02 1.628e+02 1.910e+02 2.445e+02 5.972e+02, threshold=3.820e+02, percent-clipped=3.0 2023-04-27 16:13:13,804 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=119870.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:13:37,792 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9457, 4.3572, 1.2772, 2.1089, 2.2445, 2.7936, 2.4102, 1.0026], device='cuda:4'), covar=tensor([0.1336, 0.0940, 0.1802, 0.1229, 0.1103, 0.1068, 0.1495, 0.2166], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0238, 0.0136, 0.0119, 0.0132, 0.0151, 0.0115, 0.0117], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 16:13:53,930 INFO [finetune.py:976] (4/7) Epoch 21, batch 5350, loss[loss=0.19, simple_loss=0.2576, pruned_loss=0.06125, over 4739.00 frames. ], tot_loss[loss=0.173, simple_loss=0.2466, pruned_loss=0.04975, over 955836.11 frames. ], batch size: 54, lr: 3.17e-03, grad_scale: 16.0 2023-04-27 16:14:43,496 INFO [finetune.py:976] (4/7) Epoch 21, batch 5400, loss[loss=0.1415, simple_loss=0.2137, pruned_loss=0.03469, over 4884.00 frames. ], tot_loss[loss=0.1713, simple_loss=0.244, pruned_loss=0.04932, over 956111.13 frames. ], batch size: 32, lr: 3.17e-03, grad_scale: 16.0 2023-04-27 16:14:48,954 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.105e+02 1.397e+02 1.683e+02 2.017e+02 3.693e+02, threshold=3.366e+02, percent-clipped=0.0 2023-04-27 16:15:05,937 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4815, 1.7787, 1.3786, 1.1990, 1.1807, 1.1337, 1.4083, 1.1340], device='cuda:4'), covar=tensor([0.1360, 0.1333, 0.1301, 0.1634, 0.2073, 0.1698, 0.0911, 0.1831], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0210, 0.0167, 0.0202, 0.0198, 0.0184, 0.0155, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 16:15:16,769 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-04-27 16:15:18,157 INFO [finetune.py:976] (4/7) Epoch 21, batch 5450, loss[loss=0.1672, simple_loss=0.2324, pruned_loss=0.05096, over 4903.00 frames. ], tot_loss[loss=0.1704, simple_loss=0.2425, pruned_loss=0.0492, over 956547.49 frames. ], batch size: 36, lr: 3.17e-03, grad_scale: 16.0 2023-04-27 16:15:18,881 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4927, 1.0734, 0.4100, 1.2069, 1.1281, 1.3624, 1.2884, 1.2656], device='cuda:4'), covar=tensor([0.0513, 0.0399, 0.0395, 0.0578, 0.0282, 0.0508, 0.0493, 0.0588], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0029, 0.0019, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0045, 0.0038, 0.0052, 0.0038, 0.0050, 0.0050, 0.0051], device='cuda:4') 2023-04-27 16:15:34,275 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=120021.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:15:49,962 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7271, 1.2558, 4.5871, 4.2567, 4.0006, 4.3323, 4.2222, 4.1069], device='cuda:4'), covar=tensor([0.6629, 0.6595, 0.1154, 0.1976, 0.1250, 0.2026, 0.1688, 0.1402], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0306, 0.0404, 0.0404, 0.0348, 0.0410, 0.0312, 0.0366], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 16:16:00,803 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0326, 2.0148, 1.8851, 1.6424, 2.2777, 1.7203, 2.7392, 1.7160], device='cuda:4'), covar=tensor([0.4007, 0.2052, 0.4557, 0.2938, 0.1485, 0.2618, 0.1156, 0.4234], device='cuda:4'), in_proj_covar=tensor([0.0345, 0.0352, 0.0429, 0.0356, 0.0384, 0.0379, 0.0372, 0.0421], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 16:16:02,719 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4717, 1.9818, 2.3191, 2.9302, 2.3997, 1.8926, 1.8464, 2.1976], device='cuda:4'), covar=tensor([0.3064, 0.2970, 0.1600, 0.2350, 0.2589, 0.2448, 0.3603, 0.1914], device='cuda:4'), in_proj_covar=tensor([0.0294, 0.0245, 0.0227, 0.0315, 0.0220, 0.0234, 0.0229, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 16:16:03,187 INFO [finetune.py:976] (4/7) Epoch 21, batch 5500, loss[loss=0.1442, simple_loss=0.2147, pruned_loss=0.0369, over 4901.00 frames. ], tot_loss[loss=0.1683, simple_loss=0.2398, pruned_loss=0.0484, over 956806.99 frames. ], batch size: 43, lr: 3.17e-03, grad_scale: 16.0 2023-04-27 16:16:14,183 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.019e+02 1.665e+02 1.823e+02 2.184e+02 3.265e+02, threshold=3.646e+02, percent-clipped=0.0 2023-04-27 16:17:07,730 INFO [finetune.py:976] (4/7) Epoch 21, batch 5550, loss[loss=0.1851, simple_loss=0.2564, pruned_loss=0.05692, over 4850.00 frames. ], tot_loss[loss=0.1691, simple_loss=0.2407, pruned_loss=0.04872, over 956423.18 frames. ], batch size: 44, lr: 3.17e-03, grad_scale: 16.0 2023-04-27 16:18:05,780 INFO [finetune.py:976] (4/7) Epoch 21, batch 5600, loss[loss=0.1483, simple_loss=0.2215, pruned_loss=0.03756, over 4800.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2447, pruned_loss=0.04959, over 954317.81 frames. ], batch size: 25, lr: 3.17e-03, grad_scale: 16.0 2023-04-27 16:18:06,499 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-27 16:18:11,016 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 1.536e+02 1.871e+02 2.277e+02 4.229e+02, threshold=3.743e+02, percent-clipped=5.0 2023-04-27 16:18:15,218 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=120170.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:18:36,565 INFO [finetune.py:976] (4/7) Epoch 21, batch 5650, loss[loss=0.1497, simple_loss=0.2275, pruned_loss=0.03593, over 4837.00 frames. ], tot_loss[loss=0.1732, simple_loss=0.2467, pruned_loss=0.04989, over 952216.90 frames. ], batch size: 30, lr: 3.17e-03, grad_scale: 16.0 2023-04-27 16:18:45,335 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=120218.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:18:54,552 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.73 vs. limit=2.0 2023-04-27 16:19:19,443 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8479, 1.9980, 1.9523, 2.2429, 2.1391, 2.2727, 1.8669, 3.8388], device='cuda:4'), covar=tensor([0.0458, 0.0677, 0.0635, 0.0989, 0.0493, 0.0494, 0.0612, 0.0168], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 16:19:28,169 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0348, 2.3898, 2.1340, 2.2928, 1.8947, 2.0958, 2.0376, 1.6232], device='cuda:4'), covar=tensor([0.1404, 0.0938, 0.0737, 0.0917, 0.2797, 0.1112, 0.1487, 0.2163], device='cuda:4'), in_proj_covar=tensor([0.0283, 0.0297, 0.0214, 0.0275, 0.0310, 0.0253, 0.0247, 0.0261], device='cuda:4'), out_proj_covar=tensor([1.1344e-04, 1.1763e-04, 8.4397e-05, 1.0851e-04, 1.2583e-04, 1.0022e-04, 9.9690e-05, 1.0353e-04], device='cuda:4') 2023-04-27 16:19:30,423 INFO [finetune.py:976] (4/7) Epoch 21, batch 5700, loss[loss=0.1609, simple_loss=0.2187, pruned_loss=0.05154, over 3999.00 frames. ], tot_loss[loss=0.17, simple_loss=0.242, pruned_loss=0.04894, over 932276.13 frames. ], batch size: 17, lr: 3.17e-03, grad_scale: 16.0 2023-04-27 16:19:41,630 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.050e+02 1.444e+02 1.777e+02 2.267e+02 3.469e+02, threshold=3.555e+02, percent-clipped=0.0 2023-04-27 16:19:54,263 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-04-27 16:20:17,303 INFO [finetune.py:976] (4/7) Epoch 22, batch 0, loss[loss=0.1517, simple_loss=0.2283, pruned_loss=0.03752, over 4885.00 frames. ], tot_loss[loss=0.1517, simple_loss=0.2283, pruned_loss=0.03752, over 4885.00 frames. ], batch size: 32, lr: 3.16e-03, grad_scale: 16.0 2023-04-27 16:20:17,304 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 16:20:33,673 INFO [finetune.py:1010] (4/7) Epoch 22, validation: loss=0.1546, simple_loss=0.2251, pruned_loss=0.04204, over 2265189.00 frames. 2023-04-27 16:20:33,673 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-27 16:20:37,881 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4085, 2.3144, 2.5626, 2.9445, 2.7970, 2.2209, 1.9416, 2.5459], device='cuda:4'), covar=tensor([0.0920, 0.0987, 0.0673, 0.0611, 0.0660, 0.1052, 0.0839, 0.0602], device='cuda:4'), in_proj_covar=tensor([0.0187, 0.0201, 0.0183, 0.0174, 0.0177, 0.0180, 0.0152, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 16:20:59,190 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=120321.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:21:05,119 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.91 vs. limit=2.0 2023-04-27 16:21:06,359 INFO [finetune.py:976] (4/7) Epoch 22, batch 50, loss[loss=0.1706, simple_loss=0.2445, pruned_loss=0.04837, over 4896.00 frames. ], tot_loss[loss=0.1786, simple_loss=0.2513, pruned_loss=0.05293, over 216275.28 frames. ], batch size: 46, lr: 3.16e-03, grad_scale: 16.0 2023-04-27 16:21:21,922 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9808, 1.7059, 2.1289, 2.4048, 2.0353, 1.8745, 2.0065, 1.9721], device='cuda:4'), covar=tensor([0.4742, 0.6876, 0.7210, 0.5591, 0.6002, 0.8955, 0.8874, 0.9874], device='cuda:4'), in_proj_covar=tensor([0.0429, 0.0412, 0.0505, 0.0504, 0.0457, 0.0487, 0.0495, 0.0502], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 16:21:23,744 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4222, 1.6892, 1.7817, 1.8961, 1.7372, 1.8611, 1.8803, 1.9061], device='cuda:4'), covar=tensor([0.3454, 0.4573, 0.4400, 0.3839, 0.5198, 0.6699, 0.4796, 0.4414], device='cuda:4'), in_proj_covar=tensor([0.0339, 0.0375, 0.0323, 0.0338, 0.0348, 0.0395, 0.0359, 0.0330], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 16:21:24,325 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9281, 3.9776, 0.8216, 2.1938, 2.1181, 2.7592, 2.3548, 0.8281], device='cuda:4'), covar=tensor([0.1300, 0.0974, 0.2184, 0.1250, 0.1158, 0.1031, 0.1561, 0.2335], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0238, 0.0136, 0.0118, 0.0132, 0.0151, 0.0115, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 16:21:27,313 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.523e+01 1.622e+02 1.941e+02 2.362e+02 4.057e+02, threshold=3.882e+02, percent-clipped=2.0 2023-04-27 16:21:31,537 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=120369.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:21:51,384 INFO [finetune.py:976] (4/7) Epoch 22, batch 100, loss[loss=0.1518, simple_loss=0.2174, pruned_loss=0.04312, over 4820.00 frames. ], tot_loss[loss=0.1736, simple_loss=0.2448, pruned_loss=0.05119, over 380644.66 frames. ], batch size: 33, lr: 3.16e-03, grad_scale: 16.0 2023-04-27 16:22:27,946 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.97 vs. limit=5.0 2023-04-27 16:22:55,498 INFO [finetune.py:976] (4/7) Epoch 22, batch 150, loss[loss=0.1747, simple_loss=0.2502, pruned_loss=0.0496, over 4907.00 frames. ], tot_loss[loss=0.169, simple_loss=0.2398, pruned_loss=0.04913, over 508252.64 frames. ], batch size: 43, lr: 3.16e-03, grad_scale: 16.0 2023-04-27 16:23:28,999 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.065e+02 1.459e+02 1.791e+02 2.121e+02 3.336e+02, threshold=3.582e+02, percent-clipped=0.0 2023-04-27 16:23:48,709 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.68 vs. limit=2.0 2023-04-27 16:23:58,242 INFO [finetune.py:976] (4/7) Epoch 22, batch 200, loss[loss=0.1991, simple_loss=0.2621, pruned_loss=0.0681, over 4828.00 frames. ], tot_loss[loss=0.1659, simple_loss=0.2358, pruned_loss=0.04806, over 607216.41 frames. ], batch size: 30, lr: 3.16e-03, grad_scale: 16.0 2023-04-27 16:24:23,862 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7931, 2.2158, 1.9427, 2.1121, 1.6147, 2.0006, 1.9149, 1.4659], device='cuda:4'), covar=tensor([0.2053, 0.1341, 0.0829, 0.1157, 0.3509, 0.1225, 0.1866, 0.2555], device='cuda:4'), in_proj_covar=tensor([0.0285, 0.0299, 0.0215, 0.0277, 0.0313, 0.0255, 0.0248, 0.0262], device='cuda:4'), out_proj_covar=tensor([1.1406e-04, 1.1830e-04, 8.5167e-05, 1.0937e-04, 1.2680e-04, 1.0077e-04, 1.0012e-04, 1.0386e-04], device='cuda:4') 2023-04-27 16:24:23,885 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8476, 2.3213, 1.8598, 1.6621, 1.3473, 1.3651, 1.8682, 1.3394], device='cuda:4'), covar=tensor([0.1655, 0.1165, 0.1382, 0.1634, 0.2250, 0.1933, 0.0978, 0.1960], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0209, 0.0168, 0.0202, 0.0198, 0.0184, 0.0155, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 16:24:35,497 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.8338, 2.0277, 2.0596, 2.1948, 1.9023, 2.0911, 2.1327, 2.0555], device='cuda:4'), covar=tensor([0.3624, 0.5757, 0.4588, 0.4078, 0.5610, 0.6684, 0.5788, 0.5157], device='cuda:4'), in_proj_covar=tensor([0.0338, 0.0371, 0.0322, 0.0336, 0.0347, 0.0393, 0.0357, 0.0328], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 16:25:06,005 INFO [finetune.py:976] (4/7) Epoch 22, batch 250, loss[loss=0.184, simple_loss=0.2603, pruned_loss=0.05389, over 4845.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2391, pruned_loss=0.04934, over 683163.90 frames. ], batch size: 49, lr: 3.16e-03, grad_scale: 16.0 2023-04-27 16:25:37,781 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-27 16:25:49,437 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.081e+02 1.595e+02 1.873e+02 2.274e+02 4.484e+02, threshold=3.746e+02, percent-clipped=2.0 2023-04-27 16:26:12,262 INFO [finetune.py:976] (4/7) Epoch 22, batch 300, loss[loss=0.168, simple_loss=0.2459, pruned_loss=0.04501, over 4899.00 frames. ], tot_loss[loss=0.1714, simple_loss=0.2427, pruned_loss=0.0501, over 742611.02 frames. ], batch size: 32, lr: 3.16e-03, grad_scale: 16.0 2023-04-27 16:27:06,868 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6547, 2.7253, 2.2061, 2.3103, 2.8865, 2.2769, 3.6139, 2.0635], device='cuda:4'), covar=tensor([0.3866, 0.2206, 0.4295, 0.3691, 0.1487, 0.2674, 0.1298, 0.4122], device='cuda:4'), in_proj_covar=tensor([0.0341, 0.0351, 0.0427, 0.0354, 0.0382, 0.0374, 0.0369, 0.0419], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 16:27:19,463 INFO [finetune.py:976] (4/7) Epoch 22, batch 350, loss[loss=0.2021, simple_loss=0.2728, pruned_loss=0.06568, over 4792.00 frames. ], tot_loss[loss=0.1714, simple_loss=0.2432, pruned_loss=0.0498, over 789096.18 frames. ], batch size: 29, lr: 3.16e-03, grad_scale: 16.0 2023-04-27 16:28:01,877 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.016e+02 1.528e+02 1.802e+02 2.377e+02 4.881e+02, threshold=3.603e+02, percent-clipped=3.0 2023-04-27 16:28:20,625 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=120676.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:28:24,207 INFO [finetune.py:976] (4/7) Epoch 22, batch 400, loss[loss=0.1543, simple_loss=0.2331, pruned_loss=0.03778, over 4811.00 frames. ], tot_loss[loss=0.1718, simple_loss=0.2444, pruned_loss=0.0496, over 826221.07 frames. ], batch size: 40, lr: 3.16e-03, grad_scale: 16.0 2023-04-27 16:28:44,121 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0950, 1.3400, 1.8929, 2.3282, 1.8982, 1.5072, 1.3206, 1.6989], device='cuda:4'), covar=tensor([0.3715, 0.4142, 0.1972, 0.2771, 0.2943, 0.2974, 0.4558, 0.2233], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0243, 0.0225, 0.0312, 0.0218, 0.0232, 0.0227, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 16:29:32,129 INFO [finetune.py:976] (4/7) Epoch 22, batch 450, loss[loss=0.1399, simple_loss=0.207, pruned_loss=0.03643, over 4866.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2444, pruned_loss=0.04977, over 853928.28 frames. ], batch size: 31, lr: 3.16e-03, grad_scale: 16.0 2023-04-27 16:29:40,508 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=120737.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:29:56,713 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9001, 1.7297, 1.9617, 2.2405, 2.3039, 1.9031, 1.5854, 2.0288], device='cuda:4'), covar=tensor([0.0840, 0.1106, 0.0704, 0.0652, 0.0596, 0.0839, 0.0778, 0.0594], device='cuda:4'), in_proj_covar=tensor([0.0184, 0.0198, 0.0181, 0.0172, 0.0175, 0.0177, 0.0149, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 16:30:03,215 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2388, 1.3906, 1.7496, 1.8205, 1.6971, 1.7953, 1.8059, 1.7672], device='cuda:4'), covar=tensor([0.3928, 0.5203, 0.3958, 0.4185, 0.5141, 0.6981, 0.4598, 0.4545], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0372, 0.0323, 0.0337, 0.0347, 0.0394, 0.0356, 0.0328], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 16:30:03,658 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.146e+02 1.511e+02 1.855e+02 2.296e+02 3.408e+02, threshold=3.711e+02, percent-clipped=0.0 2023-04-27 16:30:15,273 INFO [finetune.py:976] (4/7) Epoch 22, batch 500, loss[loss=0.1228, simple_loss=0.2011, pruned_loss=0.02219, over 4765.00 frames. ], tot_loss[loss=0.1704, simple_loss=0.2423, pruned_loss=0.04926, over 878905.86 frames. ], batch size: 28, lr: 3.16e-03, grad_scale: 16.0 2023-04-27 16:30:22,588 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6889, 1.2328, 1.3586, 1.3738, 1.8016, 1.4813, 1.1701, 1.3038], device='cuda:4'), covar=tensor([0.1454, 0.1539, 0.2021, 0.1311, 0.0836, 0.1546, 0.1846, 0.2215], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0310, 0.0349, 0.0287, 0.0324, 0.0308, 0.0300, 0.0370], device='cuda:4'), out_proj_covar=tensor([6.3866e-05, 6.4161e-05, 7.3787e-05, 5.7746e-05, 6.6868e-05, 6.4541e-05, 6.2753e-05, 7.8634e-05], device='cuda:4') 2023-04-27 16:30:28,372 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.66 vs. limit=2.0 2023-04-27 16:30:49,310 INFO [finetune.py:976] (4/7) Epoch 22, batch 550, loss[loss=0.1726, simple_loss=0.2444, pruned_loss=0.05037, over 4810.00 frames. ], tot_loss[loss=0.1685, simple_loss=0.2402, pruned_loss=0.04846, over 896427.31 frames. ], batch size: 51, lr: 3.16e-03, grad_scale: 16.0 2023-04-27 16:30:59,024 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2084, 1.4511, 3.8405, 3.5875, 3.3564, 3.7231, 3.6697, 3.3284], device='cuda:4'), covar=tensor([0.7337, 0.5237, 0.1195, 0.1822, 0.1265, 0.1877, 0.1601, 0.1610], device='cuda:4'), in_proj_covar=tensor([0.0314, 0.0307, 0.0409, 0.0408, 0.0350, 0.0412, 0.0314, 0.0369], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 16:31:15,725 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.143e+02 1.573e+02 1.876e+02 2.291e+02 4.467e+02, threshold=3.751e+02, percent-clipped=2.0 2023-04-27 16:31:38,961 INFO [finetune.py:976] (4/7) Epoch 22, batch 600, loss[loss=0.1777, simple_loss=0.2466, pruned_loss=0.05443, over 4897.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2402, pruned_loss=0.04882, over 909485.80 frames. ], batch size: 32, lr: 3.16e-03, grad_scale: 16.0 2023-04-27 16:32:45,425 INFO [finetune.py:976] (4/7) Epoch 22, batch 650, loss[loss=0.198, simple_loss=0.2786, pruned_loss=0.0587, over 4862.00 frames. ], tot_loss[loss=0.1725, simple_loss=0.244, pruned_loss=0.05054, over 919611.73 frames. ], batch size: 44, lr: 3.16e-03, grad_scale: 16.0 2023-04-27 16:32:55,846 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0859, 2.7387, 1.1104, 1.5803, 2.1774, 1.2577, 3.5635, 2.0689], device='cuda:4'), covar=tensor([0.0710, 0.0631, 0.0909, 0.1252, 0.0487, 0.1094, 0.0226, 0.0557], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0073, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0009, 0.0007, 0.0007, 0.0008, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-27 16:33:16,629 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3171, 2.4585, 2.4462, 2.8650, 2.7338, 2.8131, 2.4290, 4.9918], device='cuda:4'), covar=tensor([0.0427, 0.0660, 0.0660, 0.0900, 0.0469, 0.0353, 0.0564, 0.0102], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0039, 0.0038, 0.0037, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 16:33:26,706 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.051e+02 1.656e+02 1.853e+02 2.297e+02 3.999e+02, threshold=3.705e+02, percent-clipped=2.0 2023-04-27 16:33:52,001 INFO [finetune.py:976] (4/7) Epoch 22, batch 700, loss[loss=0.1628, simple_loss=0.2304, pruned_loss=0.04762, over 4792.00 frames. ], tot_loss[loss=0.1746, simple_loss=0.2462, pruned_loss=0.0515, over 927824.51 frames. ], batch size: 51, lr: 3.16e-03, grad_scale: 16.0 2023-04-27 16:34:21,525 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=121003.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:34:45,578 INFO [finetune.py:976] (4/7) Epoch 22, batch 750, loss[loss=0.2038, simple_loss=0.2729, pruned_loss=0.06739, over 4889.00 frames. ], tot_loss[loss=0.1763, simple_loss=0.248, pruned_loss=0.05232, over 933959.04 frames. ], batch size: 32, lr: 3.16e-03, grad_scale: 16.0 2023-04-27 16:34:45,656 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=121032.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:34:48,219 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9057, 0.8764, 1.0742, 1.0157, 0.8973, 0.8191, 0.9115, 0.4903], device='cuda:4'), covar=tensor([0.0518, 0.0459, 0.0440, 0.0432, 0.0712, 0.0963, 0.0441, 0.0708], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0068, 0.0066, 0.0067, 0.0074, 0.0096, 0.0072, 0.0064], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 16:35:00,245 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4433, 1.9531, 2.3606, 2.7337, 2.4014, 1.9769, 1.7234, 2.2310], device='cuda:4'), covar=tensor([0.3303, 0.2864, 0.1561, 0.2117, 0.2356, 0.2684, 0.3742, 0.1891], device='cuda:4'), in_proj_covar=tensor([0.0294, 0.0244, 0.0226, 0.0313, 0.0220, 0.0233, 0.0228, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 16:35:03,866 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.37 vs. limit=5.0 2023-04-27 16:35:04,851 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.944e+01 1.522e+02 1.865e+02 2.464e+02 7.582e+02, threshold=3.731e+02, percent-clipped=3.0 2023-04-27 16:35:05,581 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=121064.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:35:15,408 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-27 16:35:19,165 INFO [finetune.py:976] (4/7) Epoch 22, batch 800, loss[loss=0.1666, simple_loss=0.2498, pruned_loss=0.04168, over 4804.00 frames. ], tot_loss[loss=0.1741, simple_loss=0.2465, pruned_loss=0.05088, over 937736.31 frames. ], batch size: 40, lr: 3.16e-03, grad_scale: 16.0 2023-04-27 16:35:37,491 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0620, 2.5086, 1.0803, 1.4015, 2.0279, 1.2945, 3.2073, 1.7732], device='cuda:4'), covar=tensor([0.0649, 0.0600, 0.0807, 0.1158, 0.0445, 0.0924, 0.0203, 0.0598], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0052, 0.0073, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-27 16:35:52,465 INFO [finetune.py:976] (4/7) Epoch 22, batch 850, loss[loss=0.261, simple_loss=0.302, pruned_loss=0.11, over 4150.00 frames. ], tot_loss[loss=0.1733, simple_loss=0.2455, pruned_loss=0.05055, over 938973.66 frames. ], batch size: 65, lr: 3.16e-03, grad_scale: 16.0 2023-04-27 16:36:11,705 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.009e+02 1.493e+02 1.782e+02 2.235e+02 3.771e+02, threshold=3.565e+02, percent-clipped=1.0 2023-04-27 16:36:12,474 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=121164.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:36:25,801 INFO [finetune.py:976] (4/7) Epoch 22, batch 900, loss[loss=0.201, simple_loss=0.2601, pruned_loss=0.07101, over 4256.00 frames. ], tot_loss[loss=0.171, simple_loss=0.2425, pruned_loss=0.0498, over 941231.10 frames. ], batch size: 65, lr: 3.16e-03, grad_scale: 16.0 2023-04-27 16:36:29,587 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.9176, 2.2138, 2.1147, 2.3020, 1.9948, 2.1980, 2.1418, 2.0992], device='cuda:4'), covar=tensor([0.3851, 0.6091, 0.4803, 0.4491, 0.5760, 0.7075, 0.6398, 0.5371], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0372, 0.0323, 0.0337, 0.0346, 0.0392, 0.0355, 0.0328], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 16:36:32,549 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2471, 1.2909, 3.8383, 3.5624, 3.3838, 3.6988, 3.6374, 3.3634], device='cuda:4'), covar=tensor([0.6940, 0.5619, 0.1190, 0.1839, 0.1245, 0.1630, 0.1688, 0.1561], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0304, 0.0405, 0.0406, 0.0348, 0.0408, 0.0311, 0.0365], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 16:36:52,980 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=121225.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:36:59,093 INFO [finetune.py:976] (4/7) Epoch 22, batch 950, loss[loss=0.2282, simple_loss=0.2846, pruned_loss=0.08585, over 4835.00 frames. ], tot_loss[loss=0.1698, simple_loss=0.2407, pruned_loss=0.04946, over 944103.84 frames. ], batch size: 47, lr: 3.16e-03, grad_scale: 16.0 2023-04-27 16:37:35,638 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.216e+01 1.432e+02 1.893e+02 2.215e+02 5.796e+02, threshold=3.785e+02, percent-clipped=6.0 2023-04-27 16:38:01,135 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.8324, 3.7187, 2.8333, 4.4523, 3.8896, 3.8948, 1.5154, 3.8410], device='cuda:4'), covar=tensor([0.1763, 0.1536, 0.3813, 0.1455, 0.4950, 0.1785, 0.6733, 0.2743], device='cuda:4'), in_proj_covar=tensor([0.0245, 0.0216, 0.0252, 0.0305, 0.0295, 0.0246, 0.0275, 0.0273], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 16:38:01,692 INFO [finetune.py:976] (4/7) Epoch 22, batch 1000, loss[loss=0.1708, simple_loss=0.2481, pruned_loss=0.04671, over 4818.00 frames. ], tot_loss[loss=0.1725, simple_loss=0.2436, pruned_loss=0.0507, over 946287.91 frames. ], batch size: 41, lr: 3.16e-03, grad_scale: 16.0 2023-04-27 16:38:10,017 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4361, 1.5937, 1.4814, 2.0455, 1.8186, 2.0571, 1.6158, 4.4014], device='cuda:4'), covar=tensor([0.0520, 0.0831, 0.0781, 0.1193, 0.0638, 0.0507, 0.0766, 0.0078], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 16:38:41,538 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=121311.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 16:38:56,685 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6210, 1.7504, 1.4508, 1.0744, 1.2259, 1.2254, 1.4959, 1.1016], device='cuda:4'), covar=tensor([0.1730, 0.1235, 0.1513, 0.1663, 0.2343, 0.1941, 0.1011, 0.2034], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0208, 0.0167, 0.0201, 0.0198, 0.0183, 0.0154, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 16:39:01,159 INFO [finetune.py:976] (4/7) Epoch 22, batch 1050, loss[loss=0.153, simple_loss=0.2284, pruned_loss=0.03885, over 4899.00 frames. ], tot_loss[loss=0.1734, simple_loss=0.2452, pruned_loss=0.05075, over 948025.68 frames. ], batch size: 37, lr: 3.16e-03, grad_scale: 16.0 2023-04-27 16:39:01,763 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=121332.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:39:10,734 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.71 vs. limit=2.0 2023-04-27 16:39:18,529 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=121359.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:39:20,861 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.274e+02 1.706e+02 2.101e+02 2.455e+02 4.347e+02, threshold=4.202e+02, percent-clipped=2.0 2023-04-27 16:39:22,178 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4313, 2.9437, 2.4220, 2.7977, 2.0900, 2.7234, 2.7386, 2.0347], device='cuda:4'), covar=tensor([0.1958, 0.0955, 0.0841, 0.1235, 0.2861, 0.1031, 0.1657, 0.2841], device='cuda:4'), in_proj_covar=tensor([0.0287, 0.0301, 0.0217, 0.0279, 0.0316, 0.0256, 0.0250, 0.0264], device='cuda:4'), out_proj_covar=tensor([1.1511e-04, 1.1925e-04, 8.5830e-05, 1.1013e-04, 1.2812e-04, 1.0142e-04, 1.0096e-04, 1.0454e-04], device='cuda:4') 2023-04-27 16:39:22,807 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6784, 2.0211, 1.7324, 1.9884, 1.5535, 1.7968, 1.7395, 1.2973], device='cuda:4'), covar=tensor([0.1870, 0.1213, 0.0925, 0.1204, 0.3441, 0.1164, 0.1839, 0.2458], device='cuda:4'), in_proj_covar=tensor([0.0288, 0.0301, 0.0217, 0.0279, 0.0316, 0.0256, 0.0250, 0.0264], device='cuda:4'), out_proj_covar=tensor([1.1514e-04, 1.1928e-04, 8.5847e-05, 1.1016e-04, 1.2815e-04, 1.0144e-04, 1.0099e-04, 1.0457e-04], device='cuda:4') 2023-04-27 16:39:26,938 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=121372.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 16:39:31,744 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=121380.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:39:33,386 INFO [finetune.py:976] (4/7) Epoch 22, batch 1100, loss[loss=0.216, simple_loss=0.283, pruned_loss=0.0745, over 4855.00 frames. ], tot_loss[loss=0.1733, simple_loss=0.2452, pruned_loss=0.05076, over 948551.66 frames. ], batch size: 44, lr: 3.16e-03, grad_scale: 16.0 2023-04-27 16:39:40,980 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1975, 2.0755, 2.2985, 2.6125, 2.6977, 2.0915, 1.8648, 2.2968], device='cuda:4'), covar=tensor([0.0846, 0.1027, 0.0714, 0.0608, 0.0642, 0.0915, 0.0821, 0.0595], device='cuda:4'), in_proj_covar=tensor([0.0185, 0.0199, 0.0181, 0.0172, 0.0175, 0.0178, 0.0150, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 16:39:50,919 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=121407.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:40:06,341 INFO [finetune.py:976] (4/7) Epoch 22, batch 1150, loss[loss=0.1853, simple_loss=0.2572, pruned_loss=0.05672, over 4925.00 frames. ], tot_loss[loss=0.1738, simple_loss=0.246, pruned_loss=0.05087, over 950091.82 frames. ], batch size: 33, lr: 3.16e-03, grad_scale: 16.0 2023-04-27 16:40:27,721 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.115e+02 1.622e+02 1.904e+02 2.379e+02 4.461e+02, threshold=3.808e+02, percent-clipped=1.0 2023-04-27 16:40:30,896 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=121468.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:40:39,781 INFO [finetune.py:976] (4/7) Epoch 22, batch 1200, loss[loss=0.1856, simple_loss=0.2494, pruned_loss=0.06091, over 4925.00 frames. ], tot_loss[loss=0.1721, simple_loss=0.2444, pruned_loss=0.04996, over 949321.78 frames. ], batch size: 38, lr: 3.15e-03, grad_scale: 16.0 2023-04-27 16:40:45,563 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5505, 1.7302, 1.7257, 2.1435, 2.0216, 2.1128, 1.7603, 4.6068], device='cuda:4'), covar=tensor([0.0527, 0.0842, 0.0808, 0.1208, 0.0607, 0.0495, 0.0703, 0.0106], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 16:41:04,809 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-27 16:41:05,070 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=121520.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:41:12,829 INFO [finetune.py:976] (4/7) Epoch 22, batch 1250, loss[loss=0.1602, simple_loss=0.2255, pruned_loss=0.04741, over 4903.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.2415, pruned_loss=0.04874, over 952572.03 frames. ], batch size: 35, lr: 3.15e-03, grad_scale: 32.0 2023-04-27 16:41:34,047 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.134e+02 1.460e+02 1.755e+02 2.215e+02 4.561e+02, threshold=3.510e+02, percent-clipped=2.0 2023-04-27 16:41:41,148 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.63 vs. limit=2.0 2023-04-27 16:41:46,169 INFO [finetune.py:976] (4/7) Epoch 22, batch 1300, loss[loss=0.1893, simple_loss=0.2524, pruned_loss=0.06314, over 4867.00 frames. ], tot_loss[loss=0.1684, simple_loss=0.2396, pruned_loss=0.04863, over 952728.01 frames. ], batch size: 34, lr: 3.15e-03, grad_scale: 32.0 2023-04-27 16:42:19,077 INFO [finetune.py:976] (4/7) Epoch 22, batch 1350, loss[loss=0.183, simple_loss=0.2376, pruned_loss=0.06417, over 4226.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2399, pruned_loss=0.04887, over 953022.73 frames. ], batch size: 18, lr: 3.15e-03, grad_scale: 32.0 2023-04-27 16:42:55,975 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=121659.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:42:58,252 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.014e+02 1.587e+02 1.863e+02 2.182e+02 3.847e+02, threshold=3.726e+02, percent-clipped=2.0 2023-04-27 16:42:58,995 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9836, 2.0450, 1.8413, 1.6916, 2.1497, 1.8152, 2.6805, 1.5027], device='cuda:4'), covar=tensor([0.3361, 0.1639, 0.4536, 0.2528, 0.1422, 0.2112, 0.1200, 0.4793], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0346, 0.0421, 0.0349, 0.0377, 0.0371, 0.0364, 0.0414], device='cuda:4'), out_proj_covar=tensor([9.9830e-05, 1.0354e-04, 1.2777e-04, 1.0514e-04, 1.1229e-04, 1.1053e-04, 1.0690e-04, 1.2505e-04], device='cuda:4') 2023-04-27 16:43:04,719 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=121667.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 16:43:19,351 INFO [finetune.py:976] (4/7) Epoch 22, batch 1400, loss[loss=0.1854, simple_loss=0.2761, pruned_loss=0.0473, over 4816.00 frames. ], tot_loss[loss=0.1722, simple_loss=0.2443, pruned_loss=0.05006, over 952807.86 frames. ], batch size: 51, lr: 3.15e-03, grad_scale: 32.0 2023-04-27 16:43:59,858 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=121707.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:44:01,583 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-27 16:44:24,395 INFO [finetune.py:976] (4/7) Epoch 22, batch 1450, loss[loss=0.161, simple_loss=0.2429, pruned_loss=0.03953, over 4863.00 frames. ], tot_loss[loss=0.1736, simple_loss=0.2462, pruned_loss=0.05055, over 953831.05 frames. ], batch size: 31, lr: 3.15e-03, grad_scale: 32.0 2023-04-27 16:45:04,292 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.075e+02 1.748e+02 1.956e+02 2.570e+02 3.972e+02, threshold=3.912e+02, percent-clipped=2.0 2023-04-27 16:45:04,383 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=121763.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:45:27,550 INFO [finetune.py:976] (4/7) Epoch 22, batch 1500, loss[loss=0.1807, simple_loss=0.2456, pruned_loss=0.05793, over 4787.00 frames. ], tot_loss[loss=0.1755, simple_loss=0.248, pruned_loss=0.05149, over 954479.97 frames. ], batch size: 29, lr: 3.15e-03, grad_scale: 32.0 2023-04-27 16:45:58,987 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=121820.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:46:06,252 INFO [finetune.py:976] (4/7) Epoch 22, batch 1550, loss[loss=0.2034, simple_loss=0.2649, pruned_loss=0.07093, over 4923.00 frames. ], tot_loss[loss=0.1742, simple_loss=0.2469, pruned_loss=0.0508, over 954782.97 frames. ], batch size: 33, lr: 3.15e-03, grad_scale: 32.0 2023-04-27 16:46:28,045 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.267e+01 1.607e+02 1.890e+02 2.272e+02 5.935e+02, threshold=3.780e+02, percent-clipped=2.0 2023-04-27 16:46:31,191 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=121868.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:46:39,811 INFO [finetune.py:976] (4/7) Epoch 22, batch 1600, loss[loss=0.1494, simple_loss=0.2234, pruned_loss=0.03773, over 4819.00 frames. ], tot_loss[loss=0.1722, simple_loss=0.2442, pruned_loss=0.05013, over 953274.98 frames. ], batch size: 30, lr: 3.15e-03, grad_scale: 32.0 2023-04-27 16:46:55,623 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.6400, 3.6022, 2.8214, 4.2602, 3.6021, 3.6665, 1.7178, 3.7055], device='cuda:4'), covar=tensor([0.1831, 0.1389, 0.3843, 0.1467, 0.3202, 0.1728, 0.6026, 0.2182], device='cuda:4'), in_proj_covar=tensor([0.0242, 0.0215, 0.0250, 0.0304, 0.0292, 0.0244, 0.0274, 0.0270], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 16:47:13,763 INFO [finetune.py:976] (4/7) Epoch 22, batch 1650, loss[loss=0.1482, simple_loss=0.2264, pruned_loss=0.03496, over 4933.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.2409, pruned_loss=0.04908, over 953877.40 frames. ], batch size: 33, lr: 3.15e-03, grad_scale: 32.0 2023-04-27 16:47:24,855 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=121950.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:47:33,996 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.216e+02 1.538e+02 1.825e+02 2.150e+02 3.786e+02, threshold=3.649e+02, percent-clipped=1.0 2023-04-27 16:47:38,074 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=121967.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 16:47:47,186 INFO [finetune.py:976] (4/7) Epoch 22, batch 1700, loss[loss=0.1747, simple_loss=0.2345, pruned_loss=0.05744, over 4062.00 frames. ], tot_loss[loss=0.1681, simple_loss=0.2389, pruned_loss=0.0486, over 953985.73 frames. ], batch size: 65, lr: 3.15e-03, grad_scale: 32.0 2023-04-27 16:47:56,667 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-04-27 16:48:06,912 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=122011.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:48:16,071 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=122015.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 16:48:37,955 INFO [finetune.py:976] (4/7) Epoch 22, batch 1750, loss[loss=0.1591, simple_loss=0.2257, pruned_loss=0.04622, over 4798.00 frames. ], tot_loss[loss=0.169, simple_loss=0.2403, pruned_loss=0.04886, over 955203.25 frames. ], batch size: 25, lr: 3.15e-03, grad_scale: 32.0 2023-04-27 16:48:47,516 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2861, 2.9399, 0.8577, 1.5952, 1.6204, 2.1662, 1.7503, 1.0292], device='cuda:4'), covar=tensor([0.1399, 0.1129, 0.1968, 0.1355, 0.1163, 0.0952, 0.1450, 0.1817], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0240, 0.0137, 0.0120, 0.0132, 0.0152, 0.0116, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 16:48:49,233 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-04-27 16:48:58,091 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5110, 1.6786, 0.8194, 1.2398, 1.7812, 1.3942, 1.3177, 1.3723], device='cuda:4'), covar=tensor([0.0517, 0.0378, 0.0367, 0.0584, 0.0288, 0.0536, 0.0520, 0.0593], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0028, 0.0019, 0.0028, 0.0028, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0045, 0.0038, 0.0052, 0.0038, 0.0050, 0.0050, 0.0051], device='cuda:4') 2023-04-27 16:49:13,881 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.158e+02 1.616e+02 1.845e+02 2.409e+02 4.396e+02, threshold=3.689e+02, percent-clipped=3.0 2023-04-27 16:49:19,185 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=122063.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:49:43,795 INFO [finetune.py:976] (4/7) Epoch 22, batch 1800, loss[loss=0.1607, simple_loss=0.2384, pruned_loss=0.04147, over 4826.00 frames. ], tot_loss[loss=0.1705, simple_loss=0.2431, pruned_loss=0.04889, over 956148.47 frames. ], batch size: 30, lr: 3.15e-03, grad_scale: 32.0 2023-04-27 16:50:18,109 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=122111.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:50:27,677 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=122117.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:50:48,451 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=2.01 vs. limit=2.0 2023-04-27 16:50:49,858 INFO [finetune.py:976] (4/7) Epoch 22, batch 1850, loss[loss=0.1619, simple_loss=0.2332, pruned_loss=0.0453, over 4769.00 frames. ], tot_loss[loss=0.1722, simple_loss=0.2445, pruned_loss=0.04995, over 953817.64 frames. ], batch size: 28, lr: 3.15e-03, grad_scale: 32.0 2023-04-27 16:51:14,576 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-27 16:51:22,417 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4448, 1.5311, 1.3798, 1.7751, 1.6327, 1.9601, 1.3957, 3.3888], device='cuda:4'), covar=tensor([0.0592, 0.0884, 0.0869, 0.1274, 0.0684, 0.0492, 0.0833, 0.0160], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 16:51:27,102 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.233e+01 1.753e+02 1.978e+02 2.407e+02 4.460e+02, threshold=3.956e+02, percent-clipped=4.0 2023-04-27 16:51:48,576 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=122178.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 16:51:56,334 INFO [finetune.py:976] (4/7) Epoch 22, batch 1900, loss[loss=0.1749, simple_loss=0.252, pruned_loss=0.0489, over 4843.00 frames. ], tot_loss[loss=0.1727, simple_loss=0.2455, pruned_loss=0.04993, over 955190.15 frames. ], batch size: 47, lr: 3.15e-03, grad_scale: 32.0 2023-04-27 16:52:05,469 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.53 vs. limit=2.0 2023-04-27 16:52:14,270 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=122195.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:52:19,468 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.74 vs. limit=2.0 2023-04-27 16:52:40,112 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-27 16:52:40,530 INFO [finetune.py:976] (4/7) Epoch 22, batch 1950, loss[loss=0.171, simple_loss=0.2412, pruned_loss=0.05042, over 4819.00 frames. ], tot_loss[loss=0.1715, simple_loss=0.2444, pruned_loss=0.04929, over 954867.54 frames. ], batch size: 33, lr: 3.15e-03, grad_scale: 32.0 2023-04-27 16:52:47,896 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7622, 1.5463, 1.7610, 2.1105, 2.0567, 1.7689, 1.3879, 1.8770], device='cuda:4'), covar=tensor([0.0771, 0.1196, 0.0769, 0.0552, 0.0607, 0.0826, 0.0815, 0.0576], device='cuda:4'), in_proj_covar=tensor([0.0185, 0.0199, 0.0182, 0.0172, 0.0176, 0.0178, 0.0150, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 16:52:55,277 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=122256.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:52:59,361 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.104e+02 1.614e+02 1.828e+02 2.204e+02 4.650e+02, threshold=3.656e+02, percent-clipped=1.0 2023-04-27 16:53:13,160 INFO [finetune.py:976] (4/7) Epoch 22, batch 2000, loss[loss=0.172, simple_loss=0.242, pruned_loss=0.05104, over 4827.00 frames. ], tot_loss[loss=0.1696, simple_loss=0.2422, pruned_loss=0.0485, over 955130.17 frames. ], batch size: 40, lr: 3.15e-03, grad_scale: 32.0 2023-04-27 16:53:17,389 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9366, 1.9928, 1.8434, 1.6287, 2.1329, 1.7340, 2.6745, 1.6490], device='cuda:4'), covar=tensor([0.3604, 0.1810, 0.4593, 0.2716, 0.1444, 0.2264, 0.1238, 0.4202], device='cuda:4'), in_proj_covar=tensor([0.0342, 0.0353, 0.0431, 0.0356, 0.0384, 0.0378, 0.0370, 0.0424], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 16:53:20,058 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-04-27 16:53:29,047 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=122306.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:53:32,418 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.69 vs. limit=2.0 2023-04-27 16:53:37,525 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=122319.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:53:38,729 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0829, 1.7586, 1.9782, 2.4008, 2.2733, 1.9914, 1.7268, 2.0457], device='cuda:4'), covar=tensor([0.0683, 0.1065, 0.0664, 0.0457, 0.0594, 0.0758, 0.0725, 0.0546], device='cuda:4'), in_proj_covar=tensor([0.0186, 0.0200, 0.0182, 0.0173, 0.0176, 0.0179, 0.0151, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 16:53:41,148 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6955, 1.5222, 1.7131, 2.0900, 2.0487, 1.6821, 1.3654, 1.7535], device='cuda:4'), covar=tensor([0.0818, 0.1277, 0.0853, 0.0587, 0.0598, 0.0873, 0.0809, 0.0679], device='cuda:4'), in_proj_covar=tensor([0.0186, 0.0200, 0.0182, 0.0173, 0.0176, 0.0179, 0.0151, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 16:53:46,291 INFO [finetune.py:976] (4/7) Epoch 22, batch 2050, loss[loss=0.1362, simple_loss=0.1983, pruned_loss=0.03706, over 4320.00 frames. ], tot_loss[loss=0.1673, simple_loss=0.2393, pruned_loss=0.04764, over 954905.19 frames. ], batch size: 18, lr: 3.15e-03, grad_scale: 32.0 2023-04-27 16:54:12,100 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.067e+02 1.559e+02 1.860e+02 2.262e+02 4.767e+02, threshold=3.720e+02, percent-clipped=3.0 2023-04-27 16:54:35,087 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=122380.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:54:36,186 INFO [finetune.py:976] (4/7) Epoch 22, batch 2100, loss[loss=0.2201, simple_loss=0.2805, pruned_loss=0.07987, over 4868.00 frames. ], tot_loss[loss=0.1681, simple_loss=0.2402, pruned_loss=0.048, over 954055.14 frames. ], batch size: 31, lr: 3.15e-03, grad_scale: 32.0 2023-04-27 16:55:09,786 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-27 16:55:40,950 INFO [finetune.py:976] (4/7) Epoch 22, batch 2150, loss[loss=0.1711, simple_loss=0.2384, pruned_loss=0.05196, over 4877.00 frames. ], tot_loss[loss=0.1702, simple_loss=0.2428, pruned_loss=0.04882, over 953809.46 frames. ], batch size: 34, lr: 3.15e-03, grad_scale: 32.0 2023-04-27 16:56:18,253 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6285, 1.7668, 2.0134, 2.0536, 1.7887, 1.8859, 2.0608, 2.0460], device='cuda:4'), covar=tensor([0.3743, 0.5714, 0.4820, 0.4602, 0.5779, 0.7323, 0.5341, 0.5108], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0372, 0.0323, 0.0337, 0.0346, 0.0393, 0.0355, 0.0329], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 16:56:19,329 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.125e+02 1.667e+02 1.982e+02 2.424e+02 1.094e+03, threshold=3.964e+02, percent-clipped=3.0 2023-04-27 16:56:27,944 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=122468.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:56:30,962 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=122473.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 16:56:33,168 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.09 vs. limit=2.0 2023-04-27 16:56:38,930 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.73 vs. limit=2.0 2023-04-27 16:56:42,043 INFO [finetune.py:976] (4/7) Epoch 22, batch 2200, loss[loss=0.2158, simple_loss=0.2835, pruned_loss=0.07402, over 4813.00 frames. ], tot_loss[loss=0.1724, simple_loss=0.245, pruned_loss=0.04988, over 952710.27 frames. ], batch size: 38, lr: 3.15e-03, grad_scale: 32.0 2023-04-27 16:57:48,039 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=122529.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:57:49,741 INFO [finetune.py:976] (4/7) Epoch 22, batch 2250, loss[loss=0.1506, simple_loss=0.2279, pruned_loss=0.03668, over 4777.00 frames. ], tot_loss[loss=0.1735, simple_loss=0.2465, pruned_loss=0.05021, over 953417.88 frames. ], batch size: 26, lr: 3.15e-03, grad_scale: 32.0 2023-04-27 16:58:20,295 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=122551.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 16:58:34,015 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.195e+02 1.620e+02 1.922e+02 2.285e+02 4.968e+02, threshold=3.845e+02, percent-clipped=1.0 2023-04-27 16:58:56,472 INFO [finetune.py:976] (4/7) Epoch 22, batch 2300, loss[loss=0.1354, simple_loss=0.2205, pruned_loss=0.02514, over 4760.00 frames. ], tot_loss[loss=0.1729, simple_loss=0.2464, pruned_loss=0.0497, over 952470.21 frames. ], batch size: 28, lr: 3.15e-03, grad_scale: 32.0 2023-04-27 16:59:36,000 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=122606.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:00:08,781 INFO [finetune.py:976] (4/7) Epoch 22, batch 2350, loss[loss=0.1468, simple_loss=0.2132, pruned_loss=0.04024, over 4891.00 frames. ], tot_loss[loss=0.1716, simple_loss=0.2447, pruned_loss=0.04928, over 954164.06 frames. ], batch size: 32, lr: 3.15e-03, grad_scale: 32.0 2023-04-27 17:00:19,519 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6548, 1.5112, 1.6749, 2.0384, 2.0003, 1.5904, 1.2986, 1.7146], device='cuda:4'), covar=tensor([0.0748, 0.1099, 0.0753, 0.0510, 0.0581, 0.0836, 0.0809, 0.0627], device='cuda:4'), in_proj_covar=tensor([0.0186, 0.0201, 0.0182, 0.0174, 0.0176, 0.0178, 0.0152, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 17:00:40,759 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=122654.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:00:45,807 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.58 vs. limit=2.0 2023-04-27 17:00:46,226 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.405e+01 1.589e+02 1.908e+02 2.329e+02 5.605e+02, threshold=3.816e+02, percent-clipped=4.0 2023-04-27 17:00:56,256 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-04-27 17:01:02,396 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1824, 1.6934, 2.0688, 2.1090, 2.0631, 1.6395, 1.2257, 1.7021], device='cuda:4'), covar=tensor([0.3108, 0.3002, 0.1630, 0.2251, 0.2300, 0.2618, 0.3969, 0.1862], device='cuda:4'), in_proj_covar=tensor([0.0294, 0.0245, 0.0226, 0.0314, 0.0220, 0.0233, 0.0228, 0.0183], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 17:01:04,136 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=122675.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:01:14,133 INFO [finetune.py:976] (4/7) Epoch 22, batch 2400, loss[loss=0.1522, simple_loss=0.2234, pruned_loss=0.04052, over 4845.00 frames. ], tot_loss[loss=0.1703, simple_loss=0.2426, pruned_loss=0.04896, over 955756.78 frames. ], batch size: 44, lr: 3.15e-03, grad_scale: 32.0 2023-04-27 17:01:17,401 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7453, 0.9670, 1.7206, 2.1275, 1.7926, 1.6908, 1.7247, 1.7249], device='cuda:4'), covar=tensor([0.4935, 0.6719, 0.6347, 0.6372, 0.6034, 0.7810, 0.8395, 0.8089], device='cuda:4'), in_proj_covar=tensor([0.0433, 0.0416, 0.0510, 0.0509, 0.0461, 0.0492, 0.0500, 0.0507], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 17:02:21,392 INFO [finetune.py:976] (4/7) Epoch 22, batch 2450, loss[loss=0.1774, simple_loss=0.2466, pruned_loss=0.05407, over 4804.00 frames. ], tot_loss[loss=0.1681, simple_loss=0.2397, pruned_loss=0.04819, over 954679.47 frames. ], batch size: 51, lr: 3.14e-03, grad_scale: 32.0 2023-04-27 17:02:25,795 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2389, 1.6678, 2.0437, 2.2157, 2.0222, 1.6276, 1.1265, 1.6679], device='cuda:4'), covar=tensor([0.3094, 0.3196, 0.1545, 0.2282, 0.2564, 0.2577, 0.4234, 0.2074], device='cuda:4'), in_proj_covar=tensor([0.0294, 0.0245, 0.0226, 0.0314, 0.0220, 0.0234, 0.0228, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 17:02:42,827 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.071e+02 1.578e+02 1.755e+02 2.241e+02 3.458e+02, threshold=3.510e+02, percent-clipped=0.0 2023-04-27 17:02:48,966 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=122773.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 17:02:54,345 INFO [finetune.py:976] (4/7) Epoch 22, batch 2500, loss[loss=0.2539, simple_loss=0.3108, pruned_loss=0.09848, over 4870.00 frames. ], tot_loss[loss=0.1697, simple_loss=0.2411, pruned_loss=0.0491, over 955086.34 frames. ], batch size: 31, lr: 3.14e-03, grad_scale: 32.0 2023-04-27 17:02:55,100 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6573, 1.9169, 1.9847, 2.0347, 1.9032, 2.0294, 2.1035, 2.0402], device='cuda:4'), covar=tensor([0.3997, 0.5532, 0.4669, 0.4760, 0.6012, 0.7283, 0.5385, 0.5041], device='cuda:4'), in_proj_covar=tensor([0.0339, 0.0373, 0.0325, 0.0340, 0.0348, 0.0396, 0.0357, 0.0331], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 17:03:21,720 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=122821.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:03:23,545 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=122824.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:03:28,387 INFO [finetune.py:976] (4/7) Epoch 22, batch 2550, loss[loss=0.1743, simple_loss=0.2619, pruned_loss=0.04336, over 4741.00 frames. ], tot_loss[loss=0.1713, simple_loss=0.2436, pruned_loss=0.04952, over 956100.70 frames. ], batch size: 54, lr: 3.14e-03, grad_scale: 32.0 2023-04-27 17:03:40,595 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=122851.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:03:48,778 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.124e+02 1.572e+02 1.988e+02 2.306e+02 4.105e+02, threshold=3.976e+02, percent-clipped=1.0 2023-04-27 17:03:50,145 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2818, 2.8636, 2.3052, 2.8132, 2.0544, 2.5406, 2.7591, 1.9381], device='cuda:4'), covar=tensor([0.1884, 0.0928, 0.0770, 0.0982, 0.3127, 0.1065, 0.1760, 0.2693], device='cuda:4'), in_proj_covar=tensor([0.0283, 0.0301, 0.0217, 0.0276, 0.0314, 0.0255, 0.0248, 0.0262], device='cuda:4'), out_proj_covar=tensor([1.1338e-04, 1.1942e-04, 8.5567e-05, 1.0916e-04, 1.2733e-04, 1.0071e-04, 9.9885e-05, 1.0350e-04], device='cuda:4') 2023-04-27 17:04:01,407 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=122872.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:04:13,219 INFO [finetune.py:976] (4/7) Epoch 22, batch 2600, loss[loss=0.1431, simple_loss=0.2137, pruned_loss=0.03623, over 4749.00 frames. ], tot_loss[loss=0.1723, simple_loss=0.2453, pruned_loss=0.04962, over 955887.81 frames. ], batch size: 28, lr: 3.14e-03, grad_scale: 32.0 2023-04-27 17:04:34,932 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=122899.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:04:44,487 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2736, 1.7324, 1.5457, 2.0167, 1.8636, 2.0710, 1.5759, 4.3679], device='cuda:4'), covar=tensor([0.0576, 0.0787, 0.0808, 0.1168, 0.0664, 0.0506, 0.0723, 0.0097], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 17:05:19,786 INFO [finetune.py:976] (4/7) Epoch 22, batch 2650, loss[loss=0.1677, simple_loss=0.2522, pruned_loss=0.04161, over 4826.00 frames. ], tot_loss[loss=0.1718, simple_loss=0.245, pruned_loss=0.04932, over 953029.04 frames. ], batch size: 30, lr: 3.14e-03, grad_scale: 32.0 2023-04-27 17:05:20,521 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=122933.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:06:00,528 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.522e+02 1.838e+02 2.219e+02 4.134e+02, threshold=3.677e+02, percent-clipped=1.0 2023-04-27 17:06:01,813 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1968, 1.7209, 1.6048, 1.9982, 1.8267, 2.1261, 1.5602, 4.4028], device='cuda:4'), covar=tensor([0.0563, 0.0742, 0.0766, 0.1169, 0.0631, 0.0497, 0.0727, 0.0076], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 17:06:15,881 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=122975.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:06:24,437 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6646, 0.7338, 1.5300, 2.0257, 1.7359, 1.5437, 1.5861, 1.5734], device='cuda:4'), covar=tensor([0.3920, 0.5609, 0.5027, 0.5173, 0.5101, 0.6201, 0.6479, 0.7645], device='cuda:4'), in_proj_covar=tensor([0.0431, 0.0415, 0.0509, 0.0508, 0.0460, 0.0491, 0.0499, 0.0507], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 17:06:26,131 INFO [finetune.py:976] (4/7) Epoch 22, batch 2700, loss[loss=0.1639, simple_loss=0.2388, pruned_loss=0.04448, over 4859.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.243, pruned_loss=0.04792, over 951937.23 frames. ], batch size: 31, lr: 3.14e-03, grad_scale: 32.0 2023-04-27 17:06:40,959 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.88 vs. limit=2.0 2023-04-27 17:06:56,566 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=123023.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:07:02,524 INFO [finetune.py:976] (4/7) Epoch 22, batch 2750, loss[loss=0.1439, simple_loss=0.2206, pruned_loss=0.03365, over 4847.00 frames. ], tot_loss[loss=0.1683, simple_loss=0.2413, pruned_loss=0.04762, over 951668.25 frames. ], batch size: 44, lr: 3.14e-03, grad_scale: 32.0 2023-04-27 17:07:05,069 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2436, 1.5149, 1.3032, 1.5174, 1.3324, 1.3613, 1.4159, 1.0147], device='cuda:4'), covar=tensor([0.1690, 0.1156, 0.0903, 0.1195, 0.3240, 0.1094, 0.1577, 0.2231], device='cuda:4'), in_proj_covar=tensor([0.0286, 0.0304, 0.0219, 0.0279, 0.0316, 0.0257, 0.0250, 0.0264], device='cuda:4'), out_proj_covar=tensor([1.1436e-04, 1.2051e-04, 8.6384e-05, 1.1027e-04, 1.2810e-04, 1.0160e-04, 1.0079e-04, 1.0449e-04], device='cuda:4') 2023-04-27 17:07:08,724 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9457, 1.9446, 1.8688, 1.5965, 2.1169, 1.7814, 2.6756, 1.6126], device='cuda:4'), covar=tensor([0.2968, 0.1591, 0.4194, 0.2548, 0.1490, 0.2095, 0.1113, 0.4174], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0351, 0.0424, 0.0355, 0.0383, 0.0377, 0.0370, 0.0420], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 17:07:28,110 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.026e+01 1.558e+02 1.841e+02 2.200e+02 4.189e+02, threshold=3.681e+02, percent-clipped=2.0 2023-04-27 17:07:42,808 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.4769, 1.3684, 1.4108, 1.0445, 1.3134, 1.1859, 1.6337, 1.2681], device='cuda:4'), covar=tensor([0.3335, 0.1736, 0.5071, 0.2550, 0.1677, 0.2259, 0.1567, 0.4752], device='cuda:4'), in_proj_covar=tensor([0.0341, 0.0352, 0.0426, 0.0356, 0.0384, 0.0378, 0.0371, 0.0421], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 17:07:53,159 INFO [finetune.py:976] (4/7) Epoch 22, batch 2800, loss[loss=0.1583, simple_loss=0.2397, pruned_loss=0.03849, over 4792.00 frames. ], tot_loss[loss=0.1672, simple_loss=0.2395, pruned_loss=0.04751, over 952978.31 frames. ], batch size: 26, lr: 3.14e-03, grad_scale: 32.0 2023-04-27 17:08:48,178 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=123124.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:08:53,534 INFO [finetune.py:976] (4/7) Epoch 22, batch 2850, loss[loss=0.2116, simple_loss=0.2775, pruned_loss=0.07289, over 4811.00 frames. ], tot_loss[loss=0.1664, simple_loss=0.2379, pruned_loss=0.04746, over 952741.26 frames. ], batch size: 39, lr: 3.14e-03, grad_scale: 32.0 2023-04-27 17:09:12,897 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.070e+02 1.588e+02 1.846e+02 2.327e+02 3.843e+02, threshold=3.691e+02, percent-clipped=1.0 2023-04-27 17:09:19,397 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=123172.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:09:27,500 INFO [finetune.py:976] (4/7) Epoch 22, batch 2900, loss[loss=0.1955, simple_loss=0.2787, pruned_loss=0.05615, over 4787.00 frames. ], tot_loss[loss=0.1702, simple_loss=0.2418, pruned_loss=0.04924, over 953790.77 frames. ], batch size: 29, lr: 3.14e-03, grad_scale: 32.0 2023-04-27 17:10:09,343 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=123228.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:10:11,741 INFO [finetune.py:976] (4/7) Epoch 22, batch 2950, loss[loss=0.2027, simple_loss=0.2826, pruned_loss=0.0614, over 4941.00 frames. ], tot_loss[loss=0.1731, simple_loss=0.2454, pruned_loss=0.05041, over 953400.93 frames. ], batch size: 39, lr: 3.14e-03, grad_scale: 32.0 2023-04-27 17:10:55,132 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.161e+02 1.651e+02 2.036e+02 2.473e+02 7.289e+02, threshold=4.071e+02, percent-clipped=4.0 2023-04-27 17:11:18,372 INFO [finetune.py:976] (4/7) Epoch 22, batch 3000, loss[loss=0.206, simple_loss=0.2774, pruned_loss=0.06726, over 4889.00 frames. ], tot_loss[loss=0.1743, simple_loss=0.2471, pruned_loss=0.05077, over 954374.45 frames. ], batch size: 35, lr: 3.14e-03, grad_scale: 32.0 2023-04-27 17:11:18,372 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 17:11:20,598 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3163, 1.4741, 1.7780, 1.9059, 1.8426, 1.9927, 1.7945, 1.8489], device='cuda:4'), covar=tensor([0.3549, 0.5375, 0.4772, 0.4627, 0.5682, 0.7150, 0.5444, 0.4739], device='cuda:4'), in_proj_covar=tensor([0.0339, 0.0373, 0.0325, 0.0340, 0.0347, 0.0395, 0.0357, 0.0331], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 17:11:25,756 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1600, 2.4520, 1.0371, 1.3437, 2.1138, 1.2808, 3.0082, 1.7320], device='cuda:4'), covar=tensor([0.0639, 0.0548, 0.0801, 0.1256, 0.0432, 0.0957, 0.0277, 0.0589], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0073, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 17:11:29,117 INFO [finetune.py:1010] (4/7) Epoch 22, validation: loss=0.1537, simple_loss=0.2227, pruned_loss=0.04237, over 2265189.00 frames. 2023-04-27 17:11:29,117 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-27 17:11:52,874 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2407, 3.0024, 2.1641, 2.4410, 1.5337, 1.5528, 2.3265, 1.5215], device='cuda:4'), covar=tensor([0.1620, 0.1372, 0.1482, 0.1645, 0.2442, 0.2118, 0.0964, 0.2134], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0211, 0.0168, 0.0204, 0.0200, 0.0186, 0.0156, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 17:12:11,398 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=123314.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:12:33,517 INFO [finetune.py:976] (4/7) Epoch 22, batch 3050, loss[loss=0.1911, simple_loss=0.2564, pruned_loss=0.06288, over 4822.00 frames. ], tot_loss[loss=0.1741, simple_loss=0.2472, pruned_loss=0.05047, over 954403.43 frames. ], batch size: 33, lr: 3.14e-03, grad_scale: 32.0 2023-04-27 17:12:51,654 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0508, 2.9072, 2.0292, 2.2229, 1.5109, 1.5092, 2.1816, 1.4788], device='cuda:4'), covar=tensor([0.1581, 0.1251, 0.1444, 0.1637, 0.2415, 0.1956, 0.0973, 0.1979], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0213, 0.0170, 0.0205, 0.0201, 0.0187, 0.0156, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 17:13:01,592 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.350e+01 1.609e+02 1.855e+02 2.357e+02 3.763e+02, threshold=3.710e+02, percent-clipped=0.0 2023-04-27 17:13:03,079 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.68 vs. limit=5.0 2023-04-27 17:13:14,329 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=123375.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:13:24,039 INFO [finetune.py:976] (4/7) Epoch 22, batch 3100, loss[loss=0.1349, simple_loss=0.2066, pruned_loss=0.03156, over 4726.00 frames. ], tot_loss[loss=0.1721, simple_loss=0.2449, pruned_loss=0.0497, over 954245.34 frames. ], batch size: 23, lr: 3.14e-03, grad_scale: 32.0 2023-04-27 17:14:29,432 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=123428.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:14:31,788 INFO [finetune.py:976] (4/7) Epoch 22, batch 3150, loss[loss=0.1505, simple_loss=0.2217, pruned_loss=0.03966, over 4837.00 frames. ], tot_loss[loss=0.1709, simple_loss=0.2428, pruned_loss=0.04947, over 954509.39 frames. ], batch size: 47, lr: 3.14e-03, grad_scale: 32.0 2023-04-27 17:14:53,789 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.14 vs. limit=5.0 2023-04-27 17:15:16,033 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.001e+02 1.572e+02 1.779e+02 2.164e+02 3.617e+02, threshold=3.558e+02, percent-clipped=0.0 2023-04-27 17:15:24,072 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8559, 2.1611, 1.8360, 1.5328, 1.4168, 1.4435, 1.8188, 1.3670], device='cuda:4'), covar=tensor([0.1691, 0.1305, 0.1424, 0.1746, 0.2242, 0.1958, 0.1007, 0.2014], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0212, 0.0169, 0.0204, 0.0200, 0.0186, 0.0156, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 17:15:38,621 INFO [finetune.py:976] (4/7) Epoch 22, batch 3200, loss[loss=0.1327, simple_loss=0.2147, pruned_loss=0.02536, over 4773.00 frames. ], tot_loss[loss=0.1692, simple_loss=0.2403, pruned_loss=0.04911, over 957459.77 frames. ], batch size: 28, lr: 3.14e-03, grad_scale: 32.0 2023-04-27 17:15:48,873 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=123489.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:15:50,237 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-27 17:15:50,243 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.71 vs. limit=5.0 2023-04-27 17:15:56,409 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1727, 1.8081, 2.0895, 2.4753, 2.4229, 2.0020, 1.7007, 2.0666], device='cuda:4'), covar=tensor([0.0739, 0.1097, 0.0674, 0.0497, 0.0578, 0.0857, 0.0755, 0.0596], device='cuda:4'), in_proj_covar=tensor([0.0185, 0.0199, 0.0182, 0.0173, 0.0175, 0.0178, 0.0151, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 17:16:43,113 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=123528.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:16:51,254 INFO [finetune.py:976] (4/7) Epoch 22, batch 3250, loss[loss=0.2049, simple_loss=0.2672, pruned_loss=0.07129, over 4797.00 frames. ], tot_loss[loss=0.1708, simple_loss=0.2415, pruned_loss=0.05, over 956475.48 frames. ], batch size: 29, lr: 3.14e-03, grad_scale: 64.0 2023-04-27 17:17:00,995 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.8676, 2.8645, 2.1894, 3.2554, 2.8953, 2.8310, 1.0880, 2.8076], device='cuda:4'), covar=tensor([0.2479, 0.1861, 0.3752, 0.3280, 0.3683, 0.2391, 0.6240, 0.3036], device='cuda:4'), in_proj_covar=tensor([0.0245, 0.0218, 0.0254, 0.0308, 0.0298, 0.0247, 0.0278, 0.0272], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 17:17:06,450 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.5219, 4.4423, 3.0746, 5.1601, 4.5877, 4.4638, 1.6737, 4.6416], device='cuda:4'), covar=tensor([0.1435, 0.1011, 0.3282, 0.1009, 0.4225, 0.1530, 0.6488, 0.1760], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0217, 0.0253, 0.0308, 0.0298, 0.0247, 0.0278, 0.0272], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 17:17:26,390 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.204e+02 1.572e+02 1.866e+02 2.251e+02 4.300e+02, threshold=3.733e+02, percent-clipped=4.0 2023-04-27 17:17:40,417 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=123576.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:17:44,044 INFO [finetune.py:976] (4/7) Epoch 22, batch 3300, loss[loss=0.1681, simple_loss=0.255, pruned_loss=0.04057, over 4817.00 frames. ], tot_loss[loss=0.1722, simple_loss=0.2443, pruned_loss=0.05005, over 957322.91 frames. ], batch size: 39, lr: 3.14e-03, grad_scale: 64.0 2023-04-27 17:18:01,819 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.10 vs. limit=2.0 2023-04-27 17:18:17,759 INFO [finetune.py:976] (4/7) Epoch 22, batch 3350, loss[loss=0.1734, simple_loss=0.251, pruned_loss=0.04793, over 4839.00 frames. ], tot_loss[loss=0.1733, simple_loss=0.2459, pruned_loss=0.05038, over 956976.44 frames. ], batch size: 47, lr: 3.14e-03, grad_scale: 64.0 2023-04-27 17:18:25,827 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2587, 1.7483, 2.1215, 2.7687, 2.1270, 1.6220, 1.5777, 2.0325], device='cuda:4'), covar=tensor([0.3440, 0.3208, 0.1693, 0.2162, 0.2838, 0.2814, 0.3836, 0.2111], device='cuda:4'), in_proj_covar=tensor([0.0294, 0.0245, 0.0226, 0.0315, 0.0220, 0.0233, 0.0228, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 17:18:40,221 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.125e+02 1.630e+02 1.835e+02 2.174e+02 3.522e+02, threshold=3.670e+02, percent-clipped=0.0 2023-04-27 17:18:43,916 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=123670.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:18:51,588 INFO [finetune.py:976] (4/7) Epoch 22, batch 3400, loss[loss=0.1607, simple_loss=0.2333, pruned_loss=0.04406, over 4912.00 frames. ], tot_loss[loss=0.1741, simple_loss=0.2468, pruned_loss=0.05069, over 957250.88 frames. ], batch size: 33, lr: 3.14e-03, grad_scale: 32.0 2023-04-27 17:19:25,127 INFO [finetune.py:976] (4/7) Epoch 22, batch 3450, loss[loss=0.1642, simple_loss=0.2399, pruned_loss=0.04428, over 4919.00 frames. ], tot_loss[loss=0.1747, simple_loss=0.2471, pruned_loss=0.05115, over 956692.53 frames. ], batch size: 38, lr: 3.14e-03, grad_scale: 32.0 2023-04-27 17:19:57,039 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.012e+02 1.610e+02 1.907e+02 2.404e+02 3.474e+02, threshold=3.815e+02, percent-clipped=0.0 2023-04-27 17:20:09,111 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2280, 2.8379, 2.1382, 2.2193, 1.6351, 1.6257, 2.3166, 1.6363], device='cuda:4'), covar=tensor([0.1579, 0.1432, 0.1408, 0.1675, 0.2172, 0.1948, 0.0922, 0.1935], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0212, 0.0169, 0.0205, 0.0201, 0.0186, 0.0156, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 17:20:21,186 INFO [finetune.py:976] (4/7) Epoch 22, batch 3500, loss[loss=0.1634, simple_loss=0.2304, pruned_loss=0.04817, over 4833.00 frames. ], tot_loss[loss=0.1721, simple_loss=0.2441, pruned_loss=0.0501, over 955893.30 frames. ], batch size: 30, lr: 3.14e-03, grad_scale: 32.0 2023-04-27 17:20:28,349 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=123784.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:21:22,865 INFO [finetune.py:976] (4/7) Epoch 22, batch 3550, loss[loss=0.1889, simple_loss=0.2484, pruned_loss=0.06472, over 4816.00 frames. ], tot_loss[loss=0.1706, simple_loss=0.2418, pruned_loss=0.04968, over 958485.48 frames. ], batch size: 41, lr: 3.14e-03, grad_scale: 32.0 2023-04-27 17:21:42,694 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.34 vs. limit=2.0 2023-04-27 17:21:43,413 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.018e+02 1.536e+02 1.822e+02 2.195e+02 3.918e+02, threshold=3.645e+02, percent-clipped=1.0 2023-04-27 17:21:49,910 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=123872.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:21:56,773 INFO [finetune.py:976] (4/7) Epoch 22, batch 3600, loss[loss=0.1482, simple_loss=0.2209, pruned_loss=0.03774, over 4758.00 frames. ], tot_loss[loss=0.1674, simple_loss=0.2383, pruned_loss=0.04822, over 956045.94 frames. ], batch size: 26, lr: 3.14e-03, grad_scale: 32.0 2023-04-27 17:22:22,609 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6967, 2.0438, 2.0812, 2.1971, 2.0302, 2.1095, 2.1842, 2.1200], device='cuda:4'), covar=tensor([0.3807, 0.5580, 0.4526, 0.4366, 0.5587, 0.7456, 0.5085, 0.4828], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0375, 0.0327, 0.0340, 0.0349, 0.0396, 0.0358, 0.0331], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 17:22:30,656 INFO [finetune.py:976] (4/7) Epoch 22, batch 3650, loss[loss=0.1878, simple_loss=0.2525, pruned_loss=0.06151, over 4893.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2398, pruned_loss=0.04898, over 956251.50 frames. ], batch size: 35, lr: 3.14e-03, grad_scale: 32.0 2023-04-27 17:22:31,423 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=123933.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:23:14,433 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.182e+02 1.652e+02 1.968e+02 2.235e+02 6.540e+02, threshold=3.937e+02, percent-clipped=5.0 2023-04-27 17:23:23,379 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=123970.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:23:37,656 INFO [finetune.py:976] (4/7) Epoch 22, batch 3700, loss[loss=0.1461, simple_loss=0.2281, pruned_loss=0.03207, over 4743.00 frames. ], tot_loss[loss=0.1709, simple_loss=0.2428, pruned_loss=0.04955, over 955718.66 frames. ], batch size: 27, lr: 3.14e-03, grad_scale: 32.0 2023-04-27 17:24:22,560 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=124018.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:24:44,186 INFO [finetune.py:976] (4/7) Epoch 22, batch 3750, loss[loss=0.2039, simple_loss=0.2784, pruned_loss=0.06464, over 4817.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2437, pruned_loss=0.04925, over 955471.54 frames. ], batch size: 39, lr: 3.13e-03, grad_scale: 32.0 2023-04-27 17:24:52,267 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=124036.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:25:08,768 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=124062.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:25:09,864 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.228e+02 1.543e+02 1.791e+02 2.127e+02 4.626e+02, threshold=3.581e+02, percent-clipped=2.0 2023-04-27 17:25:22,236 INFO [finetune.py:976] (4/7) Epoch 22, batch 3800, loss[loss=0.1713, simple_loss=0.2519, pruned_loss=0.04539, over 4815.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2441, pruned_loss=0.04901, over 957266.21 frames. ], batch size: 38, lr: 3.13e-03, grad_scale: 32.0 2023-04-27 17:25:24,010 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=124084.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:25:32,955 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=124097.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:25:43,283 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1711, 2.7614, 2.0429, 2.2004, 1.5466, 1.5763, 2.2097, 1.4893], device='cuda:4'), covar=tensor([0.1361, 0.1335, 0.1359, 0.1609, 0.2110, 0.1824, 0.0898, 0.1913], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0211, 0.0169, 0.0205, 0.0201, 0.0186, 0.0156, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 17:25:49,444 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=124123.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:25:56,296 INFO [finetune.py:976] (4/7) Epoch 22, batch 3850, loss[loss=0.142, simple_loss=0.22, pruned_loss=0.03199, over 4829.00 frames. ], tot_loss[loss=0.1696, simple_loss=0.2424, pruned_loss=0.04838, over 954676.61 frames. ], batch size: 30, lr: 3.13e-03, grad_scale: 32.0 2023-04-27 17:25:56,354 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=124132.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:26:17,752 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.098e+02 1.596e+02 1.902e+02 2.251e+02 4.955e+02, threshold=3.804e+02, percent-clipped=1.0 2023-04-27 17:26:20,938 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6957, 1.2562, 1.7817, 2.0738, 1.7071, 1.5950, 1.6729, 1.7195], device='cuda:4'), covar=tensor([0.5150, 0.7294, 0.6701, 0.7268, 0.6582, 0.9271, 0.8822, 0.9850], device='cuda:4'), in_proj_covar=tensor([0.0432, 0.0416, 0.0512, 0.0507, 0.0463, 0.0492, 0.0498, 0.0508], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 17:26:23,281 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7585, 1.8342, 0.9912, 1.4514, 1.9641, 1.6347, 1.5418, 1.6183], device='cuda:4'), covar=tensor([0.0473, 0.0368, 0.0326, 0.0536, 0.0246, 0.0489, 0.0511, 0.0559], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0028, 0.0019, 0.0028, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0045, 0.0038, 0.0052, 0.0038, 0.0050, 0.0050, 0.0051], device='cuda:4') 2023-04-27 17:26:34,190 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.54 vs. limit=2.0 2023-04-27 17:26:35,209 INFO [finetune.py:976] (4/7) Epoch 22, batch 3900, loss[loss=0.1598, simple_loss=0.2229, pruned_loss=0.04834, over 4872.00 frames. ], tot_loss[loss=0.169, simple_loss=0.2409, pruned_loss=0.04849, over 956344.84 frames. ], batch size: 34, lr: 3.13e-03, grad_scale: 32.0 2023-04-27 17:27:04,811 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.87 vs. limit=5.0 2023-04-27 17:27:08,040 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.10 vs. limit=2.0 2023-04-27 17:27:38,095 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=124228.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:27:40,957 INFO [finetune.py:976] (4/7) Epoch 22, batch 3950, loss[loss=0.1295, simple_loss=0.1984, pruned_loss=0.0303, over 4830.00 frames. ], tot_loss[loss=0.1672, simple_loss=0.2386, pruned_loss=0.04791, over 957200.20 frames. ], batch size: 25, lr: 3.13e-03, grad_scale: 32.0 2023-04-27 17:28:07,729 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.810e+01 1.371e+02 1.651e+02 2.165e+02 4.664e+02, threshold=3.302e+02, percent-clipped=1.0 2023-04-27 17:28:13,182 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=124272.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:28:18,022 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5945, 1.4440, 4.5002, 4.2650, 3.9354, 4.2421, 4.1771, 3.9721], device='cuda:4'), covar=tensor([0.6990, 0.5763, 0.1033, 0.1569, 0.1023, 0.1511, 0.1265, 0.1583], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0305, 0.0404, 0.0405, 0.0346, 0.0408, 0.0310, 0.0363], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 17:28:19,119 INFO [finetune.py:976] (4/7) Epoch 22, batch 4000, loss[loss=0.206, simple_loss=0.269, pruned_loss=0.07151, over 4816.00 frames. ], tot_loss[loss=0.1685, simple_loss=0.2391, pruned_loss=0.04897, over 956304.92 frames. ], batch size: 33, lr: 3.13e-03, grad_scale: 32.0 2023-04-27 17:28:52,762 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7056, 1.5448, 1.9916, 2.0681, 1.5258, 1.3229, 1.6501, 1.0054], device='cuda:4'), covar=tensor([0.0564, 0.0630, 0.0381, 0.0580, 0.0672, 0.0960, 0.0535, 0.0723], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0067, 0.0066, 0.0067, 0.0074, 0.0094, 0.0072, 0.0064], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 17:29:09,164 INFO [finetune.py:976] (4/7) Epoch 22, batch 4050, loss[loss=0.1781, simple_loss=0.2466, pruned_loss=0.05473, over 4744.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.242, pruned_loss=0.05014, over 955464.13 frames. ], batch size: 28, lr: 3.13e-03, grad_scale: 32.0 2023-04-27 17:29:09,902 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=124333.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:29:38,992 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=124348.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 17:29:54,263 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.048e+02 1.579e+02 1.902e+02 2.335e+02 4.180e+02, threshold=3.804e+02, percent-clipped=3.0 2023-04-27 17:30:17,168 INFO [finetune.py:976] (4/7) Epoch 22, batch 4100, loss[loss=0.1276, simple_loss=0.1993, pruned_loss=0.02793, over 4725.00 frames. ], tot_loss[loss=0.173, simple_loss=0.2447, pruned_loss=0.05065, over 955981.21 frames. ], batch size: 23, lr: 3.13e-03, grad_scale: 32.0 2023-04-27 17:30:34,616 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=124392.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:30:58,103 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=124409.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 17:31:09,015 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=124418.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:31:27,632 INFO [finetune.py:976] (4/7) Epoch 22, batch 4150, loss[loss=0.1745, simple_loss=0.2389, pruned_loss=0.05501, over 4064.00 frames. ], tot_loss[loss=0.1746, simple_loss=0.2468, pruned_loss=0.05125, over 955526.04 frames. ], batch size: 65, lr: 3.13e-03, grad_scale: 32.0 2023-04-27 17:31:27,747 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=124432.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:32:10,275 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.606e+01 1.564e+02 1.862e+02 2.379e+02 7.363e+02, threshold=3.724e+02, percent-clipped=1.0 2023-04-27 17:32:33,359 INFO [finetune.py:976] (4/7) Epoch 22, batch 4200, loss[loss=0.1573, simple_loss=0.2347, pruned_loss=0.04001, over 4742.00 frames. ], tot_loss[loss=0.1737, simple_loss=0.2467, pruned_loss=0.05038, over 955944.01 frames. ], batch size: 54, lr: 3.13e-03, grad_scale: 32.0 2023-04-27 17:32:33,470 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2884, 1.5306, 1.4839, 1.8028, 1.7504, 1.8395, 1.4635, 3.6123], device='cuda:4'), covar=tensor([0.0614, 0.0837, 0.0808, 0.1270, 0.0660, 0.0552, 0.0772, 0.0135], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 17:32:45,619 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=124493.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:32:47,956 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7245, 1.2819, 1.3695, 1.3883, 1.8448, 1.5476, 1.2473, 1.3223], device='cuda:4'), covar=tensor([0.1476, 0.1331, 0.1805, 0.1293, 0.0940, 0.1347, 0.2019, 0.2372], device='cuda:4'), in_proj_covar=tensor([0.0315, 0.0313, 0.0352, 0.0290, 0.0327, 0.0308, 0.0302, 0.0374], device='cuda:4'), out_proj_covar=tensor([6.4658e-05, 6.4650e-05, 7.4143e-05, 5.8521e-05, 6.7436e-05, 6.4587e-05, 6.3089e-05, 7.9434e-05], device='cuda:4') 2023-04-27 17:33:09,235 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=124528.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:33:11,581 INFO [finetune.py:976] (4/7) Epoch 22, batch 4250, loss[loss=0.1604, simple_loss=0.2309, pruned_loss=0.04497, over 4849.00 frames. ], tot_loss[loss=0.1724, simple_loss=0.2451, pruned_loss=0.04983, over 957319.44 frames. ], batch size: 49, lr: 3.13e-03, grad_scale: 32.0 2023-04-27 17:33:33,183 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.121e+02 1.562e+02 1.869e+02 2.245e+02 4.302e+02, threshold=3.738e+02, percent-clipped=2.0 2023-04-27 17:33:37,054 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7219, 2.1461, 1.8184, 2.1089, 1.6123, 1.7485, 1.7987, 1.3500], device='cuda:4'), covar=tensor([0.1678, 0.0925, 0.0841, 0.1025, 0.3187, 0.1104, 0.1763, 0.2152], device='cuda:4'), in_proj_covar=tensor([0.0287, 0.0303, 0.0219, 0.0279, 0.0318, 0.0259, 0.0252, 0.0265], device='cuda:4'), out_proj_covar=tensor([1.1479e-04, 1.1995e-04, 8.6577e-05, 1.1018e-04, 1.2863e-04, 1.0230e-04, 1.0147e-04, 1.0454e-04], device='cuda:4') 2023-04-27 17:33:40,671 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=124576.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:33:44,720 INFO [finetune.py:976] (4/7) Epoch 22, batch 4300, loss[loss=0.1552, simple_loss=0.2246, pruned_loss=0.04292, over 4910.00 frames. ], tot_loss[loss=0.1706, simple_loss=0.2427, pruned_loss=0.04921, over 955675.97 frames. ], batch size: 35, lr: 3.13e-03, grad_scale: 32.0 2023-04-27 17:34:10,658 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0422, 2.7253, 2.1225, 2.2168, 1.5137, 1.5124, 2.3753, 1.4594], device='cuda:4'), covar=tensor([0.1585, 0.1452, 0.1388, 0.1607, 0.2200, 0.1847, 0.0869, 0.1983], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0212, 0.0169, 0.0205, 0.0201, 0.0186, 0.0156, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 17:34:15,438 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=124628.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:34:17,826 INFO [finetune.py:976] (4/7) Epoch 22, batch 4350, loss[loss=0.1562, simple_loss=0.2291, pruned_loss=0.04168, over 4864.00 frames. ], tot_loss[loss=0.1684, simple_loss=0.2401, pruned_loss=0.04837, over 957401.32 frames. ], batch size: 31, lr: 3.13e-03, grad_scale: 32.0 2023-04-27 17:34:22,056 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=124638.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:34:38,879 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=124663.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:34:39,361 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.058e+02 1.483e+02 1.838e+02 2.190e+02 4.498e+02, threshold=3.677e+02, percent-clipped=2.0 2023-04-27 17:34:46,929 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.48 vs. limit=2.0 2023-04-27 17:34:51,252 INFO [finetune.py:976] (4/7) Epoch 22, batch 4400, loss[loss=0.2056, simple_loss=0.2804, pruned_loss=0.06538, over 4939.00 frames. ], tot_loss[loss=0.1698, simple_loss=0.2413, pruned_loss=0.04914, over 955860.50 frames. ], batch size: 39, lr: 3.13e-03, grad_scale: 32.0 2023-04-27 17:34:56,784 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4682, 1.5500, 4.0906, 3.8166, 3.6280, 3.9664, 3.8183, 3.5466], device='cuda:4'), covar=tensor([0.7605, 0.5734, 0.1243, 0.1972, 0.1297, 0.1457, 0.2338, 0.1792], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0304, 0.0406, 0.0406, 0.0347, 0.0408, 0.0311, 0.0364], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 17:34:58,058 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=124692.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:35:02,329 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=124699.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:35:11,387 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=124704.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 17:35:33,054 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=124718.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:35:33,774 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-04-27 17:35:42,950 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=124724.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:35:53,106 INFO [finetune.py:976] (4/7) Epoch 22, batch 4450, loss[loss=0.1871, simple_loss=0.2493, pruned_loss=0.06242, over 4828.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2434, pruned_loss=0.04945, over 956133.24 frames. ], batch size: 30, lr: 3.13e-03, grad_scale: 32.0 2023-04-27 17:36:03,832 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=124740.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:36:30,399 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.087e+02 1.602e+02 1.930e+02 2.289e+02 4.578e+02, threshold=3.860e+02, percent-clipped=3.0 2023-04-27 17:36:38,225 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=124766.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:37:00,153 INFO [finetune.py:976] (4/7) Epoch 22, batch 4500, loss[loss=0.1426, simple_loss=0.2236, pruned_loss=0.03075, over 4906.00 frames. ], tot_loss[loss=0.1723, simple_loss=0.2444, pruned_loss=0.05004, over 954199.88 frames. ], batch size: 37, lr: 3.13e-03, grad_scale: 32.0 2023-04-27 17:37:03,927 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=124788.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:37:47,350 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=124820.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:38:07,812 INFO [finetune.py:976] (4/7) Epoch 22, batch 4550, loss[loss=0.1948, simple_loss=0.2689, pruned_loss=0.06035, over 4812.00 frames. ], tot_loss[loss=0.1733, simple_loss=0.2457, pruned_loss=0.05047, over 954905.45 frames. ], batch size: 45, lr: 3.13e-03, grad_scale: 32.0 2023-04-27 17:38:40,619 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.24 vs. limit=5.0 2023-04-27 17:38:49,604 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.035e+02 1.528e+02 1.862e+02 2.235e+02 3.504e+02, threshold=3.725e+02, percent-clipped=0.0 2023-04-27 17:39:13,963 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=124881.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:39:14,477 INFO [finetune.py:976] (4/7) Epoch 22, batch 4600, loss[loss=0.1675, simple_loss=0.2483, pruned_loss=0.04335, over 4820.00 frames. ], tot_loss[loss=0.1719, simple_loss=0.2447, pruned_loss=0.04952, over 955432.43 frames. ], batch size: 40, lr: 3.13e-03, grad_scale: 32.0 2023-04-27 17:39:25,734 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.73 vs. limit=2.0 2023-04-27 17:39:57,020 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1939, 1.6694, 2.0231, 2.3771, 2.0285, 1.6455, 1.3193, 1.8000], device='cuda:4'), covar=tensor([0.2940, 0.2953, 0.1556, 0.2089, 0.2424, 0.2506, 0.3827, 0.1821], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0244, 0.0225, 0.0314, 0.0221, 0.0233, 0.0227, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 17:39:57,575 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9926, 2.5774, 1.1064, 1.4236, 1.9576, 1.1898, 2.9036, 1.4551], device='cuda:4'), covar=tensor([0.0713, 0.0595, 0.0734, 0.1051, 0.0424, 0.0927, 0.0235, 0.0614], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0050, 0.0051, 0.0073, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-27 17:40:00,430 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=124928.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:40:03,270 INFO [finetune.py:976] (4/7) Epoch 22, batch 4650, loss[loss=0.1535, simple_loss=0.2263, pruned_loss=0.04032, over 4872.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2414, pruned_loss=0.0481, over 954870.96 frames. ], batch size: 34, lr: 3.13e-03, grad_scale: 32.0 2023-04-27 17:40:23,402 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.522e+01 1.501e+02 1.822e+02 2.156e+02 6.565e+02, threshold=3.645e+02, percent-clipped=1.0 2023-04-27 17:40:31,672 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=124976.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:40:36,852 INFO [finetune.py:976] (4/7) Epoch 22, batch 4700, loss[loss=0.1576, simple_loss=0.2329, pruned_loss=0.04117, over 4821.00 frames. ], tot_loss[loss=0.1683, simple_loss=0.2401, pruned_loss=0.04828, over 954532.64 frames. ], batch size: 51, lr: 3.13e-03, grad_scale: 32.0 2023-04-27 17:40:44,820 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=124994.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:40:51,151 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=125004.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 17:40:57,855 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3402, 1.7537, 2.2408, 2.6676, 2.1826, 1.8042, 1.5957, 1.9838], device='cuda:4'), covar=tensor([0.3195, 0.3089, 0.1505, 0.2202, 0.2515, 0.2437, 0.3904, 0.2051], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0243, 0.0225, 0.0312, 0.0220, 0.0232, 0.0226, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 17:41:00,859 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=125019.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:41:10,030 INFO [finetune.py:976] (4/7) Epoch 22, batch 4750, loss[loss=0.2008, simple_loss=0.2645, pruned_loss=0.06859, over 4073.00 frames. ], tot_loss[loss=0.1661, simple_loss=0.2372, pruned_loss=0.04744, over 953438.89 frames. ], batch size: 65, lr: 3.13e-03, grad_scale: 32.0 2023-04-27 17:41:23,892 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=125052.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 17:41:31,621 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.684e+01 1.607e+02 1.875e+02 2.293e+02 4.178e+02, threshold=3.749e+02, percent-clipped=1.0 2023-04-27 17:41:43,190 INFO [finetune.py:976] (4/7) Epoch 22, batch 4800, loss[loss=0.1651, simple_loss=0.2398, pruned_loss=0.0452, over 4226.00 frames. ], tot_loss[loss=0.1684, simple_loss=0.2399, pruned_loss=0.04847, over 953582.33 frames. ], batch size: 65, lr: 3.13e-03, grad_scale: 32.0 2023-04-27 17:41:48,390 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=125088.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:42:14,801 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=125128.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:42:17,145 INFO [finetune.py:976] (4/7) Epoch 22, batch 4850, loss[loss=0.2053, simple_loss=0.2808, pruned_loss=0.06493, over 4835.00 frames. ], tot_loss[loss=0.1696, simple_loss=0.2417, pruned_loss=0.04871, over 955030.52 frames. ], batch size: 47, lr: 3.13e-03, grad_scale: 32.0 2023-04-27 17:42:17,230 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5982, 1.3590, 4.0823, 3.8321, 3.5741, 3.9029, 3.7583, 3.5527], device='cuda:4'), covar=tensor([0.7283, 0.5891, 0.1082, 0.1670, 0.1212, 0.1665, 0.2164, 0.1425], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0304, 0.0404, 0.0406, 0.0346, 0.0408, 0.0312, 0.0364], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 17:42:20,148 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=125136.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:42:44,288 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.162e+02 1.525e+02 1.734e+02 2.236e+02 4.357e+02, threshold=3.469e+02, percent-clipped=1.0 2023-04-27 17:43:03,308 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=125176.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:43:13,297 INFO [finetune.py:976] (4/7) Epoch 22, batch 4900, loss[loss=0.194, simple_loss=0.2755, pruned_loss=0.05626, over 4817.00 frames. ], tot_loss[loss=0.1725, simple_loss=0.2447, pruned_loss=0.0502, over 953132.78 frames. ], batch size: 41, lr: 3.13e-03, grad_scale: 32.0 2023-04-27 17:43:18,251 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=125189.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:44:14,013 INFO [finetune.py:976] (4/7) Epoch 22, batch 4950, loss[loss=0.1724, simple_loss=0.2427, pruned_loss=0.0511, over 4723.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2448, pruned_loss=0.04959, over 953561.29 frames. ], batch size: 23, lr: 3.13e-03, grad_scale: 32.0 2023-04-27 17:44:27,576 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-27 17:44:50,838 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 1.480e+02 1.752e+02 2.103e+02 4.760e+02, threshold=3.505e+02, percent-clipped=3.0 2023-04-27 17:45:13,456 INFO [finetune.py:976] (4/7) Epoch 22, batch 5000, loss[loss=0.1656, simple_loss=0.2284, pruned_loss=0.05144, over 4896.00 frames. ], tot_loss[loss=0.1713, simple_loss=0.2438, pruned_loss=0.04939, over 955453.54 frames. ], batch size: 32, lr: 3.13e-03, grad_scale: 16.0 2023-04-27 17:45:33,323 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=125294.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:45:33,374 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7790, 1.4717, 1.4234, 1.5814, 1.9706, 1.6134, 1.3700, 1.3389], device='cuda:4'), covar=tensor([0.1557, 0.1287, 0.1907, 0.1454, 0.0808, 0.1485, 0.1820, 0.2124], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0308, 0.0348, 0.0285, 0.0323, 0.0304, 0.0297, 0.0370], device='cuda:4'), out_proj_covar=tensor([6.3642e-05, 6.3625e-05, 7.3283e-05, 5.7328e-05, 6.6611e-05, 6.3685e-05, 6.2061e-05, 7.8465e-05], device='cuda:4') 2023-04-27 17:46:05,805 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=125319.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:46:07,027 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6724, 1.7291, 0.7375, 1.3683, 1.7531, 1.5248, 1.4267, 1.5683], device='cuda:4'), covar=tensor([0.0486, 0.0331, 0.0335, 0.0511, 0.0265, 0.0456, 0.0451, 0.0563], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0028, 0.0019, 0.0028, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0045, 0.0038, 0.0052, 0.0038, 0.0050, 0.0050, 0.0051], device='cuda:4') 2023-04-27 17:46:18,222 INFO [finetune.py:976] (4/7) Epoch 22, batch 5050, loss[loss=0.2078, simple_loss=0.2755, pruned_loss=0.07008, over 4907.00 frames. ], tot_loss[loss=0.1701, simple_loss=0.2418, pruned_loss=0.04917, over 956765.13 frames. ], batch size: 36, lr: 3.13e-03, grad_scale: 16.0 2023-04-27 17:46:32,850 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=125342.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:46:43,477 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=125356.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:46:48,791 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.011e+02 1.666e+02 1.966e+02 2.426e+02 4.294e+02, threshold=3.932e+02, percent-clipped=5.0 2023-04-27 17:46:50,094 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=125367.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:46:55,667 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=125376.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:46:59,446 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.53 vs. limit=5.0 2023-04-27 17:46:59,779 INFO [finetune.py:976] (4/7) Epoch 22, batch 5100, loss[loss=0.1471, simple_loss=0.2079, pruned_loss=0.04311, over 4772.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.2389, pruned_loss=0.04826, over 956179.73 frames. ], batch size: 54, lr: 3.12e-03, grad_scale: 16.0 2023-04-27 17:47:00,625 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-04-27 17:47:24,384 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=125417.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 17:47:26,790 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=125421.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:47:29,861 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7633, 2.1941, 1.8561, 1.5770, 1.3409, 1.3537, 1.7985, 1.2436], device='cuda:4'), covar=tensor([0.1624, 0.1332, 0.1471, 0.1777, 0.2356, 0.1956, 0.0996, 0.2117], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0211, 0.0169, 0.0203, 0.0200, 0.0185, 0.0156, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 17:47:33,404 INFO [finetune.py:976] (4/7) Epoch 22, batch 5150, loss[loss=0.1855, simple_loss=0.2494, pruned_loss=0.06078, over 4914.00 frames. ], tot_loss[loss=0.1684, simple_loss=0.2394, pruned_loss=0.0487, over 956832.44 frames. ], batch size: 37, lr: 3.12e-03, grad_scale: 16.0 2023-04-27 17:47:37,061 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=125437.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:48:02,908 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.720e+01 1.588e+02 1.846e+02 2.174e+02 3.765e+02, threshold=3.692e+02, percent-clipped=0.0 2023-04-27 17:48:09,691 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=125476.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:48:13,158 INFO [finetune.py:976] (4/7) Epoch 22, batch 5200, loss[loss=0.1399, simple_loss=0.2202, pruned_loss=0.02984, over 4836.00 frames. ], tot_loss[loss=0.1712, simple_loss=0.2428, pruned_loss=0.04975, over 957913.74 frames. ], batch size: 33, lr: 3.12e-03, grad_scale: 16.0 2023-04-27 17:48:13,280 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=125482.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:48:14,967 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=125484.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:48:40,645 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5143, 3.2240, 2.6773, 2.7311, 1.9478, 1.9542, 2.8857, 1.8626], device='cuda:4'), covar=tensor([0.1550, 0.1355, 0.1257, 0.1507, 0.2202, 0.1858, 0.0824, 0.1945], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0210, 0.0168, 0.0203, 0.0199, 0.0184, 0.0155, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 17:48:42,433 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=125524.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:48:47,214 INFO [finetune.py:976] (4/7) Epoch 22, batch 5250, loss[loss=0.1767, simple_loss=0.2527, pruned_loss=0.05038, over 4768.00 frames. ], tot_loss[loss=0.1726, simple_loss=0.2445, pruned_loss=0.05029, over 958052.93 frames. ], batch size: 28, lr: 3.12e-03, grad_scale: 16.0 2023-04-27 17:48:58,220 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.7220, 4.5834, 2.9433, 5.3779, 4.7472, 4.6677, 1.9717, 4.5948], device='cuda:4'), covar=tensor([0.1568, 0.1012, 0.3642, 0.0878, 0.2759, 0.1833, 0.5430, 0.2298], device='cuda:4'), in_proj_covar=tensor([0.0243, 0.0216, 0.0250, 0.0304, 0.0295, 0.0245, 0.0274, 0.0269], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 17:49:09,788 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.141e+02 1.600e+02 1.860e+02 2.252e+02 4.468e+02, threshold=3.721e+02, percent-clipped=2.0 2023-04-27 17:49:20,592 INFO [finetune.py:976] (4/7) Epoch 22, batch 5300, loss[loss=0.1872, simple_loss=0.2565, pruned_loss=0.05898, over 4727.00 frames. ], tot_loss[loss=0.1732, simple_loss=0.2455, pruned_loss=0.05043, over 956528.02 frames. ], batch size: 23, lr: 3.12e-03, grad_scale: 16.0 2023-04-27 17:49:34,599 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=125603.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:49:54,074 INFO [finetune.py:976] (4/7) Epoch 22, batch 5350, loss[loss=0.1566, simple_loss=0.2313, pruned_loss=0.04096, over 4896.00 frames. ], tot_loss[loss=0.1727, simple_loss=0.2456, pruned_loss=0.04986, over 955307.01 frames. ], batch size: 37, lr: 3.12e-03, grad_scale: 16.0 2023-04-27 17:49:55,519 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-04-27 17:50:15,156 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=125664.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:50:16,105 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.320e+01 1.581e+02 1.868e+02 2.376e+02 4.466e+02, threshold=3.736e+02, percent-clipped=2.0 2023-04-27 17:50:38,573 INFO [finetune.py:976] (4/7) Epoch 22, batch 5400, loss[loss=0.1881, simple_loss=0.2555, pruned_loss=0.06034, over 4916.00 frames. ], tot_loss[loss=0.171, simple_loss=0.2434, pruned_loss=0.04933, over 954960.45 frames. ], batch size: 37, lr: 3.12e-03, grad_scale: 16.0 2023-04-27 17:51:00,774 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4035, 1.7881, 1.7526, 2.2966, 2.4522, 2.0196, 1.9815, 1.7844], device='cuda:4'), covar=tensor([0.1509, 0.1595, 0.1694, 0.1334, 0.1133, 0.1660, 0.1852, 0.2085], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0308, 0.0349, 0.0286, 0.0324, 0.0305, 0.0298, 0.0369], device='cuda:4'), out_proj_covar=tensor([6.3585e-05, 6.3720e-05, 7.3527e-05, 5.7440e-05, 6.6721e-05, 6.3839e-05, 6.2269e-05, 7.8363e-05], device='cuda:4') 2023-04-27 17:51:19,941 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=125712.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 17:51:40,613 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=125725.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:51:45,247 INFO [finetune.py:976] (4/7) Epoch 22, batch 5450, loss[loss=0.1715, simple_loss=0.2484, pruned_loss=0.04733, over 4820.00 frames. ], tot_loss[loss=0.1687, simple_loss=0.2403, pruned_loss=0.04855, over 955491.93 frames. ], batch size: 41, lr: 3.12e-03, grad_scale: 16.0 2023-04-27 17:51:45,331 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=125732.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:52:28,295 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.173e+01 1.432e+02 1.653e+02 1.978e+02 3.483e+02, threshold=3.306e+02, percent-clipped=0.0 2023-04-27 17:52:47,950 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=125777.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:52:51,412 INFO [finetune.py:976] (4/7) Epoch 22, batch 5500, loss[loss=0.1825, simple_loss=0.2483, pruned_loss=0.05836, over 4843.00 frames. ], tot_loss[loss=0.1657, simple_loss=0.2367, pruned_loss=0.04729, over 954876.68 frames. ], batch size: 47, lr: 3.12e-03, grad_scale: 16.0 2023-04-27 17:52:58,404 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=125784.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:52:59,618 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=125786.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:53:57,320 INFO [finetune.py:976] (4/7) Epoch 22, batch 5550, loss[loss=0.2741, simple_loss=0.3423, pruned_loss=0.103, over 4092.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.2385, pruned_loss=0.04827, over 952101.60 frames. ], batch size: 65, lr: 3.12e-03, grad_scale: 16.0 2023-04-27 17:54:03,350 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=125832.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:54:06,527 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-27 17:54:41,116 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.131e+02 1.577e+02 1.939e+02 2.274e+02 3.783e+02, threshold=3.878e+02, percent-clipped=3.0 2023-04-27 17:54:57,561 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2328, 2.8345, 1.2047, 1.5326, 2.3364, 1.3203, 3.7489, 1.9057], device='cuda:4'), covar=tensor([0.0708, 0.0660, 0.0821, 0.1360, 0.0498, 0.1092, 0.0301, 0.0642], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0065, 0.0048, 0.0046, 0.0050, 0.0052, 0.0074, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 17:55:02,079 INFO [finetune.py:976] (4/7) Epoch 22, batch 5600, loss[loss=0.1855, simple_loss=0.2789, pruned_loss=0.04599, over 4815.00 frames. ], tot_loss[loss=0.1697, simple_loss=0.2413, pruned_loss=0.049, over 949793.44 frames. ], batch size: 40, lr: 3.12e-03, grad_scale: 16.0 2023-04-27 17:55:24,417 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7099, 2.3771, 1.8101, 1.8196, 1.3010, 1.3494, 1.8850, 1.2743], device='cuda:4'), covar=tensor([0.1838, 0.1520, 0.1589, 0.1855, 0.2529, 0.2187, 0.1033, 0.2268], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0212, 0.0170, 0.0205, 0.0201, 0.0186, 0.0156, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 17:56:05,136 INFO [finetune.py:976] (4/7) Epoch 22, batch 5650, loss[loss=0.1951, simple_loss=0.2771, pruned_loss=0.05652, over 4871.00 frames. ], tot_loss[loss=0.1707, simple_loss=0.2437, pruned_loss=0.04883, over 950732.55 frames. ], batch size: 44, lr: 3.12e-03, grad_scale: 16.0 2023-04-27 17:56:31,893 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=125959.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:56:35,157 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.85 vs. limit=2.0 2023-04-27 17:56:35,408 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.900e+01 1.470e+02 1.804e+02 2.167e+02 3.376e+02, threshold=3.608e+02, percent-clipped=0.0 2023-04-27 17:56:36,116 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5122, 2.7353, 2.3987, 2.7702, 2.1213, 2.4706, 2.4795, 2.0258], device='cuda:4'), covar=tensor([0.1473, 0.1267, 0.0729, 0.1138, 0.2976, 0.0937, 0.1834, 0.2326], device='cuda:4'), in_proj_covar=tensor([0.0286, 0.0301, 0.0217, 0.0278, 0.0316, 0.0257, 0.0250, 0.0263], device='cuda:4'), out_proj_covar=tensor([1.1414e-04, 1.1914e-04, 8.5611e-05, 1.0957e-04, 1.2804e-04, 1.0145e-04, 1.0070e-04, 1.0380e-04], device='cuda:4') 2023-04-27 17:56:45,404 INFO [finetune.py:976] (4/7) Epoch 22, batch 5700, loss[loss=0.1551, simple_loss=0.2094, pruned_loss=0.05038, over 4050.00 frames. ], tot_loss[loss=0.168, simple_loss=0.2398, pruned_loss=0.04816, over 934423.82 frames. ], batch size: 17, lr: 3.12e-03, grad_scale: 16.0 2023-04-27 17:56:54,942 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=125998.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:57:25,282 INFO [finetune.py:976] (4/7) Epoch 23, batch 0, loss[loss=0.1712, simple_loss=0.2459, pruned_loss=0.04824, over 4916.00 frames. ], tot_loss[loss=0.1712, simple_loss=0.2459, pruned_loss=0.04824, over 4916.00 frames. ], batch size: 33, lr: 3.12e-03, grad_scale: 16.0 2023-04-27 17:57:25,282 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 17:57:29,682 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5536, 1.2299, 1.4035, 1.3455, 1.7324, 1.4593, 1.2549, 1.3512], device='cuda:4'), covar=tensor([0.1954, 0.1916, 0.2485, 0.1702, 0.1492, 0.1969, 0.2151, 0.2980], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0309, 0.0350, 0.0285, 0.0324, 0.0306, 0.0298, 0.0370], device='cuda:4'), out_proj_covar=tensor([6.3892e-05, 6.3932e-05, 7.3722e-05, 5.7300e-05, 6.6699e-05, 6.4144e-05, 6.2271e-05, 7.8590e-05], device='cuda:4') 2023-04-27 17:57:31,453 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3421, 1.2785, 3.8679, 3.5758, 3.4459, 3.7441, 3.7911, 3.3873], device='cuda:4'), covar=tensor([0.6939, 0.5135, 0.1216, 0.1986, 0.1255, 0.1320, 0.0717, 0.1583], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0305, 0.0404, 0.0405, 0.0346, 0.0408, 0.0313, 0.0364], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 17:57:33,337 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3661, 1.2577, 3.8358, 3.5029, 3.4303, 3.7374, 3.7996, 3.3767], device='cuda:4'), covar=tensor([0.7133, 0.5411, 0.1240, 0.2152, 0.1278, 0.1362, 0.0698, 0.1656], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0305, 0.0404, 0.0405, 0.0346, 0.0408, 0.0313, 0.0364], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 17:57:46,919 INFO [finetune.py:1010] (4/7) Epoch 23, validation: loss=0.1552, simple_loss=0.2246, pruned_loss=0.04292, over 2265189.00 frames. 2023-04-27 17:57:46,920 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-27 17:57:48,913 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=126012.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 17:57:50,796 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2131, 1.4617, 1.6857, 1.8413, 1.6843, 1.7960, 1.7620, 1.7603], device='cuda:4'), covar=tensor([0.3980, 0.4873, 0.3943, 0.3846, 0.4983, 0.6379, 0.4446, 0.4261], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0374, 0.0325, 0.0340, 0.0347, 0.0395, 0.0356, 0.0331], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 17:58:05,033 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.65 vs. limit=2.0 2023-04-27 17:58:05,543 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=126032.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:58:06,787 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1987, 1.8230, 2.0899, 2.4960, 2.3885, 1.9247, 1.9027, 2.1693], device='cuda:4'), covar=tensor([0.0807, 0.1183, 0.0728, 0.0644, 0.0603, 0.0910, 0.0740, 0.0599], device='cuda:4'), in_proj_covar=tensor([0.0186, 0.0201, 0.0182, 0.0174, 0.0176, 0.0180, 0.0151, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 17:58:16,772 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4582, 2.9795, 0.9821, 1.6150, 2.3028, 1.4171, 4.0447, 1.8096], device='cuda:4'), covar=tensor([0.0656, 0.0701, 0.0907, 0.1212, 0.0503, 0.1053, 0.0229, 0.0639], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0064, 0.0048, 0.0046, 0.0050, 0.0052, 0.0074, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 17:58:23,638 INFO [finetune.py:976] (4/7) Epoch 23, batch 50, loss[loss=0.1726, simple_loss=0.2314, pruned_loss=0.05686, over 4123.00 frames. ], tot_loss[loss=0.1709, simple_loss=0.2445, pruned_loss=0.04863, over 216550.72 frames. ], batch size: 17, lr: 3.12e-03, grad_scale: 16.0 2023-04-27 17:58:24,213 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=126059.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:58:25,268 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=126060.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:58:28,246 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.078e+02 1.502e+02 1.744e+02 2.087e+02 3.495e+02, threshold=3.488e+02, percent-clipped=0.0 2023-04-27 17:58:31,102 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.59 vs. limit=5.0 2023-04-27 17:58:41,164 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=126077.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:58:42,980 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=126080.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:58:43,056 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4838, 1.7368, 1.3881, 1.0831, 1.1644, 1.1116, 1.3758, 1.0937], device='cuda:4'), covar=tensor([0.1713, 0.1252, 0.1509, 0.1802, 0.2323, 0.1999, 0.1047, 0.2088], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0212, 0.0170, 0.0205, 0.0201, 0.0187, 0.0156, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 17:58:48,842 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=126081.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 17:59:21,047 INFO [finetune.py:976] (4/7) Epoch 23, batch 100, loss[loss=0.1299, simple_loss=0.2111, pruned_loss=0.02433, over 4779.00 frames. ], tot_loss[loss=0.164, simple_loss=0.2367, pruned_loss=0.04566, over 379359.58 frames. ], batch size: 29, lr: 3.12e-03, grad_scale: 16.0 2023-04-27 17:59:42,931 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=126125.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:00:28,539 INFO [finetune.py:976] (4/7) Epoch 23, batch 150, loss[loss=0.1652, simple_loss=0.2347, pruned_loss=0.04784, over 4870.00 frames. ], tot_loss[loss=0.1655, simple_loss=0.2359, pruned_loss=0.04755, over 507678.26 frames. ], batch size: 34, lr: 3.12e-03, grad_scale: 16.0 2023-04-27 18:00:38,440 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.878e+01 1.494e+02 1.914e+02 2.300e+02 4.167e+02, threshold=3.828e+02, percent-clipped=5.0 2023-04-27 18:00:58,751 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6635, 2.5156, 2.7942, 2.9901, 2.8922, 2.4695, 2.2000, 2.7448], device='cuda:4'), covar=tensor([0.0771, 0.0880, 0.0554, 0.0606, 0.0596, 0.0838, 0.0724, 0.0517], device='cuda:4'), in_proj_covar=tensor([0.0185, 0.0200, 0.0182, 0.0174, 0.0175, 0.0179, 0.0151, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 18:01:00,637 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5627, 1.6857, 1.4439, 0.9974, 1.1976, 1.1790, 1.4503, 1.1242], device='cuda:4'), covar=tensor([0.1763, 0.1374, 0.1540, 0.1921, 0.2406, 0.2135, 0.1034, 0.2146], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0212, 0.0170, 0.0205, 0.0201, 0.0186, 0.0156, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 18:01:34,982 INFO [finetune.py:976] (4/7) Epoch 23, batch 200, loss[loss=0.1899, simple_loss=0.2555, pruned_loss=0.06214, over 4900.00 frames. ], tot_loss[loss=0.1655, simple_loss=0.2352, pruned_loss=0.04794, over 608368.01 frames. ], batch size: 32, lr: 3.12e-03, grad_scale: 16.0 2023-04-27 18:02:08,265 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-27 18:02:19,986 INFO [finetune.py:976] (4/7) Epoch 23, batch 250, loss[loss=0.1924, simple_loss=0.2664, pruned_loss=0.05924, over 4870.00 frames. ], tot_loss[loss=0.1703, simple_loss=0.2408, pruned_loss=0.04991, over 685090.75 frames. ], batch size: 31, lr: 3.11e-03, grad_scale: 16.0 2023-04-27 18:02:20,603 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=126259.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:02:24,187 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.438e+01 1.622e+02 1.975e+02 2.331e+02 7.246e+02, threshold=3.950e+02, percent-clipped=3.0 2023-04-27 18:02:51,720 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=126307.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:02:53,383 INFO [finetune.py:976] (4/7) Epoch 23, batch 300, loss[loss=0.2161, simple_loss=0.2861, pruned_loss=0.07303, over 4924.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2437, pruned_loss=0.05012, over 743752.53 frames. ], batch size: 36, lr: 3.11e-03, grad_scale: 16.0 2023-04-27 18:03:04,143 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.28 vs. limit=5.0 2023-04-27 18:03:22,883 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=126354.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:03:26,337 INFO [finetune.py:976] (4/7) Epoch 23, batch 350, loss[loss=0.1392, simple_loss=0.2137, pruned_loss=0.03238, over 4822.00 frames. ], tot_loss[loss=0.1725, simple_loss=0.2451, pruned_loss=0.04992, over 791162.83 frames. ], batch size: 25, lr: 3.11e-03, grad_scale: 16.0 2023-04-27 18:03:30,385 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.435e+01 1.518e+02 1.849e+02 2.150e+02 3.695e+02, threshold=3.697e+02, percent-clipped=0.0 2023-04-27 18:03:41,540 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=126380.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:03:42,127 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=126381.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:03:47,288 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.77 vs. limit=5.0 2023-04-27 18:03:58,659 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0302, 1.7371, 2.1670, 2.4469, 2.0763, 1.9028, 2.0288, 2.0060], device='cuda:4'), covar=tensor([0.4966, 0.7688, 0.8034, 0.6183, 0.6638, 0.8983, 0.9694, 1.0208], device='cuda:4'), in_proj_covar=tensor([0.0434, 0.0416, 0.0509, 0.0508, 0.0463, 0.0493, 0.0499, 0.0510], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 18:03:59,739 INFO [finetune.py:976] (4/7) Epoch 23, batch 400, loss[loss=0.1661, simple_loss=0.2413, pruned_loss=0.04542, over 4822.00 frames. ], tot_loss[loss=0.1721, simple_loss=0.2451, pruned_loss=0.04957, over 827979.65 frames. ], batch size: 38, lr: 3.11e-03, grad_scale: 16.0 2023-04-27 18:04:30,796 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=126429.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:04:38,775 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.94 vs. limit=2.0 2023-04-27 18:04:42,249 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6254, 1.5208, 1.9530, 1.9463, 1.5001, 1.3847, 1.5935, 1.0750], device='cuda:4'), covar=tensor([0.0479, 0.0668, 0.0338, 0.0664, 0.0693, 0.1060, 0.0595, 0.0565], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0068, 0.0066, 0.0068, 0.0075, 0.0095, 0.0073, 0.0064], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 18:04:42,871 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=126441.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:04:48,375 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=126450.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:04:54,265 INFO [finetune.py:976] (4/7) Epoch 23, batch 450, loss[loss=0.1828, simple_loss=0.2595, pruned_loss=0.05308, over 4808.00 frames. ], tot_loss[loss=0.1713, simple_loss=0.2445, pruned_loss=0.04907, over 857872.75 frames. ], batch size: 41, lr: 3.11e-03, grad_scale: 16.0 2023-04-27 18:04:55,715 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-04-27 18:04:58,395 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.037e+02 1.511e+02 1.837e+02 2.263e+02 5.457e+02, threshold=3.673e+02, percent-clipped=4.0 2023-04-27 18:04:59,031 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9743, 2.6513, 2.0227, 2.0899, 1.4809, 1.4978, 2.1476, 1.3898], device='cuda:4'), covar=tensor([0.1692, 0.1338, 0.1424, 0.1622, 0.2260, 0.1956, 0.0996, 0.2099], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0209, 0.0168, 0.0202, 0.0198, 0.0184, 0.0154, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 18:05:27,629 INFO [finetune.py:976] (4/7) Epoch 23, batch 500, loss[loss=0.164, simple_loss=0.2399, pruned_loss=0.04406, over 4824.00 frames. ], tot_loss[loss=0.1692, simple_loss=0.2421, pruned_loss=0.04811, over 879788.87 frames. ], batch size: 38, lr: 3.11e-03, grad_scale: 16.0 2023-04-27 18:05:29,454 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=126511.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:06:06,865 INFO [finetune.py:976] (4/7) Epoch 23, batch 550, loss[loss=0.1805, simple_loss=0.2428, pruned_loss=0.05913, over 4781.00 frames. ], tot_loss[loss=0.1674, simple_loss=0.2392, pruned_loss=0.04782, over 897762.27 frames. ], batch size: 28, lr: 3.11e-03, grad_scale: 16.0 2023-04-27 18:06:16,306 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.004e+02 1.502e+02 1.815e+02 2.160e+02 5.481e+02, threshold=3.630e+02, percent-clipped=1.0 2023-04-27 18:06:59,580 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-04-27 18:07:12,072 INFO [finetune.py:976] (4/7) Epoch 23, batch 600, loss[loss=0.1965, simple_loss=0.2891, pruned_loss=0.05201, over 4891.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2405, pruned_loss=0.04857, over 912430.05 frames. ], batch size: 43, lr: 3.11e-03, grad_scale: 16.0 2023-04-27 18:07:33,780 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-04-27 18:07:55,171 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.68 vs. limit=5.0 2023-04-27 18:08:14,790 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=126654.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:08:17,792 INFO [finetune.py:976] (4/7) Epoch 23, batch 650, loss[loss=0.156, simple_loss=0.2209, pruned_loss=0.04552, over 4048.00 frames. ], tot_loss[loss=0.1709, simple_loss=0.243, pruned_loss=0.04942, over 918326.73 frames. ], batch size: 17, lr: 3.11e-03, grad_scale: 16.0 2023-04-27 18:08:26,641 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.220e+02 1.615e+02 1.965e+02 2.363e+02 5.710e+02, threshold=3.929e+02, percent-clipped=5.0 2023-04-27 18:09:01,389 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8748, 1.5246, 1.3937, 1.7637, 2.0133, 1.6335, 1.4312, 1.3814], device='cuda:4'), covar=tensor([0.1495, 0.1414, 0.1898, 0.1070, 0.0865, 0.1891, 0.2143, 0.2210], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0310, 0.0351, 0.0286, 0.0325, 0.0306, 0.0298, 0.0370], device='cuda:4'), out_proj_covar=tensor([6.4050e-05, 6.4128e-05, 7.4150e-05, 5.7500e-05, 6.7053e-05, 6.4260e-05, 6.2120e-05, 7.8485e-05], device='cuda:4') 2023-04-27 18:09:20,359 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=126702.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:09:24,630 INFO [finetune.py:976] (4/7) Epoch 23, batch 700, loss[loss=0.1306, simple_loss=0.2039, pruned_loss=0.02869, over 4035.00 frames. ], tot_loss[loss=0.1721, simple_loss=0.2448, pruned_loss=0.04967, over 926894.99 frames. ], batch size: 17, lr: 3.11e-03, grad_scale: 16.0 2023-04-27 18:10:03,669 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=126736.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:10:19,573 INFO [finetune.py:976] (4/7) Epoch 23, batch 750, loss[loss=0.1816, simple_loss=0.2565, pruned_loss=0.05336, over 4921.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.2468, pruned_loss=0.05044, over 934512.45 frames. ], batch size: 33, lr: 3.11e-03, grad_scale: 16.0 2023-04-27 18:10:23,175 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.166e+01 1.471e+02 1.836e+02 2.076e+02 2.754e+02, threshold=3.672e+02, percent-clipped=0.0 2023-04-27 18:10:51,754 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=126806.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:10:53,513 INFO [finetune.py:976] (4/7) Epoch 23, batch 800, loss[loss=0.2423, simple_loss=0.3059, pruned_loss=0.08933, over 4114.00 frames. ], tot_loss[loss=0.1733, simple_loss=0.2468, pruned_loss=0.04993, over 938716.17 frames. ], batch size: 66, lr: 3.11e-03, grad_scale: 16.0 2023-04-27 18:10:57,926 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=126816.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:10:58,555 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1551, 1.5683, 1.5199, 1.7818, 1.6697, 1.8970, 1.4186, 3.3492], device='cuda:4'), covar=tensor([0.0608, 0.0772, 0.0753, 0.1146, 0.0615, 0.0440, 0.0697, 0.0150], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0042, 0.0040, 0.0037, 0.0037, 0.0054], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 18:11:27,334 INFO [finetune.py:976] (4/7) Epoch 23, batch 850, loss[loss=0.1922, simple_loss=0.2582, pruned_loss=0.06312, over 4929.00 frames. ], tot_loss[loss=0.1722, simple_loss=0.2449, pruned_loss=0.04973, over 943430.55 frames. ], batch size: 33, lr: 3.11e-03, grad_scale: 16.0 2023-04-27 18:11:29,894 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2103, 1.3538, 1.6552, 1.7756, 1.7050, 1.7885, 1.6966, 1.7005], device='cuda:4'), covar=tensor([0.3796, 0.5406, 0.4667, 0.4716, 0.5588, 0.7300, 0.4992, 0.4893], device='cuda:4'), in_proj_covar=tensor([0.0338, 0.0373, 0.0323, 0.0338, 0.0345, 0.0394, 0.0356, 0.0330], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 18:11:30,954 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.555e+01 1.428e+02 1.632e+02 2.009e+02 4.490e+02, threshold=3.264e+02, percent-clipped=1.0 2023-04-27 18:11:44,278 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=126877.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:12:28,956 INFO [finetune.py:976] (4/7) Epoch 23, batch 900, loss[loss=0.1324, simple_loss=0.2069, pruned_loss=0.02893, over 4792.00 frames. ], tot_loss[loss=0.1696, simple_loss=0.2419, pruned_loss=0.04865, over 946735.09 frames. ], batch size: 26, lr: 3.11e-03, grad_scale: 16.0 2023-04-27 18:12:30,275 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3713, 1.5212, 4.1458, 3.9177, 3.6497, 4.0065, 3.9339, 3.6776], device='cuda:4'), covar=tensor([0.7039, 0.5294, 0.1097, 0.1556, 0.1073, 0.1601, 0.1411, 0.1474], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0308, 0.0406, 0.0408, 0.0349, 0.0411, 0.0314, 0.0364], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 18:13:36,983 INFO [finetune.py:976] (4/7) Epoch 23, batch 950, loss[loss=0.2125, simple_loss=0.2841, pruned_loss=0.07043, over 4812.00 frames. ], tot_loss[loss=0.1692, simple_loss=0.2405, pruned_loss=0.04893, over 949684.33 frames. ], batch size: 40, lr: 3.11e-03, grad_scale: 16.0 2023-04-27 18:13:40,660 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.057e+02 1.507e+02 1.824e+02 2.265e+02 3.979e+02, threshold=3.647e+02, percent-clipped=2.0 2023-04-27 18:14:11,163 INFO [finetune.py:976] (4/7) Epoch 23, batch 1000, loss[loss=0.2085, simple_loss=0.2735, pruned_loss=0.07169, over 4936.00 frames. ], tot_loss[loss=0.1707, simple_loss=0.2423, pruned_loss=0.04957, over 951065.47 frames. ], batch size: 38, lr: 3.11e-03, grad_scale: 16.0 2023-04-27 18:14:28,603 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=127036.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:14:42,829 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5686, 1.3949, 0.6074, 1.2829, 1.6170, 1.4287, 1.3670, 1.3832], device='cuda:4'), covar=tensor([0.0519, 0.0389, 0.0367, 0.0569, 0.0285, 0.0521, 0.0507, 0.0576], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0029, 0.0019, 0.0028, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0046, 0.0038, 0.0052, 0.0039, 0.0050, 0.0050, 0.0052], device='cuda:4') 2023-04-27 18:14:45,072 INFO [finetune.py:976] (4/7) Epoch 23, batch 1050, loss[loss=0.2086, simple_loss=0.2707, pruned_loss=0.07324, over 4928.00 frames. ], tot_loss[loss=0.1734, simple_loss=0.2457, pruned_loss=0.05054, over 952091.98 frames. ], batch size: 38, lr: 3.11e-03, grad_scale: 16.0 2023-04-27 18:14:48,720 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.212e+02 1.574e+02 1.858e+02 2.309e+02 5.197e+02, threshold=3.716e+02, percent-clipped=5.0 2023-04-27 18:15:00,298 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=127084.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:15:13,720 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.56 vs. limit=5.0 2023-04-27 18:15:16,226 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=127106.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:15:18,458 INFO [finetune.py:976] (4/7) Epoch 23, batch 1100, loss[loss=0.2021, simple_loss=0.2773, pruned_loss=0.06342, over 4851.00 frames. ], tot_loss[loss=0.1743, simple_loss=0.2467, pruned_loss=0.051, over 952045.68 frames. ], batch size: 44, lr: 3.11e-03, grad_scale: 16.0 2023-04-27 18:15:35,422 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8016, 4.1215, 0.8129, 2.2613, 2.3274, 2.6804, 2.3625, 0.8579], device='cuda:4'), covar=tensor([0.1380, 0.1006, 0.2103, 0.1207, 0.0996, 0.1177, 0.1527, 0.2249], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0238, 0.0137, 0.0119, 0.0131, 0.0150, 0.0116, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 18:15:47,875 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=127154.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:15:51,854 INFO [finetune.py:976] (4/7) Epoch 23, batch 1150, loss[loss=0.1369, simple_loss=0.2101, pruned_loss=0.03183, over 4737.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.2465, pruned_loss=0.05068, over 950983.72 frames. ], batch size: 23, lr: 3.11e-03, grad_scale: 16.0 2023-04-27 18:15:56,437 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.040e+02 1.449e+02 1.782e+02 2.216e+02 4.823e+02, threshold=3.563e+02, percent-clipped=1.0 2023-04-27 18:15:59,072 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-27 18:16:00,193 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8709, 2.3159, 0.8630, 1.2424, 1.5231, 1.1953, 2.4320, 1.3532], device='cuda:4'), covar=tensor([0.0644, 0.0572, 0.0698, 0.1204, 0.0470, 0.0995, 0.0321, 0.0687], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0073, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 18:16:00,786 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=127172.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:16:25,337 INFO [finetune.py:976] (4/7) Epoch 23, batch 1200, loss[loss=0.1473, simple_loss=0.2189, pruned_loss=0.03788, over 4751.00 frames. ], tot_loss[loss=0.1718, simple_loss=0.2441, pruned_loss=0.04969, over 951652.52 frames. ], batch size: 23, lr: 3.11e-03, grad_scale: 16.0 2023-04-27 18:16:34,739 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6685, 1.6597, 0.9328, 1.3764, 1.7768, 1.5108, 1.4550, 1.5315], device='cuda:4'), covar=tensor([0.0456, 0.0343, 0.0311, 0.0513, 0.0266, 0.0487, 0.0458, 0.0533], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0028, 0.0019, 0.0028, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0046, 0.0038, 0.0052, 0.0038, 0.0050, 0.0050, 0.0051], device='cuda:4') 2023-04-27 18:16:52,554 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=127249.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:16:52,604 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1230, 1.9307, 2.2729, 2.4652, 2.1974, 1.9363, 2.0935, 2.0658], device='cuda:4'), covar=tensor([0.4252, 0.5791, 0.5714, 0.4908, 0.5498, 0.8001, 0.7931, 0.9165], device='cuda:4'), in_proj_covar=tensor([0.0434, 0.0416, 0.0509, 0.0508, 0.0463, 0.0493, 0.0498, 0.0509], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 18:17:03,695 INFO [finetune.py:976] (4/7) Epoch 23, batch 1250, loss[loss=0.1433, simple_loss=0.2049, pruned_loss=0.04081, over 4836.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.2413, pruned_loss=0.04889, over 951483.43 frames. ], batch size: 30, lr: 3.11e-03, grad_scale: 16.0 2023-04-27 18:17:14,002 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.661e+01 1.481e+02 1.752e+02 2.149e+02 4.730e+02, threshold=3.504e+02, percent-clipped=1.0 2023-04-27 18:18:08,734 INFO [finetune.py:976] (4/7) Epoch 23, batch 1300, loss[loss=0.1506, simple_loss=0.2163, pruned_loss=0.04244, over 4101.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.2388, pruned_loss=0.04828, over 953121.79 frames. ], batch size: 65, lr: 3.11e-03, grad_scale: 32.0 2023-04-27 18:18:09,512 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=127310.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:18:49,602 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9018, 2.2854, 0.8665, 1.2645, 1.6137, 1.2067, 2.4740, 1.3337], device='cuda:4'), covar=tensor([0.0684, 0.0532, 0.0661, 0.1239, 0.0462, 0.1034, 0.0282, 0.0716], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0065, 0.0048, 0.0046, 0.0050, 0.0051, 0.0073, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 18:18:49,611 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0594, 1.3498, 1.2973, 1.6404, 1.5147, 1.4617, 1.3669, 2.4890], device='cuda:4'), covar=tensor([0.0618, 0.0838, 0.0836, 0.1218, 0.0653, 0.0513, 0.0711, 0.0199], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0040, 0.0038, 0.0037, 0.0054], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 18:19:13,550 INFO [finetune.py:976] (4/7) Epoch 23, batch 1350, loss[loss=0.1348, simple_loss=0.2047, pruned_loss=0.03249, over 4214.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.24, pruned_loss=0.04895, over 954287.44 frames. ], batch size: 18, lr: 3.11e-03, grad_scale: 32.0 2023-04-27 18:19:23,394 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.776e+01 1.488e+02 1.772e+02 2.253e+02 3.450e+02, threshold=3.544e+02, percent-clipped=0.0 2023-04-27 18:20:19,976 INFO [finetune.py:976] (4/7) Epoch 23, batch 1400, loss[loss=0.2072, simple_loss=0.2902, pruned_loss=0.06211, over 4841.00 frames. ], tot_loss[loss=0.1707, simple_loss=0.2428, pruned_loss=0.04927, over 955950.68 frames. ], batch size: 49, lr: 3.11e-03, grad_scale: 32.0 2023-04-27 18:20:43,125 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=127427.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 18:21:02,766 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1418, 1.4772, 1.3616, 1.7485, 1.6678, 1.7194, 1.4232, 3.0841], device='cuda:4'), covar=tensor([0.0635, 0.0780, 0.0789, 0.1172, 0.0587, 0.0482, 0.0705, 0.0146], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0042, 0.0040, 0.0037, 0.0037, 0.0054], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 18:21:20,824 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8613, 2.5528, 1.9426, 1.8362, 1.4051, 1.3755, 1.9909, 1.3886], device='cuda:4'), covar=tensor([0.1541, 0.1232, 0.1293, 0.1604, 0.2213, 0.1816, 0.0936, 0.1945], device='cuda:4'), in_proj_covar=tensor([0.0195, 0.0209, 0.0167, 0.0202, 0.0198, 0.0184, 0.0154, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 18:21:24,259 INFO [finetune.py:976] (4/7) Epoch 23, batch 1450, loss[loss=0.1804, simple_loss=0.2521, pruned_loss=0.05434, over 4792.00 frames. ], tot_loss[loss=0.1704, simple_loss=0.2434, pruned_loss=0.04872, over 956379.04 frames. ], batch size: 26, lr: 3.11e-03, grad_scale: 32.0 2023-04-27 18:21:34,077 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.128e+02 1.543e+02 1.916e+02 2.299e+02 4.494e+02, threshold=3.833e+02, percent-clipped=8.0 2023-04-27 18:21:43,896 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=127472.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:21:44,015 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.44 vs. limit=2.0 2023-04-27 18:22:06,023 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=127488.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 18:22:29,897 INFO [finetune.py:976] (4/7) Epoch 23, batch 1500, loss[loss=0.1964, simple_loss=0.2652, pruned_loss=0.06376, over 4880.00 frames. ], tot_loss[loss=0.171, simple_loss=0.2441, pruned_loss=0.04896, over 954342.75 frames. ], batch size: 32, lr: 3.11e-03, grad_scale: 32.0 2023-04-27 18:22:36,465 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1325, 2.6326, 2.1331, 2.5318, 1.6643, 2.1376, 2.4471, 1.6361], device='cuda:4'), covar=tensor([0.2123, 0.1294, 0.0898, 0.1235, 0.3737, 0.1335, 0.1995, 0.2955], device='cuda:4'), in_proj_covar=tensor([0.0281, 0.0295, 0.0212, 0.0273, 0.0309, 0.0252, 0.0245, 0.0259], device='cuda:4'), out_proj_covar=tensor([1.1212e-04, 1.1697e-04, 8.3515e-05, 1.0760e-04, 1.2488e-04, 9.9605e-05, 9.8769e-05, 1.0214e-04], device='cuda:4') 2023-04-27 18:22:37,634 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=127520.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:22:39,507 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3466, 3.0463, 0.9529, 1.6716, 2.1858, 1.2174, 3.8115, 1.7692], device='cuda:4'), covar=tensor([0.0667, 0.0787, 0.0940, 0.1217, 0.0514, 0.1033, 0.0234, 0.0627], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0065, 0.0047, 0.0046, 0.0050, 0.0051, 0.0073, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 18:22:43,216 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=127529.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:23:19,694 INFO [finetune.py:976] (4/7) Epoch 23, batch 1550, loss[loss=0.1471, simple_loss=0.2244, pruned_loss=0.03487, over 4895.00 frames. ], tot_loss[loss=0.1709, simple_loss=0.2442, pruned_loss=0.04882, over 954162.12 frames. ], batch size: 32, lr: 3.10e-03, grad_scale: 32.0 2023-04-27 18:23:23,843 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.013e+02 1.538e+02 1.819e+02 2.091e+02 4.341e+02, threshold=3.638e+02, percent-clipped=1.0 2023-04-27 18:23:41,773 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=127590.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:23:50,931 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=127605.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:23:53,256 INFO [finetune.py:976] (4/7) Epoch 23, batch 1600, loss[loss=0.1529, simple_loss=0.2276, pruned_loss=0.03913, over 4790.00 frames. ], tot_loss[loss=0.1683, simple_loss=0.2414, pruned_loss=0.04764, over 955300.43 frames. ], batch size: 29, lr: 3.10e-03, grad_scale: 32.0 2023-04-27 18:24:11,071 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.9172, 2.1626, 2.1330, 2.3208, 2.0514, 2.1690, 2.2829, 2.1771], device='cuda:4'), covar=tensor([0.3623, 0.5674, 0.4875, 0.4294, 0.5637, 0.7120, 0.5370, 0.5185], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0376, 0.0326, 0.0340, 0.0349, 0.0396, 0.0358, 0.0332], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 18:24:13,979 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.2210, 2.6717, 2.5901, 2.7846, 2.4851, 2.6005, 2.7384, 2.6861], device='cuda:4'), covar=tensor([0.3532, 0.4893, 0.4365, 0.3905, 0.5309, 0.6335, 0.4942, 0.4290], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0376, 0.0326, 0.0340, 0.0349, 0.0396, 0.0358, 0.0332], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 18:24:20,715 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.59 vs. limit=2.0 2023-04-27 18:24:26,893 INFO [finetune.py:976] (4/7) Epoch 23, batch 1650, loss[loss=0.1897, simple_loss=0.2585, pruned_loss=0.06039, over 4707.00 frames. ], tot_loss[loss=0.1663, simple_loss=0.2383, pruned_loss=0.0472, over 954213.98 frames. ], batch size: 59, lr: 3.10e-03, grad_scale: 32.0 2023-04-27 18:24:28,867 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=127662.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:24:31,046 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.001e+02 1.420e+02 1.883e+02 2.201e+02 3.441e+02, threshold=3.766e+02, percent-clipped=0.0 2023-04-27 18:25:00,252 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3060, 1.2218, 3.7795, 3.3375, 3.4706, 3.5307, 3.5387, 3.1755], device='cuda:4'), covar=tensor([0.9347, 0.8364, 0.2008, 0.3677, 0.2078, 0.3719, 0.2603, 0.3375], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0306, 0.0404, 0.0407, 0.0347, 0.0409, 0.0314, 0.0364], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 18:25:00,769 INFO [finetune.py:976] (4/7) Epoch 23, batch 1700, loss[loss=0.1633, simple_loss=0.2251, pruned_loss=0.05071, over 4778.00 frames. ], tot_loss[loss=0.166, simple_loss=0.2378, pruned_loss=0.04712, over 954877.98 frames. ], batch size: 28, lr: 3.10e-03, grad_scale: 32.0 2023-04-27 18:25:10,334 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=127723.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:25:34,585 INFO [finetune.py:976] (4/7) Epoch 23, batch 1750, loss[loss=0.1771, simple_loss=0.2504, pruned_loss=0.05188, over 4899.00 frames. ], tot_loss[loss=0.1683, simple_loss=0.2402, pruned_loss=0.04818, over 954429.80 frames. ], batch size: 35, lr: 3.10e-03, grad_scale: 32.0 2023-04-27 18:25:38,230 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.809e+01 1.624e+02 1.849e+02 2.197e+02 5.063e+02, threshold=3.698e+02, percent-clipped=3.0 2023-04-27 18:25:40,859 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1054, 2.4866, 2.1028, 1.9347, 1.6537, 1.6598, 2.1210, 1.5831], device='cuda:4'), covar=tensor([0.1415, 0.1350, 0.1293, 0.1663, 0.2129, 0.1729, 0.0901, 0.1893], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0209, 0.0167, 0.0202, 0.0197, 0.0184, 0.0154, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 18:25:46,631 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8289, 1.8045, 4.4192, 4.1191, 3.9476, 4.1460, 4.1621, 3.8728], device='cuda:4'), covar=tensor([0.6620, 0.4883, 0.1177, 0.1789, 0.1082, 0.1670, 0.1094, 0.1632], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0307, 0.0405, 0.0408, 0.0348, 0.0410, 0.0314, 0.0365], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 18:25:48,421 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=127779.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:25:50,843 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=127783.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 18:25:56,104 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5268, 1.4838, 1.8562, 1.8722, 1.4321, 1.1709, 1.4851, 0.9095], device='cuda:4'), covar=tensor([0.0526, 0.0543, 0.0341, 0.0501, 0.0715, 0.1339, 0.0586, 0.0717], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0068, 0.0067, 0.0068, 0.0075, 0.0096, 0.0073, 0.0065], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 18:26:14,169 INFO [finetune.py:976] (4/7) Epoch 23, batch 1800, loss[loss=0.1661, simple_loss=0.247, pruned_loss=0.04256, over 4865.00 frames. ], tot_loss[loss=0.1702, simple_loss=0.2429, pruned_loss=0.04875, over 954418.58 frames. ], batch size: 31, lr: 3.10e-03, grad_scale: 32.0 2023-04-27 18:26:57,475 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=127840.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:27:09,564 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=127850.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:27:21,426 INFO [finetune.py:976] (4/7) Epoch 23, batch 1850, loss[loss=0.193, simple_loss=0.2587, pruned_loss=0.0637, over 4756.00 frames. ], tot_loss[loss=0.1715, simple_loss=0.2444, pruned_loss=0.04927, over 955622.35 frames. ], batch size: 54, lr: 3.10e-03, grad_scale: 32.0 2023-04-27 18:27:31,371 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.092e+02 1.632e+02 1.910e+02 2.193e+02 3.676e+02, threshold=3.820e+02, percent-clipped=0.0 2023-04-27 18:27:32,667 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0702, 1.4843, 4.5446, 4.2376, 3.9416, 4.2006, 4.0672, 4.0188], device='cuda:4'), covar=tensor([0.6890, 0.5712, 0.1051, 0.1889, 0.1221, 0.1471, 0.2238, 0.1615], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0307, 0.0406, 0.0408, 0.0350, 0.0411, 0.0316, 0.0366], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 18:27:40,524 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=127872.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:27:44,259 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-27 18:27:49,850 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=127885.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:28:04,009 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=127905.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:28:06,380 INFO [finetune.py:976] (4/7) Epoch 23, batch 1900, loss[loss=0.1685, simple_loss=0.2529, pruned_loss=0.04209, over 4828.00 frames. ], tot_loss[loss=0.1727, simple_loss=0.2459, pruned_loss=0.04976, over 954510.07 frames. ], batch size: 47, lr: 3.10e-03, grad_scale: 32.0 2023-04-27 18:28:07,708 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=127911.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:28:32,847 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=127933.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:28:57,593 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=127953.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:29:07,612 INFO [finetune.py:976] (4/7) Epoch 23, batch 1950, loss[loss=0.1714, simple_loss=0.2428, pruned_loss=0.05005, over 4813.00 frames. ], tot_loss[loss=0.1712, simple_loss=0.2444, pruned_loss=0.04898, over 953983.28 frames. ], batch size: 40, lr: 3.10e-03, grad_scale: 32.0 2023-04-27 18:29:16,663 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.063e+02 1.493e+02 1.855e+02 2.257e+02 4.055e+02, threshold=3.710e+02, percent-clipped=1.0 2023-04-27 18:30:00,664 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.8271, 4.9294, 3.2275, 5.6490, 4.9426, 4.9292, 2.1675, 4.8899], device='cuda:4'), covar=tensor([0.1810, 0.1059, 0.2879, 0.0895, 0.4585, 0.1813, 0.6395, 0.2046], device='cuda:4'), in_proj_covar=tensor([0.0247, 0.0219, 0.0255, 0.0307, 0.0300, 0.0248, 0.0276, 0.0274], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 18:30:02,325 INFO [finetune.py:976] (4/7) Epoch 23, batch 2000, loss[loss=0.1489, simple_loss=0.2169, pruned_loss=0.04047, over 4821.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2421, pruned_loss=0.04834, over 953889.17 frames. ], batch size: 30, lr: 3.10e-03, grad_scale: 32.0 2023-04-27 18:30:07,961 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=128018.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:30:15,989 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1479, 2.3167, 1.0752, 1.4301, 1.7799, 1.2528, 2.9088, 1.6188], device='cuda:4'), covar=tensor([0.0591, 0.0667, 0.0731, 0.1115, 0.0473, 0.0916, 0.0266, 0.0589], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0050, 0.0051, 0.0073, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 18:30:35,022 INFO [finetune.py:976] (4/7) Epoch 23, batch 2050, loss[loss=0.1515, simple_loss=0.2193, pruned_loss=0.04182, over 4832.00 frames. ], tot_loss[loss=0.1663, simple_loss=0.2383, pruned_loss=0.04719, over 954806.21 frames. ], batch size: 33, lr: 3.10e-03, grad_scale: 32.0 2023-04-27 18:30:39,641 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.902e+01 1.493e+02 1.754e+02 2.070e+02 4.110e+02, threshold=3.508e+02, percent-clipped=2.0 2023-04-27 18:30:50,719 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=128083.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 18:30:58,443 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.34 vs. limit=5.0 2023-04-27 18:31:08,824 INFO [finetune.py:976] (4/7) Epoch 23, batch 2100, loss[loss=0.1823, simple_loss=0.2516, pruned_loss=0.05645, over 4831.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.2388, pruned_loss=0.04767, over 953347.84 frames. ], batch size: 39, lr: 3.10e-03, grad_scale: 32.0 2023-04-27 18:31:22,861 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=128131.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 18:31:25,347 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=128135.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:31:42,238 INFO [finetune.py:976] (4/7) Epoch 23, batch 2150, loss[loss=0.184, simple_loss=0.2605, pruned_loss=0.05374, over 4900.00 frames. ], tot_loss[loss=0.1693, simple_loss=0.2416, pruned_loss=0.04849, over 953516.31 frames. ], batch size: 36, lr: 3.10e-03, grad_scale: 32.0 2023-04-27 18:31:46,841 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.015e+02 1.634e+02 1.958e+02 2.421e+02 4.804e+02, threshold=3.916e+02, percent-clipped=1.0 2023-04-27 18:31:50,529 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.9662, 3.9563, 2.8585, 4.5388, 3.9287, 3.9658, 1.5151, 3.9318], device='cuda:4'), covar=tensor([0.1701, 0.1306, 0.3052, 0.1560, 0.3634, 0.1906, 0.6263, 0.2360], device='cuda:4'), in_proj_covar=tensor([0.0245, 0.0218, 0.0253, 0.0306, 0.0297, 0.0246, 0.0275, 0.0272], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 18:31:50,871 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-27 18:31:59,134 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=128185.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:32:02,650 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3992, 1.2199, 1.6109, 1.5942, 1.3039, 1.2063, 1.2365, 0.6226], device='cuda:4'), covar=tensor([0.0549, 0.0681, 0.0364, 0.0557, 0.0712, 0.1062, 0.0526, 0.0653], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0068, 0.0067, 0.0068, 0.0075, 0.0096, 0.0073, 0.0065], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 18:32:09,957 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7681, 1.3737, 1.8473, 2.2815, 1.8918, 1.7472, 1.7797, 1.7177], device='cuda:4'), covar=tensor([0.4164, 0.6772, 0.5928, 0.5095, 0.5508, 0.7208, 0.7931, 0.8502], device='cuda:4'), in_proj_covar=tensor([0.0434, 0.0418, 0.0509, 0.0506, 0.0464, 0.0494, 0.0499, 0.0509], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 18:32:13,545 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=128206.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:32:15,339 INFO [finetune.py:976] (4/7) Epoch 23, batch 2200, loss[loss=0.1594, simple_loss=0.2452, pruned_loss=0.03681, over 4807.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2442, pruned_loss=0.04896, over 954400.32 frames. ], batch size: 25, lr: 3.10e-03, grad_scale: 32.0 2023-04-27 18:32:20,371 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-04-27 18:32:39,413 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=128228.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:32:42,389 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=128233.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:32:48,375 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.76 vs. limit=5.0 2023-04-27 18:33:12,840 INFO [finetune.py:976] (4/7) Epoch 23, batch 2250, loss[loss=0.1877, simple_loss=0.2379, pruned_loss=0.06876, over 4801.00 frames. ], tot_loss[loss=0.1724, simple_loss=0.2454, pruned_loss=0.04965, over 956265.47 frames. ], batch size: 25, lr: 3.10e-03, grad_scale: 32.0 2023-04-27 18:33:22,484 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.137e+02 1.532e+02 1.818e+02 2.293e+02 4.621e+02, threshold=3.635e+02, percent-clipped=2.0 2023-04-27 18:33:57,139 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7301, 2.0102, 1.0493, 1.5104, 2.2387, 1.5652, 1.5492, 1.6554], device='cuda:4'), covar=tensor([0.0484, 0.0341, 0.0282, 0.0552, 0.0222, 0.0497, 0.0470, 0.0539], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0024, 0.0022, 0.0028, 0.0019, 0.0028, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0046, 0.0038, 0.0052, 0.0038, 0.0050, 0.0050, 0.0051], device='cuda:4') 2023-04-27 18:34:26,342 INFO [finetune.py:976] (4/7) Epoch 23, batch 2300, loss[loss=0.1656, simple_loss=0.2502, pruned_loss=0.04049, over 4902.00 frames. ], tot_loss[loss=0.1719, simple_loss=0.2449, pruned_loss=0.04943, over 954347.00 frames. ], batch size: 35, lr: 3.10e-03, grad_scale: 32.0 2023-04-27 18:34:37,303 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=128318.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:34:38,442 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4373, 1.2655, 1.6024, 1.6087, 1.3575, 1.2673, 1.3080, 0.7049], device='cuda:4'), covar=tensor([0.0506, 0.0697, 0.0387, 0.0555, 0.0765, 0.1138, 0.0522, 0.0643], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0068, 0.0067, 0.0068, 0.0075, 0.0095, 0.0073, 0.0064], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 18:34:39,027 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=128320.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:35:34,257 INFO [finetune.py:976] (4/7) Epoch 23, batch 2350, loss[loss=0.1746, simple_loss=0.2536, pruned_loss=0.04778, over 4854.00 frames. ], tot_loss[loss=0.1698, simple_loss=0.2424, pruned_loss=0.04859, over 952393.58 frames. ], batch size: 44, lr: 3.10e-03, grad_scale: 32.0 2023-04-27 18:35:37,976 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.014e+02 1.560e+02 1.785e+02 2.201e+02 3.854e+02, threshold=3.569e+02, percent-clipped=2.0 2023-04-27 18:35:38,688 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=128366.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:36:01,143 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=128381.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:36:31,443 INFO [finetune.py:976] (4/7) Epoch 23, batch 2400, loss[loss=0.1607, simple_loss=0.2301, pruned_loss=0.04567, over 4809.00 frames. ], tot_loss[loss=0.167, simple_loss=0.2389, pruned_loss=0.0476, over 954102.38 frames. ], batch size: 25, lr: 3.10e-03, grad_scale: 32.0 2023-04-27 18:36:34,509 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=128413.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:36:58,493 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5730, 3.6118, 1.0918, 1.9581, 2.1063, 2.5765, 2.0319, 1.2616], device='cuda:4'), covar=tensor([0.1616, 0.1693, 0.2214, 0.1513, 0.1196, 0.1320, 0.1701, 0.1931], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0238, 0.0137, 0.0119, 0.0131, 0.0150, 0.0117, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 18:37:01,393 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=128435.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:37:29,233 INFO [finetune.py:976] (4/7) Epoch 23, batch 2450, loss[loss=0.1427, simple_loss=0.223, pruned_loss=0.03123, over 4908.00 frames. ], tot_loss[loss=0.1651, simple_loss=0.2365, pruned_loss=0.0469, over 954717.20 frames. ], batch size: 43, lr: 3.10e-03, grad_scale: 32.0 2023-04-27 18:37:35,178 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.697e+01 1.581e+02 1.977e+02 2.297e+02 4.168e+02, threshold=3.955e+02, percent-clipped=2.0 2023-04-27 18:37:46,735 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=128474.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:37:56,946 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=128483.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:38:27,892 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=128506.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:38:29,613 INFO [finetune.py:976] (4/7) Epoch 23, batch 2500, loss[loss=0.1276, simple_loss=0.2028, pruned_loss=0.02616, over 4776.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.239, pruned_loss=0.04802, over 956195.85 frames. ], batch size: 28, lr: 3.10e-03, grad_scale: 32.0 2023-04-27 18:38:53,798 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=128528.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:39:29,768 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=128554.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:39:32,694 INFO [finetune.py:976] (4/7) Epoch 23, batch 2550, loss[loss=0.1354, simple_loss=0.2223, pruned_loss=0.0242, over 4903.00 frames. ], tot_loss[loss=0.1706, simple_loss=0.2425, pruned_loss=0.04932, over 955946.16 frames. ], batch size: 43, lr: 3.10e-03, grad_scale: 32.0 2023-04-27 18:39:42,964 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.076e+02 1.730e+02 1.875e+02 2.334e+02 4.772e+02, threshold=3.751e+02, percent-clipped=1.0 2023-04-27 18:39:55,800 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=128576.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:40:40,367 INFO [finetune.py:976] (4/7) Epoch 23, batch 2600, loss[loss=0.1636, simple_loss=0.2543, pruned_loss=0.03638, over 4817.00 frames. ], tot_loss[loss=0.1716, simple_loss=0.2442, pruned_loss=0.04955, over 951454.97 frames. ], batch size: 45, lr: 3.10e-03, grad_scale: 32.0 2023-04-27 18:41:01,970 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-04-27 18:41:16,252 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5100, 1.9724, 2.2688, 2.8795, 2.3550, 1.9031, 1.9401, 2.2165], device='cuda:4'), covar=tensor([0.2862, 0.3024, 0.1593, 0.2391, 0.2647, 0.2421, 0.3363, 0.2006], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0245, 0.0227, 0.0314, 0.0219, 0.0234, 0.0227, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 18:41:44,294 INFO [finetune.py:976] (4/7) Epoch 23, batch 2650, loss[loss=0.1619, simple_loss=0.2382, pruned_loss=0.04275, over 4768.00 frames. ], tot_loss[loss=0.1712, simple_loss=0.2441, pruned_loss=0.04913, over 951952.28 frames. ], batch size: 26, lr: 3.10e-03, grad_scale: 32.0 2023-04-27 18:41:53,486 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.191e+02 1.651e+02 1.881e+02 2.286e+02 5.160e+02, threshold=3.761e+02, percent-clipped=1.0 2023-04-27 18:41:54,212 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8552, 2.2003, 1.8665, 2.1215, 1.6752, 1.8437, 1.7749, 1.4569], device='cuda:4'), covar=tensor([0.1497, 0.1093, 0.0769, 0.1038, 0.3209, 0.1002, 0.1635, 0.2197], device='cuda:4'), in_proj_covar=tensor([0.0283, 0.0299, 0.0213, 0.0274, 0.0313, 0.0254, 0.0247, 0.0261], device='cuda:4'), out_proj_covar=tensor([1.1309e-04, 1.1826e-04, 8.4090e-05, 1.0823e-04, 1.2639e-04, 1.0057e-04, 9.9559e-05, 1.0314e-04], device='cuda:4') 2023-04-27 18:42:11,813 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=128676.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:42:55,228 INFO [finetune.py:976] (4/7) Epoch 23, batch 2700, loss[loss=0.2015, simple_loss=0.2638, pruned_loss=0.06963, over 4886.00 frames. ], tot_loss[loss=0.1706, simple_loss=0.2437, pruned_loss=0.0488, over 951822.05 frames. ], batch size: 35, lr: 3.10e-03, grad_scale: 32.0 2023-04-27 18:42:59,916 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-27 18:43:52,313 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9872, 0.9865, 1.6070, 1.6740, 1.6170, 1.8017, 1.6182, 1.6086], device='cuda:4'), covar=tensor([0.3819, 0.5279, 0.4582, 0.4529, 0.5473, 0.7211, 0.4618, 0.4727], device='cuda:4'), in_proj_covar=tensor([0.0339, 0.0374, 0.0326, 0.0338, 0.0347, 0.0396, 0.0357, 0.0330], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 18:44:00,968 INFO [finetune.py:976] (4/7) Epoch 23, batch 2750, loss[loss=0.1541, simple_loss=0.225, pruned_loss=0.04155, over 4759.00 frames. ], tot_loss[loss=0.1685, simple_loss=0.241, pruned_loss=0.048, over 951192.83 frames. ], batch size: 28, lr: 3.10e-03, grad_scale: 32.0 2023-04-27 18:44:04,649 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.255e+01 1.547e+02 1.823e+02 2.138e+02 3.437e+02, threshold=3.646e+02, percent-clipped=0.0 2023-04-27 18:44:12,834 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=128769.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:44:24,736 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6716, 3.5382, 1.1966, 1.9022, 2.0219, 2.6500, 1.9119, 1.1729], device='cuda:4'), covar=tensor([0.1382, 0.1046, 0.1905, 0.1317, 0.1104, 0.0940, 0.1578, 0.1762], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0239, 0.0137, 0.0119, 0.0131, 0.0151, 0.0117, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 18:44:32,559 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=128781.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:45:09,449 INFO [finetune.py:976] (4/7) Epoch 23, batch 2800, loss[loss=0.1601, simple_loss=0.2311, pruned_loss=0.04459, over 4859.00 frames. ], tot_loss[loss=0.1665, simple_loss=0.2386, pruned_loss=0.04717, over 954001.55 frames. ], batch size: 31, lr: 3.10e-03, grad_scale: 32.0 2023-04-27 18:45:55,096 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=128842.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:46:15,232 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-04-27 18:46:17,709 INFO [finetune.py:976] (4/7) Epoch 23, batch 2850, loss[loss=0.2174, simple_loss=0.2856, pruned_loss=0.07459, over 4861.00 frames. ], tot_loss[loss=0.1665, simple_loss=0.2379, pruned_loss=0.04757, over 955577.49 frames. ], batch size: 47, lr: 3.10e-03, grad_scale: 32.0 2023-04-27 18:46:22,646 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.077e+02 1.467e+02 1.826e+02 2.202e+02 4.194e+02, threshold=3.652e+02, percent-clipped=1.0 2023-04-27 18:46:32,688 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7302, 2.1031, 1.6746, 1.3703, 1.2930, 1.2794, 1.7032, 1.2057], device='cuda:4'), covar=tensor([0.1745, 0.1285, 0.1449, 0.1802, 0.2279, 0.2033, 0.1036, 0.2034], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0210, 0.0169, 0.0205, 0.0199, 0.0186, 0.0156, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 18:46:43,956 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.8292, 2.8078, 2.1719, 3.2421, 2.8130, 2.8024, 1.2036, 2.8154], device='cuda:4'), covar=tensor([0.2339, 0.1947, 0.3599, 0.3172, 0.3936, 0.2322, 0.6100, 0.2911], device='cuda:4'), in_proj_covar=tensor([0.0243, 0.0218, 0.0251, 0.0304, 0.0295, 0.0245, 0.0273, 0.0271], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 18:46:50,227 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=128890.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 18:47:11,863 INFO [finetune.py:976] (4/7) Epoch 23, batch 2900, loss[loss=0.1675, simple_loss=0.2601, pruned_loss=0.0374, over 4808.00 frames. ], tot_loss[loss=0.1683, simple_loss=0.2409, pruned_loss=0.04786, over 954410.66 frames. ], batch size: 51, lr: 3.09e-03, grad_scale: 32.0 2023-04-27 18:47:17,823 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1654, 2.1024, 1.7048, 1.8205, 2.2044, 1.8185, 2.6957, 1.5166], device='cuda:4'), covar=tensor([0.3622, 0.1932, 0.4706, 0.3021, 0.1639, 0.2436, 0.1351, 0.4422], device='cuda:4'), in_proj_covar=tensor([0.0338, 0.0348, 0.0426, 0.0351, 0.0377, 0.0375, 0.0368, 0.0419], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 18:47:36,793 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.77 vs. limit=5.0 2023-04-27 18:47:56,522 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=128951.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 18:48:04,747 INFO [finetune.py:976] (4/7) Epoch 23, batch 2950, loss[loss=0.1577, simple_loss=0.2335, pruned_loss=0.04091, over 4813.00 frames. ], tot_loss[loss=0.1691, simple_loss=0.2427, pruned_loss=0.04772, over 956222.36 frames. ], batch size: 33, lr: 3.09e-03, grad_scale: 32.0 2023-04-27 18:48:10,254 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.058e+02 1.560e+02 1.806e+02 2.158e+02 5.289e+02, threshold=3.612e+02, percent-clipped=5.0 2023-04-27 18:48:22,824 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=128976.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:48:48,717 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3171, 1.5439, 1.8877, 1.9705, 1.9032, 1.9679, 1.8888, 1.9090], device='cuda:4'), covar=tensor([0.3735, 0.4898, 0.4058, 0.3973, 0.4871, 0.6481, 0.4612, 0.4233], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0372, 0.0324, 0.0336, 0.0346, 0.0392, 0.0355, 0.0328], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 18:49:01,104 INFO [finetune.py:976] (4/7) Epoch 23, batch 3000, loss[loss=0.218, simple_loss=0.2881, pruned_loss=0.07399, over 4721.00 frames. ], tot_loss[loss=0.1712, simple_loss=0.2447, pruned_loss=0.04884, over 955192.49 frames. ], batch size: 59, lr: 3.09e-03, grad_scale: 32.0 2023-04-27 18:49:01,105 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 18:49:17,637 INFO [finetune.py:1010] (4/7) Epoch 23, validation: loss=0.1527, simple_loss=0.2222, pruned_loss=0.04158, over 2265189.00 frames. 2023-04-27 18:49:17,638 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6489MB 2023-04-27 18:49:31,983 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=129024.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:49:51,611 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5127, 3.4424, 0.9096, 1.8432, 1.8931, 2.3732, 1.8914, 1.0097], device='cuda:4'), covar=tensor([0.1392, 0.1081, 0.2073, 0.1228, 0.1088, 0.1013, 0.1541, 0.1909], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0236, 0.0136, 0.0119, 0.0130, 0.0150, 0.0116, 0.0117], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 18:50:16,854 INFO [finetune.py:976] (4/7) Epoch 23, batch 3050, loss[loss=0.1659, simple_loss=0.246, pruned_loss=0.04292, over 4769.00 frames. ], tot_loss[loss=0.1714, simple_loss=0.2452, pruned_loss=0.04879, over 952383.96 frames. ], batch size: 26, lr: 3.09e-03, grad_scale: 32.0 2023-04-27 18:50:25,142 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.041e+02 1.682e+02 1.957e+02 2.360e+02 4.958e+02, threshold=3.913e+02, percent-clipped=5.0 2023-04-27 18:50:34,466 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=129069.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:51:09,075 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.7213, 1.9258, 1.7462, 1.4545, 1.7146, 1.4847, 2.2664, 1.3808], device='cuda:4'), covar=tensor([0.3684, 0.1675, 0.4368, 0.2669, 0.1827, 0.2529, 0.1466, 0.5210], device='cuda:4'), in_proj_covar=tensor([0.0339, 0.0348, 0.0426, 0.0351, 0.0378, 0.0374, 0.0367, 0.0420], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 18:51:30,732 INFO [finetune.py:976] (4/7) Epoch 23, batch 3100, loss[loss=0.1442, simple_loss=0.2132, pruned_loss=0.03763, over 4018.00 frames. ], tot_loss[loss=0.1693, simple_loss=0.2425, pruned_loss=0.04806, over 952512.03 frames. ], batch size: 17, lr: 3.09e-03, grad_scale: 16.0 2023-04-27 18:51:41,684 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=129117.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:52:07,894 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=129137.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:52:38,082 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5185, 1.4025, 1.7844, 1.8115, 1.3643, 1.2205, 1.4950, 0.8922], device='cuda:4'), covar=tensor([0.0563, 0.0667, 0.0361, 0.0539, 0.0759, 0.1144, 0.0570, 0.0629], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0068, 0.0066, 0.0067, 0.0074, 0.0095, 0.0073, 0.0064], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 18:52:39,818 INFO [finetune.py:976] (4/7) Epoch 23, batch 3150, loss[loss=0.2018, simple_loss=0.2525, pruned_loss=0.07556, over 4820.00 frames. ], tot_loss[loss=0.1676, simple_loss=0.2398, pruned_loss=0.04768, over 951962.63 frames. ], batch size: 39, lr: 3.09e-03, grad_scale: 16.0 2023-04-27 18:52:50,085 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.045e+02 1.492e+02 1.784e+02 2.330e+02 3.707e+02, threshold=3.568e+02, percent-clipped=0.0 2023-04-27 18:52:51,109 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-27 18:53:12,401 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4410, 1.3284, 1.6595, 1.6965, 1.3308, 1.2716, 1.3577, 0.8684], device='cuda:4'), covar=tensor([0.0524, 0.0510, 0.0341, 0.0479, 0.0759, 0.1007, 0.0541, 0.0550], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0067, 0.0066, 0.0067, 0.0074, 0.0095, 0.0073, 0.0064], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 18:53:46,921 INFO [finetune.py:976] (4/7) Epoch 23, batch 3200, loss[loss=0.1317, simple_loss=0.1985, pruned_loss=0.03245, over 4703.00 frames. ], tot_loss[loss=0.1641, simple_loss=0.2361, pruned_loss=0.04606, over 953646.91 frames. ], batch size: 23, lr: 3.09e-03, grad_scale: 16.0 2023-04-27 18:53:57,280 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5051, 1.6636, 1.5823, 1.8972, 1.7936, 2.1216, 1.4436, 3.7185], device='cuda:4'), covar=tensor([0.0545, 0.0775, 0.0756, 0.1150, 0.0601, 0.0442, 0.0735, 0.0111], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 18:54:21,034 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-27 18:54:42,093 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=129246.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 18:54:55,916 INFO [finetune.py:976] (4/7) Epoch 23, batch 3250, loss[loss=0.2247, simple_loss=0.2881, pruned_loss=0.08068, over 4167.00 frames. ], tot_loss[loss=0.1648, simple_loss=0.2368, pruned_loss=0.04636, over 952834.33 frames. ], batch size: 65, lr: 3.09e-03, grad_scale: 16.0 2023-04-27 18:55:05,241 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1538, 2.1657, 1.7340, 1.7983, 2.2404, 1.8417, 2.6248, 1.6333], device='cuda:4'), covar=tensor([0.3377, 0.1507, 0.4576, 0.2558, 0.1489, 0.2139, 0.1280, 0.4013], device='cuda:4'), in_proj_covar=tensor([0.0336, 0.0346, 0.0423, 0.0347, 0.0374, 0.0371, 0.0364, 0.0416], device='cuda:4'), out_proj_covar=tensor([9.9490e-05, 1.0344e-04, 1.2837e-04, 1.0443e-04, 1.1110e-04, 1.1066e-04, 1.0686e-04, 1.2541e-04], device='cuda:4') 2023-04-27 18:55:06,949 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.754e+01 1.533e+02 1.786e+02 2.239e+02 3.821e+02, threshold=3.572e+02, percent-clipped=2.0 2023-04-27 18:55:51,326 INFO [finetune.py:976] (4/7) Epoch 23, batch 3300, loss[loss=0.2609, simple_loss=0.3328, pruned_loss=0.09448, over 4864.00 frames. ], tot_loss[loss=0.1674, simple_loss=0.2399, pruned_loss=0.04741, over 953442.03 frames. ], batch size: 34, lr: 3.09e-03, grad_scale: 16.0 2023-04-27 18:56:10,057 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2725, 1.2321, 1.3026, 1.5964, 1.6409, 1.2835, 0.9655, 1.4445], device='cuda:4'), covar=tensor([0.0807, 0.1241, 0.0872, 0.0577, 0.0580, 0.0805, 0.0788, 0.0578], device='cuda:4'), in_proj_covar=tensor([0.0186, 0.0202, 0.0183, 0.0172, 0.0177, 0.0180, 0.0150, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 18:56:49,845 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.9259, 3.9356, 2.9135, 4.4976, 3.9843, 3.7583, 1.9383, 3.9711], device='cuda:4'), covar=tensor([0.1770, 0.1304, 0.2873, 0.1446, 0.3194, 0.1944, 0.5432, 0.2208], device='cuda:4'), in_proj_covar=tensor([0.0245, 0.0218, 0.0252, 0.0305, 0.0296, 0.0246, 0.0273, 0.0273], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 18:56:51,619 INFO [finetune.py:976] (4/7) Epoch 23, batch 3350, loss[loss=0.1643, simple_loss=0.2328, pruned_loss=0.04792, over 4699.00 frames. ], tot_loss[loss=0.1686, simple_loss=0.2416, pruned_loss=0.04778, over 949897.67 frames. ], batch size: 23, lr: 3.09e-03, grad_scale: 16.0 2023-04-27 18:57:02,541 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.057e+02 1.708e+02 1.958e+02 2.263e+02 4.173e+02, threshold=3.917e+02, percent-clipped=1.0 2023-04-27 18:57:20,721 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.9030, 4.8225, 3.3549, 5.5165, 4.9035, 4.7205, 2.0835, 4.9276], device='cuda:4'), covar=tensor([0.1391, 0.1014, 0.2495, 0.0730, 0.3518, 0.1525, 0.5619, 0.1807], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0218, 0.0252, 0.0305, 0.0296, 0.0245, 0.0273, 0.0273], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 18:57:58,594 INFO [finetune.py:976] (4/7) Epoch 23, batch 3400, loss[loss=0.204, simple_loss=0.2713, pruned_loss=0.06834, over 4895.00 frames. ], tot_loss[loss=0.1691, simple_loss=0.2425, pruned_loss=0.04786, over 950571.71 frames. ], batch size: 36, lr: 3.09e-03, grad_scale: 16.0 2023-04-27 18:58:06,857 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=129413.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 18:58:40,738 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=129437.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 18:59:06,102 INFO [finetune.py:976] (4/7) Epoch 23, batch 3450, loss[loss=0.1344, simple_loss=0.2032, pruned_loss=0.03286, over 4800.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2425, pruned_loss=0.04766, over 952491.48 frames. ], batch size: 25, lr: 3.09e-03, grad_scale: 16.0 2023-04-27 18:59:15,873 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.992e+01 1.559e+02 1.867e+02 2.155e+02 3.707e+02, threshold=3.734e+02, percent-clipped=0.0 2023-04-27 18:59:26,956 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=129474.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 18:59:40,157 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=129485.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:00:07,512 INFO [finetune.py:976] (4/7) Epoch 23, batch 3500, loss[loss=0.1372, simple_loss=0.2087, pruned_loss=0.03289, over 4846.00 frames. ], tot_loss[loss=0.1682, simple_loss=0.2409, pruned_loss=0.04774, over 953696.62 frames. ], batch size: 49, lr: 3.09e-03, grad_scale: 16.0 2023-04-27 19:00:46,496 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=129546.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 19:01:00,755 INFO [finetune.py:976] (4/7) Epoch 23, batch 3550, loss[loss=0.1534, simple_loss=0.2253, pruned_loss=0.04072, over 4900.00 frames. ], tot_loss[loss=0.167, simple_loss=0.2388, pruned_loss=0.04765, over 955089.06 frames. ], batch size: 32, lr: 3.09e-03, grad_scale: 16.0 2023-04-27 19:01:08,388 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.192e+01 1.610e+02 1.893e+02 2.308e+02 5.470e+02, threshold=3.785e+02, percent-clipped=3.0 2023-04-27 19:01:14,619 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0155, 3.8204, 1.4041, 2.0849, 2.3893, 2.7251, 2.3590, 1.4347], device='cuda:4'), covar=tensor([0.1226, 0.1206, 0.1720, 0.1164, 0.0901, 0.1006, 0.1241, 0.1626], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0239, 0.0136, 0.0120, 0.0132, 0.0151, 0.0117, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 19:01:38,378 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=129594.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 19:01:38,403 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0873, 2.6012, 1.2799, 1.5109, 2.1670, 1.4248, 3.1077, 1.6829], device='cuda:4'), covar=tensor([0.0680, 0.1004, 0.0822, 0.1143, 0.0442, 0.0912, 0.0242, 0.0607], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0072, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0009, 0.0007, 0.0007, 0.0008, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-27 19:01:57,767 INFO [finetune.py:976] (4/7) Epoch 23, batch 3600, loss[loss=0.1689, simple_loss=0.2383, pruned_loss=0.04972, over 4905.00 frames. ], tot_loss[loss=0.1655, simple_loss=0.2363, pruned_loss=0.04731, over 954720.38 frames. ], batch size: 35, lr: 3.09e-03, grad_scale: 16.0 2023-04-27 19:01:59,683 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=129612.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:02:47,465 INFO [finetune.py:976] (4/7) Epoch 23, batch 3650, loss[loss=0.1506, simple_loss=0.2328, pruned_loss=0.03418, over 4914.00 frames. ], tot_loss[loss=0.1682, simple_loss=0.2396, pruned_loss=0.04839, over 955452.89 frames. ], batch size: 36, lr: 3.09e-03, grad_scale: 16.0 2023-04-27 19:02:51,895 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.776e+01 1.523e+02 1.795e+02 2.282e+02 4.473e+02, threshold=3.590e+02, percent-clipped=4.0 2023-04-27 19:03:02,083 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=129673.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:03:03,962 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9217, 2.5406, 1.9712, 1.7827, 1.4494, 1.4269, 1.9369, 1.3904], device='cuda:4'), covar=tensor([0.1712, 0.1333, 0.1341, 0.1679, 0.2280, 0.1959, 0.0992, 0.2020], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0211, 0.0169, 0.0204, 0.0199, 0.0186, 0.0156, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 19:03:46,496 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2926, 2.1097, 2.3321, 2.9292, 2.8092, 2.2889, 2.0378, 2.4356], device='cuda:4'), covar=tensor([0.0785, 0.0942, 0.0651, 0.0478, 0.0553, 0.0755, 0.0661, 0.0544], device='cuda:4'), in_proj_covar=tensor([0.0187, 0.0202, 0.0184, 0.0172, 0.0176, 0.0180, 0.0150, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 19:03:55,980 INFO [finetune.py:976] (4/7) Epoch 23, batch 3700, loss[loss=0.1953, simple_loss=0.2738, pruned_loss=0.05836, over 4764.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2439, pruned_loss=0.04918, over 954781.03 frames. ], batch size: 28, lr: 3.09e-03, grad_scale: 16.0 2023-04-27 19:05:03,362 INFO [finetune.py:976] (4/7) Epoch 23, batch 3750, loss[loss=0.1902, simple_loss=0.2563, pruned_loss=0.06205, over 4798.00 frames. ], tot_loss[loss=0.1736, simple_loss=0.246, pruned_loss=0.05058, over 952546.47 frames. ], batch size: 51, lr: 3.09e-03, grad_scale: 16.0 2023-04-27 19:05:12,868 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.105e+02 1.590e+02 1.958e+02 2.395e+02 5.557e+02, threshold=3.915e+02, percent-clipped=3.0 2023-04-27 19:05:15,279 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=129769.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 19:05:45,762 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=129791.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:06:08,151 INFO [finetune.py:976] (4/7) Epoch 23, batch 3800, loss[loss=0.1574, simple_loss=0.2368, pruned_loss=0.03903, over 4912.00 frames. ], tot_loss[loss=0.175, simple_loss=0.2482, pruned_loss=0.05096, over 955356.88 frames. ], batch size: 33, lr: 3.09e-03, grad_scale: 16.0 2023-04-27 19:06:21,479 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.4917, 3.4566, 2.9629, 4.0636, 3.3872, 3.4510, 1.8563, 3.5437], device='cuda:4'), covar=tensor([0.1675, 0.1320, 0.3644, 0.1353, 0.3111, 0.1657, 0.4544, 0.2323], device='cuda:4'), in_proj_covar=tensor([0.0245, 0.0217, 0.0253, 0.0304, 0.0295, 0.0245, 0.0273, 0.0273], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 19:07:04,473 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=129852.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:07:12,848 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6927, 2.0499, 1.7737, 2.0413, 1.5130, 1.8216, 1.7379, 1.4935], device='cuda:4'), covar=tensor([0.1595, 0.0920, 0.0710, 0.0890, 0.3073, 0.0891, 0.1542, 0.1941], device='cuda:4'), in_proj_covar=tensor([0.0281, 0.0298, 0.0212, 0.0273, 0.0310, 0.0253, 0.0245, 0.0259], device='cuda:4'), out_proj_covar=tensor([1.1230e-04, 1.1774e-04, 8.3541e-05, 1.0752e-04, 1.2537e-04, 9.9845e-05, 9.8668e-05, 1.0231e-04], device='cuda:4') 2023-04-27 19:07:12,976 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.97 vs. limit=5.0 2023-04-27 19:07:13,927 INFO [finetune.py:976] (4/7) Epoch 23, batch 3850, loss[loss=0.1811, simple_loss=0.2449, pruned_loss=0.05869, over 4820.00 frames. ], tot_loss[loss=0.173, simple_loss=0.2458, pruned_loss=0.05012, over 956461.19 frames. ], batch size: 33, lr: 3.09e-03, grad_scale: 16.0 2023-04-27 19:07:23,111 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8860, 2.2417, 1.9733, 2.2897, 1.6772, 2.0735, 1.9465, 1.6420], device='cuda:4'), covar=tensor([0.1864, 0.1169, 0.0748, 0.1059, 0.3134, 0.1036, 0.1815, 0.2201], device='cuda:4'), in_proj_covar=tensor([0.0281, 0.0298, 0.0212, 0.0273, 0.0310, 0.0253, 0.0245, 0.0259], device='cuda:4'), out_proj_covar=tensor([1.1229e-04, 1.1777e-04, 8.3541e-05, 1.0752e-04, 1.2542e-04, 9.9865e-05, 9.8670e-05, 1.0232e-04], device='cuda:4') 2023-04-27 19:07:24,864 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.155e+02 1.706e+02 1.886e+02 2.157e+02 5.437e+02, threshold=3.771e+02, percent-clipped=4.0 2023-04-27 19:08:15,872 INFO [finetune.py:976] (4/7) Epoch 23, batch 3900, loss[loss=0.1395, simple_loss=0.2179, pruned_loss=0.03053, over 4813.00 frames. ], tot_loss[loss=0.1701, simple_loss=0.2427, pruned_loss=0.04876, over 956825.58 frames. ], batch size: 38, lr: 3.09e-03, grad_scale: 16.0 2023-04-27 19:09:17,305 INFO [finetune.py:976] (4/7) Epoch 23, batch 3950, loss[loss=0.1254, simple_loss=0.2007, pruned_loss=0.02507, over 4904.00 frames. ], tot_loss[loss=0.1676, simple_loss=0.2391, pruned_loss=0.04808, over 957794.32 frames. ], batch size: 35, lr: 3.09e-03, grad_scale: 16.0 2023-04-27 19:09:25,099 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7427, 1.7031, 4.6965, 4.4297, 4.1292, 4.4975, 4.3034, 4.0699], device='cuda:4'), covar=tensor([0.7366, 0.5487, 0.0915, 0.1339, 0.1087, 0.2097, 0.1288, 0.1661], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0306, 0.0406, 0.0406, 0.0346, 0.0407, 0.0315, 0.0366], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 19:09:27,939 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.540e+01 1.458e+02 1.749e+02 1.978e+02 4.242e+02, threshold=3.497e+02, percent-clipped=1.0 2023-04-27 19:09:35,336 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=129968.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:10:16,061 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=130000.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:10:22,145 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=130002.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:10:26,044 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=2.01 vs. limit=2.0 2023-04-27 19:10:33,621 INFO [finetune.py:976] (4/7) Epoch 23, batch 4000, loss[loss=0.167, simple_loss=0.2557, pruned_loss=0.03915, over 4842.00 frames. ], tot_loss[loss=0.1679, simple_loss=0.2391, pruned_loss=0.04835, over 954880.81 frames. ], batch size: 49, lr: 3.09e-03, grad_scale: 16.0 2023-04-27 19:10:59,433 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.14 vs. limit=5.0 2023-04-27 19:11:22,122 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5936, 1.4228, 1.2393, 1.5271, 1.8398, 1.5088, 1.3397, 1.1433], device='cuda:4'), covar=tensor([0.1793, 0.1497, 0.1815, 0.1409, 0.0957, 0.1869, 0.2362, 0.2344], device='cuda:4'), in_proj_covar=tensor([0.0316, 0.0312, 0.0354, 0.0288, 0.0327, 0.0311, 0.0302, 0.0374], device='cuda:4'), out_proj_covar=tensor([6.4810e-05, 6.4517e-05, 7.4575e-05, 5.7745e-05, 6.7261e-05, 6.5219e-05, 6.2956e-05, 7.9494e-05], device='cuda:4') 2023-04-27 19:11:42,777 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4629, 1.3869, 1.8550, 1.8542, 1.3366, 1.2835, 1.5661, 0.9424], device='cuda:4'), covar=tensor([0.0593, 0.0674, 0.0336, 0.0564, 0.0805, 0.1077, 0.0530, 0.0638], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0068, 0.0067, 0.0068, 0.0075, 0.0096, 0.0073, 0.0064], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 19:11:43,268 INFO [finetune.py:976] (4/7) Epoch 23, batch 4050, loss[loss=0.1728, simple_loss=0.251, pruned_loss=0.04728, over 4895.00 frames. ], tot_loss[loss=0.1687, simple_loss=0.2403, pruned_loss=0.04851, over 953545.45 frames. ], batch size: 37, lr: 3.09e-03, grad_scale: 16.0 2023-04-27 19:11:44,626 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=130061.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:11:45,895 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=130063.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 19:11:53,610 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.121e+02 1.688e+02 1.937e+02 2.332e+02 4.574e+02, threshold=3.874e+02, percent-clipped=1.0 2023-04-27 19:11:56,081 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=130069.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 19:12:52,034 INFO [finetune.py:976] (4/7) Epoch 23, batch 4100, loss[loss=0.2473, simple_loss=0.3181, pruned_loss=0.08825, over 4141.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2437, pruned_loss=0.04924, over 953573.86 frames. ], batch size: 65, lr: 3.09e-03, grad_scale: 16.0 2023-04-27 19:13:01,824 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=130117.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 19:13:32,158 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=130147.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:13:51,930 INFO [finetune.py:976] (4/7) Epoch 23, batch 4150, loss[loss=0.173, simple_loss=0.2562, pruned_loss=0.04489, over 4797.00 frames. ], tot_loss[loss=0.1721, simple_loss=0.2452, pruned_loss=0.04948, over 953901.88 frames. ], batch size: 45, lr: 3.09e-03, grad_scale: 16.0 2023-04-27 19:14:02,701 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.039e+02 1.725e+02 1.972e+02 2.342e+02 5.183e+02, threshold=3.944e+02, percent-clipped=4.0 2023-04-27 19:14:48,093 INFO [finetune.py:976] (4/7) Epoch 23, batch 4200, loss[loss=0.159, simple_loss=0.2403, pruned_loss=0.03886, over 4815.00 frames. ], tot_loss[loss=0.1703, simple_loss=0.2438, pruned_loss=0.04839, over 954802.59 frames. ], batch size: 38, lr: 3.09e-03, grad_scale: 16.0 2023-04-27 19:15:09,576 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1918, 3.2970, 0.7308, 1.5588, 1.6861, 2.3092, 1.8249, 1.0665], device='cuda:4'), covar=tensor([0.1785, 0.1485, 0.2610, 0.1730, 0.1293, 0.1381, 0.1686, 0.2083], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0238, 0.0136, 0.0119, 0.0131, 0.0151, 0.0117, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 19:15:20,598 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7190, 1.6741, 2.0810, 2.1967, 1.6420, 1.4901, 1.7989, 1.1339], device='cuda:4'), covar=tensor([0.0641, 0.0699, 0.0552, 0.0843, 0.0767, 0.1096, 0.0777, 0.0769], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0068, 0.0067, 0.0068, 0.0075, 0.0096, 0.0073, 0.0064], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 19:15:51,649 INFO [finetune.py:976] (4/7) Epoch 23, batch 4250, loss[loss=0.1363, simple_loss=0.2056, pruned_loss=0.03354, over 4866.00 frames. ], tot_loss[loss=0.1687, simple_loss=0.2416, pruned_loss=0.04789, over 955576.29 frames. ], batch size: 34, lr: 3.08e-03, grad_scale: 16.0 2023-04-27 19:15:53,502 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.8225, 2.7576, 2.9840, 3.3694, 3.1857, 2.9170, 2.2306, 2.9937], device='cuda:4'), covar=tensor([0.0761, 0.0811, 0.0550, 0.0534, 0.0566, 0.0711, 0.0720, 0.0532], device='cuda:4'), in_proj_covar=tensor([0.0186, 0.0202, 0.0183, 0.0173, 0.0176, 0.0179, 0.0149, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 19:16:01,962 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.003e+02 1.509e+02 1.751e+02 2.292e+02 3.450e+02, threshold=3.503e+02, percent-clipped=0.0 2023-04-27 19:16:03,326 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=130268.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:16:55,314 INFO [finetune.py:976] (4/7) Epoch 23, batch 4300, loss[loss=0.1706, simple_loss=0.2392, pruned_loss=0.05097, over 4909.00 frames. ], tot_loss[loss=0.167, simple_loss=0.2392, pruned_loss=0.0474, over 954390.21 frames. ], batch size: 37, lr: 3.08e-03, grad_scale: 16.0 2023-04-27 19:17:02,226 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.41 vs. limit=5.0 2023-04-27 19:17:05,108 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=130316.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:17:58,667 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=130356.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:18:05,095 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=130358.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 19:18:05,660 INFO [finetune.py:976] (4/7) Epoch 23, batch 4350, loss[loss=0.1626, simple_loss=0.2356, pruned_loss=0.04483, over 4812.00 frames. ], tot_loss[loss=0.1652, simple_loss=0.2367, pruned_loss=0.04691, over 954919.41 frames. ], batch size: 25, lr: 3.08e-03, grad_scale: 16.0 2023-04-27 19:18:10,117 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.083e+02 1.587e+02 1.809e+02 2.203e+02 4.398e+02, threshold=3.619e+02, percent-clipped=3.0 2023-04-27 19:18:29,071 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=130384.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:19:02,658 INFO [finetune.py:976] (4/7) Epoch 23, batch 4400, loss[loss=0.1847, simple_loss=0.2647, pruned_loss=0.05239, over 4811.00 frames. ], tot_loss[loss=0.1666, simple_loss=0.2382, pruned_loss=0.04748, over 955057.65 frames. ], batch size: 51, lr: 3.08e-03, grad_scale: 16.0 2023-04-27 19:19:34,785 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=130432.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:19:45,163 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=130438.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:19:55,343 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=130445.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:19:56,558 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=130447.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:20:16,250 INFO [finetune.py:976] (4/7) Epoch 23, batch 4450, loss[loss=0.1573, simple_loss=0.2333, pruned_loss=0.04066, over 4748.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2415, pruned_loss=0.04816, over 957298.23 frames. ], batch size: 27, lr: 3.08e-03, grad_scale: 16.0 2023-04-27 19:20:25,828 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.347e+01 1.587e+02 1.803e+02 2.328e+02 4.838e+02, threshold=3.606e+02, percent-clipped=2.0 2023-04-27 19:21:01,293 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=130493.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:21:02,959 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=130495.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:21:10,490 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.55 vs. limit=2.0 2023-04-27 19:21:11,850 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=130499.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:21:23,654 INFO [finetune.py:976] (4/7) Epoch 23, batch 4500, loss[loss=0.1697, simple_loss=0.2329, pruned_loss=0.05325, over 4233.00 frames. ], tot_loss[loss=0.1699, simple_loss=0.2429, pruned_loss=0.04839, over 956690.61 frames. ], batch size: 65, lr: 3.08e-03, grad_scale: 16.0 2023-04-27 19:21:43,546 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-04-27 19:22:17,956 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=130548.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:22:31,479 INFO [finetune.py:976] (4/7) Epoch 23, batch 4550, loss[loss=0.1891, simple_loss=0.2646, pruned_loss=0.05678, over 4789.00 frames. ], tot_loss[loss=0.1704, simple_loss=0.2442, pruned_loss=0.04836, over 955790.80 frames. ], batch size: 51, lr: 3.08e-03, grad_scale: 16.0 2023-04-27 19:22:42,285 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=130565.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:22:42,754 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.159e+02 1.614e+02 1.836e+02 2.049e+02 5.292e+02, threshold=3.672e+02, percent-clipped=1.0 2023-04-27 19:23:36,954 INFO [finetune.py:976] (4/7) Epoch 23, batch 4600, loss[loss=0.1598, simple_loss=0.2223, pruned_loss=0.04861, over 4767.00 frames. ], tot_loss[loss=0.1703, simple_loss=0.2441, pruned_loss=0.04826, over 956171.69 frames. ], batch size: 27, lr: 3.08e-03, grad_scale: 16.0 2023-04-27 19:23:37,098 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=130609.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 19:23:54,132 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5308, 1.3023, 4.3084, 4.0159, 3.7818, 4.1002, 4.0400, 3.8133], device='cuda:4'), covar=tensor([0.7128, 0.6094, 0.1033, 0.1666, 0.1180, 0.1514, 0.1484, 0.1458], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0309, 0.0408, 0.0410, 0.0349, 0.0412, 0.0318, 0.0369], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 19:23:56,101 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=130626.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:24:36,990 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=130656.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:24:38,757 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=130658.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 19:24:39,265 INFO [finetune.py:976] (4/7) Epoch 23, batch 4650, loss[loss=0.1643, simple_loss=0.2327, pruned_loss=0.04797, over 4719.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.2424, pruned_loss=0.04825, over 955809.84 frames. ], batch size: 54, lr: 3.08e-03, grad_scale: 16.0 2023-04-27 19:24:48,332 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.298e+01 1.478e+02 1.669e+02 2.060e+02 3.844e+02, threshold=3.337e+02, percent-clipped=2.0 2023-04-27 19:25:26,884 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.85 vs. limit=5.0 2023-04-27 19:25:39,304 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=130704.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:25:40,552 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=130706.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:25:47,205 INFO [finetune.py:976] (4/7) Epoch 23, batch 4700, loss[loss=0.1399, simple_loss=0.2196, pruned_loss=0.0301, over 4761.00 frames. ], tot_loss[loss=0.169, simple_loss=0.2409, pruned_loss=0.0485, over 956806.04 frames. ], batch size: 26, lr: 3.08e-03, grad_scale: 16.0 2023-04-27 19:26:22,720 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=130740.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:26:50,255 INFO [finetune.py:976] (4/7) Epoch 23, batch 4750, loss[loss=0.1726, simple_loss=0.2523, pruned_loss=0.04647, over 4811.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.2382, pruned_loss=0.04775, over 954789.29 frames. ], batch size: 38, lr: 3.08e-03, grad_scale: 16.0 2023-04-27 19:27:00,090 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.073e+02 1.517e+02 1.838e+02 2.137e+02 3.767e+02, threshold=3.677e+02, percent-clipped=3.0 2023-04-27 19:27:24,059 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=130788.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:27:33,676 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=130794.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:27:54,320 INFO [finetune.py:976] (4/7) Epoch 23, batch 4800, loss[loss=0.2087, simple_loss=0.2897, pruned_loss=0.06381, over 4835.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.2413, pruned_loss=0.04884, over 955410.56 frames. ], batch size: 49, lr: 3.08e-03, grad_scale: 16.0 2023-04-27 19:27:55,747 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-27 19:28:32,635 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-04-27 19:28:54,604 INFO [finetune.py:976] (4/7) Epoch 23, batch 4850, loss[loss=0.1487, simple_loss=0.2263, pruned_loss=0.0355, over 4751.00 frames. ], tot_loss[loss=0.1721, simple_loss=0.2448, pruned_loss=0.04969, over 954403.02 frames. ], batch size: 59, lr: 3.08e-03, grad_scale: 16.0 2023-04-27 19:28:55,410 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4099, 1.6463, 1.6187, 1.8720, 1.8441, 2.0708, 1.4537, 4.2777], device='cuda:4'), covar=tensor([0.0548, 0.0821, 0.0776, 0.1192, 0.0622, 0.0559, 0.0769, 0.0125], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 19:29:04,838 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.28 vs. limit=5.0 2023-04-27 19:29:05,296 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.174e+02 1.564e+02 1.952e+02 2.266e+02 5.605e+02, threshold=3.905e+02, percent-clipped=3.0 2023-04-27 19:30:00,409 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=130904.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 19:30:03,975 INFO [finetune.py:976] (4/7) Epoch 23, batch 4900, loss[loss=0.1914, simple_loss=0.2595, pruned_loss=0.06164, over 4793.00 frames. ], tot_loss[loss=0.1721, simple_loss=0.2452, pruned_loss=0.04945, over 951859.50 frames. ], batch size: 51, lr: 3.08e-03, grad_scale: 16.0 2023-04-27 19:30:23,697 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=130921.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:30:25,595 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9293, 2.3363, 1.0002, 1.2318, 1.6978, 1.1371, 2.8316, 1.3453], device='cuda:4'), covar=tensor([0.0657, 0.0624, 0.0761, 0.1198, 0.0510, 0.1012, 0.0252, 0.0682], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0072, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0009, 0.0007, 0.0007, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 19:30:56,225 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-27 19:31:10,177 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-04-27 19:31:11,883 INFO [finetune.py:976] (4/7) Epoch 23, batch 4950, loss[loss=0.1687, simple_loss=0.2539, pruned_loss=0.04173, over 4812.00 frames. ], tot_loss[loss=0.1734, simple_loss=0.2465, pruned_loss=0.0501, over 951465.27 frames. ], batch size: 38, lr: 3.08e-03, grad_scale: 16.0 2023-04-27 19:31:21,965 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.168e+02 1.568e+02 1.814e+02 2.171e+02 6.365e+02, threshold=3.628e+02, percent-clipped=1.0 2023-04-27 19:32:04,080 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-27 19:32:26,083 INFO [finetune.py:976] (4/7) Epoch 23, batch 5000, loss[loss=0.1748, simple_loss=0.2288, pruned_loss=0.0604, over 4833.00 frames. ], tot_loss[loss=0.1714, simple_loss=0.2441, pruned_loss=0.04938, over 952088.07 frames. ], batch size: 25, lr: 3.08e-03, grad_scale: 16.0 2023-04-27 19:32:59,700 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2720, 2.6712, 2.1885, 2.5866, 1.9962, 2.2760, 2.3040, 1.8240], device='cuda:4'), covar=tensor([0.1870, 0.1175, 0.0822, 0.1068, 0.2966, 0.1252, 0.2041, 0.2420], device='cuda:4'), in_proj_covar=tensor([0.0283, 0.0300, 0.0214, 0.0274, 0.0315, 0.0255, 0.0248, 0.0261], device='cuda:4'), out_proj_covar=tensor([1.1333e-04, 1.1877e-04, 8.4374e-05, 1.0794e-04, 1.2706e-04, 1.0080e-04, 9.9857e-05, 1.0308e-04], device='cuda:4') 2023-04-27 19:33:10,302 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=131040.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:33:35,491 INFO [finetune.py:976] (4/7) Epoch 23, batch 5050, loss[loss=0.1583, simple_loss=0.2248, pruned_loss=0.04592, over 4870.00 frames. ], tot_loss[loss=0.1687, simple_loss=0.2409, pruned_loss=0.04826, over 951761.63 frames. ], batch size: 49, lr: 3.08e-03, grad_scale: 16.0 2023-04-27 19:33:41,158 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.129e+02 1.462e+02 1.838e+02 2.219e+02 3.974e+02, threshold=3.675e+02, percent-clipped=2.0 2023-04-27 19:33:41,996 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7834, 2.1242, 1.8273, 2.0695, 1.6477, 1.7982, 1.7798, 1.4454], device='cuda:4'), covar=tensor([0.1590, 0.0976, 0.0781, 0.0808, 0.3216, 0.1089, 0.1695, 0.2194], device='cuda:4'), in_proj_covar=tensor([0.0284, 0.0301, 0.0215, 0.0275, 0.0316, 0.0256, 0.0248, 0.0263], device='cuda:4'), out_proj_covar=tensor([1.1374e-04, 1.1927e-04, 8.4736e-05, 1.0847e-04, 1.2757e-04, 1.0125e-04, 1.0007e-04, 1.0363e-04], device='cuda:4') 2023-04-27 19:34:14,398 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=131088.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:34:14,425 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=131088.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:34:15,337 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.14 vs. limit=5.0 2023-04-27 19:34:18,298 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=131094.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:34:38,867 INFO [finetune.py:976] (4/7) Epoch 23, batch 5100, loss[loss=0.1521, simple_loss=0.2191, pruned_loss=0.04261, over 4784.00 frames. ], tot_loss[loss=0.1659, simple_loss=0.2376, pruned_loss=0.0471, over 951667.16 frames. ], batch size: 26, lr: 3.08e-03, grad_scale: 32.0 2023-04-27 19:35:11,741 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=131136.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:35:21,311 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=131142.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:35:41,502 INFO [finetune.py:976] (4/7) Epoch 23, batch 5150, loss[loss=0.1928, simple_loss=0.2598, pruned_loss=0.06288, over 4873.00 frames. ], tot_loss[loss=0.1686, simple_loss=0.2397, pruned_loss=0.04871, over 949050.33 frames. ], batch size: 34, lr: 3.08e-03, grad_scale: 32.0 2023-04-27 19:35:51,238 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.008e+02 1.590e+02 1.900e+02 2.291e+02 6.669e+02, threshold=3.800e+02, percent-clipped=5.0 2023-04-27 19:36:02,295 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=131174.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:36:11,467 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4503, 1.5896, 1.5206, 1.7088, 1.7107, 2.0594, 1.4116, 3.6910], device='cuda:4'), covar=tensor([0.0542, 0.0785, 0.0770, 0.1229, 0.0614, 0.0470, 0.0721, 0.0170], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 19:36:32,811 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9143, 2.3790, 1.0809, 1.3360, 1.9374, 1.1970, 3.1628, 1.5494], device='cuda:4'), covar=tensor([0.0692, 0.0585, 0.0778, 0.1227, 0.0507, 0.1002, 0.0219, 0.0655], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0050, 0.0072, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0009, 0.0007, 0.0007, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 19:36:42,371 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=131204.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 19:36:45,367 INFO [finetune.py:976] (4/7) Epoch 23, batch 5200, loss[loss=0.1986, simple_loss=0.2741, pruned_loss=0.06158, over 4816.00 frames. ], tot_loss[loss=0.1704, simple_loss=0.242, pruned_loss=0.04938, over 948638.90 frames. ], batch size: 45, lr: 3.08e-03, grad_scale: 32.0 2023-04-27 19:37:02,218 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-27 19:37:04,428 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=131221.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:37:22,656 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=131235.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 19:37:43,652 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=131252.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:37:52,695 INFO [finetune.py:976] (4/7) Epoch 23, batch 5250, loss[loss=0.1718, simple_loss=0.2481, pruned_loss=0.0477, over 4819.00 frames. ], tot_loss[loss=0.1724, simple_loss=0.2444, pruned_loss=0.0502, over 947530.19 frames. ], batch size: 33, lr: 3.08e-03, grad_scale: 32.0 2023-04-27 19:37:57,555 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.170e+02 1.647e+02 1.953e+02 2.357e+02 4.289e+02, threshold=3.906e+02, percent-clipped=1.0 2023-04-27 19:38:04,768 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=131269.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:38:50,170 INFO [finetune.py:976] (4/7) Epoch 23, batch 5300, loss[loss=0.1788, simple_loss=0.2595, pruned_loss=0.04901, over 4897.00 frames. ], tot_loss[loss=0.1737, simple_loss=0.2461, pruned_loss=0.05066, over 948535.88 frames. ], batch size: 32, lr: 3.08e-03, grad_scale: 32.0 2023-04-27 19:39:44,646 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2585, 1.6419, 1.5898, 1.9194, 1.9421, 2.1184, 1.4730, 4.2456], device='cuda:4'), covar=tensor([0.0564, 0.0824, 0.0786, 0.1223, 0.0606, 0.0514, 0.0739, 0.0111], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 19:39:56,070 INFO [finetune.py:976] (4/7) Epoch 23, batch 5350, loss[loss=0.1503, simple_loss=0.2117, pruned_loss=0.04439, over 4252.00 frames. ], tot_loss[loss=0.1739, simple_loss=0.2467, pruned_loss=0.05057, over 947784.99 frames. ], batch size: 18, lr: 3.08e-03, grad_scale: 32.0 2023-04-27 19:40:05,445 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7640, 2.1581, 1.7761, 2.0552, 1.5325, 1.8325, 1.7537, 1.4542], device='cuda:4'), covar=tensor([0.2150, 0.1579, 0.1166, 0.1454, 0.3633, 0.1385, 0.2194, 0.2453], device='cuda:4'), in_proj_covar=tensor([0.0282, 0.0298, 0.0213, 0.0273, 0.0312, 0.0254, 0.0246, 0.0260], device='cuda:4'), out_proj_covar=tensor([1.1273e-04, 1.1780e-04, 8.3955e-05, 1.0770e-04, 1.2615e-04, 1.0038e-04, 9.9100e-05, 1.0264e-04], device='cuda:4') 2023-04-27 19:40:07,008 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.812e+01 1.537e+02 1.790e+02 2.239e+02 4.084e+02, threshold=3.580e+02, percent-clipped=2.0 2023-04-27 19:40:52,094 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.53 vs. limit=2.0 2023-04-27 19:41:03,583 INFO [finetune.py:976] (4/7) Epoch 23, batch 5400, loss[loss=0.1647, simple_loss=0.2453, pruned_loss=0.04201, over 4788.00 frames. ], tot_loss[loss=0.1721, simple_loss=0.2447, pruned_loss=0.0498, over 947618.11 frames. ], batch size: 51, lr: 3.08e-03, grad_scale: 32.0 2023-04-27 19:41:03,703 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=131409.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:41:12,757 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=131416.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:42:12,940 INFO [finetune.py:976] (4/7) Epoch 23, batch 5450, loss[loss=0.1436, simple_loss=0.2204, pruned_loss=0.03339, over 4845.00 frames. ], tot_loss[loss=0.1709, simple_loss=0.2424, pruned_loss=0.04964, over 950528.14 frames. ], batch size: 30, lr: 3.08e-03, grad_scale: 32.0 2023-04-27 19:42:22,593 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.036e+02 1.575e+02 1.987e+02 2.572e+02 5.206e+02, threshold=3.973e+02, percent-clipped=5.0 2023-04-27 19:42:31,924 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=131470.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:42:43,231 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=131477.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:43:05,104 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=131492.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:43:22,051 INFO [finetune.py:976] (4/7) Epoch 23, batch 5500, loss[loss=0.1868, simple_loss=0.2716, pruned_loss=0.05106, over 4906.00 frames. ], tot_loss[loss=0.167, simple_loss=0.238, pruned_loss=0.04802, over 950349.51 frames. ], batch size: 43, lr: 3.08e-03, grad_scale: 32.0 2023-04-27 19:43:52,914 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=131530.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 19:44:16,336 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=131553.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:44:21,723 INFO [finetune.py:976] (4/7) Epoch 23, batch 5550, loss[loss=0.149, simple_loss=0.2125, pruned_loss=0.04274, over 4222.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.2387, pruned_loss=0.04816, over 952374.07 frames. ], batch size: 18, lr: 3.08e-03, grad_scale: 32.0 2023-04-27 19:44:32,380 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.039e+02 1.589e+02 1.849e+02 2.302e+02 7.246e+02, threshold=3.698e+02, percent-clipped=1.0 2023-04-27 19:44:39,033 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5305, 0.6948, 1.5488, 1.9413, 1.6187, 1.4663, 1.5237, 1.5104], device='cuda:4'), covar=tensor([0.3781, 0.5696, 0.4839, 0.4940, 0.4907, 0.6321, 0.6349, 0.6793], device='cuda:4'), in_proj_covar=tensor([0.0434, 0.0418, 0.0512, 0.0507, 0.0463, 0.0496, 0.0499, 0.0512], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 19:45:11,842 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9973, 2.2935, 1.1439, 1.3533, 1.8209, 1.1454, 3.1190, 1.5377], device='cuda:4'), covar=tensor([0.0659, 0.0708, 0.0742, 0.1137, 0.0487, 0.0993, 0.0206, 0.0600], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0072, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0007, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 19:45:32,466 INFO [finetune.py:976] (4/7) Epoch 23, batch 5600, loss[loss=0.156, simple_loss=0.2292, pruned_loss=0.04144, over 4822.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2428, pruned_loss=0.04969, over 953086.70 frames. ], batch size: 25, lr: 3.07e-03, grad_scale: 32.0 2023-04-27 19:46:28,599 INFO [finetune.py:976] (4/7) Epoch 23, batch 5650, loss[loss=0.1712, simple_loss=0.2604, pruned_loss=0.04102, over 4826.00 frames. ], tot_loss[loss=0.1712, simple_loss=0.2439, pruned_loss=0.04927, over 954073.38 frames. ], batch size: 33, lr: 3.07e-03, grad_scale: 32.0 2023-04-27 19:46:38,006 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.000e+01 1.549e+02 1.818e+02 2.173e+02 4.178e+02, threshold=3.635e+02, percent-clipped=2.0 2023-04-27 19:47:29,410 INFO [finetune.py:976] (4/7) Epoch 23, batch 5700, loss[loss=0.1395, simple_loss=0.1929, pruned_loss=0.04308, over 4161.00 frames. ], tot_loss[loss=0.1703, simple_loss=0.2421, pruned_loss=0.04928, over 939640.92 frames. ], batch size: 18, lr: 3.07e-03, grad_scale: 32.0 2023-04-27 19:47:39,609 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4318, 1.2525, 1.5830, 1.5800, 1.3019, 1.1995, 1.3077, 0.8237], device='cuda:4'), covar=tensor([0.0426, 0.0546, 0.0445, 0.0420, 0.0604, 0.0944, 0.0462, 0.0604], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0067, 0.0066, 0.0068, 0.0074, 0.0094, 0.0072, 0.0064], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 19:47:50,142 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3267, 1.3632, 3.5310, 3.3220, 3.1997, 3.5153, 3.4444, 3.1451], device='cuda:4'), covar=tensor([0.6431, 0.5150, 0.1357, 0.1757, 0.1129, 0.1493, 0.1124, 0.1518], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0308, 0.0408, 0.0408, 0.0348, 0.0410, 0.0316, 0.0368], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 19:47:50,808 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3261, 2.8622, 2.2343, 2.1922, 1.8199, 1.8491, 2.3453, 1.7890], device='cuda:4'), covar=tensor([0.1440, 0.1242, 0.1380, 0.1619, 0.2229, 0.1797, 0.0911, 0.1908], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0211, 0.0168, 0.0205, 0.0200, 0.0186, 0.0156, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 19:48:21,856 INFO [finetune.py:976] (4/7) Epoch 24, batch 0, loss[loss=0.165, simple_loss=0.2278, pruned_loss=0.05107, over 4777.00 frames. ], tot_loss[loss=0.165, simple_loss=0.2278, pruned_loss=0.05107, over 4777.00 frames. ], batch size: 26, lr: 3.07e-03, grad_scale: 32.0 2023-04-27 19:48:21,856 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 19:48:25,624 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3587, 1.5478, 1.9399, 2.0787, 2.0460, 2.1057, 1.9386, 2.0461], device='cuda:4'), covar=tensor([0.3486, 0.5430, 0.4690, 0.4601, 0.5233, 0.6849, 0.5329, 0.4435], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0375, 0.0328, 0.0339, 0.0349, 0.0394, 0.0357, 0.0329], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 19:48:31,973 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2738, 1.4641, 1.8376, 1.9628, 1.8964, 1.9546, 1.8777, 1.9264], device='cuda:4'), covar=tensor([0.4177, 0.5697, 0.4736, 0.4965, 0.5505, 0.7296, 0.5745, 0.4708], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0375, 0.0328, 0.0339, 0.0349, 0.0394, 0.0357, 0.0329], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 19:48:42,355 INFO [finetune.py:1010] (4/7) Epoch 24, validation: loss=0.1552, simple_loss=0.2243, pruned_loss=0.04308, over 2265189.00 frames. 2023-04-27 19:48:42,356 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6529MB 2023-04-27 19:48:45,628 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1521, 1.3649, 1.2598, 1.6538, 1.5097, 1.5457, 1.2942, 2.4855], device='cuda:4'), covar=tensor([0.0608, 0.0879, 0.0856, 0.1262, 0.0688, 0.0494, 0.0773, 0.0247], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 19:49:12,806 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=131765.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:49:13,313 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.101e+01 1.500e+02 1.790e+02 2.328e+02 5.462e+02, threshold=3.579e+02, percent-clipped=4.0 2023-04-27 19:49:21,991 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=131772.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:49:37,728 INFO [finetune.py:976] (4/7) Epoch 24, batch 50, loss[loss=0.1707, simple_loss=0.2433, pruned_loss=0.04899, over 4820.00 frames. ], tot_loss[loss=0.1712, simple_loss=0.2429, pruned_loss=0.04972, over 215614.39 frames. ], batch size: 30, lr: 3.07e-03, grad_scale: 32.0 2023-04-27 19:50:07,969 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7094, 2.0234, 0.8839, 1.4648, 1.9143, 1.5830, 1.5175, 1.6256], device='cuda:4'), covar=tensor([0.0504, 0.0350, 0.0338, 0.0565, 0.0265, 0.0532, 0.0502, 0.0593], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0028, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0046, 0.0038, 0.0052, 0.0038, 0.0050, 0.0050, 0.0051], device='cuda:4') 2023-04-27 19:50:39,523 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=131830.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:50:48,615 INFO [finetune.py:976] (4/7) Epoch 24, batch 100, loss[loss=0.152, simple_loss=0.2215, pruned_loss=0.04126, over 4844.00 frames. ], tot_loss[loss=0.1641, simple_loss=0.2353, pruned_loss=0.04646, over 379918.13 frames. ], batch size: 47, lr: 3.07e-03, grad_scale: 32.0 2023-04-27 19:50:56,378 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=131848.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:51:01,830 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=131857.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:51:07,129 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.089e+01 1.524e+02 1.795e+02 2.090e+02 4.072e+02, threshold=3.590e+02, percent-clipped=1.0 2023-04-27 19:51:23,459 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=131878.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:51:35,580 INFO [finetune.py:976] (4/7) Epoch 24, batch 150, loss[loss=0.1385, simple_loss=0.2138, pruned_loss=0.03155, over 4761.00 frames. ], tot_loss[loss=0.1613, simple_loss=0.2316, pruned_loss=0.04549, over 507654.91 frames. ], batch size: 27, lr: 3.07e-03, grad_scale: 32.0 2023-04-27 19:52:18,771 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=131918.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:52:44,204 INFO [finetune.py:976] (4/7) Epoch 24, batch 200, loss[loss=0.1714, simple_loss=0.2265, pruned_loss=0.05819, over 4678.00 frames. ], tot_loss[loss=0.166, simple_loss=0.2354, pruned_loss=0.04829, over 607998.88 frames. ], batch size: 23, lr: 3.07e-03, grad_scale: 32.0 2023-04-27 19:53:12,910 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.093e+02 1.534e+02 1.797e+02 2.198e+02 4.905e+02, threshold=3.594e+02, percent-clipped=2.0 2023-04-27 19:53:27,659 INFO [finetune.py:976] (4/7) Epoch 24, batch 250, loss[loss=0.2229, simple_loss=0.2978, pruned_loss=0.07399, over 4813.00 frames. ], tot_loss[loss=0.1692, simple_loss=0.2396, pruned_loss=0.04937, over 682893.53 frames. ], batch size: 41, lr: 3.07e-03, grad_scale: 32.0 2023-04-27 19:53:46,821 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0774, 2.5986, 1.0912, 1.4584, 2.1083, 1.2682, 3.6242, 1.7122], device='cuda:4'), covar=tensor([0.0670, 0.0636, 0.0798, 0.1316, 0.0507, 0.1053, 0.0269, 0.0686], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0063, 0.0047, 0.0045, 0.0049, 0.0050, 0.0072, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0009, 0.0007, 0.0007, 0.0008, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-27 19:53:57,450 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=132031.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:54:01,449 INFO [finetune.py:976] (4/7) Epoch 24, batch 300, loss[loss=0.1664, simple_loss=0.2403, pruned_loss=0.04625, over 4895.00 frames. ], tot_loss[loss=0.17, simple_loss=0.2411, pruned_loss=0.04939, over 742240.87 frames. ], batch size: 43, lr: 3.07e-03, grad_scale: 32.0 2023-04-27 19:54:24,634 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=132054.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:54:36,476 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=132065.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:54:36,980 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.215e+02 1.704e+02 1.925e+02 2.353e+02 6.924e+02, threshold=3.849e+02, percent-clipped=2.0 2023-04-27 19:54:46,361 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=132072.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:54:47,635 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9871, 0.9841, 1.1153, 1.1619, 0.9742, 0.9194, 0.9452, 0.5374], device='cuda:4'), covar=tensor([0.0523, 0.0623, 0.0441, 0.0538, 0.0713, 0.1042, 0.0459, 0.0594], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0067, 0.0066, 0.0068, 0.0074, 0.0095, 0.0072, 0.0064], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 19:55:00,440 INFO [finetune.py:976] (4/7) Epoch 24, batch 350, loss[loss=0.1892, simple_loss=0.2561, pruned_loss=0.06119, over 4735.00 frames. ], tot_loss[loss=0.172, simple_loss=0.2438, pruned_loss=0.05008, over 790444.93 frames. ], batch size: 54, lr: 3.07e-03, grad_scale: 32.0 2023-04-27 19:55:10,398 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=132092.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:55:31,175 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2086, 2.5549, 0.8375, 1.5868, 1.5992, 1.8460, 1.7243, 0.8624], device='cuda:4'), covar=tensor([0.1347, 0.1163, 0.1714, 0.1252, 0.1111, 0.0969, 0.1579, 0.1616], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0236, 0.0136, 0.0119, 0.0131, 0.0151, 0.0116, 0.0117], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 19:55:40,637 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=132113.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:55:41,951 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=132115.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 19:55:44,972 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=132120.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:56:06,855 INFO [finetune.py:976] (4/7) Epoch 24, batch 400, loss[loss=0.2195, simple_loss=0.2854, pruned_loss=0.07674, over 4818.00 frames. ], tot_loss[loss=0.1732, simple_loss=0.2458, pruned_loss=0.05031, over 829008.17 frames. ], batch size: 38, lr: 3.07e-03, grad_scale: 32.0 2023-04-27 19:56:13,354 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0589, 1.3988, 5.3446, 5.0233, 4.7127, 5.1994, 4.7816, 4.7491], device='cuda:4'), covar=tensor([0.6722, 0.5984, 0.0835, 0.1440, 0.0867, 0.1096, 0.1097, 0.1587], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0310, 0.0410, 0.0410, 0.0349, 0.0412, 0.0318, 0.0370], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 19:56:14,646 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8102, 2.0693, 1.6824, 1.5261, 1.3215, 1.3333, 1.6813, 1.2835], device='cuda:4'), covar=tensor([0.1665, 0.1309, 0.1509, 0.1694, 0.2328, 0.1967, 0.1086, 0.2045], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0209, 0.0168, 0.0203, 0.0199, 0.0185, 0.0155, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 19:56:25,735 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=132148.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:56:35,572 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.40 vs. limit=5.0 2023-04-27 19:56:49,858 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.121e+02 1.575e+02 1.808e+02 2.256e+02 3.614e+02, threshold=3.616e+02, percent-clipped=0.0 2023-04-27 19:57:04,663 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4425, 1.9622, 2.3536, 2.9865, 2.3841, 1.9328, 1.9239, 2.3585], device='cuda:4'), covar=tensor([0.2991, 0.3017, 0.1712, 0.2266, 0.2677, 0.2650, 0.3796, 0.1826], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0245, 0.0228, 0.0313, 0.0221, 0.0235, 0.0226, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 19:57:08,717 INFO [finetune.py:976] (4/7) Epoch 24, batch 450, loss[loss=0.1597, simple_loss=0.2411, pruned_loss=0.03916, over 4762.00 frames. ], tot_loss[loss=0.1715, simple_loss=0.2442, pruned_loss=0.04945, over 858415.65 frames. ], batch size: 28, lr: 3.07e-03, grad_scale: 32.0 2023-04-27 19:57:11,534 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-27 19:57:15,196 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=132196.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:57:27,640 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=132213.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:57:30,977 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.64 vs. limit=2.0 2023-04-27 19:57:42,216 INFO [finetune.py:976] (4/7) Epoch 24, batch 500, loss[loss=0.1525, simple_loss=0.2255, pruned_loss=0.03976, over 4747.00 frames. ], tot_loss[loss=0.1701, simple_loss=0.2419, pruned_loss=0.04915, over 880800.39 frames. ], batch size: 27, lr: 3.07e-03, grad_scale: 32.0 2023-04-27 19:57:59,190 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-04-27 19:58:00,953 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8343, 1.3266, 1.8915, 2.3271, 1.9649, 1.8231, 1.8782, 1.7858], device='cuda:4'), covar=tensor([0.4314, 0.6704, 0.5728, 0.5134, 0.5416, 0.7279, 0.7657, 0.8507], device='cuda:4'), in_proj_covar=tensor([0.0433, 0.0417, 0.0510, 0.0505, 0.0462, 0.0494, 0.0499, 0.0511], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 19:58:03,239 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.098e+02 1.544e+02 1.768e+02 2.170e+02 4.737e+02, threshold=3.537e+02, percent-clipped=2.0 2023-04-27 19:58:16,037 INFO [finetune.py:976] (4/7) Epoch 24, batch 550, loss[loss=0.1439, simple_loss=0.2075, pruned_loss=0.04018, over 4833.00 frames. ], tot_loss[loss=0.1683, simple_loss=0.239, pruned_loss=0.04877, over 895672.28 frames. ], batch size: 25, lr: 3.07e-03, grad_scale: 32.0 2023-04-27 19:58:31,298 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1965, 2.6804, 1.0486, 1.5358, 2.0998, 1.3054, 3.6221, 1.8400], device='cuda:4'), covar=tensor([0.0622, 0.0620, 0.0795, 0.1184, 0.0504, 0.0951, 0.0224, 0.0617], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0050, 0.0072, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0009, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 19:58:41,618 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-27 19:58:49,839 INFO [finetune.py:976] (4/7) Epoch 24, batch 600, loss[loss=0.1665, simple_loss=0.2378, pruned_loss=0.04763, over 4917.00 frames. ], tot_loss[loss=0.169, simple_loss=0.2399, pruned_loss=0.04901, over 910782.74 frames. ], batch size: 37, lr: 3.07e-03, grad_scale: 32.0 2023-04-27 19:58:57,678 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.9233, 3.9337, 2.7046, 4.5508, 4.0431, 3.8768, 1.7413, 3.9264], device='cuda:4'), covar=tensor([0.1802, 0.1142, 0.3536, 0.1395, 0.3719, 0.1801, 0.6010, 0.2379], device='cuda:4'), in_proj_covar=tensor([0.0246, 0.0221, 0.0254, 0.0305, 0.0296, 0.0248, 0.0274, 0.0273], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 19:59:10,244 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.026e+02 1.689e+02 1.942e+02 2.477e+02 4.504e+02, threshold=3.885e+02, percent-clipped=3.0 2023-04-27 19:59:22,997 INFO [finetune.py:976] (4/7) Epoch 24, batch 650, loss[loss=0.1526, simple_loss=0.2295, pruned_loss=0.03784, over 4920.00 frames. ], tot_loss[loss=0.1714, simple_loss=0.243, pruned_loss=0.04992, over 919392.55 frames. ], batch size: 36, lr: 3.07e-03, grad_scale: 32.0 2023-04-27 19:59:23,065 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=132387.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 19:59:39,615 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=132410.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 20:00:08,102 INFO [finetune.py:976] (4/7) Epoch 24, batch 700, loss[loss=0.1455, simple_loss=0.2331, pruned_loss=0.02897, over 4785.00 frames. ], tot_loss[loss=0.1722, simple_loss=0.2446, pruned_loss=0.04991, over 928687.96 frames. ], batch size: 29, lr: 3.07e-03, grad_scale: 32.0 2023-04-27 20:00:42,189 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6745, 1.3521, 1.8089, 1.9907, 1.7371, 1.6303, 1.7133, 1.7341], device='cuda:4'), covar=tensor([0.5985, 0.7704, 0.8223, 0.8706, 0.7012, 1.0176, 1.0062, 1.1684], device='cuda:4'), in_proj_covar=tensor([0.0432, 0.0415, 0.0509, 0.0503, 0.0461, 0.0492, 0.0497, 0.0508], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 20:00:50,346 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.129e+02 1.633e+02 1.958e+02 2.360e+02 4.192e+02, threshold=3.915e+02, percent-clipped=4.0 2023-04-27 20:01:15,826 INFO [finetune.py:976] (4/7) Epoch 24, batch 750, loss[loss=0.1467, simple_loss=0.2223, pruned_loss=0.03554, over 4807.00 frames. ], tot_loss[loss=0.1725, simple_loss=0.2457, pruned_loss=0.04966, over 936186.20 frames. ], batch size: 25, lr: 3.06e-03, grad_scale: 32.0 2023-04-27 20:01:48,575 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=132513.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:02:21,499 INFO [finetune.py:976] (4/7) Epoch 24, batch 800, loss[loss=0.1739, simple_loss=0.2391, pruned_loss=0.05432, over 4703.00 frames. ], tot_loss[loss=0.1716, simple_loss=0.2451, pruned_loss=0.04902, over 940246.15 frames. ], batch size: 59, lr: 3.06e-03, grad_scale: 32.0 2023-04-27 20:02:30,512 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-27 20:02:53,600 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=132561.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:03:01,812 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.951e+01 1.478e+02 1.835e+02 2.240e+02 6.510e+02, threshold=3.671e+02, percent-clipped=1.0 2023-04-27 20:03:27,812 INFO [finetune.py:976] (4/7) Epoch 24, batch 850, loss[loss=0.1703, simple_loss=0.2342, pruned_loss=0.05318, over 4823.00 frames. ], tot_loss[loss=0.1696, simple_loss=0.2425, pruned_loss=0.04836, over 943479.16 frames. ], batch size: 38, lr: 3.06e-03, grad_scale: 16.0 2023-04-27 20:04:34,501 INFO [finetune.py:976] (4/7) Epoch 24, batch 900, loss[loss=0.1875, simple_loss=0.2404, pruned_loss=0.06735, over 4058.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.2397, pruned_loss=0.04721, over 946158.12 frames. ], batch size: 65, lr: 3.06e-03, grad_scale: 16.0 2023-04-27 20:04:42,290 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-27 20:05:02,099 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=132660.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:05:12,900 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.989e+01 1.487e+02 1.792e+02 2.042e+02 3.425e+02, threshold=3.585e+02, percent-clipped=0.0 2023-04-27 20:05:44,310 INFO [finetune.py:976] (4/7) Epoch 24, batch 950, loss[loss=0.1466, simple_loss=0.212, pruned_loss=0.04055, over 4353.00 frames. ], tot_loss[loss=0.1654, simple_loss=0.2378, pruned_loss=0.04649, over 949421.44 frames. ], batch size: 65, lr: 3.06e-03, grad_scale: 16.0 2023-04-27 20:05:44,408 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=132687.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:06:08,253 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=132710.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 20:06:09,479 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7419, 2.2870, 1.8776, 2.1824, 1.6033, 1.9005, 1.7367, 1.5235], device='cuda:4'), covar=tensor([0.1982, 0.1072, 0.0814, 0.1117, 0.3219, 0.0992, 0.1852, 0.2373], device='cuda:4'), in_proj_covar=tensor([0.0287, 0.0303, 0.0216, 0.0275, 0.0315, 0.0256, 0.0248, 0.0264], device='cuda:4'), out_proj_covar=tensor([1.1471e-04, 1.1983e-04, 8.5092e-05, 1.0849e-04, 1.2720e-04, 1.0095e-04, 1.0021e-04, 1.0439e-04], device='cuda:4') 2023-04-27 20:06:27,927 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=132721.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:06:43,787 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=132735.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:06:44,412 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.6938, 3.5679, 2.7734, 4.2524, 3.6796, 3.6546, 1.6099, 3.6456], device='cuda:4'), covar=tensor([0.1617, 0.1313, 0.3612, 0.1282, 0.2963, 0.1663, 0.5424, 0.2355], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0218, 0.0252, 0.0302, 0.0293, 0.0245, 0.0272, 0.0270], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 20:06:49,487 INFO [finetune.py:976] (4/7) Epoch 24, batch 1000, loss[loss=0.2135, simple_loss=0.2886, pruned_loss=0.06915, over 4798.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.2402, pruned_loss=0.04762, over 950209.29 frames. ], batch size: 51, lr: 3.06e-03, grad_scale: 16.0 2023-04-27 20:06:55,149 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=132746.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:07:02,344 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=132758.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:07:07,765 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.158e+02 1.549e+02 1.832e+02 2.153e+02 4.117e+02, threshold=3.664e+02, percent-clipped=1.0 2023-04-27 20:07:22,249 INFO [finetune.py:976] (4/7) Epoch 24, batch 1050, loss[loss=0.1996, simple_loss=0.2764, pruned_loss=0.06144, over 4816.00 frames. ], tot_loss[loss=0.1699, simple_loss=0.2434, pruned_loss=0.04825, over 951883.86 frames. ], batch size: 38, lr: 3.06e-03, grad_scale: 16.0 2023-04-27 20:07:35,151 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=132807.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:07:56,050 INFO [finetune.py:976] (4/7) Epoch 24, batch 1100, loss[loss=0.1974, simple_loss=0.2594, pruned_loss=0.06772, over 4898.00 frames. ], tot_loss[loss=0.1712, simple_loss=0.2447, pruned_loss=0.0489, over 952676.60 frames. ], batch size: 32, lr: 3.06e-03, grad_scale: 16.0 2023-04-27 20:08:14,647 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.617e+01 1.589e+02 1.846e+02 2.225e+02 4.028e+02, threshold=3.692e+02, percent-clipped=2.0 2023-04-27 20:08:28,715 INFO [finetune.py:976] (4/7) Epoch 24, batch 1150, loss[loss=0.1535, simple_loss=0.232, pruned_loss=0.03744, over 4886.00 frames. ], tot_loss[loss=0.1719, simple_loss=0.2455, pruned_loss=0.04913, over 954464.51 frames. ], batch size: 35, lr: 3.06e-03, grad_scale: 16.0 2023-04-27 20:09:11,141 INFO [finetune.py:976] (4/7) Epoch 24, batch 1200, loss[loss=0.2147, simple_loss=0.2762, pruned_loss=0.07658, over 4179.00 frames. ], tot_loss[loss=0.1706, simple_loss=0.2437, pruned_loss=0.04876, over 954407.60 frames. ], batch size: 65, lr: 3.06e-03, grad_scale: 16.0 2023-04-27 20:09:45,985 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5198, 1.3129, 4.1417, 3.9137, 3.5845, 3.9194, 3.8614, 3.6680], device='cuda:4'), covar=tensor([0.6665, 0.5497, 0.1039, 0.1627, 0.1160, 0.1418, 0.1813, 0.1377], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0307, 0.0405, 0.0407, 0.0347, 0.0409, 0.0316, 0.0365], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 20:09:48,368 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.050e+02 1.568e+02 1.953e+02 2.291e+02 3.731e+02, threshold=3.906e+02, percent-clipped=1.0 2023-04-27 20:09:56,244 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2310, 1.6803, 2.0699, 2.5099, 2.0737, 1.6684, 1.3745, 1.8968], device='cuda:4'), covar=tensor([0.2925, 0.3030, 0.1481, 0.1899, 0.2609, 0.2367, 0.4003, 0.1840], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0243, 0.0225, 0.0311, 0.0220, 0.0232, 0.0225, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 20:10:17,850 INFO [finetune.py:976] (4/7) Epoch 24, batch 1250, loss[loss=0.1814, simple_loss=0.2516, pruned_loss=0.05566, over 4816.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2414, pruned_loss=0.04814, over 954961.71 frames. ], batch size: 41, lr: 3.06e-03, grad_scale: 16.0 2023-04-27 20:10:19,385 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.58 vs. limit=5.0 2023-04-27 20:10:53,172 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=133016.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:11:03,101 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2694, 1.8118, 2.1255, 2.6227, 2.1885, 1.7205, 1.5898, 1.9883], device='cuda:4'), covar=tensor([0.2667, 0.2753, 0.1434, 0.2064, 0.2381, 0.2370, 0.3963, 0.1917], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0243, 0.0226, 0.0311, 0.0220, 0.0233, 0.0225, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 20:11:17,185 INFO [finetune.py:976] (4/7) Epoch 24, batch 1300, loss[loss=0.1366, simple_loss=0.207, pruned_loss=0.03312, over 4787.00 frames. ], tot_loss[loss=0.1663, simple_loss=0.2381, pruned_loss=0.04721, over 954098.47 frames. ], batch size: 26, lr: 3.06e-03, grad_scale: 16.0 2023-04-27 20:11:37,888 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.023e+02 1.644e+02 1.911e+02 2.318e+02 6.826e+02, threshold=3.821e+02, percent-clipped=2.0 2023-04-27 20:11:50,602 INFO [finetune.py:976] (4/7) Epoch 24, batch 1350, loss[loss=0.1786, simple_loss=0.2403, pruned_loss=0.05845, over 4894.00 frames. ], tot_loss[loss=0.1657, simple_loss=0.2374, pruned_loss=0.04695, over 955659.29 frames. ], batch size: 35, lr: 3.06e-03, grad_scale: 16.0 2023-04-27 20:12:07,604 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=133102.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:12:28,951 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.41 vs. limit=5.0 2023-04-27 20:12:52,452 INFO [finetune.py:976] (4/7) Epoch 24, batch 1400, loss[loss=0.1717, simple_loss=0.2429, pruned_loss=0.05025, over 4909.00 frames. ], tot_loss[loss=0.1698, simple_loss=0.2421, pruned_loss=0.04872, over 956148.01 frames. ], batch size: 36, lr: 3.06e-03, grad_scale: 16.0 2023-04-27 20:13:02,841 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5070, 1.3960, 0.4999, 1.2102, 1.3704, 1.3529, 1.2486, 1.2915], device='cuda:4'), covar=tensor([0.0510, 0.0389, 0.0408, 0.0583, 0.0314, 0.0539, 0.0538, 0.0596], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0027, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0045, 0.0038, 0.0052, 0.0038, 0.0050, 0.0050, 0.0051], device='cuda:4') 2023-04-27 20:13:29,245 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.160e+02 1.717e+02 2.018e+02 2.257e+02 3.714e+02, threshold=4.037e+02, percent-clipped=0.0 2023-04-27 20:13:41,458 INFO [finetune.py:976] (4/7) Epoch 24, batch 1450, loss[loss=0.1342, simple_loss=0.2065, pruned_loss=0.03096, over 4775.00 frames. ], tot_loss[loss=0.1708, simple_loss=0.2435, pruned_loss=0.04904, over 957488.21 frames. ], batch size: 26, lr: 3.06e-03, grad_scale: 16.0 2023-04-27 20:13:55,453 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9748, 2.4314, 2.0402, 2.3509, 1.6586, 2.1237, 2.0213, 1.6061], device='cuda:4'), covar=tensor([0.1966, 0.1226, 0.0852, 0.1177, 0.3445, 0.1106, 0.1666, 0.2391], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0305, 0.0218, 0.0278, 0.0318, 0.0259, 0.0251, 0.0267], device='cuda:4'), out_proj_covar=tensor([1.1571e-04, 1.2047e-04, 8.6004e-05, 1.0954e-04, 1.2842e-04, 1.0220e-04, 1.0141e-04, 1.0549e-04], device='cuda:4') 2023-04-27 20:14:31,505 INFO [finetune.py:976] (4/7) Epoch 24, batch 1500, loss[loss=0.1813, simple_loss=0.2548, pruned_loss=0.05388, over 4820.00 frames. ], tot_loss[loss=0.1702, simple_loss=0.2434, pruned_loss=0.04849, over 955792.08 frames. ], batch size: 30, lr: 3.06e-03, grad_scale: 16.0 2023-04-27 20:14:37,103 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5501, 1.2714, 4.3872, 4.1279, 3.8154, 4.2081, 4.0197, 3.8460], device='cuda:4'), covar=tensor([0.7076, 0.6260, 0.0987, 0.1556, 0.1080, 0.1531, 0.1636, 0.1564], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0307, 0.0405, 0.0408, 0.0349, 0.0409, 0.0317, 0.0367], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 20:14:37,334 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.75 vs. limit=5.0 2023-04-27 20:15:14,564 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.127e+02 1.609e+02 1.860e+02 2.213e+02 4.460e+02, threshold=3.721e+02, percent-clipped=1.0 2023-04-27 20:15:32,673 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1695, 1.6461, 2.0156, 2.4600, 2.0110, 1.5810, 1.2504, 1.8222], device='cuda:4'), covar=tensor([0.3335, 0.3303, 0.1813, 0.2242, 0.2650, 0.2768, 0.4311, 0.2171], device='cuda:4'), in_proj_covar=tensor([0.0294, 0.0244, 0.0227, 0.0313, 0.0220, 0.0234, 0.0226, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 20:15:44,451 INFO [finetune.py:976] (4/7) Epoch 24, batch 1550, loss[loss=0.1643, simple_loss=0.236, pruned_loss=0.04632, over 4824.00 frames. ], tot_loss[loss=0.1697, simple_loss=0.2435, pruned_loss=0.04792, over 954707.57 frames. ], batch size: 38, lr: 3.06e-03, grad_scale: 16.0 2023-04-27 20:16:10,949 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=133316.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:16:34,614 INFO [finetune.py:976] (4/7) Epoch 24, batch 1600, loss[loss=0.182, simple_loss=0.2546, pruned_loss=0.05468, over 4903.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2425, pruned_loss=0.04814, over 955753.35 frames. ], batch size: 43, lr: 3.06e-03, grad_scale: 16.0 2023-04-27 20:17:17,457 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=133364.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:17:19,229 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.014e+02 1.451e+02 1.794e+02 2.296e+02 6.131e+02, threshold=3.588e+02, percent-clipped=2.0 2023-04-27 20:17:21,165 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0985, 0.7881, 0.9124, 0.8229, 1.2399, 0.9606, 0.9035, 0.9543], device='cuda:4'), covar=tensor([0.1637, 0.1274, 0.1766, 0.1497, 0.0910, 0.1395, 0.1571, 0.2223], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0308, 0.0351, 0.0285, 0.0328, 0.0307, 0.0300, 0.0374], device='cuda:4'), out_proj_covar=tensor([6.4072e-05, 6.3518e-05, 7.3807e-05, 5.7175e-05, 6.7553e-05, 6.4337e-05, 6.2528e-05, 7.9295e-05], device='cuda:4') 2023-04-27 20:17:29,330 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=133374.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 20:17:30,013 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.63 vs. limit=5.0 2023-04-27 20:17:30,579 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=133376.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:17:42,836 INFO [finetune.py:976] (4/7) Epoch 24, batch 1650, loss[loss=0.1401, simple_loss=0.2122, pruned_loss=0.03401, over 4874.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.24, pruned_loss=0.04711, over 955657.49 frames. ], batch size: 31, lr: 3.06e-03, grad_scale: 16.0 2023-04-27 20:17:52,082 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.7308, 2.4709, 2.7051, 3.2660, 3.0951, 2.4308, 2.2739, 2.7974], device='cuda:4'), covar=tensor([0.0701, 0.0936, 0.0598, 0.0494, 0.0503, 0.0881, 0.0681, 0.0519], device='cuda:4'), in_proj_covar=tensor([0.0187, 0.0205, 0.0186, 0.0173, 0.0178, 0.0180, 0.0151, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 20:18:02,182 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=133402.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:18:48,093 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=133435.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 20:18:49,165 INFO [finetune.py:976] (4/7) Epoch 24, batch 1700, loss[loss=0.1871, simple_loss=0.2652, pruned_loss=0.05453, over 4908.00 frames. ], tot_loss[loss=0.1653, simple_loss=0.2379, pruned_loss=0.04633, over 956654.29 frames. ], batch size: 43, lr: 3.06e-03, grad_scale: 16.0 2023-04-27 20:18:49,284 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=133437.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:18:55,759 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4858, 3.5338, 0.9394, 1.7942, 2.0661, 2.4314, 1.9355, 0.9354], device='cuda:4'), covar=tensor([0.1522, 0.0851, 0.2237, 0.1386, 0.1094, 0.1097, 0.1548, 0.2059], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0239, 0.0137, 0.0121, 0.0133, 0.0152, 0.0117, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 20:19:08,015 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=133450.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:19:31,796 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.276e+01 1.467e+02 1.697e+02 2.216e+02 4.193e+02, threshold=3.394e+02, percent-clipped=4.0 2023-04-27 20:19:33,882 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-04-27 20:19:41,274 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=133474.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:19:55,049 INFO [finetune.py:976] (4/7) Epoch 24, batch 1750, loss[loss=0.1724, simple_loss=0.2451, pruned_loss=0.04984, over 4119.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.2401, pruned_loss=0.04748, over 954970.31 frames. ], batch size: 65, lr: 3.06e-03, grad_scale: 16.0 2023-04-27 20:20:15,581 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1834, 2.5057, 1.1847, 1.4645, 2.1373, 1.2803, 3.3855, 1.6532], device='cuda:4'), covar=tensor([0.0598, 0.0621, 0.0739, 0.1306, 0.0466, 0.0977, 0.0225, 0.0641], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0063, 0.0047, 0.0046, 0.0049, 0.0051, 0.0072, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0009, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-27 20:21:01,623 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=133535.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:21:08,176 INFO [finetune.py:976] (4/7) Epoch 24, batch 1800, loss[loss=0.1377, simple_loss=0.2046, pruned_loss=0.03536, over 4703.00 frames. ], tot_loss[loss=0.1705, simple_loss=0.2432, pruned_loss=0.0489, over 953558.92 frames. ], batch size: 23, lr: 3.06e-03, grad_scale: 16.0 2023-04-27 20:21:12,584 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3234, 1.5843, 2.0377, 2.8446, 2.0197, 1.5720, 1.7037, 1.9630], device='cuda:4'), covar=tensor([0.4259, 0.4467, 0.2313, 0.2993, 0.3482, 0.3317, 0.4621, 0.2550], device='cuda:4'), in_proj_covar=tensor([0.0295, 0.0246, 0.0228, 0.0315, 0.0222, 0.0236, 0.0228, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 20:21:13,207 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3117, 1.7674, 2.1943, 2.6733, 2.1790, 1.6510, 1.5748, 2.0340], device='cuda:4'), covar=tensor([0.3162, 0.3111, 0.1562, 0.2189, 0.2696, 0.2649, 0.3887, 0.2048], device='cuda:4'), in_proj_covar=tensor([0.0295, 0.0246, 0.0228, 0.0315, 0.0222, 0.0236, 0.0228, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 20:21:44,189 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.423e+01 1.601e+02 1.910e+02 2.304e+02 3.946e+02, threshold=3.820e+02, percent-clipped=1.0 2023-04-27 20:22:08,895 INFO [finetune.py:976] (4/7) Epoch 24, batch 1850, loss[loss=0.1835, simple_loss=0.2557, pruned_loss=0.05567, over 4730.00 frames. ], tot_loss[loss=0.1735, simple_loss=0.2458, pruned_loss=0.05064, over 952862.09 frames. ], batch size: 54, lr: 3.06e-03, grad_scale: 16.0 2023-04-27 20:22:42,122 INFO [finetune.py:976] (4/7) Epoch 24, batch 1900, loss[loss=0.2079, simple_loss=0.2817, pruned_loss=0.06704, over 4800.00 frames. ], tot_loss[loss=0.1757, simple_loss=0.2481, pruned_loss=0.05166, over 952344.32 frames. ], batch size: 41, lr: 3.06e-03, grad_scale: 16.0 2023-04-27 20:22:51,335 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=133652.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:23:12,125 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.071e+02 1.575e+02 1.806e+02 2.357e+02 5.107e+02, threshold=3.612e+02, percent-clipped=3.0 2023-04-27 20:23:26,868 INFO [finetune.py:976] (4/7) Epoch 24, batch 1950, loss[loss=0.2118, simple_loss=0.2746, pruned_loss=0.07449, over 4818.00 frames. ], tot_loss[loss=0.1727, simple_loss=0.2452, pruned_loss=0.05014, over 952700.93 frames. ], batch size: 39, lr: 3.06e-03, grad_scale: 16.0 2023-04-27 20:23:35,139 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.63 vs. limit=2.0 2023-04-27 20:23:40,406 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9350, 1.5553, 1.5524, 1.7355, 2.1235, 1.6981, 1.5137, 1.4652], device='cuda:4'), covar=tensor([0.1618, 0.1267, 0.1814, 0.1109, 0.0804, 0.1522, 0.2203, 0.2198], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0306, 0.0348, 0.0282, 0.0326, 0.0304, 0.0297, 0.0371], device='cuda:4'), out_proj_covar=tensor([6.3538e-05, 6.2987e-05, 7.3261e-05, 5.6637e-05, 6.7111e-05, 6.3727e-05, 6.1837e-05, 7.8675e-05], device='cuda:4') 2023-04-27 20:23:43,320 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=133713.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:23:56,164 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=133730.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 20:23:57,386 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=133732.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:24:00,819 INFO [finetune.py:976] (4/7) Epoch 24, batch 2000, loss[loss=0.1491, simple_loss=0.2255, pruned_loss=0.03636, over 4866.00 frames. ], tot_loss[loss=0.1699, simple_loss=0.2418, pruned_loss=0.04899, over 953456.25 frames. ], batch size: 34, lr: 3.06e-03, grad_scale: 16.0 2023-04-27 20:24:25,089 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.642e+01 1.427e+02 1.749e+02 2.000e+02 3.515e+02, threshold=3.497e+02, percent-clipped=0.0 2023-04-27 20:24:55,311 INFO [finetune.py:976] (4/7) Epoch 24, batch 2050, loss[loss=0.1836, simple_loss=0.2508, pruned_loss=0.05818, over 4823.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.238, pruned_loss=0.04783, over 954694.70 frames. ], batch size: 40, lr: 3.06e-03, grad_scale: 16.0 2023-04-27 20:25:15,052 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.90 vs. limit=2.0 2023-04-27 20:25:48,582 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=133830.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:25:58,916 INFO [finetune.py:976] (4/7) Epoch 24, batch 2100, loss[loss=0.1656, simple_loss=0.2345, pruned_loss=0.04834, over 4379.00 frames. ], tot_loss[loss=0.1667, simple_loss=0.2378, pruned_loss=0.04779, over 953881.86 frames. ], batch size: 19, lr: 3.06e-03, grad_scale: 16.0 2023-04-27 20:26:18,187 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3616, 2.3335, 2.6164, 2.9268, 2.8578, 2.1043, 2.0315, 2.3512], device='cuda:4'), covar=tensor([0.0838, 0.0993, 0.0626, 0.0594, 0.0608, 0.1032, 0.0765, 0.0664], device='cuda:4'), in_proj_covar=tensor([0.0189, 0.0207, 0.0189, 0.0175, 0.0181, 0.0183, 0.0152, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 20:26:22,912 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.079e+02 1.658e+02 2.018e+02 2.431e+02 5.100e+02, threshold=4.036e+02, percent-clipped=6.0 2023-04-27 20:26:37,513 INFO [finetune.py:976] (4/7) Epoch 24, batch 2150, loss[loss=0.1796, simple_loss=0.26, pruned_loss=0.04964, over 4738.00 frames. ], tot_loss[loss=0.1699, simple_loss=0.2417, pruned_loss=0.04908, over 954719.83 frames. ], batch size: 59, lr: 3.05e-03, grad_scale: 16.0 2023-04-27 20:26:47,176 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9844, 2.6744, 1.9813, 2.0420, 1.4022, 1.4414, 2.0750, 1.3944], device='cuda:4'), covar=tensor([0.1700, 0.1282, 0.1462, 0.1614, 0.2282, 0.1954, 0.0988, 0.2007], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0211, 0.0169, 0.0205, 0.0199, 0.0186, 0.0156, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 20:27:09,765 INFO [finetune.py:976] (4/7) Epoch 24, batch 2200, loss[loss=0.217, simple_loss=0.2789, pruned_loss=0.07753, over 4835.00 frames. ], tot_loss[loss=0.1717, simple_loss=0.2441, pruned_loss=0.04965, over 954604.41 frames. ], batch size: 47, lr: 3.05e-03, grad_scale: 16.0 2023-04-27 20:27:37,979 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0636, 1.4793, 1.3475, 1.6497, 1.5714, 1.6533, 1.3146, 3.0635], device='cuda:4'), covar=tensor([0.0648, 0.0794, 0.0754, 0.1189, 0.0658, 0.0520, 0.0728, 0.0163], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0038, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 20:27:41,111 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2879, 1.7198, 2.1792, 2.6260, 2.1306, 1.6559, 1.5601, 1.9550], device='cuda:4'), covar=tensor([0.3211, 0.3167, 0.1593, 0.2242, 0.2678, 0.2859, 0.4065, 0.2073], device='cuda:4'), in_proj_covar=tensor([0.0294, 0.0245, 0.0228, 0.0315, 0.0221, 0.0235, 0.0228, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 20:27:48,467 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.369e+01 1.555e+02 1.795e+02 2.184e+02 7.045e+02, threshold=3.590e+02, percent-clipped=2.0 2023-04-27 20:28:06,398 INFO [finetune.py:976] (4/7) Epoch 24, batch 2250, loss[loss=0.1603, simple_loss=0.2385, pruned_loss=0.04102, over 4829.00 frames. ], tot_loss[loss=0.1723, simple_loss=0.2452, pruned_loss=0.04973, over 954694.18 frames. ], batch size: 33, lr: 3.05e-03, grad_scale: 16.0 2023-04-27 20:28:07,885 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.64 vs. limit=2.0 2023-04-27 20:28:14,992 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=133997.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:28:23,010 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=134008.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:28:33,437 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7979, 2.2120, 1.8853, 2.1277, 1.6588, 1.9055, 1.9072, 1.4819], device='cuda:4'), covar=tensor([0.1753, 0.1190, 0.0873, 0.1155, 0.3155, 0.1121, 0.1834, 0.2409], device='cuda:4'), in_proj_covar=tensor([0.0287, 0.0302, 0.0217, 0.0278, 0.0315, 0.0257, 0.0251, 0.0266], device='cuda:4'), out_proj_covar=tensor([1.1488e-04, 1.1943e-04, 8.5287e-05, 1.0950e-04, 1.2717e-04, 1.0136e-04, 1.0140e-04, 1.0500e-04], device='cuda:4') 2023-04-27 20:28:36,963 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=134030.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 20:28:38,182 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=134032.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:28:41,586 INFO [finetune.py:976] (4/7) Epoch 24, batch 2300, loss[loss=0.1662, simple_loss=0.2351, pruned_loss=0.04865, over 4717.00 frames. ], tot_loss[loss=0.1707, simple_loss=0.244, pruned_loss=0.04874, over 953471.72 frames. ], batch size: 59, lr: 3.05e-03, grad_scale: 16.0 2023-04-27 20:28:55,890 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=134058.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:29:00,805 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7183, 2.1353, 1.6419, 1.4488, 1.2735, 1.2986, 1.6837, 1.2202], device='cuda:4'), covar=tensor([0.1499, 0.1133, 0.1503, 0.1663, 0.2130, 0.1800, 0.0948, 0.1911], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0209, 0.0168, 0.0204, 0.0198, 0.0184, 0.0155, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 20:29:01,260 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.437e+01 1.459e+02 1.887e+02 2.167e+02 3.394e+02, threshold=3.773e+02, percent-clipped=1.0 2023-04-27 20:29:08,092 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=134078.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 20:29:09,324 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=134080.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:29:09,463 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.54 vs. limit=2.0 2023-04-27 20:29:14,456 INFO [finetune.py:976] (4/7) Epoch 24, batch 2350, loss[loss=0.1779, simple_loss=0.2573, pruned_loss=0.04918, over 4829.00 frames. ], tot_loss[loss=0.1682, simple_loss=0.2411, pruned_loss=0.04769, over 953387.31 frames. ], batch size: 38, lr: 3.05e-03, grad_scale: 16.0 2023-04-27 20:29:25,733 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6684, 1.7085, 0.7590, 1.3797, 1.6589, 1.5412, 1.4494, 1.5055], device='cuda:4'), covar=tensor([0.0497, 0.0351, 0.0332, 0.0551, 0.0260, 0.0474, 0.0447, 0.0560], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0028, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0046, 0.0038, 0.0052, 0.0039, 0.0050, 0.0050, 0.0052], device='cuda:4') 2023-04-27 20:29:30,573 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1898, 2.1720, 1.8532, 1.8617, 2.2747, 1.8879, 2.8488, 1.6365], device='cuda:4'), covar=tensor([0.3473, 0.1929, 0.4835, 0.2838, 0.1485, 0.2576, 0.1129, 0.4609], device='cuda:4'), in_proj_covar=tensor([0.0333, 0.0347, 0.0420, 0.0345, 0.0374, 0.0370, 0.0363, 0.0417], device='cuda:4'), out_proj_covar=tensor([9.8575e-05, 1.0367e-04, 1.2714e-04, 1.0366e-04, 1.1098e-04, 1.1022e-04, 1.0657e-04, 1.2533e-04], device='cuda:4') 2023-04-27 20:29:31,801 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3873, 1.7129, 1.5400, 2.2051, 2.3601, 1.9111, 1.8722, 1.6545], device='cuda:4'), covar=tensor([0.1538, 0.1568, 0.2036, 0.1773, 0.1022, 0.1888, 0.1952, 0.2148], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0305, 0.0347, 0.0281, 0.0324, 0.0304, 0.0296, 0.0369], device='cuda:4'), out_proj_covar=tensor([6.3245e-05, 6.2876e-05, 7.2983e-05, 5.6454e-05, 6.6815e-05, 6.3609e-05, 6.1564e-05, 7.8197e-05], device='cuda:4') 2023-04-27 20:29:36,602 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1643, 2.6339, 1.0729, 1.6637, 2.1043, 1.2598, 3.3603, 1.7015], device='cuda:4'), covar=tensor([0.0594, 0.0695, 0.0792, 0.1062, 0.0474, 0.0845, 0.0186, 0.0553], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0073, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 20:29:42,046 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=134130.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:29:46,679 INFO [finetune.py:976] (4/7) Epoch 24, batch 2400, loss[loss=0.1408, simple_loss=0.2185, pruned_loss=0.03159, over 4843.00 frames. ], tot_loss[loss=0.1676, simple_loss=0.24, pruned_loss=0.04757, over 953795.40 frames. ], batch size: 30, lr: 3.05e-03, grad_scale: 16.0 2023-04-27 20:30:06,809 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.831e+01 1.496e+02 1.764e+02 2.072e+02 3.760e+02, threshold=3.528e+02, percent-clipped=0.0 2023-04-27 20:30:13,577 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=134178.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:30:19,541 INFO [finetune.py:976] (4/7) Epoch 24, batch 2450, loss[loss=0.168, simple_loss=0.2426, pruned_loss=0.04673, over 4895.00 frames. ], tot_loss[loss=0.1646, simple_loss=0.2367, pruned_loss=0.04627, over 955109.02 frames. ], batch size: 35, lr: 3.05e-03, grad_scale: 16.0 2023-04-27 20:31:07,210 INFO [finetune.py:976] (4/7) Epoch 24, batch 2500, loss[loss=0.1844, simple_loss=0.2456, pruned_loss=0.06157, over 4380.00 frames. ], tot_loss[loss=0.168, simple_loss=0.2401, pruned_loss=0.04798, over 952773.39 frames. ], batch size: 19, lr: 3.05e-03, grad_scale: 16.0 2023-04-27 20:31:39,320 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.116e+02 1.581e+02 1.827e+02 2.358e+02 4.176e+02, threshold=3.655e+02, percent-clipped=4.0 2023-04-27 20:31:43,077 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9290, 1.9596, 1.9256, 1.6231, 2.0344, 1.6833, 2.6784, 1.6001], device='cuda:4'), covar=tensor([0.3431, 0.1849, 0.4528, 0.2474, 0.1683, 0.2524, 0.1250, 0.4610], device='cuda:4'), in_proj_covar=tensor([0.0331, 0.0345, 0.0418, 0.0344, 0.0372, 0.0368, 0.0361, 0.0415], device='cuda:4'), out_proj_covar=tensor([9.8064e-05, 1.0304e-04, 1.2661e-04, 1.0325e-04, 1.1041e-04, 1.0956e-04, 1.0606e-04, 1.2478e-04], device='cuda:4') 2023-04-27 20:32:02,532 INFO [finetune.py:976] (4/7) Epoch 24, batch 2550, loss[loss=0.1563, simple_loss=0.233, pruned_loss=0.03978, over 4806.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.244, pruned_loss=0.04911, over 954096.57 frames. ], batch size: 39, lr: 3.05e-03, grad_scale: 16.0 2023-04-27 20:32:26,936 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=134304.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:32:34,602 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=134308.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:32:35,225 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=134309.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:33:08,807 INFO [finetune.py:976] (4/7) Epoch 24, batch 2600, loss[loss=0.1469, simple_loss=0.2042, pruned_loss=0.04483, over 4228.00 frames. ], tot_loss[loss=0.1715, simple_loss=0.2445, pruned_loss=0.04921, over 953374.94 frames. ], batch size: 18, lr: 3.05e-03, grad_scale: 16.0 2023-04-27 20:33:27,037 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=134353.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:33:30,512 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=134356.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:33:42,153 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=134365.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:33:43,228 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.125e+02 1.588e+02 1.844e+02 2.194e+02 4.309e+02, threshold=3.689e+02, percent-clipped=2.0 2023-04-27 20:33:45,771 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=134370.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:33:54,003 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8727, 4.0585, 0.8162, 2.2639, 2.4948, 2.8814, 2.4477, 0.9582], device='cuda:4'), covar=tensor([0.1244, 0.0871, 0.2138, 0.1159, 0.0888, 0.0921, 0.1341, 0.2137], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0239, 0.0136, 0.0120, 0.0133, 0.0152, 0.0116, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 20:34:07,320 INFO [finetune.py:976] (4/7) Epoch 24, batch 2650, loss[loss=0.1657, simple_loss=0.2392, pruned_loss=0.04609, over 4792.00 frames. ], tot_loss[loss=0.1729, simple_loss=0.2465, pruned_loss=0.04964, over 954514.36 frames. ], batch size: 51, lr: 3.05e-03, grad_scale: 16.0 2023-04-27 20:34:38,317 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.96 vs. limit=5.0 2023-04-27 20:35:02,817 INFO [finetune.py:976] (4/7) Epoch 24, batch 2700, loss[loss=0.1631, simple_loss=0.2328, pruned_loss=0.04674, over 4895.00 frames. ], tot_loss[loss=0.1704, simple_loss=0.244, pruned_loss=0.04844, over 954610.39 frames. ], batch size: 32, lr: 3.05e-03, grad_scale: 16.0 2023-04-27 20:35:23,806 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.513e+01 1.459e+02 1.776e+02 2.011e+02 3.451e+02, threshold=3.553e+02, percent-clipped=0.0 2023-04-27 20:35:36,505 INFO [finetune.py:976] (4/7) Epoch 24, batch 2750, loss[loss=0.1792, simple_loss=0.2379, pruned_loss=0.06026, over 4813.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.2404, pruned_loss=0.04733, over 954407.90 frames. ], batch size: 25, lr: 3.05e-03, grad_scale: 16.0 2023-04-27 20:36:43,197 INFO [finetune.py:976] (4/7) Epoch 24, batch 2800, loss[loss=0.1736, simple_loss=0.2307, pruned_loss=0.05827, over 4934.00 frames. ], tot_loss[loss=0.1643, simple_loss=0.2366, pruned_loss=0.04598, over 953867.15 frames. ], batch size: 33, lr: 3.05e-03, grad_scale: 16.0 2023-04-27 20:37:13,595 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4596, 2.5690, 2.1020, 2.3000, 2.5620, 2.2498, 3.3255, 1.9875], device='cuda:4'), covar=tensor([0.3687, 0.2434, 0.4508, 0.3339, 0.2032, 0.2589, 0.1905, 0.4057], device='cuda:4'), in_proj_covar=tensor([0.0335, 0.0348, 0.0422, 0.0346, 0.0375, 0.0370, 0.0366, 0.0417], device='cuda:4'), out_proj_covar=tensor([9.9106e-05, 1.0382e-04, 1.2778e-04, 1.0397e-04, 1.1144e-04, 1.1029e-04, 1.0736e-04, 1.2549e-04], device='cuda:4') 2023-04-27 20:37:21,525 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9335, 1.2638, 3.3057, 3.0891, 2.9519, 3.2670, 3.2355, 2.9244], device='cuda:4'), covar=tensor([0.7356, 0.5505, 0.1503, 0.2122, 0.1473, 0.2284, 0.1566, 0.1828], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0304, 0.0402, 0.0406, 0.0346, 0.0407, 0.0315, 0.0365], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 20:37:23,288 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.232e+01 1.527e+02 1.797e+02 2.170e+02 3.863e+02, threshold=3.594e+02, percent-clipped=3.0 2023-04-27 20:37:48,299 INFO [finetune.py:976] (4/7) Epoch 24, batch 2850, loss[loss=0.1615, simple_loss=0.2369, pruned_loss=0.043, over 4868.00 frames. ], tot_loss[loss=0.1632, simple_loss=0.235, pruned_loss=0.04572, over 952881.14 frames. ], batch size: 31, lr: 3.05e-03, grad_scale: 32.0 2023-04-27 20:38:06,502 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9529, 1.2673, 4.7460, 4.4894, 4.1181, 4.4059, 4.1977, 4.1627], device='cuda:4'), covar=tensor([0.6994, 0.6076, 0.1054, 0.1723, 0.1166, 0.1529, 0.2150, 0.1875], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0304, 0.0401, 0.0405, 0.0346, 0.0406, 0.0315, 0.0365], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 20:38:27,673 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5198, 1.4089, 1.8107, 1.8381, 1.3921, 1.2415, 1.4950, 0.9463], device='cuda:4'), covar=tensor([0.0489, 0.0629, 0.0365, 0.0511, 0.0670, 0.0996, 0.0577, 0.0556], device='cuda:4'), in_proj_covar=tensor([0.0068, 0.0067, 0.0065, 0.0067, 0.0073, 0.0094, 0.0072, 0.0063], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 20:38:59,964 INFO [finetune.py:976] (4/7) Epoch 24, batch 2900, loss[loss=0.2205, simple_loss=0.3001, pruned_loss=0.07047, over 4863.00 frames. ], tot_loss[loss=0.1669, simple_loss=0.2389, pruned_loss=0.04743, over 953049.29 frames. ], batch size: 44, lr: 3.05e-03, grad_scale: 32.0 2023-04-27 20:39:04,406 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5889, 2.1154, 1.6087, 1.3266, 1.2086, 1.2176, 1.6171, 1.1687], device='cuda:4'), covar=tensor([0.1797, 0.1218, 0.1547, 0.1789, 0.2349, 0.1999, 0.1092, 0.2109], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0209, 0.0168, 0.0204, 0.0198, 0.0184, 0.0155, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 20:39:10,049 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.1469, 4.0733, 3.0344, 4.7873, 4.1793, 4.0705, 1.7492, 4.0515], device='cuda:4'), covar=tensor([0.1616, 0.1091, 0.3936, 0.0944, 0.2622, 0.1514, 0.5620, 0.2205], device='cuda:4'), in_proj_covar=tensor([0.0246, 0.0221, 0.0255, 0.0308, 0.0299, 0.0249, 0.0276, 0.0277], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 20:39:15,074 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=134653.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:39:24,552 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=134660.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:39:33,044 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=134665.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:39:34,170 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.095e+02 1.576e+02 1.923e+02 2.277e+02 4.210e+02, threshold=3.845e+02, percent-clipped=2.0 2023-04-27 20:40:04,276 INFO [finetune.py:976] (4/7) Epoch 24, batch 2950, loss[loss=0.1735, simple_loss=0.2494, pruned_loss=0.04877, over 4882.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.2406, pruned_loss=0.04739, over 953348.14 frames. ], batch size: 35, lr: 3.05e-03, grad_scale: 32.0 2023-04-27 20:40:17,333 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1525, 1.4932, 1.4127, 1.6966, 1.5789, 1.6810, 1.3663, 3.0392], device='cuda:4'), covar=tensor([0.0634, 0.0825, 0.0760, 0.1164, 0.0659, 0.0522, 0.0737, 0.0162], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0040, 0.0043, 0.0040, 0.0038, 0.0038, 0.0056], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0014, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 20:40:18,493 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=134701.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:41:09,548 INFO [finetune.py:976] (4/7) Epoch 24, batch 3000, loss[loss=0.1827, simple_loss=0.2571, pruned_loss=0.05411, over 4858.00 frames. ], tot_loss[loss=0.1696, simple_loss=0.2427, pruned_loss=0.04818, over 955645.82 frames. ], batch size: 49, lr: 3.05e-03, grad_scale: 32.0 2023-04-27 20:41:09,548 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 20:41:25,508 INFO [finetune.py:1010] (4/7) Epoch 24, validation: loss=0.1526, simple_loss=0.2221, pruned_loss=0.04154, over 2265189.00 frames. 2023-04-27 20:41:25,509 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6529MB 2023-04-27 20:41:44,219 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.083e+02 1.565e+02 1.917e+02 2.257e+02 3.857e+02, threshold=3.833e+02, percent-clipped=1.0 2023-04-27 20:41:50,479 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4910, 0.9931, 1.2611, 1.1040, 1.5848, 1.3053, 1.0173, 1.2013], device='cuda:4'), covar=tensor([0.1352, 0.1416, 0.1770, 0.1370, 0.0823, 0.1477, 0.1889, 0.2173], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0307, 0.0351, 0.0283, 0.0327, 0.0306, 0.0299, 0.0373], device='cuda:4'), out_proj_covar=tensor([6.3776e-05, 6.3354e-05, 7.3858e-05, 5.6670e-05, 6.7416e-05, 6.4116e-05, 6.2175e-05, 7.9076e-05], device='cuda:4') 2023-04-27 20:41:57,442 INFO [finetune.py:976] (4/7) Epoch 24, batch 3050, loss[loss=0.1354, simple_loss=0.2023, pruned_loss=0.03427, over 4703.00 frames. ], tot_loss[loss=0.1712, simple_loss=0.2445, pruned_loss=0.04891, over 955463.44 frames. ], batch size: 23, lr: 3.05e-03, grad_scale: 32.0 2023-04-27 20:42:06,489 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-27 20:42:30,085 INFO [finetune.py:976] (4/7) Epoch 24, batch 3100, loss[loss=0.1797, simple_loss=0.2502, pruned_loss=0.05458, over 4919.00 frames. ], tot_loss[loss=0.1697, simple_loss=0.2426, pruned_loss=0.04837, over 954403.51 frames. ], batch size: 33, lr: 3.05e-03, grad_scale: 32.0 2023-04-27 20:42:47,234 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=2.06 vs. limit=2.0 2023-04-27 20:42:47,425 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=134862.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:42:50,371 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.815e+01 1.536e+02 1.798e+02 2.119e+02 3.796e+02, threshold=3.595e+02, percent-clipped=0.0 2023-04-27 20:42:56,025 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8408, 2.3807, 0.8004, 1.1827, 1.5168, 1.0806, 2.4611, 1.2762], device='cuda:4'), covar=tensor([0.0770, 0.0609, 0.0751, 0.1524, 0.0543, 0.1191, 0.0369, 0.0853], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0073, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 20:43:02,615 INFO [finetune.py:976] (4/7) Epoch 24, batch 3150, loss[loss=0.1269, simple_loss=0.2033, pruned_loss=0.02522, over 4758.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.2402, pruned_loss=0.04764, over 953809.67 frames. ], batch size: 26, lr: 3.05e-03, grad_scale: 32.0 2023-04-27 20:43:18,289 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1883, 2.6532, 2.2898, 2.4464, 1.8996, 2.2777, 2.2823, 1.8299], device='cuda:4'), covar=tensor([0.1966, 0.1393, 0.0836, 0.1458, 0.3478, 0.1174, 0.1989, 0.2702], device='cuda:4'), in_proj_covar=tensor([0.0285, 0.0301, 0.0216, 0.0277, 0.0313, 0.0255, 0.0251, 0.0265], device='cuda:4'), out_proj_covar=tensor([1.1366e-04, 1.1888e-04, 8.4815e-05, 1.0922e-04, 1.2624e-04, 1.0049e-04, 1.0129e-04, 1.0483e-04], device='cuda:4') 2023-04-27 20:43:28,086 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=134923.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:43:29,995 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8093, 1.2677, 1.8774, 2.3056, 1.9143, 1.7489, 1.8193, 1.7522], device='cuda:4'), covar=tensor([0.4356, 0.6564, 0.5925, 0.5279, 0.5749, 0.7517, 0.7405, 0.8434], device='cuda:4'), in_proj_covar=tensor([0.0434, 0.0418, 0.0510, 0.0505, 0.0464, 0.0497, 0.0501, 0.0511], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 20:43:36,610 INFO [finetune.py:976] (4/7) Epoch 24, batch 3200, loss[loss=0.1613, simple_loss=0.233, pruned_loss=0.04478, over 4776.00 frames. ], tot_loss[loss=0.165, simple_loss=0.2369, pruned_loss=0.04656, over 954239.70 frames. ], batch size: 28, lr: 3.05e-03, grad_scale: 32.0 2023-04-27 20:43:53,553 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=134960.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:43:56,584 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=134965.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:43:57,659 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.048e+02 1.515e+02 1.832e+02 2.291e+02 6.391e+02, threshold=3.663e+02, percent-clipped=4.0 2023-04-27 20:44:10,013 INFO [finetune.py:976] (4/7) Epoch 24, batch 3250, loss[loss=0.1796, simple_loss=0.2378, pruned_loss=0.06071, over 4657.00 frames. ], tot_loss[loss=0.1652, simple_loss=0.2372, pruned_loss=0.04657, over 954421.69 frames. ], batch size: 23, lr: 3.05e-03, grad_scale: 32.0 2023-04-27 20:44:25,317 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=135008.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:44:28,802 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=135013.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:44:39,224 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9223, 2.3630, 1.3583, 1.5835, 2.2893, 1.7377, 1.6108, 1.7955], device='cuda:4'), covar=tensor([0.0451, 0.0330, 0.0280, 0.0544, 0.0236, 0.0495, 0.0489, 0.0517], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0028, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0046, 0.0038, 0.0052, 0.0039, 0.0050, 0.0050, 0.0052], device='cuda:4') 2023-04-27 20:44:43,394 INFO [finetune.py:976] (4/7) Epoch 24, batch 3300, loss[loss=0.2046, simple_loss=0.276, pruned_loss=0.06661, over 4847.00 frames. ], tot_loss[loss=0.1686, simple_loss=0.241, pruned_loss=0.04807, over 954925.09 frames. ], batch size: 47, lr: 3.05e-03, grad_scale: 32.0 2023-04-27 20:44:52,593 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.58 vs. limit=5.0 2023-04-27 20:45:15,631 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.097e+02 1.494e+02 1.834e+02 2.222e+02 6.565e+02, threshold=3.668e+02, percent-clipped=2.0 2023-04-27 20:45:44,559 INFO [finetune.py:976] (4/7) Epoch 24, batch 3350, loss[loss=0.1615, simple_loss=0.2436, pruned_loss=0.03969, over 4809.00 frames. ], tot_loss[loss=0.1706, simple_loss=0.2439, pruned_loss=0.04863, over 954518.76 frames. ], batch size: 45, lr: 3.05e-03, grad_scale: 32.0 2023-04-27 20:46:49,139 INFO [finetune.py:976] (4/7) Epoch 24, batch 3400, loss[loss=0.1833, simple_loss=0.2506, pruned_loss=0.05796, over 4915.00 frames. ], tot_loss[loss=0.1718, simple_loss=0.2452, pruned_loss=0.04921, over 954662.53 frames. ], batch size: 38, lr: 3.05e-03, grad_scale: 32.0 2023-04-27 20:46:51,727 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6311, 3.9763, 0.7281, 2.0751, 2.2288, 2.9724, 2.2835, 0.9629], device='cuda:4'), covar=tensor([0.1480, 0.1181, 0.2321, 0.1362, 0.1012, 0.0979, 0.1555, 0.2048], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0236, 0.0134, 0.0119, 0.0131, 0.0151, 0.0116, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 20:47:25,515 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.127e+02 1.583e+02 1.888e+02 2.345e+02 4.447e+02, threshold=3.777e+02, percent-clipped=3.0 2023-04-27 20:47:33,955 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=135171.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:47:45,332 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0981, 0.7720, 0.9557, 0.8297, 1.1574, 0.9609, 0.8272, 0.9628], device='cuda:4'), covar=tensor([0.1865, 0.1435, 0.2358, 0.1583, 0.1076, 0.1620, 0.1804, 0.2637], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0305, 0.0349, 0.0281, 0.0325, 0.0303, 0.0297, 0.0370], device='cuda:4'), out_proj_covar=tensor([6.3512e-05, 6.2820e-05, 7.3514e-05, 5.6325e-05, 6.6875e-05, 6.3517e-05, 6.1745e-05, 7.8342e-05], device='cuda:4') 2023-04-27 20:47:54,156 INFO [finetune.py:976] (4/7) Epoch 24, batch 3450, loss[loss=0.142, simple_loss=0.2091, pruned_loss=0.03745, over 4726.00 frames. ], tot_loss[loss=0.1703, simple_loss=0.2438, pruned_loss=0.04843, over 954224.44 frames. ], batch size: 23, lr: 3.05e-03, grad_scale: 32.0 2023-04-27 20:47:56,818 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-27 20:48:27,446 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=135218.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:48:36,541 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=135232.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:48:39,438 INFO [finetune.py:976] (4/7) Epoch 24, batch 3500, loss[loss=0.2185, simple_loss=0.2816, pruned_loss=0.07776, over 4248.00 frames. ], tot_loss[loss=0.1686, simple_loss=0.2412, pruned_loss=0.04799, over 951191.81 frames. ], batch size: 66, lr: 3.04e-03, grad_scale: 32.0 2023-04-27 20:48:59,253 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.073e+02 1.566e+02 1.818e+02 2.142e+02 4.699e+02, threshold=3.637e+02, percent-clipped=2.0 2023-04-27 20:49:13,318 INFO [finetune.py:976] (4/7) Epoch 24, batch 3550, loss[loss=0.1344, simple_loss=0.2113, pruned_loss=0.0287, over 4762.00 frames. ], tot_loss[loss=0.166, simple_loss=0.2385, pruned_loss=0.04677, over 953061.10 frames. ], batch size: 26, lr: 3.04e-03, grad_scale: 32.0 2023-04-27 20:49:17,154 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2039, 1.6155, 2.0261, 2.2163, 2.0208, 1.6539, 1.0929, 1.7110], device='cuda:4'), covar=tensor([0.3515, 0.3401, 0.1851, 0.2510, 0.2777, 0.2834, 0.4447, 0.2078], device='cuda:4'), in_proj_covar=tensor([0.0296, 0.0246, 0.0229, 0.0316, 0.0222, 0.0236, 0.0230, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 20:49:47,283 INFO [finetune.py:976] (4/7) Epoch 24, batch 3600, loss[loss=0.1455, simple_loss=0.2161, pruned_loss=0.0375, over 4914.00 frames. ], tot_loss[loss=0.1661, simple_loss=0.2373, pruned_loss=0.04739, over 952650.71 frames. ], batch size: 37, lr: 3.04e-03, grad_scale: 32.0 2023-04-27 20:50:01,506 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-27 20:50:05,964 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.034e+02 1.602e+02 1.884e+02 2.235e+02 3.277e+02, threshold=3.769e+02, percent-clipped=0.0 2023-04-27 20:50:20,186 INFO [finetune.py:976] (4/7) Epoch 24, batch 3650, loss[loss=0.1188, simple_loss=0.1941, pruned_loss=0.02172, over 4746.00 frames. ], tot_loss[loss=0.1662, simple_loss=0.2379, pruned_loss=0.04726, over 952817.44 frames. ], batch size: 23, lr: 3.04e-03, grad_scale: 32.0 2023-04-27 20:50:53,967 INFO [finetune.py:976] (4/7) Epoch 24, batch 3700, loss[loss=0.1509, simple_loss=0.2271, pruned_loss=0.03732, over 4680.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2417, pruned_loss=0.04855, over 953447.11 frames. ], batch size: 23, lr: 3.04e-03, grad_scale: 32.0 2023-04-27 20:51:12,081 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.75 vs. limit=2.0 2023-04-27 20:51:12,479 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.055e+02 1.617e+02 1.866e+02 2.198e+02 4.179e+02, threshold=3.733e+02, percent-clipped=1.0 2023-04-27 20:51:27,090 INFO [finetune.py:976] (4/7) Epoch 24, batch 3750, loss[loss=0.1742, simple_loss=0.2491, pruned_loss=0.04963, over 4736.00 frames. ], tot_loss[loss=0.1722, simple_loss=0.2444, pruned_loss=0.04997, over 952452.03 frames. ], batch size: 54, lr: 3.04e-03, grad_scale: 32.0 2023-04-27 20:51:44,417 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5621, 3.4059, 0.7593, 1.8612, 1.9341, 2.4791, 1.9543, 1.0103], device='cuda:4'), covar=tensor([0.1197, 0.0866, 0.2037, 0.1112, 0.0968, 0.0960, 0.1400, 0.2021], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0236, 0.0135, 0.0120, 0.0131, 0.0151, 0.0116, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 20:52:05,322 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=135518.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:52:15,494 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3060, 1.8564, 2.1236, 2.6476, 2.3074, 1.8116, 1.6996, 2.0886], device='cuda:4'), covar=tensor([0.2534, 0.2546, 0.1348, 0.1949, 0.2056, 0.2139, 0.3713, 0.1848], device='cuda:4'), in_proj_covar=tensor([0.0296, 0.0246, 0.0229, 0.0317, 0.0222, 0.0236, 0.0230, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 20:52:16,497 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=135527.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:52:29,077 INFO [finetune.py:976] (4/7) Epoch 24, batch 3800, loss[loss=0.1442, simple_loss=0.2265, pruned_loss=0.03094, over 4756.00 frames. ], tot_loss[loss=0.1716, simple_loss=0.2445, pruned_loss=0.04938, over 954008.19 frames. ], batch size: 27, lr: 3.04e-03, grad_scale: 32.0 2023-04-27 20:53:08,363 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=135566.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:53:08,921 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.795e+01 1.517e+02 1.799e+02 2.182e+02 4.922e+02, threshold=3.597e+02, percent-clipped=3.0 2023-04-27 20:53:32,879 INFO [finetune.py:976] (4/7) Epoch 24, batch 3850, loss[loss=0.1518, simple_loss=0.2174, pruned_loss=0.04312, over 4883.00 frames. ], tot_loss[loss=0.1691, simple_loss=0.242, pruned_loss=0.0481, over 953282.96 frames. ], batch size: 32, lr: 3.04e-03, grad_scale: 32.0 2023-04-27 20:54:28,827 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=135630.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:54:33,844 INFO [finetune.py:976] (4/7) Epoch 24, batch 3900, loss[loss=0.1482, simple_loss=0.218, pruned_loss=0.03922, over 4698.00 frames. ], tot_loss[loss=0.1678, simple_loss=0.2401, pruned_loss=0.04773, over 953157.20 frames. ], batch size: 23, lr: 3.04e-03, grad_scale: 32.0 2023-04-27 20:55:15,856 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.029e+02 1.536e+02 1.810e+02 2.224e+02 3.785e+02, threshold=3.619e+02, percent-clipped=2.0 2023-04-27 20:55:44,888 INFO [finetune.py:976] (4/7) Epoch 24, batch 3950, loss[loss=0.1638, simple_loss=0.2377, pruned_loss=0.04495, over 4819.00 frames. ], tot_loss[loss=0.1654, simple_loss=0.2375, pruned_loss=0.04671, over 954094.79 frames. ], batch size: 41, lr: 3.04e-03, grad_scale: 32.0 2023-04-27 20:55:47,449 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=135691.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:56:44,335 INFO [finetune.py:976] (4/7) Epoch 24, batch 4000, loss[loss=0.1251, simple_loss=0.2027, pruned_loss=0.02374, over 4685.00 frames. ], tot_loss[loss=0.1643, simple_loss=0.2364, pruned_loss=0.04615, over 952244.49 frames. ], batch size: 23, lr: 3.04e-03, grad_scale: 32.0 2023-04-27 20:57:22,846 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.079e+02 1.579e+02 1.856e+02 2.321e+02 4.338e+02, threshold=3.712e+02, percent-clipped=1.0 2023-04-27 20:57:52,041 INFO [finetune.py:976] (4/7) Epoch 24, batch 4050, loss[loss=0.1817, simple_loss=0.2589, pruned_loss=0.0522, over 4870.00 frames. ], tot_loss[loss=0.1672, simple_loss=0.2399, pruned_loss=0.0472, over 952059.33 frames. ], batch size: 34, lr: 3.04e-03, grad_scale: 32.0 2023-04-27 20:58:07,475 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2902, 1.9784, 2.4337, 2.6886, 2.2869, 2.2217, 2.3248, 2.3296], device='cuda:4'), covar=tensor([0.4630, 0.7230, 0.7263, 0.5610, 0.6482, 0.8603, 0.8509, 0.9604], device='cuda:4'), in_proj_covar=tensor([0.0433, 0.0416, 0.0508, 0.0504, 0.0464, 0.0495, 0.0500, 0.0512], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 20:58:25,452 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8297, 1.6728, 1.9341, 2.1645, 2.2048, 1.7674, 1.4766, 1.9660], device='cuda:4'), covar=tensor([0.0891, 0.1309, 0.0866, 0.0726, 0.0669, 0.0941, 0.0856, 0.0678], device='cuda:4'), in_proj_covar=tensor([0.0189, 0.0206, 0.0189, 0.0176, 0.0181, 0.0182, 0.0153, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 20:58:25,477 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8429, 2.3657, 1.8667, 1.7608, 1.3101, 1.3330, 1.9009, 1.3040], device='cuda:4'), covar=tensor([0.1791, 0.1244, 0.1389, 0.1640, 0.2317, 0.2050, 0.0979, 0.2042], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0210, 0.0169, 0.0205, 0.0200, 0.0185, 0.0157, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 20:58:39,906 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=135827.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:58:52,538 INFO [finetune.py:976] (4/7) Epoch 24, batch 4100, loss[loss=0.1902, simple_loss=0.2441, pruned_loss=0.06811, over 4832.00 frames. ], tot_loss[loss=0.1693, simple_loss=0.2428, pruned_loss=0.04795, over 952239.93 frames. ], batch size: 30, lr: 3.04e-03, grad_scale: 32.0 2023-04-27 20:59:02,539 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-04-27 20:59:20,815 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.99 vs. limit=5.0 2023-04-27 20:59:21,926 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.49 vs. limit=2.0 2023-04-27 20:59:24,281 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7413, 1.3141, 1.7963, 2.2436, 1.8064, 1.6977, 1.7640, 1.6937], device='cuda:4'), covar=tensor([0.4386, 0.6657, 0.6616, 0.6031, 0.5612, 0.7922, 0.7941, 0.8914], device='cuda:4'), in_proj_covar=tensor([0.0434, 0.0417, 0.0509, 0.0506, 0.0465, 0.0497, 0.0501, 0.0513], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 20:59:28,243 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.032e+02 1.687e+02 1.984e+02 2.360e+02 5.004e+02, threshold=3.968e+02, percent-clipped=3.0 2023-04-27 20:59:33,115 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=135875.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 20:59:40,397 INFO [finetune.py:976] (4/7) Epoch 24, batch 4150, loss[loss=0.1657, simple_loss=0.2374, pruned_loss=0.04698, over 4808.00 frames. ], tot_loss[loss=0.1703, simple_loss=0.2439, pruned_loss=0.04841, over 952710.39 frames. ], batch size: 33, lr: 3.04e-03, grad_scale: 32.0 2023-04-27 20:59:46,038 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-27 20:59:54,941 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6526, 1.9115, 1.8326, 2.5245, 2.6224, 2.1896, 2.0725, 1.8015], device='cuda:4'), covar=tensor([0.1463, 0.1767, 0.1860, 0.1148, 0.0964, 0.1970, 0.1973, 0.2354], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0303, 0.0345, 0.0279, 0.0321, 0.0300, 0.0293, 0.0366], device='cuda:4'), out_proj_covar=tensor([6.2972e-05, 6.2362e-05, 7.2536e-05, 5.5936e-05, 6.6052e-05, 6.2847e-05, 6.1019e-05, 7.7620e-05], device='cuda:4') 2023-04-27 21:00:07,660 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1325, 1.8345, 2.3182, 2.6157, 2.1477, 2.0819, 2.2069, 2.1468], device='cuda:4'), covar=tensor([0.4769, 0.7415, 0.7408, 0.5720, 0.6506, 0.8370, 0.8455, 0.9646], device='cuda:4'), in_proj_covar=tensor([0.0436, 0.0420, 0.0511, 0.0508, 0.0468, 0.0499, 0.0503, 0.0516], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 21:00:14,127 INFO [finetune.py:976] (4/7) Epoch 24, batch 4200, loss[loss=0.1533, simple_loss=0.2416, pruned_loss=0.03247, over 4862.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2435, pruned_loss=0.0476, over 952073.02 frames. ], batch size: 34, lr: 3.04e-03, grad_scale: 32.0 2023-04-27 21:00:35,176 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.056e+02 1.525e+02 1.774e+02 2.091e+02 5.050e+02, threshold=3.548e+02, percent-clipped=3.0 2023-04-27 21:00:44,751 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=135982.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:00:47,168 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=135986.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:00:47,703 INFO [finetune.py:976] (4/7) Epoch 24, batch 4250, loss[loss=0.1569, simple_loss=0.2301, pruned_loss=0.04183, over 4826.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.241, pruned_loss=0.04715, over 955214.87 frames. ], batch size: 38, lr: 3.04e-03, grad_scale: 32.0 2023-04-27 21:00:52,096 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=135994.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:01:22,184 INFO [finetune.py:976] (4/7) Epoch 24, batch 4300, loss[loss=0.1713, simple_loss=0.2427, pruned_loss=0.04999, over 4765.00 frames. ], tot_loss[loss=0.1654, simple_loss=0.238, pruned_loss=0.04636, over 955463.98 frames. ], batch size: 27, lr: 3.04e-03, grad_scale: 32.0 2023-04-27 21:01:25,997 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=136043.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:01:34,850 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=136055.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:01:42,944 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.901e+01 1.479e+02 1.753e+02 2.087e+02 3.589e+02, threshold=3.506e+02, percent-clipped=1.0 2023-04-27 21:01:46,130 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5563, 1.1033, 0.3963, 1.2468, 1.1097, 1.4393, 1.3096, 1.2809], device='cuda:4'), covar=tensor([0.0478, 0.0400, 0.0420, 0.0563, 0.0310, 0.0505, 0.0516, 0.0564], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0028, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0046, 0.0038, 0.0052, 0.0039, 0.0050, 0.0050, 0.0052], device='cuda:4') 2023-04-27 21:01:55,531 INFO [finetune.py:976] (4/7) Epoch 24, batch 4350, loss[loss=0.1659, simple_loss=0.2503, pruned_loss=0.04072, over 4812.00 frames. ], tot_loss[loss=0.1638, simple_loss=0.2361, pruned_loss=0.04581, over 956573.48 frames. ], batch size: 39, lr: 3.04e-03, grad_scale: 32.0 2023-04-27 21:02:04,165 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0581, 1.0118, 1.2291, 1.1615, 1.0477, 0.9638, 1.0333, 0.6107], device='cuda:4'), covar=tensor([0.0469, 0.0561, 0.0401, 0.0508, 0.0639, 0.1118, 0.0453, 0.0547], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0067, 0.0065, 0.0067, 0.0073, 0.0094, 0.0072, 0.0063], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 21:02:05,390 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-04-27 21:02:33,975 INFO [finetune.py:976] (4/7) Epoch 24, batch 4400, loss[loss=0.2266, simple_loss=0.2887, pruned_loss=0.08221, over 4808.00 frames. ], tot_loss[loss=0.1656, simple_loss=0.2377, pruned_loss=0.04674, over 954509.80 frames. ], batch size: 51, lr: 3.04e-03, grad_scale: 32.0 2023-04-27 21:02:42,606 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=136142.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:02:45,432 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.83 vs. limit=5.0 2023-04-27 21:02:54,278 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.60 vs. limit=2.0 2023-04-27 21:03:16,196 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.001e+02 1.622e+02 1.920e+02 2.374e+02 4.375e+02, threshold=3.840e+02, percent-clipped=3.0 2023-04-27 21:03:40,891 INFO [finetune.py:976] (4/7) Epoch 24, batch 4450, loss[loss=0.1743, simple_loss=0.252, pruned_loss=0.04832, over 4849.00 frames. ], tot_loss[loss=0.169, simple_loss=0.2416, pruned_loss=0.04822, over 954322.16 frames. ], batch size: 47, lr: 3.04e-03, grad_scale: 32.0 2023-04-27 21:04:02,271 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=136203.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:04:53,817 INFO [finetune.py:976] (4/7) Epoch 24, batch 4500, loss[loss=0.2316, simple_loss=0.2875, pruned_loss=0.08782, over 4792.00 frames. ], tot_loss[loss=0.1704, simple_loss=0.243, pruned_loss=0.04892, over 952978.69 frames. ], batch size: 51, lr: 3.04e-03, grad_scale: 32.0 2023-04-27 21:05:29,629 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.760e+01 1.501e+02 1.833e+02 2.147e+02 4.216e+02, threshold=3.666e+02, percent-clipped=1.0 2023-04-27 21:05:58,020 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=136286.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:05:58,525 INFO [finetune.py:976] (4/7) Epoch 24, batch 4550, loss[loss=0.2531, simple_loss=0.3027, pruned_loss=0.1018, over 4261.00 frames. ], tot_loss[loss=0.1722, simple_loss=0.2448, pruned_loss=0.04978, over 953203.30 frames. ], batch size: 65, lr: 3.04e-03, grad_scale: 32.0 2023-04-27 21:06:56,735 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=136334.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:06:57,402 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=136335.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:07:04,275 INFO [finetune.py:976] (4/7) Epoch 24, batch 4600, loss[loss=0.2216, simple_loss=0.2813, pruned_loss=0.08093, over 4820.00 frames. ], tot_loss[loss=0.1724, simple_loss=0.2453, pruned_loss=0.04972, over 952878.98 frames. ], batch size: 33, lr: 3.04e-03, grad_scale: 32.0 2023-04-27 21:07:04,984 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=136338.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:07:18,413 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=136350.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:07:40,424 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.023e+02 1.579e+02 1.858e+02 2.127e+02 3.942e+02, threshold=3.716e+02, percent-clipped=1.0 2023-04-27 21:08:03,892 INFO [finetune.py:976] (4/7) Epoch 24, batch 4650, loss[loss=0.1474, simple_loss=0.2215, pruned_loss=0.03663, over 4808.00 frames. ], tot_loss[loss=0.1703, simple_loss=0.2425, pruned_loss=0.04904, over 951845.46 frames. ], batch size: 25, lr: 3.04e-03, grad_scale: 32.0 2023-04-27 21:08:20,796 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=136396.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:09:13,810 INFO [finetune.py:976] (4/7) Epoch 24, batch 4700, loss[loss=0.123, simple_loss=0.1881, pruned_loss=0.02893, over 4841.00 frames. ], tot_loss[loss=0.1669, simple_loss=0.2388, pruned_loss=0.04752, over 952663.45 frames. ], batch size: 49, lr: 3.04e-03, grad_scale: 32.0 2023-04-27 21:09:27,628 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-27 21:09:45,099 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.039e+02 1.527e+02 1.865e+02 2.204e+02 5.324e+02, threshold=3.729e+02, percent-clipped=2.0 2023-04-27 21:09:51,226 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6757, 1.7592, 0.8180, 1.3888, 1.8236, 1.5576, 1.4643, 1.5457], device='cuda:4'), covar=tensor([0.0501, 0.0360, 0.0356, 0.0563, 0.0276, 0.0519, 0.0489, 0.0556], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0028, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0045, 0.0038, 0.0052, 0.0039, 0.0050, 0.0050, 0.0052], device='cuda:4') 2023-04-27 21:09:58,845 INFO [finetune.py:976] (4/7) Epoch 24, batch 4750, loss[loss=0.1443, simple_loss=0.2276, pruned_loss=0.03046, over 4744.00 frames. ], tot_loss[loss=0.1658, simple_loss=0.2373, pruned_loss=0.04709, over 953288.39 frames. ], batch size: 27, lr: 3.04e-03, grad_scale: 32.0 2023-04-27 21:10:07,141 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=136498.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:10:47,891 INFO [finetune.py:976] (4/7) Epoch 24, batch 4800, loss[loss=0.1879, simple_loss=0.2623, pruned_loss=0.05679, over 4815.00 frames. ], tot_loss[loss=0.1667, simple_loss=0.2382, pruned_loss=0.04761, over 953077.68 frames. ], batch size: 45, lr: 3.04e-03, grad_scale: 32.0 2023-04-27 21:11:18,274 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.72 vs. limit=5.0 2023-04-27 21:11:29,602 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.193e+02 1.561e+02 1.788e+02 2.083e+02 3.546e+02, threshold=3.576e+02, percent-clipped=0.0 2023-04-27 21:11:42,425 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9795, 1.6875, 1.9643, 2.2782, 2.3202, 1.8218, 1.5924, 2.0749], device='cuda:4'), covar=tensor([0.0851, 0.1288, 0.0794, 0.0632, 0.0555, 0.0849, 0.0782, 0.0596], device='cuda:4'), in_proj_covar=tensor([0.0188, 0.0204, 0.0187, 0.0175, 0.0179, 0.0179, 0.0152, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 21:11:53,505 INFO [finetune.py:976] (4/7) Epoch 24, batch 4850, loss[loss=0.1833, simple_loss=0.2514, pruned_loss=0.05757, over 4779.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2413, pruned_loss=0.0483, over 953555.24 frames. ], batch size: 25, lr: 3.04e-03, grad_scale: 64.0 2023-04-27 21:12:26,524 INFO [finetune.py:976] (4/7) Epoch 24, batch 4900, loss[loss=0.1709, simple_loss=0.2404, pruned_loss=0.05075, over 4784.00 frames. ], tot_loss[loss=0.1706, simple_loss=0.2432, pruned_loss=0.04901, over 952327.15 frames. ], batch size: 51, lr: 3.03e-03, grad_scale: 32.0 2023-04-27 21:12:27,226 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=136638.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:12:30,260 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=136643.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:12:35,594 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=136650.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:12:46,890 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.047e+02 1.604e+02 1.891e+02 2.229e+02 4.573e+02, threshold=3.781e+02, percent-clipped=1.0 2023-04-27 21:12:50,610 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.4257, 3.4381, 2.5947, 4.0394, 3.5095, 3.4529, 1.5790, 3.4193], device='cuda:4'), covar=tensor([0.1962, 0.1513, 0.3783, 0.1982, 0.2653, 0.1938, 0.5640, 0.2466], device='cuda:4'), in_proj_covar=tensor([0.0245, 0.0220, 0.0253, 0.0308, 0.0297, 0.0249, 0.0275, 0.0272], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 21:12:58,275 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=136686.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:12:58,837 INFO [finetune.py:976] (4/7) Epoch 24, batch 4950, loss[loss=0.1572, simple_loss=0.2285, pruned_loss=0.04294, over 4884.00 frames. ], tot_loss[loss=0.1714, simple_loss=0.2441, pruned_loss=0.04935, over 953138.56 frames. ], batch size: 32, lr: 3.03e-03, grad_scale: 32.0 2023-04-27 21:13:02,166 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=136691.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:13:07,455 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=136698.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:13:09,025 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.83 vs. limit=2.0 2023-04-27 21:13:11,671 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=136704.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:13:30,264 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3887, 2.8937, 2.3579, 2.2778, 1.6902, 1.6698, 2.4241, 1.7315], device='cuda:4'), covar=tensor([0.1562, 0.1334, 0.1225, 0.1659, 0.2210, 0.1811, 0.0897, 0.1859], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0211, 0.0169, 0.0205, 0.0200, 0.0186, 0.0157, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 21:13:32,451 INFO [finetune.py:976] (4/7) Epoch 24, batch 5000, loss[loss=0.1548, simple_loss=0.2223, pruned_loss=0.04363, over 4790.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2436, pruned_loss=0.04931, over 954482.62 frames. ], batch size: 29, lr: 3.03e-03, grad_scale: 32.0 2023-04-27 21:13:53,647 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.026e+02 1.571e+02 1.815e+02 2.179e+02 3.304e+02, threshold=3.630e+02, percent-clipped=0.0 2023-04-27 21:13:57,960 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4282, 1.2926, 1.6239, 1.6424, 1.3732, 1.2760, 1.2474, 0.6556], device='cuda:4'), covar=tensor([0.0517, 0.0602, 0.0374, 0.0523, 0.0677, 0.1115, 0.0601, 0.0607], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0067, 0.0066, 0.0068, 0.0074, 0.0095, 0.0073, 0.0063], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 21:14:01,067 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4077, 1.7480, 1.6100, 1.8794, 1.8106, 1.9085, 1.6114, 3.1001], device='cuda:4'), covar=tensor([0.0575, 0.0626, 0.0632, 0.0953, 0.0501, 0.0633, 0.0620, 0.0180], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0038, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 21:14:05,878 INFO [finetune.py:976] (4/7) Epoch 24, batch 5050, loss[loss=0.1692, simple_loss=0.2444, pruned_loss=0.04703, over 4830.00 frames. ], tot_loss[loss=0.1706, simple_loss=0.2421, pruned_loss=0.04955, over 952134.55 frames. ], batch size: 40, lr: 3.03e-03, grad_scale: 32.0 2023-04-27 21:14:09,214 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.69 vs. limit=5.0 2023-04-27 21:14:18,800 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=136798.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:14:29,847 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=136806.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:14:30,433 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8362, 1.1654, 4.7822, 4.5147, 4.1814, 4.4931, 4.2698, 4.1208], device='cuda:4'), covar=tensor([0.7157, 0.6175, 0.0983, 0.1744, 0.1078, 0.1346, 0.1975, 0.1786], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0308, 0.0408, 0.0410, 0.0351, 0.0412, 0.0321, 0.0370], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 21:14:58,799 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9541, 0.9936, 1.2282, 1.2080, 1.0300, 0.8936, 1.0002, 0.7009], device='cuda:4'), covar=tensor([0.0561, 0.0590, 0.0444, 0.0471, 0.0664, 0.1274, 0.0459, 0.0626], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0067, 0.0066, 0.0068, 0.0074, 0.0095, 0.0073, 0.0063], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 21:15:01,715 INFO [finetune.py:976] (4/7) Epoch 24, batch 5100, loss[loss=0.1721, simple_loss=0.237, pruned_loss=0.05354, over 4831.00 frames. ], tot_loss[loss=0.167, simple_loss=0.2382, pruned_loss=0.04793, over 952049.53 frames. ], batch size: 30, lr: 3.03e-03, grad_scale: 32.0 2023-04-27 21:15:07,749 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=136846.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:15:22,622 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=136867.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:15:23,112 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.056e+02 1.537e+02 1.873e+02 2.196e+02 3.846e+02, threshold=3.746e+02, percent-clipped=1.0 2023-04-27 21:15:28,012 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8638, 2.5620, 0.8932, 1.2709, 1.7422, 1.0689, 3.2723, 1.4202], device='cuda:4'), covar=tensor([0.0857, 0.0857, 0.0974, 0.1580, 0.0654, 0.1299, 0.0320, 0.0887], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0063, 0.0047, 0.0046, 0.0048, 0.0051, 0.0072, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0009, 0.0007, 0.0007, 0.0007, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-27 21:15:35,187 INFO [finetune.py:976] (4/7) Epoch 24, batch 5150, loss[loss=0.1744, simple_loss=0.253, pruned_loss=0.04792, over 4813.00 frames. ], tot_loss[loss=0.168, simple_loss=0.239, pruned_loss=0.04847, over 952995.40 frames. ], batch size: 51, lr: 3.03e-03, grad_scale: 32.0 2023-04-27 21:16:29,345 INFO [finetune.py:976] (4/7) Epoch 24, batch 5200, loss[loss=0.1812, simple_loss=0.2584, pruned_loss=0.05202, over 4913.00 frames. ], tot_loss[loss=0.1697, simple_loss=0.2414, pruned_loss=0.04902, over 951902.18 frames. ], batch size: 36, lr: 3.03e-03, grad_scale: 32.0 2023-04-27 21:16:59,707 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.230e+02 1.748e+02 2.026e+02 2.510e+02 4.483e+02, threshold=4.051e+02, percent-clipped=3.0 2023-04-27 21:17:22,717 INFO [finetune.py:976] (4/7) Epoch 24, batch 5250, loss[loss=0.1492, simple_loss=0.2302, pruned_loss=0.03404, over 4821.00 frames. ], tot_loss[loss=0.1696, simple_loss=0.2419, pruned_loss=0.04867, over 952165.03 frames. ], batch size: 38, lr: 3.03e-03, grad_scale: 32.0 2023-04-27 21:17:25,257 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=136991.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:17:36,066 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=136999.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:18:10,123 INFO [finetune.py:976] (4/7) Epoch 24, batch 5300, loss[loss=0.1553, simple_loss=0.2397, pruned_loss=0.03548, over 4831.00 frames. ], tot_loss[loss=0.1704, simple_loss=0.2433, pruned_loss=0.04878, over 952101.16 frames. ], batch size: 47, lr: 3.03e-03, grad_scale: 32.0 2023-04-27 21:18:11,867 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=137039.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:18:30,928 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.102e+02 1.584e+02 1.901e+02 2.207e+02 4.060e+02, threshold=3.801e+02, percent-clipped=1.0 2023-04-27 21:18:42,918 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0519, 2.4135, 1.0466, 1.3682, 1.8707, 1.2108, 3.0118, 1.5966], device='cuda:4'), covar=tensor([0.0667, 0.0499, 0.0731, 0.1343, 0.0521, 0.1071, 0.0389, 0.0668], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0072, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 21:18:43,523 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9442, 2.5302, 1.0559, 1.4232, 1.9523, 1.1483, 3.2689, 1.5811], device='cuda:4'), covar=tensor([0.0696, 0.0602, 0.0788, 0.1166, 0.0499, 0.1028, 0.0250, 0.0635], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0048, 0.0051, 0.0072, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 21:18:44,060 INFO [finetune.py:976] (4/7) Epoch 24, batch 5350, loss[loss=0.1648, simple_loss=0.2416, pruned_loss=0.04396, over 4833.00 frames. ], tot_loss[loss=0.1706, simple_loss=0.2439, pruned_loss=0.04861, over 952291.51 frames. ], batch size: 49, lr: 3.03e-03, grad_scale: 32.0 2023-04-27 21:19:16,911 INFO [finetune.py:976] (4/7) Epoch 24, batch 5400, loss[loss=0.1935, simple_loss=0.2611, pruned_loss=0.06299, over 4821.00 frames. ], tot_loss[loss=0.1682, simple_loss=0.2414, pruned_loss=0.04748, over 952703.24 frames. ], batch size: 30, lr: 3.03e-03, grad_scale: 32.0 2023-04-27 21:19:33,099 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=137162.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:19:37,583 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.035e+02 1.430e+02 1.764e+02 2.146e+02 4.768e+02, threshold=3.527e+02, percent-clipped=1.0 2023-04-27 21:19:50,714 INFO [finetune.py:976] (4/7) Epoch 24, batch 5450, loss[loss=0.134, simple_loss=0.2093, pruned_loss=0.02936, over 4824.00 frames. ], tot_loss[loss=0.1664, simple_loss=0.2384, pruned_loss=0.04723, over 953538.41 frames. ], batch size: 38, lr: 3.03e-03, grad_scale: 32.0 2023-04-27 21:20:00,003 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7261, 2.5012, 1.6390, 1.8195, 1.2661, 1.2875, 1.6789, 1.2055], device='cuda:4'), covar=tensor([0.1857, 0.1248, 0.1679, 0.1732, 0.2361, 0.2208, 0.1059, 0.2185], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0210, 0.0167, 0.0203, 0.0198, 0.0184, 0.0155, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 21:20:05,494 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9564, 2.3825, 1.9914, 2.2941, 1.7553, 1.9680, 1.9629, 1.4577], device='cuda:4'), covar=tensor([0.1813, 0.1022, 0.0886, 0.0976, 0.3463, 0.1023, 0.1971, 0.2810], device='cuda:4'), in_proj_covar=tensor([0.0282, 0.0297, 0.0214, 0.0274, 0.0310, 0.0254, 0.0248, 0.0263], device='cuda:4'), out_proj_covar=tensor([1.1265e-04, 1.1735e-04, 8.4302e-05, 1.0826e-04, 1.2520e-04, 9.9931e-05, 1.0021e-04, 1.0387e-04], device='cuda:4') 2023-04-27 21:20:19,557 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.81 vs. limit=2.0 2023-04-27 21:20:24,728 INFO [finetune.py:976] (4/7) Epoch 24, batch 5500, loss[loss=0.1554, simple_loss=0.2168, pruned_loss=0.04698, over 4126.00 frames. ], tot_loss[loss=0.1655, simple_loss=0.2367, pruned_loss=0.04713, over 956082.69 frames. ], batch size: 65, lr: 3.03e-03, grad_scale: 32.0 2023-04-27 21:20:44,124 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.686e+01 1.504e+02 1.776e+02 2.176e+02 3.969e+02, threshold=3.551e+02, percent-clipped=4.0 2023-04-27 21:20:51,852 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2185, 1.6432, 2.0941, 2.6256, 2.2341, 1.6479, 1.6142, 2.0027], device='cuda:4'), covar=tensor([0.2966, 0.3102, 0.1589, 0.2166, 0.2235, 0.2539, 0.3905, 0.1826], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0245, 0.0227, 0.0312, 0.0221, 0.0234, 0.0227, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 21:20:57,635 INFO [finetune.py:976] (4/7) Epoch 24, batch 5550, loss[loss=0.153, simple_loss=0.2281, pruned_loss=0.03897, over 4832.00 frames. ], tot_loss[loss=0.167, simple_loss=0.2386, pruned_loss=0.04767, over 956918.86 frames. ], batch size: 30, lr: 3.03e-03, grad_scale: 32.0 2023-04-27 21:21:05,049 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=137299.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:21:22,235 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3323, 2.1453, 1.7076, 1.8693, 2.2364, 1.8441, 2.7713, 1.5505], device='cuda:4'), covar=tensor([0.3738, 0.2197, 0.5584, 0.3137, 0.1864, 0.2591, 0.1424, 0.4906], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0350, 0.0424, 0.0348, 0.0376, 0.0373, 0.0365, 0.0420], device='cuda:4'), out_proj_covar=tensor([9.9794e-05, 1.0457e-04, 1.2829e-04, 1.0451e-04, 1.1148e-04, 1.1100e-04, 1.0708e-04, 1.2630e-04], device='cuda:4') 2023-04-27 21:21:34,763 INFO [finetune.py:976] (4/7) Epoch 24, batch 5600, loss[loss=0.1623, simple_loss=0.2345, pruned_loss=0.04512, over 4827.00 frames. ], tot_loss[loss=0.1685, simple_loss=0.2409, pruned_loss=0.04806, over 955713.91 frames. ], batch size: 33, lr: 3.03e-03, grad_scale: 32.0 2023-04-27 21:21:46,037 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=137347.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:21:46,097 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9450, 1.8427, 1.9713, 2.3601, 2.3842, 1.9311, 1.5435, 2.1359], device='cuda:4'), covar=tensor([0.0871, 0.1056, 0.0844, 0.0584, 0.0599, 0.0873, 0.0817, 0.0568], device='cuda:4'), in_proj_covar=tensor([0.0188, 0.0203, 0.0188, 0.0175, 0.0180, 0.0180, 0.0153, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 21:22:14,764 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.035e+02 1.541e+02 1.789e+02 2.150e+02 5.679e+02, threshold=3.578e+02, percent-clipped=2.0 2023-04-27 21:22:37,496 INFO [finetune.py:976] (4/7) Epoch 24, batch 5650, loss[loss=0.1748, simple_loss=0.2688, pruned_loss=0.04041, over 4814.00 frames. ], tot_loss[loss=0.1708, simple_loss=0.2441, pruned_loss=0.04873, over 957409.35 frames. ], batch size: 38, lr: 3.03e-03, grad_scale: 32.0 2023-04-27 21:22:49,123 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1215, 1.5408, 1.9509, 2.2062, 2.0064, 1.5180, 1.0940, 1.6568], device='cuda:4'), covar=tensor([0.2846, 0.2981, 0.1469, 0.1966, 0.2192, 0.2484, 0.3985, 0.1907], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0244, 0.0227, 0.0312, 0.0221, 0.0233, 0.0226, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 21:23:10,726 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1891, 1.4126, 1.2151, 1.3759, 1.1975, 1.1849, 1.2044, 0.9368], device='cuda:4'), covar=tensor([0.1390, 0.1071, 0.0750, 0.0888, 0.3205, 0.0991, 0.1359, 0.1902], device='cuda:4'), in_proj_covar=tensor([0.0283, 0.0297, 0.0214, 0.0275, 0.0311, 0.0253, 0.0248, 0.0263], device='cuda:4'), out_proj_covar=tensor([1.1309e-04, 1.1748e-04, 8.4161e-05, 1.0828e-04, 1.2541e-04, 9.9787e-05, 1.0019e-04, 1.0383e-04], device='cuda:4') 2023-04-27 21:23:28,787 INFO [finetune.py:976] (4/7) Epoch 24, batch 5700, loss[loss=0.1457, simple_loss=0.2072, pruned_loss=0.0421, over 4225.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.2402, pruned_loss=0.04761, over 940332.37 frames. ], batch size: 18, lr: 3.03e-03, grad_scale: 32.0 2023-04-27 21:23:43,839 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=137462.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:23:57,748 INFO [finetune.py:976] (4/7) Epoch 25, batch 0, loss[loss=0.1594, simple_loss=0.2409, pruned_loss=0.03895, over 4818.00 frames. ], tot_loss[loss=0.1594, simple_loss=0.2409, pruned_loss=0.03895, over 4818.00 frames. ], batch size: 47, lr: 3.03e-03, grad_scale: 32.0 2023-04-27 21:23:57,748 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 21:24:03,457 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.4806, 3.5335, 2.5653, 3.9131, 3.5988, 3.5199, 1.5946, 3.4891], device='cuda:4'), covar=tensor([0.1698, 0.1423, 0.3232, 0.2057, 0.2781, 0.1682, 0.5170, 0.2370], device='cuda:4'), in_proj_covar=tensor([0.0245, 0.0220, 0.0253, 0.0308, 0.0299, 0.0248, 0.0277, 0.0273], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 21:24:08,111 INFO [finetune.py:1010] (4/7) Epoch 25, validation: loss=0.155, simple_loss=0.224, pruned_loss=0.04295, over 2265189.00 frames. 2023-04-27 21:24:08,112 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6529MB 2023-04-27 21:24:09,920 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.060e+01 1.467e+02 1.816e+02 2.337e+02 4.208e+02, threshold=3.632e+02, percent-clipped=2.0 2023-04-27 21:24:37,395 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=137510.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:24:40,447 INFO [finetune.py:976] (4/7) Epoch 25, batch 50, loss[loss=0.1421, simple_loss=0.2093, pruned_loss=0.03747, over 4729.00 frames. ], tot_loss[loss=0.1709, simple_loss=0.2431, pruned_loss=0.04929, over 216456.42 frames. ], batch size: 54, lr: 3.02e-03, grad_scale: 32.0 2023-04-27 21:24:42,248 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6985, 1.6768, 0.8590, 1.4142, 1.7433, 1.5331, 1.4730, 1.5717], device='cuda:4'), covar=tensor([0.0472, 0.0370, 0.0343, 0.0549, 0.0256, 0.0530, 0.0501, 0.0536], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0028, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0046, 0.0038, 0.0052, 0.0038, 0.0050, 0.0050, 0.0052], device='cuda:4') 2023-04-27 21:24:44,275 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.87 vs. limit=5.0 2023-04-27 21:24:55,469 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3989, 2.7918, 1.2462, 1.7270, 2.3446, 1.4998, 3.9084, 2.1601], device='cuda:4'), covar=tensor([0.0578, 0.0556, 0.0764, 0.1217, 0.0467, 0.0967, 0.0266, 0.0564], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0063, 0.0047, 0.0046, 0.0048, 0.0051, 0.0072, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0009, 0.0007, 0.0007, 0.0007, 0.0008, 0.0010, 0.0007], device='cuda:4') 2023-04-27 21:24:57,774 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1356, 3.3484, 0.8974, 1.6677, 1.6060, 2.2569, 1.9138, 1.0565], device='cuda:4'), covar=tensor([0.1959, 0.1332, 0.2400, 0.1747, 0.1440, 0.1385, 0.1777, 0.2110], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0236, 0.0135, 0.0119, 0.0131, 0.0151, 0.0117, 0.0117], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 21:24:59,607 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5483, 1.1587, 1.3178, 1.2910, 1.6748, 1.3807, 1.2335, 1.2699], device='cuda:4'), covar=tensor([0.1481, 0.1311, 0.1880, 0.1338, 0.0892, 0.1357, 0.1658, 0.2248], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0304, 0.0348, 0.0282, 0.0324, 0.0302, 0.0296, 0.0369], device='cuda:4'), out_proj_covar=tensor([6.3359e-05, 6.2663e-05, 7.3322e-05, 5.6520e-05, 6.6681e-05, 6.3163e-05, 6.1477e-05, 7.8147e-05], device='cuda:4') 2023-04-27 21:25:13,440 INFO [finetune.py:976] (4/7) Epoch 25, batch 100, loss[loss=0.2035, simple_loss=0.279, pruned_loss=0.06399, over 4710.00 frames. ], tot_loss[loss=0.1687, simple_loss=0.2402, pruned_loss=0.04858, over 379941.69 frames. ], batch size: 54, lr: 3.02e-03, grad_scale: 32.0 2023-04-27 21:25:15,237 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.734e+01 1.567e+02 1.856e+02 2.214e+02 3.687e+02, threshold=3.711e+02, percent-clipped=3.0 2023-04-27 21:25:46,405 INFO [finetune.py:976] (4/7) Epoch 25, batch 150, loss[loss=0.1595, simple_loss=0.2312, pruned_loss=0.04389, over 4753.00 frames. ], tot_loss[loss=0.1656, simple_loss=0.2368, pruned_loss=0.0472, over 508018.11 frames. ], batch size: 27, lr: 3.02e-03, grad_scale: 32.0 2023-04-27 21:26:14,484 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.63 vs. limit=2.0 2023-04-27 21:26:20,278 INFO [finetune.py:976] (4/7) Epoch 25, batch 200, loss[loss=0.1811, simple_loss=0.2522, pruned_loss=0.05497, over 4824.00 frames. ], tot_loss[loss=0.1639, simple_loss=0.2349, pruned_loss=0.04645, over 607865.81 frames. ], batch size: 40, lr: 3.02e-03, grad_scale: 32.0 2023-04-27 21:26:22,072 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.158e+02 1.527e+02 1.776e+02 2.195e+02 3.612e+02, threshold=3.551e+02, percent-clipped=0.0 2023-04-27 21:27:09,383 INFO [finetune.py:976] (4/7) Epoch 25, batch 250, loss[loss=0.2088, simple_loss=0.2758, pruned_loss=0.07095, over 4914.00 frames. ], tot_loss[loss=0.167, simple_loss=0.2381, pruned_loss=0.04793, over 685339.64 frames. ], batch size: 42, lr: 3.02e-03, grad_scale: 32.0 2023-04-27 21:28:04,799 INFO [finetune.py:976] (4/7) Epoch 25, batch 300, loss[loss=0.1171, simple_loss=0.1909, pruned_loss=0.02163, over 4722.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.2393, pruned_loss=0.04782, over 744607.84 frames. ], batch size: 23, lr: 3.02e-03, grad_scale: 32.0 2023-04-27 21:28:06,622 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.160e+02 1.558e+02 1.872e+02 2.368e+02 4.247e+02, threshold=3.743e+02, percent-clipped=2.0 2023-04-27 21:28:20,934 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6849, 2.4560, 2.6343, 3.1677, 2.9862, 2.5975, 2.2389, 2.8225], device='cuda:4'), covar=tensor([0.0831, 0.1071, 0.0691, 0.0588, 0.0634, 0.0838, 0.0719, 0.0529], device='cuda:4'), in_proj_covar=tensor([0.0189, 0.0205, 0.0188, 0.0175, 0.0181, 0.0181, 0.0153, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 21:28:43,570 INFO [finetune.py:976] (4/7) Epoch 25, batch 350, loss[loss=0.1956, simple_loss=0.2753, pruned_loss=0.05794, over 4910.00 frames. ], tot_loss[loss=0.17, simple_loss=0.2427, pruned_loss=0.04859, over 792258.92 frames. ], batch size: 37, lr: 3.02e-03, grad_scale: 32.0 2023-04-27 21:29:21,934 INFO [finetune.py:976] (4/7) Epoch 25, batch 400, loss[loss=0.1705, simple_loss=0.2398, pruned_loss=0.05057, over 4831.00 frames. ], tot_loss[loss=0.1707, simple_loss=0.2439, pruned_loss=0.04877, over 827457.73 frames. ], batch size: 30, lr: 3.02e-03, grad_scale: 32.0 2023-04-27 21:29:29,042 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.177e+02 1.617e+02 1.875e+02 2.197e+02 4.707e+02, threshold=3.750e+02, percent-clipped=1.0 2023-04-27 21:29:59,949 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6190, 1.6575, 0.7863, 1.3218, 1.7690, 1.4814, 1.3958, 1.4400], device='cuda:4'), covar=tensor([0.0441, 0.0337, 0.0331, 0.0512, 0.0264, 0.0469, 0.0446, 0.0511], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0027, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0046, 0.0038, 0.0052, 0.0039, 0.0050, 0.0050, 0.0052], device='cuda:4') 2023-04-27 21:30:01,128 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=137892.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:30:02,950 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=137895.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:30:26,716 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8109, 2.1866, 1.9091, 1.6401, 1.3742, 1.4326, 2.1279, 1.3339], device='cuda:4'), covar=tensor([0.1749, 0.1566, 0.1426, 0.1917, 0.2313, 0.1957, 0.0866, 0.2050], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0210, 0.0169, 0.0204, 0.0199, 0.0186, 0.0156, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 21:30:27,175 INFO [finetune.py:976] (4/7) Epoch 25, batch 450, loss[loss=0.1575, simple_loss=0.2424, pruned_loss=0.03636, over 4830.00 frames. ], tot_loss[loss=0.1687, simple_loss=0.242, pruned_loss=0.04769, over 856455.85 frames. ], batch size: 40, lr: 3.02e-03, grad_scale: 32.0 2023-04-27 21:30:29,587 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6407, 1.3481, 4.3894, 4.1056, 3.8090, 4.2104, 4.1485, 3.8377], device='cuda:4'), covar=tensor([0.7059, 0.6196, 0.1050, 0.1725, 0.1071, 0.1391, 0.1134, 0.1555], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0307, 0.0407, 0.0409, 0.0349, 0.0409, 0.0320, 0.0366], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 21:31:16,218 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2405, 1.6438, 2.1608, 2.3927, 2.0628, 1.6970, 1.2653, 1.7662], device='cuda:4'), covar=tensor([0.3078, 0.3113, 0.1531, 0.2221, 0.2337, 0.2551, 0.4163, 0.1961], device='cuda:4'), in_proj_covar=tensor([0.0295, 0.0247, 0.0229, 0.0315, 0.0222, 0.0236, 0.0229, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 21:31:27,312 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=137953.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:31:33,999 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=137956.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:31:35,743 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2951, 1.7691, 2.1724, 2.6228, 2.1509, 1.7279, 1.5292, 1.9326], device='cuda:4'), covar=tensor([0.2749, 0.2926, 0.1464, 0.1914, 0.2386, 0.2530, 0.3896, 0.1931], device='cuda:4'), in_proj_covar=tensor([0.0295, 0.0247, 0.0229, 0.0315, 0.0222, 0.0236, 0.0229, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 21:31:44,821 INFO [finetune.py:976] (4/7) Epoch 25, batch 500, loss[loss=0.2339, simple_loss=0.2822, pruned_loss=0.09279, over 4813.00 frames. ], tot_loss[loss=0.166, simple_loss=0.2388, pruned_loss=0.0466, over 878822.99 frames. ], batch size: 38, lr: 3.02e-03, grad_scale: 32.0 2023-04-27 21:31:46,604 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.991e+01 1.587e+02 1.886e+02 2.282e+02 3.867e+02, threshold=3.771e+02, percent-clipped=1.0 2023-04-27 21:32:30,091 INFO [finetune.py:976] (4/7) Epoch 25, batch 550, loss[loss=0.2216, simple_loss=0.2759, pruned_loss=0.08362, over 4826.00 frames. ], tot_loss[loss=0.1649, simple_loss=0.2368, pruned_loss=0.04644, over 894847.86 frames. ], batch size: 40, lr: 3.02e-03, grad_scale: 32.0 2023-04-27 21:33:20,386 INFO [finetune.py:976] (4/7) Epoch 25, batch 600, loss[loss=0.2355, simple_loss=0.3017, pruned_loss=0.08471, over 4726.00 frames. ], tot_loss[loss=0.1649, simple_loss=0.2366, pruned_loss=0.04666, over 908109.15 frames. ], batch size: 54, lr: 3.02e-03, grad_scale: 32.0 2023-04-27 21:33:22,209 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.059e+02 1.469e+02 1.786e+02 2.001e+02 2.950e+02, threshold=3.571e+02, percent-clipped=0.0 2023-04-27 21:33:53,151 INFO [finetune.py:976] (4/7) Epoch 25, batch 650, loss[loss=0.192, simple_loss=0.2607, pruned_loss=0.06162, over 4818.00 frames. ], tot_loss[loss=0.1682, simple_loss=0.2408, pruned_loss=0.04782, over 917681.33 frames. ], batch size: 51, lr: 3.02e-03, grad_scale: 32.0 2023-04-27 21:34:26,499 INFO [finetune.py:976] (4/7) Epoch 25, batch 700, loss[loss=0.1975, simple_loss=0.2566, pruned_loss=0.0692, over 4816.00 frames. ], tot_loss[loss=0.1702, simple_loss=0.2435, pruned_loss=0.04845, over 926328.50 frames. ], batch size: 39, lr: 3.02e-03, grad_scale: 32.0 2023-04-27 21:34:27,268 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8854, 1.4837, 2.0302, 2.3848, 2.0060, 1.9507, 1.9989, 1.9020], device='cuda:4'), covar=tensor([0.4376, 0.6484, 0.5929, 0.5331, 0.5835, 0.7919, 0.7609, 0.8291], device='cuda:4'), in_proj_covar=tensor([0.0434, 0.0419, 0.0510, 0.0506, 0.0464, 0.0498, 0.0501, 0.0513], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 21:34:28,315 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.080e+02 1.605e+02 1.838e+02 2.214e+02 4.494e+02, threshold=3.677e+02, percent-clipped=3.0 2023-04-27 21:34:55,474 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.74 vs. limit=2.0 2023-04-27 21:35:26,623 INFO [finetune.py:976] (4/7) Epoch 25, batch 750, loss[loss=0.1721, simple_loss=0.2398, pruned_loss=0.05222, over 4866.00 frames. ], tot_loss[loss=0.1705, simple_loss=0.244, pruned_loss=0.04853, over 934058.60 frames. ], batch size: 44, lr: 3.02e-03, grad_scale: 32.0 2023-04-27 21:35:45,327 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1146, 2.5828, 2.1622, 2.5276, 1.8072, 2.1733, 2.2029, 1.6417], device='cuda:4'), covar=tensor([0.1930, 0.1161, 0.0798, 0.0968, 0.3084, 0.1038, 0.2082, 0.2871], device='cuda:4'), in_proj_covar=tensor([0.0281, 0.0297, 0.0212, 0.0274, 0.0309, 0.0253, 0.0247, 0.0261], device='cuda:4'), out_proj_covar=tensor([1.1242e-04, 1.1726e-04, 8.3483e-05, 1.0803e-04, 1.2450e-04, 9.9599e-05, 9.9507e-05, 1.0327e-04], device='cuda:4') 2023-04-27 21:36:09,244 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=138248.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:36:16,833 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=138251.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:36:25,762 INFO [finetune.py:976] (4/7) Epoch 25, batch 800, loss[loss=0.1846, simple_loss=0.2558, pruned_loss=0.05668, over 4900.00 frames. ], tot_loss[loss=0.1708, simple_loss=0.2447, pruned_loss=0.04841, over 937971.38 frames. ], batch size: 43, lr: 3.02e-03, grad_scale: 32.0 2023-04-27 21:36:27,567 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.037e+01 1.577e+02 1.890e+02 2.276e+02 6.092e+02, threshold=3.780e+02, percent-clipped=1.0 2023-04-27 21:36:28,896 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=138270.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:37:09,965 INFO [finetune.py:976] (4/7) Epoch 25, batch 850, loss[loss=0.1561, simple_loss=0.2319, pruned_loss=0.04012, over 4826.00 frames. ], tot_loss[loss=0.1692, simple_loss=0.2427, pruned_loss=0.04785, over 942397.71 frames. ], batch size: 30, lr: 3.02e-03, grad_scale: 32.0 2023-04-27 21:37:14,304 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=138322.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:37:24,979 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=138331.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:37:34,693 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7518, 2.0382, 1.7916, 2.0087, 1.5931, 1.7944, 1.7741, 1.4648], device='cuda:4'), covar=tensor([0.1319, 0.0870, 0.0650, 0.0783, 0.2532, 0.0831, 0.1448, 0.1743], device='cuda:4'), in_proj_covar=tensor([0.0279, 0.0294, 0.0211, 0.0272, 0.0306, 0.0250, 0.0245, 0.0259], device='cuda:4'), out_proj_covar=tensor([1.1152e-04, 1.1605e-04, 8.2915e-05, 1.0729e-04, 1.2323e-04, 9.8727e-05, 9.8712e-05, 1.0245e-04], device='cuda:4') 2023-04-27 21:37:38,184 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4889, 1.6962, 1.8923, 2.0197, 1.9030, 1.9705, 1.9624, 1.9133], device='cuda:4'), covar=tensor([0.3859, 0.5433, 0.4393, 0.4347, 0.5492, 0.6800, 0.4891, 0.4684], device='cuda:4'), in_proj_covar=tensor([0.0341, 0.0374, 0.0326, 0.0340, 0.0349, 0.0393, 0.0357, 0.0331], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 21:37:38,757 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3669, 1.3771, 1.4280, 1.6213, 1.6664, 1.3473, 0.9860, 1.5485], device='cuda:4'), covar=tensor([0.0850, 0.1341, 0.0943, 0.0605, 0.0710, 0.0892, 0.0866, 0.0611], device='cuda:4'), in_proj_covar=tensor([0.0187, 0.0202, 0.0187, 0.0173, 0.0179, 0.0178, 0.0151, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 21:38:00,619 INFO [finetune.py:976] (4/7) Epoch 25, batch 900, loss[loss=0.1509, simple_loss=0.2165, pruned_loss=0.04269, over 4760.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.2401, pruned_loss=0.04699, over 946567.81 frames. ], batch size: 54, lr: 3.02e-03, grad_scale: 32.0 2023-04-27 21:38:02,478 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.023e+01 1.492e+02 1.759e+02 2.095e+02 3.711e+02, threshold=3.518e+02, percent-clipped=0.0 2023-04-27 21:38:22,013 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=138383.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:38:54,304 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5075, 1.7373, 1.4259, 1.3112, 1.2505, 1.1838, 1.4716, 1.1435], device='cuda:4'), covar=tensor([0.1421, 0.1145, 0.1334, 0.1444, 0.1929, 0.1756, 0.0903, 0.1839], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0209, 0.0168, 0.0203, 0.0199, 0.0185, 0.0156, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 21:39:01,962 INFO [finetune.py:976] (4/7) Epoch 25, batch 950, loss[loss=0.1473, simple_loss=0.2175, pruned_loss=0.03853, over 4818.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.2395, pruned_loss=0.04737, over 950062.36 frames. ], batch size: 38, lr: 3.02e-03, grad_scale: 32.0 2023-04-27 21:39:13,739 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5530, 1.4993, 1.9737, 1.9316, 1.4396, 1.2955, 1.5461, 1.0169], device='cuda:4'), covar=tensor([0.0533, 0.0631, 0.0372, 0.0490, 0.0774, 0.1194, 0.0558, 0.0694], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0068, 0.0066, 0.0068, 0.0075, 0.0096, 0.0073, 0.0064], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 21:40:06,148 INFO [finetune.py:976] (4/7) Epoch 25, batch 1000, loss[loss=0.1308, simple_loss=0.2009, pruned_loss=0.03036, over 4697.00 frames. ], tot_loss[loss=0.1691, simple_loss=0.2412, pruned_loss=0.0485, over 950593.35 frames. ], batch size: 23, lr: 3.02e-03, grad_scale: 32.0 2023-04-27 21:40:07,971 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.116e+02 1.516e+02 1.776e+02 2.068e+02 3.891e+02, threshold=3.551e+02, percent-clipped=2.0 2023-04-27 21:41:09,820 INFO [finetune.py:976] (4/7) Epoch 25, batch 1050, loss[loss=0.1679, simple_loss=0.2408, pruned_loss=0.04749, over 4770.00 frames. ], tot_loss[loss=0.1707, simple_loss=0.2439, pruned_loss=0.04877, over 948585.72 frames. ], batch size: 28, lr: 3.02e-03, grad_scale: 32.0 2023-04-27 21:41:21,497 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6699, 1.5752, 1.7729, 2.0740, 2.0849, 1.6744, 1.4050, 1.9174], device='cuda:4'), covar=tensor([0.0913, 0.1315, 0.0874, 0.0652, 0.0706, 0.0979, 0.0828, 0.0615], device='cuda:4'), in_proj_covar=tensor([0.0187, 0.0203, 0.0187, 0.0174, 0.0179, 0.0179, 0.0152, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 21:41:52,774 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=138548.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:41:53,395 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6946, 2.0333, 1.7893, 1.9871, 1.4802, 1.6782, 1.7321, 1.3922], device='cuda:4'), covar=tensor([0.1653, 0.1216, 0.0760, 0.1019, 0.3250, 0.1073, 0.1504, 0.2214], device='cuda:4'), in_proj_covar=tensor([0.0284, 0.0299, 0.0213, 0.0276, 0.0311, 0.0255, 0.0248, 0.0263], device='cuda:4'), out_proj_covar=tensor([1.1338e-04, 1.1795e-04, 8.3949e-05, 1.0884e-04, 1.2539e-04, 1.0041e-04, 9.9919e-05, 1.0392e-04], device='cuda:4') 2023-04-27 21:41:54,997 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=138551.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:42:14,665 INFO [finetune.py:976] (4/7) Epoch 25, batch 1100, loss[loss=0.157, simple_loss=0.2377, pruned_loss=0.03814, over 4792.00 frames. ], tot_loss[loss=0.1714, simple_loss=0.2448, pruned_loss=0.04897, over 949399.81 frames. ], batch size: 45, lr: 3.02e-03, grad_scale: 32.0 2023-04-27 21:42:16,462 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.782e+01 1.591e+02 1.867e+02 2.331e+02 5.511e+02, threshold=3.734e+02, percent-clipped=3.0 2023-04-27 21:42:55,897 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=138596.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:42:58,182 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=138599.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:43:18,593 INFO [finetune.py:976] (4/7) Epoch 25, batch 1150, loss[loss=0.2005, simple_loss=0.2716, pruned_loss=0.06465, over 4789.00 frames. ], tot_loss[loss=0.1721, simple_loss=0.2458, pruned_loss=0.04917, over 951633.40 frames. ], batch size: 29, lr: 3.02e-03, grad_scale: 32.0 2023-04-27 21:43:38,294 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=138626.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:44:23,366 INFO [finetune.py:976] (4/7) Epoch 25, batch 1200, loss[loss=0.1589, simple_loss=0.2266, pruned_loss=0.04558, over 4821.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2442, pruned_loss=0.04899, over 953299.97 frames. ], batch size: 40, lr: 3.02e-03, grad_scale: 64.0 2023-04-27 21:44:26,074 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.080e+02 1.537e+02 1.836e+02 2.355e+02 3.778e+02, threshold=3.672e+02, percent-clipped=1.0 2023-04-27 21:44:43,177 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=138678.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:44:44,974 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=138681.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:44:53,095 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=138686.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:45:04,527 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1376, 0.7048, 0.9381, 0.7543, 1.2356, 0.9842, 0.9044, 0.9578], device='cuda:4'), covar=tensor([0.1718, 0.1465, 0.2039, 0.1506, 0.0989, 0.1392, 0.1439, 0.2220], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0307, 0.0350, 0.0284, 0.0325, 0.0305, 0.0299, 0.0371], device='cuda:4'), out_proj_covar=tensor([6.3804e-05, 6.3071e-05, 7.3778e-05, 5.6973e-05, 6.6833e-05, 6.3873e-05, 6.2012e-05, 7.8693e-05], device='cuda:4') 2023-04-27 21:45:05,135 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=138697.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:45:26,952 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1055, 1.7888, 2.2737, 2.4459, 2.1313, 2.0152, 2.1703, 2.0754], device='cuda:4'), covar=tensor([0.4814, 0.7243, 0.7031, 0.5779, 0.6088, 0.8920, 0.8582, 0.9492], device='cuda:4'), in_proj_covar=tensor([0.0437, 0.0421, 0.0512, 0.0507, 0.0466, 0.0501, 0.0502, 0.0516], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 21:45:29,241 INFO [finetune.py:976] (4/7) Epoch 25, batch 1250, loss[loss=0.1583, simple_loss=0.2301, pruned_loss=0.04318, over 4822.00 frames. ], tot_loss[loss=0.1691, simple_loss=0.2419, pruned_loss=0.04822, over 954852.05 frames. ], batch size: 38, lr: 3.02e-03, grad_scale: 64.0 2023-04-27 21:46:08,002 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=138742.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:46:11,076 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=138747.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:46:23,990 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=138758.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:46:33,339 INFO [finetune.py:976] (4/7) Epoch 25, batch 1300, loss[loss=0.1916, simple_loss=0.2487, pruned_loss=0.06726, over 4840.00 frames. ], tot_loss[loss=0.1661, simple_loss=0.2381, pruned_loss=0.04707, over 955761.35 frames. ], batch size: 30, lr: 3.02e-03, grad_scale: 64.0 2023-04-27 21:46:40,295 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.044e+02 1.540e+02 1.806e+02 2.216e+02 3.604e+02, threshold=3.612e+02, percent-clipped=0.0 2023-04-27 21:47:44,177 INFO [finetune.py:976] (4/7) Epoch 25, batch 1350, loss[loss=0.1195, simple_loss=0.1893, pruned_loss=0.02484, over 4216.00 frames. ], tot_loss[loss=0.1659, simple_loss=0.2378, pruned_loss=0.04696, over 955801.65 frames. ], batch size: 18, lr: 3.02e-03, grad_scale: 64.0 2023-04-27 21:48:48,892 INFO [finetune.py:976] (4/7) Epoch 25, batch 1400, loss[loss=0.168, simple_loss=0.2467, pruned_loss=0.04466, over 4905.00 frames. ], tot_loss[loss=0.1651, simple_loss=0.238, pruned_loss=0.04608, over 955415.00 frames. ], batch size: 43, lr: 3.02e-03, grad_scale: 64.0 2023-04-27 21:48:50,720 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.057e+02 1.518e+02 1.853e+02 2.271e+02 4.217e+02, threshold=3.705e+02, percent-clipped=3.0 2023-04-27 21:49:44,974 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=138909.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:49:54,330 INFO [finetune.py:976] (4/7) Epoch 25, batch 1450, loss[loss=0.1502, simple_loss=0.2221, pruned_loss=0.03914, over 4866.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.2405, pruned_loss=0.04658, over 956268.64 frames. ], batch size: 31, lr: 3.01e-03, grad_scale: 64.0 2023-04-27 21:50:07,865 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=138926.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:50:15,756 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3603, 3.3348, 0.9600, 1.5166, 1.6821, 2.3731, 1.8326, 0.9408], device='cuda:4'), covar=tensor([0.1943, 0.1646, 0.2561, 0.2073, 0.1418, 0.1419, 0.2023, 0.2617], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0240, 0.0137, 0.0121, 0.0132, 0.0153, 0.0118, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 21:50:53,069 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5059, 1.4129, 1.3235, 1.6832, 1.5600, 1.9533, 1.3079, 3.4811], device='cuda:4'), covar=tensor([0.0524, 0.0855, 0.0828, 0.1290, 0.0678, 0.0477, 0.0785, 0.0150], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 21:51:01,099 INFO [finetune.py:976] (4/7) Epoch 25, batch 1500, loss[loss=0.1866, simple_loss=0.2515, pruned_loss=0.06091, over 4820.00 frames. ], tot_loss[loss=0.1696, simple_loss=0.243, pruned_loss=0.04805, over 955145.80 frames. ], batch size: 33, lr: 3.01e-03, grad_scale: 64.0 2023-04-27 21:51:03,880 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.140e+02 1.504e+02 1.740e+02 2.143e+02 4.195e+02, threshold=3.481e+02, percent-clipped=2.0 2023-04-27 21:51:10,081 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=138970.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 21:51:12,536 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=138974.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:51:19,560 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=138978.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:51:21,394 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.59 vs. limit=5.0 2023-04-27 21:52:02,854 INFO [finetune.py:976] (4/7) Epoch 25, batch 1550, loss[loss=0.1984, simple_loss=0.2654, pruned_loss=0.0657, over 4926.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2424, pruned_loss=0.04757, over 956469.34 frames. ], batch size: 33, lr: 3.01e-03, grad_scale: 64.0 2023-04-27 21:52:11,166 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=139026.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:52:11,842 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6786, 1.5260, 0.6694, 1.3139, 1.8175, 1.4994, 1.3709, 1.4996], device='cuda:4'), covar=tensor([0.0486, 0.0409, 0.0348, 0.0553, 0.0274, 0.0505, 0.0498, 0.0532], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0028, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0046, 0.0038, 0.0053, 0.0039, 0.0050, 0.0050, 0.0052], device='cuda:4') 2023-04-27 21:52:17,672 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7068, 3.5090, 0.8795, 1.7259, 1.9784, 2.4612, 1.9414, 0.9988], device='cuda:4'), covar=tensor([0.1250, 0.0828, 0.2114, 0.1315, 0.1013, 0.1047, 0.1501, 0.2027], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0239, 0.0137, 0.0121, 0.0132, 0.0153, 0.0117, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 21:52:18,856 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=139037.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:52:22,338 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=139042.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:52:23,911 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-04-27 21:52:29,084 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=139053.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:52:41,930 INFO [finetune.py:976] (4/7) Epoch 25, batch 1600, loss[loss=0.1506, simple_loss=0.2218, pruned_loss=0.0397, over 4801.00 frames. ], tot_loss[loss=0.1673, simple_loss=0.2402, pruned_loss=0.04719, over 956427.87 frames. ], batch size: 29, lr: 3.01e-03, grad_scale: 64.0 2023-04-27 21:52:44,233 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.006e+02 1.576e+02 1.832e+02 2.139e+02 4.092e+02, threshold=3.665e+02, percent-clipped=3.0 2023-04-27 21:53:24,796 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2726, 1.7979, 2.1746, 2.4639, 2.1837, 1.7092, 1.3596, 1.9715], device='cuda:4'), covar=tensor([0.2961, 0.2784, 0.1496, 0.1871, 0.2154, 0.2470, 0.3892, 0.1758], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0244, 0.0227, 0.0311, 0.0221, 0.0234, 0.0226, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 21:53:47,320 INFO [finetune.py:976] (4/7) Epoch 25, batch 1650, loss[loss=0.155, simple_loss=0.2274, pruned_loss=0.0413, over 4695.00 frames. ], tot_loss[loss=0.166, simple_loss=0.2387, pruned_loss=0.04663, over 956873.96 frames. ], batch size: 23, lr: 3.01e-03, grad_scale: 64.0 2023-04-27 21:54:33,488 INFO [finetune.py:976] (4/7) Epoch 25, batch 1700, loss[loss=0.1437, simple_loss=0.2087, pruned_loss=0.03939, over 4344.00 frames. ], tot_loss[loss=0.1658, simple_loss=0.2379, pruned_loss=0.04687, over 957080.24 frames. ], batch size: 19, lr: 3.01e-03, grad_scale: 64.0 2023-04-27 21:54:35,334 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.004e+02 1.477e+02 1.739e+02 2.189e+02 4.097e+02, threshold=3.477e+02, percent-clipped=1.0 2023-04-27 21:54:40,449 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-27 21:55:07,056 INFO [finetune.py:976] (4/7) Epoch 25, batch 1750, loss[loss=0.1985, simple_loss=0.2721, pruned_loss=0.06247, over 4264.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.2402, pruned_loss=0.04758, over 955235.41 frames. ], batch size: 65, lr: 3.01e-03, grad_scale: 64.0 2023-04-27 21:55:23,345 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=139238.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:55:29,617 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-27 21:55:37,050 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3980, 1.8544, 1.8292, 1.9283, 1.8386, 1.9503, 1.9168, 1.8613], device='cuda:4'), covar=tensor([0.3572, 0.4688, 0.4136, 0.4189, 0.5060, 0.6356, 0.4740, 0.4675], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0375, 0.0327, 0.0340, 0.0350, 0.0394, 0.0358, 0.0332], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 21:55:40,696 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3957, 1.3668, 1.7552, 1.7252, 1.3678, 1.2011, 1.3910, 0.8979], device='cuda:4'), covar=tensor([0.0490, 0.0576, 0.0330, 0.0558, 0.0691, 0.1068, 0.0461, 0.0571], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0067, 0.0067, 0.0068, 0.0074, 0.0095, 0.0073, 0.0064], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 21:55:41,183 INFO [finetune.py:976] (4/7) Epoch 25, batch 1800, loss[loss=0.1994, simple_loss=0.2753, pruned_loss=0.0618, over 4818.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.243, pruned_loss=0.04802, over 953535.95 frames. ], batch size: 41, lr: 3.01e-03, grad_scale: 64.0 2023-04-27 21:55:41,247 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=139265.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 21:55:42,982 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.657e+01 1.551e+02 1.902e+02 2.363e+02 3.513e+02, threshold=3.803e+02, percent-clipped=1.0 2023-04-27 21:55:44,922 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5377, 3.5067, 0.9846, 1.7325, 1.9353, 2.5232, 1.8429, 0.9998], device='cuda:4'), covar=tensor([0.1463, 0.0941, 0.2110, 0.1432, 0.1132, 0.1015, 0.1763, 0.1949], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0237, 0.0136, 0.0120, 0.0131, 0.0152, 0.0117, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 21:55:48,004 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=139276.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:55:57,490 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=139289.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:56:04,607 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=139299.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:56:07,556 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0191, 1.4667, 1.4589, 1.6817, 1.6088, 1.8354, 1.3219, 3.6587], device='cuda:4'), covar=tensor([0.0701, 0.0922, 0.0892, 0.1377, 0.0751, 0.0608, 0.0886, 0.0183], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 21:56:14,568 INFO [finetune.py:976] (4/7) Epoch 25, batch 1850, loss[loss=0.1568, simple_loss=0.2286, pruned_loss=0.04252, over 4832.00 frames. ], tot_loss[loss=0.17, simple_loss=0.2438, pruned_loss=0.04803, over 955165.22 frames. ], batch size: 30, lr: 3.01e-03, grad_scale: 64.0 2023-04-27 21:56:25,776 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9209, 1.1974, 3.3006, 3.0627, 2.9749, 3.2405, 3.2369, 2.9013], device='cuda:4'), covar=tensor([0.7693, 0.5566, 0.1569, 0.2240, 0.1428, 0.2180, 0.1656, 0.1832], device='cuda:4'), in_proj_covar=tensor([0.0314, 0.0309, 0.0408, 0.0412, 0.0351, 0.0414, 0.0321, 0.0370], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 21:56:29,752 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=139337.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:56:29,775 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=139337.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 21:56:33,189 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=139342.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:56:43,881 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=139350.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:56:51,968 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=139353.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:57:05,212 INFO [finetune.py:976] (4/7) Epoch 25, batch 1900, loss[loss=0.1708, simple_loss=0.2447, pruned_loss=0.04842, over 4885.00 frames. ], tot_loss[loss=0.1696, simple_loss=0.2433, pruned_loss=0.04799, over 953648.89 frames. ], batch size: 43, lr: 3.01e-03, grad_scale: 32.0 2023-04-27 21:57:07,684 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.239e+02 1.510e+02 1.790e+02 2.163e+02 4.429e+02, threshold=3.581e+02, percent-clipped=1.0 2023-04-27 21:57:25,556 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-04-27 21:57:28,903 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=139385.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:57:36,790 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=139390.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:57:49,512 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=139401.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 21:57:56,055 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.9034, 3.8539, 2.8459, 4.4909, 3.8289, 3.8980, 1.5661, 3.8189], device='cuda:4'), covar=tensor([0.1772, 0.1259, 0.2846, 0.1423, 0.2481, 0.1650, 0.6149, 0.2382], device='cuda:4'), in_proj_covar=tensor([0.0242, 0.0217, 0.0249, 0.0303, 0.0295, 0.0246, 0.0272, 0.0271], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 21:58:09,752 INFO [finetune.py:976] (4/7) Epoch 25, batch 1950, loss[loss=0.1392, simple_loss=0.2096, pruned_loss=0.03437, over 4021.00 frames. ], tot_loss[loss=0.1684, simple_loss=0.2415, pruned_loss=0.04763, over 953023.39 frames. ], batch size: 17, lr: 3.01e-03, grad_scale: 32.0 2023-04-27 21:59:13,359 INFO [finetune.py:976] (4/7) Epoch 25, batch 2000, loss[loss=0.1639, simple_loss=0.2349, pruned_loss=0.04648, over 4817.00 frames. ], tot_loss[loss=0.1658, simple_loss=0.2384, pruned_loss=0.04657, over 953074.77 frames. ], batch size: 33, lr: 3.01e-03, grad_scale: 32.0 2023-04-27 21:59:13,468 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3153, 1.7744, 1.5957, 2.0189, 2.0642, 2.1604, 1.5794, 3.9999], device='cuda:4'), covar=tensor([0.0519, 0.0722, 0.0708, 0.1041, 0.0534, 0.0551, 0.0661, 0.0107], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 21:59:15,798 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.085e+02 1.572e+02 1.799e+02 2.251e+02 3.942e+02, threshold=3.599e+02, percent-clipped=2.0 2023-04-27 21:59:25,650 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.65 vs. limit=2.0 2023-04-27 21:59:56,676 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.56 vs. limit=2.0 2023-04-27 22:00:17,505 INFO [finetune.py:976] (4/7) Epoch 25, batch 2050, loss[loss=0.1373, simple_loss=0.2119, pruned_loss=0.03131, over 4860.00 frames. ], tot_loss[loss=0.164, simple_loss=0.2359, pruned_loss=0.04606, over 952693.86 frames. ], batch size: 31, lr: 3.01e-03, grad_scale: 32.0 2023-04-27 22:00:50,451 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-04-27 22:01:10,392 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.43 vs. limit=2.0 2023-04-27 22:01:21,674 INFO [finetune.py:976] (4/7) Epoch 25, batch 2100, loss[loss=0.1619, simple_loss=0.2397, pruned_loss=0.04212, over 4930.00 frames. ], tot_loss[loss=0.1644, simple_loss=0.236, pruned_loss=0.04636, over 953809.53 frames. ], batch size: 33, lr: 3.01e-03, grad_scale: 32.0 2023-04-27 22:01:21,781 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=139565.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 22:01:24,121 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.088e+01 1.541e+02 1.795e+02 2.157e+02 3.840e+02, threshold=3.589e+02, percent-clipped=1.0 2023-04-27 22:01:30,508 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=139571.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:01:37,246 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=139582.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:01:44,575 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=139594.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:01:57,941 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=139613.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:01:59,117 INFO [finetune.py:976] (4/7) Epoch 25, batch 2150, loss[loss=0.2471, simple_loss=0.3127, pruned_loss=0.0907, over 4791.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.2396, pruned_loss=0.04725, over 953498.16 frames. ], batch size: 51, lr: 3.01e-03, grad_scale: 32.0 2023-04-27 22:02:10,545 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=139632.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 22:02:10,578 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=139632.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:02:17,309 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=139643.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:02:18,461 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=139645.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:02:27,361 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5725, 2.3840, 2.6887, 3.0544, 3.0908, 2.4606, 2.1330, 2.7132], device='cuda:4'), covar=tensor([0.0771, 0.0994, 0.0605, 0.0540, 0.0554, 0.0814, 0.0709, 0.0529], device='cuda:4'), in_proj_covar=tensor([0.0187, 0.0204, 0.0188, 0.0174, 0.0180, 0.0179, 0.0152, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 22:02:31,935 INFO [finetune.py:976] (4/7) Epoch 25, batch 2200, loss[loss=0.1979, simple_loss=0.2682, pruned_loss=0.06383, over 4796.00 frames. ], tot_loss[loss=0.1697, simple_loss=0.2427, pruned_loss=0.04837, over 953856.68 frames. ], batch size: 51, lr: 3.01e-03, grad_scale: 32.0 2023-04-27 22:02:34,822 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.151e+02 1.529e+02 1.754e+02 2.258e+02 5.468e+02, threshold=3.509e+02, percent-clipped=4.0 2023-04-27 22:02:43,409 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-27 22:02:45,213 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7683, 2.1000, 1.7822, 2.0339, 1.5857, 1.6599, 1.7522, 1.3097], device='cuda:4'), covar=tensor([0.1711, 0.1220, 0.0810, 0.1031, 0.3250, 0.1130, 0.1743, 0.2454], device='cuda:4'), in_proj_covar=tensor([0.0286, 0.0301, 0.0214, 0.0277, 0.0313, 0.0256, 0.0249, 0.0264], device='cuda:4'), out_proj_covar=tensor([1.1425e-04, 1.1905e-04, 8.4313e-05, 1.0939e-04, 1.2612e-04, 1.0098e-04, 1.0034e-04, 1.0428e-04], device='cuda:4') 2023-04-27 22:03:04,779 INFO [finetune.py:976] (4/7) Epoch 25, batch 2250, loss[loss=0.1971, simple_loss=0.2645, pruned_loss=0.06486, over 4895.00 frames. ], tot_loss[loss=0.1713, simple_loss=0.2442, pruned_loss=0.04922, over 954969.38 frames. ], batch size: 36, lr: 3.01e-03, grad_scale: 32.0 2023-04-27 22:03:23,990 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3564, 2.8284, 1.1901, 1.7086, 2.3682, 1.4017, 3.7711, 2.2370], device='cuda:4'), covar=tensor([0.0587, 0.0501, 0.0709, 0.1223, 0.0443, 0.0988, 0.0205, 0.0509], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0073, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 22:03:38,679 INFO [finetune.py:976] (4/7) Epoch 25, batch 2300, loss[loss=0.1397, simple_loss=0.2241, pruned_loss=0.02768, over 4778.00 frames. ], tot_loss[loss=0.1691, simple_loss=0.2427, pruned_loss=0.04775, over 955285.08 frames. ], batch size: 29, lr: 3.01e-03, grad_scale: 32.0 2023-04-27 22:03:41,540 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.026e+02 1.526e+02 1.763e+02 2.328e+02 5.368e+02, threshold=3.526e+02, percent-clipped=6.0 2023-04-27 22:03:41,680 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=139769.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:03:54,761 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=139789.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:04:03,989 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=139804.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:04:10,829 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=139806.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:04:23,019 INFO [finetune.py:976] (4/7) Epoch 25, batch 2350, loss[loss=0.1545, simple_loss=0.2134, pruned_loss=0.04776, over 4042.00 frames. ], tot_loss[loss=0.1678, simple_loss=0.2406, pruned_loss=0.04747, over 953139.22 frames. ], batch size: 17, lr: 3.01e-03, grad_scale: 32.0 2023-04-27 22:04:44,831 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=139830.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:05:06,883 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=139850.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:05:06,892 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=139850.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:05:09,122 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-27 22:05:17,200 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6162, 3.1460, 2.6576, 2.9603, 2.2921, 2.5840, 2.8348, 2.0666], device='cuda:4'), covar=tensor([0.1781, 0.1141, 0.0706, 0.1128, 0.3096, 0.1196, 0.1679, 0.2651], device='cuda:4'), in_proj_covar=tensor([0.0284, 0.0300, 0.0213, 0.0277, 0.0311, 0.0255, 0.0248, 0.0263], device='cuda:4'), out_proj_covar=tensor([1.1344e-04, 1.1850e-04, 8.4008e-05, 1.0923e-04, 1.2556e-04, 1.0058e-04, 1.0005e-04, 1.0379e-04], device='cuda:4') 2023-04-27 22:05:17,688 INFO [finetune.py:976] (4/7) Epoch 25, batch 2400, loss[loss=0.1408, simple_loss=0.2081, pruned_loss=0.03677, over 4760.00 frames. ], tot_loss[loss=0.1659, simple_loss=0.2383, pruned_loss=0.0468, over 955184.91 frames. ], batch size: 28, lr: 3.01e-03, grad_scale: 32.0 2023-04-27 22:05:17,808 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=139865.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:05:19,537 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=139867.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:05:20,605 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.574e+01 1.558e+02 1.905e+02 2.229e+02 5.557e+02, threshold=3.809e+02, percent-clipped=2.0 2023-04-27 22:05:23,149 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0808, 1.9652, 2.2163, 2.5886, 2.5715, 1.9801, 1.7206, 2.3312], device='cuda:4'), covar=tensor([0.0840, 0.1093, 0.0676, 0.0570, 0.0576, 0.0911, 0.0748, 0.0527], device='cuda:4'), in_proj_covar=tensor([0.0188, 0.0206, 0.0189, 0.0175, 0.0181, 0.0180, 0.0153, 0.0181], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 22:05:37,260 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=139894.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:05:53,351 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=139911.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:05:55,626 INFO [finetune.py:976] (4/7) Epoch 25, batch 2450, loss[loss=0.1582, simple_loss=0.2275, pruned_loss=0.04442, over 4768.00 frames. ], tot_loss[loss=0.164, simple_loss=0.2358, pruned_loss=0.0461, over 955447.26 frames. ], batch size: 26, lr: 3.01e-03, grad_scale: 16.0 2023-04-27 22:06:05,131 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-27 22:06:15,194 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=139927.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:06:18,757 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=139932.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 22:06:27,878 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=139938.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:06:35,934 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=139942.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:06:37,755 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=139945.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:06:59,675 INFO [finetune.py:976] (4/7) Epoch 25, batch 2500, loss[loss=0.2032, simple_loss=0.2787, pruned_loss=0.06387, over 4861.00 frames. ], tot_loss[loss=0.1676, simple_loss=0.2387, pruned_loss=0.04819, over 953052.64 frames. ], batch size: 44, lr: 3.01e-03, grad_scale: 16.0 2023-04-27 22:07:01,350 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-27 22:07:04,983 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.044e+02 1.582e+02 1.792e+02 2.204e+02 3.647e+02, threshold=3.585e+02, percent-clipped=0.0 2023-04-27 22:07:11,621 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=139980.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:07:21,150 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=139993.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:07:35,747 INFO [finetune.py:976] (4/7) Epoch 25, batch 2550, loss[loss=0.1691, simple_loss=0.2522, pruned_loss=0.04298, over 4823.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2409, pruned_loss=0.04841, over 954347.38 frames. ], batch size: 39, lr: 3.01e-03, grad_scale: 16.0 2023-04-27 22:07:49,723 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1366, 2.4407, 1.3303, 1.7310, 2.5110, 1.8905, 1.8561, 1.9275], device='cuda:4'), covar=tensor([0.0445, 0.0302, 0.0266, 0.0503, 0.0215, 0.0482, 0.0457, 0.0517], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0028, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0045, 0.0038, 0.0052, 0.0039, 0.0050, 0.0050, 0.0052], device='cuda:4') 2023-04-27 22:08:09,498 INFO [finetune.py:976] (4/7) Epoch 25, batch 2600, loss[loss=0.191, simple_loss=0.28, pruned_loss=0.051, over 4821.00 frames. ], tot_loss[loss=0.1701, simple_loss=0.2429, pruned_loss=0.04859, over 954990.10 frames. ], batch size: 38, lr: 3.01e-03, grad_scale: 16.0 2023-04-27 22:08:09,632 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0904, 0.7045, 0.9371, 0.8586, 1.2100, 1.0310, 0.9058, 1.0059], device='cuda:4'), covar=tensor([0.2360, 0.2208, 0.2529, 0.2041, 0.1558, 0.1866, 0.2106, 0.3103], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0309, 0.0351, 0.0286, 0.0329, 0.0306, 0.0299, 0.0374], device='cuda:4'), out_proj_covar=tensor([6.3815e-05, 6.3671e-05, 7.3925e-05, 5.7355e-05, 6.7640e-05, 6.4118e-05, 6.1933e-05, 7.9229e-05], device='cuda:4') 2023-04-27 22:08:12,529 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.025e+02 1.514e+02 1.836e+02 2.215e+02 4.899e+02, threshold=3.672e+02, percent-clipped=3.0 2023-04-27 22:08:13,353 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-04-27 22:08:33,239 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8810, 1.6597, 2.0137, 2.2510, 1.9584, 1.7929, 1.9334, 1.8298], device='cuda:4'), covar=tensor([0.3989, 0.6048, 0.5776, 0.4920, 0.5109, 0.7616, 0.6932, 0.8413], device='cuda:4'), in_proj_covar=tensor([0.0438, 0.0421, 0.0512, 0.0509, 0.0466, 0.0499, 0.0504, 0.0518], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 22:08:35,644 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=140103.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:08:39,912 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=140110.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:08:42,878 INFO [finetune.py:976] (4/7) Epoch 25, batch 2650, loss[loss=0.1758, simple_loss=0.2491, pruned_loss=0.05121, over 4885.00 frames. ], tot_loss[loss=0.1702, simple_loss=0.2438, pruned_loss=0.0483, over 955001.42 frames. ], batch size: 43, lr: 3.01e-03, grad_scale: 16.0 2023-04-27 22:08:49,913 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=140125.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:09:04,026 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=140145.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:09:13,156 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=140160.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:09:14,364 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=140162.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:09:15,620 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=140164.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:09:16,092 INFO [finetune.py:976] (4/7) Epoch 25, batch 2700, loss[loss=0.1761, simple_loss=0.2447, pruned_loss=0.05374, over 4842.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2423, pruned_loss=0.04768, over 955166.06 frames. ], batch size: 47, lr: 3.01e-03, grad_scale: 16.0 2023-04-27 22:09:19,147 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.968e+01 1.589e+02 1.838e+02 2.247e+02 3.608e+02, threshold=3.675e+02, percent-clipped=0.0 2023-04-27 22:09:19,895 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=140171.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:09:55,085 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=140206.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:10:00,543 INFO [finetune.py:976] (4/7) Epoch 25, batch 2750, loss[loss=0.1225, simple_loss=0.1958, pruned_loss=0.02456, over 4760.00 frames. ], tot_loss[loss=0.1661, simple_loss=0.239, pruned_loss=0.04659, over 955801.30 frames. ], batch size: 23, lr: 3.01e-03, grad_scale: 16.0 2023-04-27 22:10:18,554 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=140227.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:10:32,153 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=140238.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:11:02,661 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9789, 2.2366, 1.9529, 2.1272, 1.5936, 1.9002, 1.8841, 1.5578], device='cuda:4'), covar=tensor([0.1704, 0.1294, 0.0772, 0.1080, 0.3412, 0.1035, 0.1866, 0.2462], device='cuda:4'), in_proj_covar=tensor([0.0283, 0.0299, 0.0212, 0.0275, 0.0310, 0.0253, 0.0247, 0.0262], device='cuda:4'), out_proj_covar=tensor([1.1324e-04, 1.1813e-04, 8.3532e-05, 1.0851e-04, 1.2495e-04, 9.9874e-05, 9.9577e-05, 1.0353e-04], device='cuda:4') 2023-04-27 22:11:03,255 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=140259.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:11:12,263 INFO [finetune.py:976] (4/7) Epoch 25, batch 2800, loss[loss=0.1709, simple_loss=0.2433, pruned_loss=0.04927, over 4692.00 frames. ], tot_loss[loss=0.1644, simple_loss=0.2366, pruned_loss=0.04607, over 954489.84 frames. ], batch size: 23, lr: 3.01e-03, grad_scale: 16.0 2023-04-27 22:11:15,317 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.076e+02 1.465e+02 1.747e+02 2.235e+02 4.062e+02, threshold=3.495e+02, percent-clipped=1.0 2023-04-27 22:11:20,366 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=140275.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:11:21,687 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1200, 0.7563, 0.9212, 0.8226, 1.2404, 1.0258, 0.9522, 0.9254], device='cuda:4'), covar=tensor([0.1770, 0.1542, 0.1900, 0.1511, 0.0934, 0.1503, 0.1601, 0.2226], device='cuda:4'), in_proj_covar=tensor([0.0314, 0.0310, 0.0353, 0.0288, 0.0330, 0.0308, 0.0300, 0.0375], device='cuda:4'), out_proj_covar=tensor([6.4163e-05, 6.3754e-05, 7.4252e-05, 5.7767e-05, 6.7786e-05, 6.4487e-05, 6.2280e-05, 7.9517e-05], device='cuda:4') 2023-04-27 22:11:32,995 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=140286.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:12:14,090 INFO [finetune.py:976] (4/7) Epoch 25, batch 2850, loss[loss=0.149, simple_loss=0.2295, pruned_loss=0.03423, over 4817.00 frames. ], tot_loss[loss=0.1633, simple_loss=0.2351, pruned_loss=0.04577, over 953626.93 frames. ], batch size: 38, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:12:17,264 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=140320.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:12:48,658 INFO [finetune.py:976] (4/7) Epoch 25, batch 2900, loss[loss=0.1868, simple_loss=0.2694, pruned_loss=0.05216, over 4836.00 frames. ], tot_loss[loss=0.1663, simple_loss=0.2387, pruned_loss=0.04697, over 955182.42 frames. ], batch size: 47, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:12:51,723 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.049e+02 1.618e+02 1.938e+02 2.286e+02 3.478e+02, threshold=3.877e+02, percent-clipped=0.0 2023-04-27 22:13:00,452 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4797, 1.7420, 1.9526, 2.0654, 1.9587, 1.9950, 2.0596, 1.9650], device='cuda:4'), covar=tensor([0.3501, 0.4814, 0.4265, 0.4179, 0.5184, 0.6784, 0.4564, 0.4593], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0374, 0.0326, 0.0339, 0.0349, 0.0393, 0.0358, 0.0332], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 22:13:03,191 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=140387.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:13:20,923 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4487, 1.7482, 2.2023, 2.9417, 2.3621, 1.8325, 1.8438, 2.2441], device='cuda:4'), covar=tensor([0.3712, 0.3884, 0.1896, 0.2700, 0.2850, 0.2832, 0.4180, 0.2176], device='cuda:4'), in_proj_covar=tensor([0.0294, 0.0246, 0.0228, 0.0314, 0.0222, 0.0235, 0.0228, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 22:13:22,623 INFO [finetune.py:976] (4/7) Epoch 25, batch 2950, loss[loss=0.1863, simple_loss=0.2706, pruned_loss=0.05102, over 4733.00 frames. ], tot_loss[loss=0.1701, simple_loss=0.2428, pruned_loss=0.04864, over 953564.47 frames. ], batch size: 54, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:13:28,750 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=140425.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:13:42,311 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=140445.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:13:44,120 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=140448.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:13:52,226 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=140459.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:13:52,866 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=140460.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:13:53,453 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.4391, 3.3145, 2.6561, 3.9261, 3.4831, 3.3860, 1.5378, 3.4005], device='cuda:4'), covar=tensor([0.2027, 0.1482, 0.2755, 0.1958, 0.2393, 0.1925, 0.5337, 0.2574], device='cuda:4'), in_proj_covar=tensor([0.0245, 0.0218, 0.0251, 0.0306, 0.0298, 0.0248, 0.0273, 0.0273], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 22:13:54,552 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=140462.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:13:56,276 INFO [finetune.py:976] (4/7) Epoch 25, batch 3000, loss[loss=0.1892, simple_loss=0.258, pruned_loss=0.06018, over 4815.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2441, pruned_loss=0.04901, over 953163.88 frames. ], batch size: 38, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:13:56,276 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 22:14:07,208 INFO [finetune.py:1010] (4/7) Epoch 25, validation: loss=0.1531, simple_loss=0.2225, pruned_loss=0.04184, over 2265189.00 frames. 2023-04-27 22:14:07,209 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6529MB 2023-04-27 22:14:07,902 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=140466.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:14:08,029 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.56 vs. limit=2.0 2023-04-27 22:14:10,181 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.175e+02 1.715e+02 2.080e+02 2.596e+02 3.715e+02, threshold=4.161e+02, percent-clipped=0.0 2023-04-27 22:14:12,096 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=140473.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:14:21,878 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=140489.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:14:24,276 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=140493.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:14:27,525 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.10 vs. limit=2.0 2023-04-27 22:14:33,157 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=140506.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:14:34,806 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=140508.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:14:35,498 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=140509.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:14:36,052 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=140510.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:14:39,578 INFO [finetune.py:976] (4/7) Epoch 25, batch 3050, loss[loss=0.1617, simple_loss=0.244, pruned_loss=0.03974, over 4906.00 frames. ], tot_loss[loss=0.1713, simple_loss=0.2451, pruned_loss=0.04876, over 953934.75 frames. ], batch size: 37, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:15:02,562 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=140550.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:15:05,406 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=140554.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:15:12,557 INFO [finetune.py:976] (4/7) Epoch 25, batch 3100, loss[loss=0.1437, simple_loss=0.2139, pruned_loss=0.03672, over 4846.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.243, pruned_loss=0.04805, over 955704.61 frames. ], batch size: 47, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:15:16,047 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.054e+02 1.508e+02 1.726e+02 2.242e+02 3.598e+02, threshold=3.452e+02, percent-clipped=0.0 2023-04-27 22:15:16,199 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=140570.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:15:21,480 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8753, 1.0768, 4.7601, 4.4053, 4.1102, 4.4784, 4.2518, 4.1597], device='cuda:4'), covar=tensor([0.6862, 0.6468, 0.1118, 0.1985, 0.1117, 0.1337, 0.1919, 0.1682], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0308, 0.0408, 0.0411, 0.0348, 0.0413, 0.0319, 0.0369], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 22:16:07,368 INFO [finetune.py:976] (4/7) Epoch 25, batch 3150, loss[loss=0.1325, simple_loss=0.2148, pruned_loss=0.02511, over 4774.00 frames. ], tot_loss[loss=0.1673, simple_loss=0.2403, pruned_loss=0.04714, over 956511.21 frames. ], batch size: 28, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:16:07,435 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=140615.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:16:19,072 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9381, 2.2930, 1.9737, 2.1816, 1.7555, 1.8893, 1.8582, 1.5243], device='cuda:4'), covar=tensor([0.1752, 0.1183, 0.0752, 0.1095, 0.3193, 0.1286, 0.1902, 0.2403], device='cuda:4'), in_proj_covar=tensor([0.0285, 0.0301, 0.0213, 0.0277, 0.0311, 0.0255, 0.0249, 0.0264], device='cuda:4'), out_proj_covar=tensor([1.1390e-04, 1.1894e-04, 8.3926e-05, 1.0907e-04, 1.2545e-04, 1.0054e-04, 1.0041e-04, 1.0440e-04], device='cuda:4') 2023-04-27 22:17:13,347 INFO [finetune.py:976] (4/7) Epoch 25, batch 3200, loss[loss=0.173, simple_loss=0.2406, pruned_loss=0.05273, over 4865.00 frames. ], tot_loss[loss=0.1645, simple_loss=0.2368, pruned_loss=0.04604, over 957057.92 frames. ], batch size: 31, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:17:21,873 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.047e+02 1.528e+02 1.799e+02 2.210e+02 5.000e+02, threshold=3.599e+02, percent-clipped=4.0 2023-04-27 22:17:34,515 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7961, 1.2737, 1.7764, 2.2861, 1.8861, 1.7293, 1.7866, 1.7067], device='cuda:4'), covar=tensor([0.4141, 0.6554, 0.5530, 0.5212, 0.5129, 0.7159, 0.7344, 0.9416], device='cuda:4'), in_proj_covar=tensor([0.0442, 0.0424, 0.0517, 0.0513, 0.0470, 0.0504, 0.0508, 0.0521], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 22:18:19,770 INFO [finetune.py:976] (4/7) Epoch 25, batch 3250, loss[loss=0.1849, simple_loss=0.2528, pruned_loss=0.05853, over 4917.00 frames. ], tot_loss[loss=0.1665, simple_loss=0.2389, pruned_loss=0.04704, over 957997.85 frames. ], batch size: 37, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:18:26,958 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=140723.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:18:55,185 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.95 vs. limit=5.0 2023-04-27 22:18:55,996 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=140743.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:19:16,146 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=140759.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:19:20,223 INFO [finetune.py:976] (4/7) Epoch 25, batch 3300, loss[loss=0.1543, simple_loss=0.2409, pruned_loss=0.03383, over 4796.00 frames. ], tot_loss[loss=0.1669, simple_loss=0.2401, pruned_loss=0.04685, over 958445.24 frames. ], batch size: 29, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:19:21,438 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=140766.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:19:27,805 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.140e+02 1.701e+02 1.963e+02 2.275e+02 6.006e+02, threshold=3.926e+02, percent-clipped=4.0 2023-04-27 22:19:48,010 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=140784.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:20:19,459 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=140807.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:20:20,762 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=140809.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:20:22,058 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.45 vs. limit=5.0 2023-04-27 22:20:23,777 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=140814.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:20:24,323 INFO [finetune.py:976] (4/7) Epoch 25, batch 3350, loss[loss=0.1589, simple_loss=0.2482, pruned_loss=0.03478, over 4818.00 frames. ], tot_loss[loss=0.1687, simple_loss=0.2423, pruned_loss=0.04761, over 957608.54 frames. ], batch size: 45, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:20:46,743 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=140845.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:20:49,467 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.88 vs. limit=5.0 2023-04-27 22:20:59,015 INFO [finetune.py:976] (4/7) Epoch 25, batch 3400, loss[loss=0.2101, simple_loss=0.2775, pruned_loss=0.07133, over 4819.00 frames. ], tot_loss[loss=0.1709, simple_loss=0.2447, pruned_loss=0.04857, over 958329.34 frames. ], batch size: 39, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:20:59,098 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=140865.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:21:02,010 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.099e+02 1.526e+02 1.810e+02 2.121e+02 3.949e+02, threshold=3.620e+02, percent-clipped=1.0 2023-04-27 22:21:02,145 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=140870.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:21:19,243 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.6229, 3.5722, 2.6741, 4.1945, 3.6795, 3.5942, 1.3938, 3.5901], device='cuda:4'), covar=tensor([0.1817, 0.1166, 0.3002, 0.1948, 0.2425, 0.1827, 0.5916, 0.2464], device='cuda:4'), in_proj_covar=tensor([0.0247, 0.0220, 0.0255, 0.0309, 0.0301, 0.0250, 0.0276, 0.0276], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 22:21:30,710 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6332, 0.6961, 1.5284, 1.9991, 1.6847, 1.4994, 1.5418, 1.5578], device='cuda:4'), covar=tensor([0.4030, 0.6635, 0.5567, 0.5347, 0.5353, 0.7147, 0.6965, 0.8510], device='cuda:4'), in_proj_covar=tensor([0.0438, 0.0420, 0.0511, 0.0508, 0.0466, 0.0500, 0.0504, 0.0516], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 22:21:32,399 INFO [finetune.py:976] (4/7) Epoch 25, batch 3450, loss[loss=0.1262, simple_loss=0.2094, pruned_loss=0.02155, over 4761.00 frames. ], tot_loss[loss=0.1703, simple_loss=0.2444, pruned_loss=0.04811, over 957379.32 frames. ], batch size: 27, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:21:32,510 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=140915.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:21:33,741 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=140917.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:21:55,390 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=140947.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:22:05,110 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=140963.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:22:06,244 INFO [finetune.py:976] (4/7) Epoch 25, batch 3500, loss[loss=0.1877, simple_loss=0.2547, pruned_loss=0.06034, over 4930.00 frames. ], tot_loss[loss=0.1678, simple_loss=0.2409, pruned_loss=0.04734, over 957892.71 frames. ], batch size: 43, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:22:09,318 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.074e+02 1.472e+02 1.838e+02 2.111e+02 1.137e+03, threshold=3.676e+02, percent-clipped=2.0 2023-04-27 22:22:14,756 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=140978.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:22:35,399 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=141008.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:22:39,555 INFO [finetune.py:976] (4/7) Epoch 25, batch 3550, loss[loss=0.1182, simple_loss=0.1988, pruned_loss=0.01875, over 4869.00 frames. ], tot_loss[loss=0.1664, simple_loss=0.2387, pruned_loss=0.04708, over 955843.91 frames. ], batch size: 31, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:22:58,168 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=141043.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:23:13,445 INFO [finetune.py:976] (4/7) Epoch 25, batch 3600, loss[loss=0.1743, simple_loss=0.2348, pruned_loss=0.05691, over 4902.00 frames. ], tot_loss[loss=0.165, simple_loss=0.2365, pruned_loss=0.04676, over 952128.04 frames. ], batch size: 36, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:23:16,482 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.061e+02 1.519e+02 1.759e+02 2.110e+02 6.340e+02, threshold=3.519e+02, percent-clipped=2.0 2023-04-27 22:23:32,936 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=141079.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:23:46,971 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=141091.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:23:47,606 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7565, 1.8107, 0.8255, 1.4483, 1.8803, 1.6109, 1.5461, 1.6246], device='cuda:4'), covar=tensor([0.0490, 0.0376, 0.0343, 0.0559, 0.0261, 0.0522, 0.0536, 0.0559], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0027, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0046, 0.0038, 0.0053, 0.0039, 0.0050, 0.0050, 0.0052], device='cuda:4') 2023-04-27 22:24:19,426 INFO [finetune.py:976] (4/7) Epoch 25, batch 3650, loss[loss=0.1709, simple_loss=0.2513, pruned_loss=0.04525, over 4810.00 frames. ], tot_loss[loss=0.1673, simple_loss=0.2393, pruned_loss=0.04765, over 951815.37 frames. ], batch size: 45, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:24:38,185 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.87 vs. limit=2.0 2023-04-27 22:24:42,748 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=141143.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:24:43,922 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=141145.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:24:57,936 INFO [finetune.py:976] (4/7) Epoch 25, batch 3700, loss[loss=0.2361, simple_loss=0.3014, pruned_loss=0.08535, over 4852.00 frames. ], tot_loss[loss=0.1704, simple_loss=0.2427, pruned_loss=0.04902, over 952581.82 frames. ], batch size: 44, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:24:58,013 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=141165.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:24:58,046 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=141165.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:25:00,977 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.179e+02 1.634e+02 1.921e+02 2.262e+02 4.366e+02, threshold=3.843e+02, percent-clipped=2.0 2023-04-27 22:25:04,208 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4229, 1.7467, 1.8412, 1.9500, 1.7935, 1.8523, 1.9184, 1.8516], device='cuda:4'), covar=tensor([0.4671, 0.5509, 0.4556, 0.4779, 0.6007, 0.7475, 0.5341, 0.5068], device='cuda:4'), in_proj_covar=tensor([0.0343, 0.0376, 0.0329, 0.0341, 0.0350, 0.0394, 0.0360, 0.0334], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 22:25:15,812 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=141193.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:25:22,667 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=141204.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:25:27,286 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7214, 1.1961, 1.8060, 2.1480, 1.7320, 1.6590, 1.7109, 1.6916], device='cuda:4'), covar=tensor([0.4330, 0.6541, 0.5978, 0.5615, 0.5633, 0.7748, 0.7504, 0.9199], device='cuda:4'), in_proj_covar=tensor([0.0438, 0.0420, 0.0513, 0.0508, 0.0466, 0.0501, 0.0504, 0.0517], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 22:25:30,031 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=141213.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:25:31,151 INFO [finetune.py:976] (4/7) Epoch 25, batch 3750, loss[loss=0.1909, simple_loss=0.2583, pruned_loss=0.06173, over 4741.00 frames. ], tot_loss[loss=0.1706, simple_loss=0.2433, pruned_loss=0.04891, over 952477.55 frames. ], batch size: 54, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:26:37,438 INFO [finetune.py:976] (4/7) Epoch 25, batch 3800, loss[loss=0.1412, simple_loss=0.2237, pruned_loss=0.0294, over 4761.00 frames. ], tot_loss[loss=0.1719, simple_loss=0.2449, pruned_loss=0.04943, over 951265.49 frames. ], batch size: 28, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:26:45,872 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.008e+02 1.561e+02 1.876e+02 2.402e+02 4.277e+02, threshold=3.753e+02, percent-clipped=1.0 2023-04-27 22:26:47,825 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=141273.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:27:06,132 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.61 vs. limit=2.0 2023-04-27 22:27:28,697 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=141303.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:27:42,126 INFO [finetune.py:976] (4/7) Epoch 25, batch 3850, loss[loss=0.1919, simple_loss=0.2615, pruned_loss=0.0611, over 4871.00 frames. ], tot_loss[loss=0.1703, simple_loss=0.243, pruned_loss=0.04879, over 950173.80 frames. ], batch size: 34, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:28:02,215 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=141329.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:28:10,367 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0324, 1.5543, 1.8386, 1.7959, 1.8373, 1.5412, 0.8315, 1.5072], device='cuda:4'), covar=tensor([0.3035, 0.3139, 0.1679, 0.2027, 0.2525, 0.2568, 0.3998, 0.2004], device='cuda:4'), in_proj_covar=tensor([0.0291, 0.0245, 0.0227, 0.0312, 0.0220, 0.0233, 0.0226, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 22:28:11,265 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.82 vs. limit=5.0 2023-04-27 22:28:27,562 INFO [finetune.py:976] (4/7) Epoch 25, batch 3900, loss[loss=0.1881, simple_loss=0.2464, pruned_loss=0.06491, over 4904.00 frames. ], tot_loss[loss=0.1684, simple_loss=0.2404, pruned_loss=0.04824, over 950746.75 frames. ], batch size: 43, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:28:27,636 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.8033, 3.6965, 2.7167, 4.4451, 3.8753, 3.8433, 1.4998, 3.8503], device='cuda:4'), covar=tensor([0.1755, 0.1311, 0.2995, 0.1660, 0.3137, 0.2094, 0.6203, 0.2136], device='cuda:4'), in_proj_covar=tensor([0.0247, 0.0220, 0.0254, 0.0307, 0.0300, 0.0251, 0.0275, 0.0275], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 22:28:31,461 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.108e+01 1.538e+02 1.758e+02 2.132e+02 4.697e+02, threshold=3.516e+02, percent-clipped=1.0 2023-04-27 22:28:37,535 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=141379.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:28:44,170 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=141390.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:28:56,257 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8871, 1.2760, 5.3744, 5.1042, 4.6737, 5.2614, 4.5935, 4.7210], device='cuda:4'), covar=tensor([0.7517, 0.7032, 0.1239, 0.1934, 0.0994, 0.1504, 0.1232, 0.1697], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0308, 0.0408, 0.0409, 0.0350, 0.0412, 0.0318, 0.0368], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 22:29:00,434 INFO [finetune.py:976] (4/7) Epoch 25, batch 3950, loss[loss=0.1474, simple_loss=0.221, pruned_loss=0.03691, over 4829.00 frames. ], tot_loss[loss=0.165, simple_loss=0.2368, pruned_loss=0.04659, over 953847.35 frames. ], batch size: 38, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:29:01,789 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=141417.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 22:29:09,300 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=141427.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:29:12,090 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-27 22:29:17,851 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=141441.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:29:33,819 INFO [finetune.py:976] (4/7) Epoch 25, batch 4000, loss[loss=0.1197, simple_loss=0.1968, pruned_loss=0.02133, over 4780.00 frames. ], tot_loss[loss=0.1626, simple_loss=0.2344, pruned_loss=0.04538, over 954127.96 frames. ], batch size: 28, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:29:33,907 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=141465.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:29:36,894 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.301e+01 1.493e+02 1.709e+02 2.038e+02 4.561e+02, threshold=3.417e+02, percent-clipped=1.0 2023-04-27 22:29:43,314 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=141478.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 22:29:55,963 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=141499.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:29:55,998 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2472, 1.4681, 1.4283, 1.6755, 1.6582, 1.7964, 1.3646, 3.5769], device='cuda:4'), covar=tensor([0.0547, 0.0782, 0.0755, 0.1200, 0.0612, 0.0481, 0.0705, 0.0126], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0042, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 22:29:57,849 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=141502.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:30:05,461 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=141513.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:30:06,625 INFO [finetune.py:976] (4/7) Epoch 25, batch 4050, loss[loss=0.1748, simple_loss=0.249, pruned_loss=0.05033, over 4826.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.2395, pruned_loss=0.04704, over 954902.49 frames. ], batch size: 30, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:30:39,500 INFO [finetune.py:976] (4/7) Epoch 25, batch 4100, loss[loss=0.1742, simple_loss=0.2577, pruned_loss=0.04539, over 4896.00 frames. ], tot_loss[loss=0.1684, simple_loss=0.2417, pruned_loss=0.04753, over 954036.77 frames. ], batch size: 37, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:30:42,488 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.442e+01 1.586e+02 1.949e+02 2.336e+02 4.544e+02, threshold=3.898e+02, percent-clipped=7.0 2023-04-27 22:30:44,336 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=141573.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:31:00,513 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.64 vs. limit=5.0 2023-04-27 22:31:04,714 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=141603.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:31:17,880 INFO [finetune.py:976] (4/7) Epoch 25, batch 4150, loss[loss=0.1771, simple_loss=0.2418, pruned_loss=0.05625, over 4865.00 frames. ], tot_loss[loss=0.1698, simple_loss=0.2431, pruned_loss=0.04822, over 954182.58 frames. ], batch size: 34, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:31:19,954 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.82 vs. limit=2.0 2023-04-27 22:31:21,627 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=141621.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:31:30,701 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7953, 1.6329, 1.9859, 2.3397, 1.5684, 1.2338, 1.6502, 0.9094], device='cuda:4'), covar=tensor([0.0703, 0.0670, 0.0511, 0.0640, 0.0796, 0.1623, 0.0784, 0.0879], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0068, 0.0067, 0.0068, 0.0075, 0.0095, 0.0073, 0.0064], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 22:32:02,963 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=141651.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:32:23,211 INFO [finetune.py:976] (4/7) Epoch 25, batch 4200, loss[loss=0.153, simple_loss=0.2267, pruned_loss=0.0396, over 4875.00 frames. ], tot_loss[loss=0.1713, simple_loss=0.245, pruned_loss=0.04881, over 954860.45 frames. ], batch size: 32, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:32:25,170 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8091, 1.6978, 2.0237, 2.3043, 1.6182, 1.3255, 1.5947, 0.8310], device='cuda:4'), covar=tensor([0.0607, 0.0605, 0.0486, 0.0694, 0.0718, 0.1481, 0.0790, 0.0857], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0068, 0.0067, 0.0068, 0.0074, 0.0094, 0.0072, 0.0064], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 22:32:26,264 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.107e+02 1.638e+02 1.894e+02 2.201e+02 5.051e+02, threshold=3.789e+02, percent-clipped=1.0 2023-04-27 22:32:47,757 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=141685.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:33:28,594 INFO [finetune.py:976] (4/7) Epoch 25, batch 4250, loss[loss=0.1668, simple_loss=0.2355, pruned_loss=0.04902, over 4817.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2425, pruned_loss=0.04768, over 954425.22 frames. ], batch size: 30, lr: 3.00e-03, grad_scale: 16.0 2023-04-27 22:33:38,480 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8952, 1.9656, 1.1059, 1.6103, 2.1166, 1.6966, 1.7062, 1.7020], device='cuda:4'), covar=tensor([0.0437, 0.0292, 0.0285, 0.0480, 0.0235, 0.0435, 0.0426, 0.0489], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0027, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0045, 0.0038, 0.0052, 0.0038, 0.0050, 0.0050, 0.0052], device='cuda:4') 2023-04-27 22:34:30,331 INFO [finetune.py:976] (4/7) Epoch 25, batch 4300, loss[loss=0.1258, simple_loss=0.203, pruned_loss=0.02432, over 4774.00 frames. ], tot_loss[loss=0.1673, simple_loss=0.2397, pruned_loss=0.04747, over 954960.77 frames. ], batch size: 28, lr: 2.99e-03, grad_scale: 16.0 2023-04-27 22:34:39,434 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.343e+01 1.507e+02 1.747e+02 2.193e+02 4.425e+02, threshold=3.494e+02, percent-clipped=2.0 2023-04-27 22:34:41,378 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=141773.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 22:35:15,326 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=141797.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:35:16,568 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=141799.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:35:36,800 INFO [finetune.py:976] (4/7) Epoch 25, batch 4350, loss[loss=0.1403, simple_loss=0.223, pruned_loss=0.02877, over 4773.00 frames. ], tot_loss[loss=0.1654, simple_loss=0.2372, pruned_loss=0.04677, over 955029.13 frames. ], batch size: 28, lr: 2.99e-03, grad_scale: 16.0 2023-04-27 22:35:38,801 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0409, 1.6529, 1.9888, 2.5537, 2.5201, 2.1168, 1.8266, 2.2585], device='cuda:4'), covar=tensor([0.0852, 0.1334, 0.0791, 0.0531, 0.0597, 0.0766, 0.0686, 0.0573], device='cuda:4'), in_proj_covar=tensor([0.0185, 0.0203, 0.0186, 0.0173, 0.0179, 0.0178, 0.0151, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 22:35:45,076 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8243, 2.3476, 1.9657, 2.1465, 1.4627, 1.8730, 1.9576, 1.5519], device='cuda:4'), covar=tensor([0.2152, 0.1375, 0.1042, 0.1551, 0.3669, 0.1446, 0.1902, 0.2525], device='cuda:4'), in_proj_covar=tensor([0.0287, 0.0303, 0.0216, 0.0279, 0.0315, 0.0258, 0.0251, 0.0267], device='cuda:4'), out_proj_covar=tensor([1.1477e-04, 1.1963e-04, 8.4870e-05, 1.0987e-04, 1.2702e-04, 1.0171e-04, 1.0100e-04, 1.0557e-04], device='cuda:4') 2023-04-27 22:36:20,236 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=141847.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:36:30,303 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.67 vs. limit=2.0 2023-04-27 22:36:40,454 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6294, 1.4062, 4.4738, 4.1587, 3.8258, 4.2846, 4.0634, 3.9425], device='cuda:4'), covar=tensor([0.7620, 0.6442, 0.1125, 0.1978, 0.1316, 0.1840, 0.1517, 0.1659], device='cuda:4'), in_proj_covar=tensor([0.0315, 0.0311, 0.0413, 0.0413, 0.0352, 0.0416, 0.0322, 0.0372], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 22:36:42,225 INFO [finetune.py:976] (4/7) Epoch 25, batch 4400, loss[loss=0.1543, simple_loss=0.2387, pruned_loss=0.03492, over 4930.00 frames. ], tot_loss[loss=0.1674, simple_loss=0.2389, pruned_loss=0.04796, over 956035.37 frames. ], batch size: 33, lr: 2.99e-03, grad_scale: 16.0 2023-04-27 22:36:43,536 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7916, 2.4910, 0.9457, 1.1528, 1.8815, 1.1550, 3.1385, 1.2638], device='cuda:4'), covar=tensor([0.0973, 0.1199, 0.1126, 0.1935, 0.0681, 0.1466, 0.0438, 0.1116], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0050, 0.0072, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 22:36:50,565 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.176e+02 1.541e+02 1.847e+02 2.232e+02 3.824e+02, threshold=3.693e+02, percent-clipped=5.0 2023-04-27 22:37:26,207 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.6986, 4.7618, 3.2119, 5.4953, 4.8458, 4.7187, 1.8160, 4.7124], device='cuda:4'), covar=tensor([0.1684, 0.1122, 0.3336, 0.1050, 0.5150, 0.2055, 0.6770, 0.2119], device='cuda:4'), in_proj_covar=tensor([0.0247, 0.0220, 0.0253, 0.0307, 0.0301, 0.0250, 0.0275, 0.0275], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 22:37:46,056 INFO [finetune.py:976] (4/7) Epoch 25, batch 4450, loss[loss=0.1699, simple_loss=0.2258, pruned_loss=0.057, over 4756.00 frames. ], tot_loss[loss=0.1702, simple_loss=0.2426, pruned_loss=0.04892, over 957609.57 frames. ], batch size: 23, lr: 2.99e-03, grad_scale: 32.0 2023-04-27 22:37:46,803 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=141916.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:38:05,411 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=141928.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:38:19,869 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-04-27 22:38:50,098 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-27 22:38:51,131 INFO [finetune.py:976] (4/7) Epoch 25, batch 4500, loss[loss=0.1836, simple_loss=0.2537, pruned_loss=0.05678, over 4867.00 frames. ], tot_loss[loss=0.1706, simple_loss=0.2436, pruned_loss=0.04877, over 956900.22 frames. ], batch size: 34, lr: 2.99e-03, grad_scale: 32.0 2023-04-27 22:38:59,406 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.003e+02 1.537e+02 1.854e+02 2.228e+02 3.851e+02, threshold=3.709e+02, percent-clipped=1.0 2023-04-27 22:39:09,406 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=141977.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:39:21,468 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=141985.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:39:23,961 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=141989.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:39:57,185 INFO [finetune.py:976] (4/7) Epoch 25, batch 4550, loss[loss=0.1594, simple_loss=0.2435, pruned_loss=0.03764, over 4866.00 frames. ], tot_loss[loss=0.171, simple_loss=0.2444, pruned_loss=0.04885, over 956784.47 frames. ], batch size: 34, lr: 2.99e-03, grad_scale: 32.0 2023-04-27 22:40:19,390 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=142033.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:40:38,473 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.56 vs. limit=2.0 2023-04-27 22:41:01,792 INFO [finetune.py:976] (4/7) Epoch 25, batch 4600, loss[loss=0.1543, simple_loss=0.2304, pruned_loss=0.0391, over 4917.00 frames. ], tot_loss[loss=0.1702, simple_loss=0.2433, pruned_loss=0.04855, over 954063.61 frames. ], batch size: 38, lr: 2.99e-03, grad_scale: 32.0 2023-04-27 22:41:10,130 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.365e+01 1.636e+02 1.871e+02 2.331e+02 4.472e+02, threshold=3.743e+02, percent-clipped=1.0 2023-04-27 22:41:12,058 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=142073.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 22:41:18,936 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5140, 1.4113, 4.3134, 4.0136, 3.7598, 4.1601, 3.9969, 3.7484], device='cuda:4'), covar=tensor([0.7498, 0.6235, 0.1084, 0.1857, 0.1174, 0.1906, 0.1410, 0.1870], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0309, 0.0409, 0.0411, 0.0350, 0.0415, 0.0320, 0.0369], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 22:41:23,280 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2723, 2.9367, 2.3897, 2.5065, 1.7174, 1.6464, 2.5433, 1.6505], device='cuda:4'), covar=tensor([0.1519, 0.1269, 0.1200, 0.1367, 0.2028, 0.1820, 0.0788, 0.1841], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0209, 0.0168, 0.0203, 0.0199, 0.0185, 0.0156, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 22:41:42,619 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6196, 1.1959, 1.3797, 1.2606, 1.7045, 1.4219, 1.2220, 1.3402], device='cuda:4'), covar=tensor([0.1547, 0.1575, 0.1968, 0.1510, 0.0971, 0.1432, 0.1742, 0.2281], device='cuda:4'), in_proj_covar=tensor([0.0314, 0.0308, 0.0352, 0.0287, 0.0329, 0.0305, 0.0299, 0.0374], device='cuda:4'), out_proj_covar=tensor([6.4179e-05, 6.3358e-05, 7.3927e-05, 5.7444e-05, 6.7412e-05, 6.3730e-05, 6.1901e-05, 7.9210e-05], device='cuda:4') 2023-04-27 22:41:43,178 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=142097.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:42:06,627 INFO [finetune.py:976] (4/7) Epoch 25, batch 4650, loss[loss=0.1378, simple_loss=0.2108, pruned_loss=0.03245, over 4901.00 frames. ], tot_loss[loss=0.1687, simple_loss=0.2407, pruned_loss=0.04836, over 953941.80 frames. ], batch size: 43, lr: 2.99e-03, grad_scale: 32.0 2023-04-27 22:42:15,789 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=142121.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 22:42:26,864 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6845, 0.7576, 1.4925, 1.9800, 1.7470, 1.5529, 1.5335, 1.5506], device='cuda:4'), covar=tensor([0.4318, 0.6154, 0.5825, 0.5669, 0.5755, 0.7114, 0.7188, 0.7010], device='cuda:4'), in_proj_covar=tensor([0.0438, 0.0419, 0.0513, 0.0507, 0.0466, 0.0500, 0.0503, 0.0516], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 22:42:28,035 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4133, 3.0658, 0.9409, 1.8029, 1.7518, 2.1772, 1.8721, 0.9326], device='cuda:4'), covar=tensor([0.1360, 0.1020, 0.1812, 0.1168, 0.1065, 0.0953, 0.1392, 0.1815], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0242, 0.0138, 0.0122, 0.0133, 0.0154, 0.0119, 0.0121], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 22:42:47,433 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=142145.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:43:11,598 INFO [finetune.py:976] (4/7) Epoch 25, batch 4700, loss[loss=0.1508, simple_loss=0.2224, pruned_loss=0.03956, over 4932.00 frames. ], tot_loss[loss=0.1666, simple_loss=0.238, pruned_loss=0.04764, over 955753.12 frames. ], batch size: 33, lr: 2.99e-03, grad_scale: 32.0 2023-04-27 22:43:19,815 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.009e+02 1.574e+02 1.870e+02 2.251e+02 4.397e+02, threshold=3.741e+02, percent-clipped=1.0 2023-04-27 22:43:30,489 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2750, 1.2817, 1.5120, 1.6328, 1.2612, 1.0445, 1.2526, 0.8157], device='cuda:4'), covar=tensor([0.0623, 0.0580, 0.0425, 0.0522, 0.0700, 0.1491, 0.0643, 0.0634], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0068, 0.0067, 0.0069, 0.0075, 0.0095, 0.0073, 0.0064], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 22:43:32,248 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0063, 1.3102, 5.1667, 4.7810, 4.4357, 4.9483, 4.5013, 4.5440], device='cuda:4'), covar=tensor([0.6652, 0.6366, 0.0987, 0.1727, 0.0962, 0.1346, 0.1358, 0.1662], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0307, 0.0407, 0.0409, 0.0348, 0.0413, 0.0318, 0.0367], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 22:43:53,127 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3088, 1.2488, 1.3271, 1.6460, 1.6439, 1.3272, 1.0896, 1.5058], device='cuda:4'), covar=tensor([0.0861, 0.1549, 0.1082, 0.0589, 0.0715, 0.0880, 0.0838, 0.0664], device='cuda:4'), in_proj_covar=tensor([0.0185, 0.0201, 0.0185, 0.0171, 0.0177, 0.0177, 0.0150, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 22:44:17,215 INFO [finetune.py:976] (4/7) Epoch 25, batch 4750, loss[loss=0.1444, simple_loss=0.2237, pruned_loss=0.03256, over 4843.00 frames. ], tot_loss[loss=0.1667, simple_loss=0.2374, pruned_loss=0.04804, over 954795.67 frames. ], batch size: 47, lr: 2.99e-03, grad_scale: 32.0 2023-04-27 22:45:29,867 INFO [finetune.py:976] (4/7) Epoch 25, batch 4800, loss[loss=0.1329, simple_loss=0.209, pruned_loss=0.02841, over 4680.00 frames. ], tot_loss[loss=0.1679, simple_loss=0.2393, pruned_loss=0.04826, over 955303.84 frames. ], batch size: 23, lr: 2.99e-03, grad_scale: 32.0 2023-04-27 22:45:34,011 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.041e+02 1.626e+02 1.868e+02 2.200e+02 5.189e+02, threshold=3.736e+02, percent-clipped=2.0 2023-04-27 22:45:40,862 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=142272.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:45:43,746 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4024, 1.3078, 1.6735, 1.6734, 1.2921, 1.1828, 1.3831, 0.8566], device='cuda:4'), covar=tensor([0.0473, 0.0589, 0.0342, 0.0459, 0.0731, 0.1037, 0.0515, 0.0560], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0068, 0.0067, 0.0069, 0.0075, 0.0095, 0.0073, 0.0064], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 22:45:54,145 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=142284.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:46:35,527 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0473, 1.9008, 1.7201, 1.6250, 2.1270, 1.6860, 2.4744, 1.5781], device='cuda:4'), covar=tensor([0.3093, 0.1752, 0.4485, 0.2565, 0.1407, 0.2139, 0.1452, 0.4112], device='cuda:4'), in_proj_covar=tensor([0.0339, 0.0353, 0.0427, 0.0352, 0.0383, 0.0375, 0.0370, 0.0424], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 22:46:36,615 INFO [finetune.py:976] (4/7) Epoch 25, batch 4850, loss[loss=0.1716, simple_loss=0.2379, pruned_loss=0.05264, over 4807.00 frames. ], tot_loss[loss=0.1709, simple_loss=0.2435, pruned_loss=0.04915, over 956130.88 frames. ], batch size: 25, lr: 2.99e-03, grad_scale: 32.0 2023-04-27 22:47:30,112 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3823, 2.6815, 1.2769, 1.6850, 2.1248, 1.6160, 3.4696, 2.0403], device='cuda:4'), covar=tensor([0.0603, 0.0715, 0.0817, 0.1310, 0.0515, 0.0977, 0.0321, 0.0664], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0072, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 22:47:42,707 INFO [finetune.py:976] (4/7) Epoch 25, batch 4900, loss[loss=0.1994, simple_loss=0.2665, pruned_loss=0.06614, over 4211.00 frames. ], tot_loss[loss=0.1722, simple_loss=0.2447, pruned_loss=0.0498, over 951038.27 frames. ], batch size: 65, lr: 2.99e-03, grad_scale: 32.0 2023-04-27 22:47:51,896 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.944e+01 1.647e+02 1.933e+02 2.315e+02 7.407e+02, threshold=3.866e+02, percent-clipped=3.0 2023-04-27 22:47:56,014 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5864, 1.9801, 1.7740, 2.4772, 2.5991, 2.1392, 2.1680, 1.8301], device='cuda:4'), covar=tensor([0.1635, 0.1495, 0.2012, 0.1456, 0.0954, 0.1684, 0.1801, 0.2175], device='cuda:4'), in_proj_covar=tensor([0.0314, 0.0308, 0.0352, 0.0286, 0.0328, 0.0305, 0.0299, 0.0374], device='cuda:4'), out_proj_covar=tensor([6.4206e-05, 6.3317e-05, 7.4080e-05, 5.7307e-05, 6.7155e-05, 6.3749e-05, 6.1951e-05, 7.9251e-05], device='cuda:4') 2023-04-27 22:48:49,020 INFO [finetune.py:976] (4/7) Epoch 25, batch 4950, loss[loss=0.1881, simple_loss=0.2446, pruned_loss=0.06584, over 4285.00 frames. ], tot_loss[loss=0.1724, simple_loss=0.2452, pruned_loss=0.04982, over 950217.00 frames. ], batch size: 66, lr: 2.99e-03, grad_scale: 32.0 2023-04-27 22:49:36,597 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.6408, 3.5455, 2.7463, 4.1748, 3.5933, 3.6410, 1.3153, 3.6228], device='cuda:4'), covar=tensor([0.1948, 0.1450, 0.3196, 0.2012, 0.3412, 0.2004, 0.6149, 0.2253], device='cuda:4'), in_proj_covar=tensor([0.0245, 0.0218, 0.0250, 0.0304, 0.0298, 0.0248, 0.0273, 0.0273], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 22:49:47,822 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9865, 1.7348, 1.5370, 1.7659, 2.1928, 1.7844, 1.6111, 1.4821], device='cuda:4'), covar=tensor([0.1699, 0.1519, 0.2190, 0.1433, 0.0991, 0.1656, 0.1953, 0.2434], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0307, 0.0350, 0.0285, 0.0326, 0.0304, 0.0298, 0.0372], device='cuda:4'), out_proj_covar=tensor([6.3848e-05, 6.3093e-05, 7.3589e-05, 5.7179e-05, 6.6926e-05, 6.3471e-05, 6.1719e-05, 7.8832e-05], device='cuda:4') 2023-04-27 22:49:57,922 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7144, 3.6303, 0.9836, 2.1370, 2.0517, 2.4631, 2.0019, 0.9787], device='cuda:4'), covar=tensor([0.1272, 0.0810, 0.1807, 0.1087, 0.0962, 0.1004, 0.1483, 0.2035], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0239, 0.0136, 0.0121, 0.0131, 0.0153, 0.0117, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 22:49:58,431 INFO [finetune.py:976] (4/7) Epoch 25, batch 5000, loss[loss=0.1713, simple_loss=0.2386, pruned_loss=0.05207, over 4904.00 frames. ], tot_loss[loss=0.1705, simple_loss=0.2432, pruned_loss=0.04885, over 950822.22 frames. ], batch size: 36, lr: 2.99e-03, grad_scale: 32.0 2023-04-27 22:50:01,484 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.151e+02 1.520e+02 1.784e+02 2.172e+02 4.625e+02, threshold=3.567e+02, percent-clipped=1.0 2023-04-27 22:51:03,147 INFO [finetune.py:976] (4/7) Epoch 25, batch 5050, loss[loss=0.144, simple_loss=0.2152, pruned_loss=0.03646, over 4719.00 frames. ], tot_loss[loss=0.1693, simple_loss=0.241, pruned_loss=0.04874, over 951553.30 frames. ], batch size: 23, lr: 2.99e-03, grad_scale: 32.0 2023-04-27 22:51:48,457 INFO [finetune.py:976] (4/7) Epoch 25, batch 5100, loss[loss=0.1818, simple_loss=0.2497, pruned_loss=0.05696, over 4824.00 frames. ], tot_loss[loss=0.1657, simple_loss=0.237, pruned_loss=0.04721, over 952995.92 frames. ], batch size: 41, lr: 2.99e-03, grad_scale: 32.0 2023-04-27 22:51:51,978 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.162e+02 1.504e+02 1.862e+02 2.404e+02 6.312e+02, threshold=3.723e+02, percent-clipped=2.0 2023-04-27 22:51:53,293 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=142572.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:51:53,947 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=142573.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:52:01,672 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=142584.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:52:21,630 INFO [finetune.py:976] (4/7) Epoch 25, batch 5150, loss[loss=0.203, simple_loss=0.267, pruned_loss=0.06953, over 4833.00 frames. ], tot_loss[loss=0.1673, simple_loss=0.2383, pruned_loss=0.04818, over 952692.94 frames. ], batch size: 33, lr: 2.99e-03, grad_scale: 32.0 2023-04-27 22:52:26,239 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=142620.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:52:33,588 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=142632.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:52:35,349 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=142634.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:52:38,361 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.82 vs. limit=5.0 2023-04-27 22:52:56,102 INFO [finetune.py:976] (4/7) Epoch 25, batch 5200, loss[loss=0.2216, simple_loss=0.3005, pruned_loss=0.07137, over 4807.00 frames. ], tot_loss[loss=0.1693, simple_loss=0.2411, pruned_loss=0.04875, over 952919.30 frames. ], batch size: 51, lr: 2.99e-03, grad_scale: 32.0 2023-04-27 22:53:00,174 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.071e+02 1.581e+02 1.903e+02 2.348e+02 3.515e+02, threshold=3.805e+02, percent-clipped=0.0 2023-04-27 22:53:05,837 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5690, 2.1138, 1.4275, 1.4240, 1.1932, 1.2232, 1.4348, 1.1078], device='cuda:4'), covar=tensor([0.1951, 0.1258, 0.1746, 0.1762, 0.2532, 0.2272, 0.1106, 0.2245], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0211, 0.0168, 0.0204, 0.0200, 0.0186, 0.0157, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 22:53:17,681 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6288, 1.4824, 4.3811, 4.0800, 3.7818, 4.1935, 4.1059, 3.8401], device='cuda:4'), covar=tensor([0.7020, 0.5766, 0.1034, 0.1782, 0.1248, 0.1658, 0.1312, 0.1557], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0309, 0.0409, 0.0411, 0.0350, 0.0415, 0.0320, 0.0368], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 22:53:29,659 INFO [finetune.py:976] (4/7) Epoch 25, batch 5250, loss[loss=0.1903, simple_loss=0.2505, pruned_loss=0.06504, over 4863.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2417, pruned_loss=0.04856, over 952918.87 frames. ], batch size: 31, lr: 2.99e-03, grad_scale: 32.0 2023-04-27 22:53:36,252 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=142724.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:53:57,551 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0996, 1.4501, 1.9460, 2.2154, 2.0544, 1.5100, 1.2364, 1.7183], device='cuda:4'), covar=tensor([0.3687, 0.4037, 0.2121, 0.2692, 0.2655, 0.3001, 0.4492, 0.2159], device='cuda:4'), in_proj_covar=tensor([0.0288, 0.0243, 0.0225, 0.0310, 0.0218, 0.0231, 0.0224, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 22:54:07,460 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2895, 1.3891, 3.7876, 3.5045, 3.3600, 3.6126, 3.6180, 3.3585], device='cuda:4'), covar=tensor([0.7259, 0.5489, 0.1197, 0.1924, 0.1273, 0.1561, 0.1886, 0.1621], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0308, 0.0407, 0.0410, 0.0349, 0.0414, 0.0318, 0.0367], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 22:54:10,551 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0550, 2.7007, 1.0625, 1.4084, 1.9019, 1.1999, 3.3650, 1.7118], device='cuda:4'), covar=tensor([0.0675, 0.0649, 0.0841, 0.1219, 0.0536, 0.1034, 0.0228, 0.0601], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0073, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 22:54:19,563 INFO [finetune.py:976] (4/7) Epoch 25, batch 5300, loss[loss=0.188, simple_loss=0.2498, pruned_loss=0.06314, over 4722.00 frames. ], tot_loss[loss=0.1698, simple_loss=0.2426, pruned_loss=0.04852, over 955013.91 frames. ], batch size: 23, lr: 2.99e-03, grad_scale: 32.0 2023-04-27 22:54:22,608 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.008e+02 1.572e+02 1.803e+02 2.243e+02 5.104e+02, threshold=3.606e+02, percent-clipped=2.0 2023-04-27 22:54:22,760 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4248, 1.3730, 1.7309, 1.7938, 1.3654, 1.0739, 1.3138, 0.7302], device='cuda:4'), covar=tensor([0.0591, 0.0631, 0.0365, 0.0516, 0.0709, 0.1581, 0.0645, 0.0763], device='cuda:4'), in_proj_covar=tensor([0.0071, 0.0068, 0.0066, 0.0068, 0.0074, 0.0095, 0.0073, 0.0064], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 22:54:44,812 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=142785.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:55:27,672 INFO [finetune.py:976] (4/7) Epoch 25, batch 5350, loss[loss=0.1407, simple_loss=0.2276, pruned_loss=0.02689, over 4885.00 frames. ], tot_loss[loss=0.1681, simple_loss=0.2418, pruned_loss=0.04718, over 954246.20 frames. ], batch size: 35, lr: 2.99e-03, grad_scale: 32.0 2023-04-27 22:55:27,809 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7414, 1.6681, 1.8795, 2.1168, 2.1219, 1.6426, 1.3712, 1.9110], device='cuda:4'), covar=tensor([0.0782, 0.1261, 0.0763, 0.0544, 0.0634, 0.0842, 0.0759, 0.0573], device='cuda:4'), in_proj_covar=tensor([0.0183, 0.0200, 0.0183, 0.0170, 0.0175, 0.0176, 0.0149, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 22:56:34,014 INFO [finetune.py:976] (4/7) Epoch 25, batch 5400, loss[loss=0.1203, simple_loss=0.1964, pruned_loss=0.0221, over 4835.00 frames. ], tot_loss[loss=0.1663, simple_loss=0.2396, pruned_loss=0.04656, over 955671.41 frames. ], batch size: 47, lr: 2.99e-03, grad_scale: 32.0 2023-04-27 22:56:42,410 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.018e+02 1.482e+02 1.784e+02 2.095e+02 4.679e+02, threshold=3.568e+02, percent-clipped=3.0 2023-04-27 22:57:18,643 INFO [finetune.py:976] (4/7) Epoch 25, batch 5450, loss[loss=0.1323, simple_loss=0.1954, pruned_loss=0.03462, over 4872.00 frames. ], tot_loss[loss=0.1644, simple_loss=0.2369, pruned_loss=0.04593, over 956083.29 frames. ], batch size: 44, lr: 2.99e-03, grad_scale: 32.0 2023-04-27 22:57:27,223 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=142929.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:57:49,159 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6072, 1.4919, 4.3828, 4.1059, 3.8669, 4.1452, 4.1525, 3.8474], device='cuda:4'), covar=tensor([0.7142, 0.5748, 0.1214, 0.2051, 0.1161, 0.2357, 0.1170, 0.1879], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0306, 0.0406, 0.0407, 0.0347, 0.0412, 0.0317, 0.0365], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-27 22:57:52,210 INFO [finetune.py:976] (4/7) Epoch 25, batch 5500, loss[loss=0.1521, simple_loss=0.209, pruned_loss=0.04754, over 4037.00 frames. ], tot_loss[loss=0.1628, simple_loss=0.2347, pruned_loss=0.04544, over 955537.70 frames. ], batch size: 65, lr: 2.99e-03, grad_scale: 32.0 2023-04-27 22:57:55,646 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.918e+01 1.380e+02 1.724e+02 2.155e+02 4.005e+02, threshold=3.448e+02, percent-clipped=1.0 2023-04-27 22:58:26,144 INFO [finetune.py:976] (4/7) Epoch 25, batch 5550, loss[loss=0.1859, simple_loss=0.2569, pruned_loss=0.05748, over 4861.00 frames. ], tot_loss[loss=0.165, simple_loss=0.2368, pruned_loss=0.04664, over 953256.46 frames. ], batch size: 31, lr: 2.99e-03, grad_scale: 32.0 2023-04-27 22:58:50,206 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8191, 2.3783, 1.8710, 1.6593, 1.3747, 1.3603, 1.9807, 1.3328], device='cuda:4'), covar=tensor([0.1768, 0.1382, 0.1424, 0.1815, 0.2216, 0.1929, 0.0962, 0.2075], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0209, 0.0168, 0.0203, 0.0199, 0.0184, 0.0156, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 22:58:57,675 INFO [finetune.py:976] (4/7) Epoch 25, batch 5600, loss[loss=0.1818, simple_loss=0.2533, pruned_loss=0.05521, over 4935.00 frames. ], tot_loss[loss=0.1667, simple_loss=0.2399, pruned_loss=0.04674, over 954262.83 frames. ], batch size: 33, lr: 2.99e-03, grad_scale: 32.0 2023-04-27 22:59:00,544 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.011e+02 1.590e+02 1.837e+02 2.169e+02 3.781e+02, threshold=3.675e+02, percent-clipped=1.0 2023-04-27 22:59:06,465 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=143080.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 22:59:27,655 INFO [finetune.py:976] (4/7) Epoch 25, batch 5650, loss[loss=0.1504, simple_loss=0.2369, pruned_loss=0.03193, over 4892.00 frames. ], tot_loss[loss=0.1687, simple_loss=0.2428, pruned_loss=0.04729, over 955920.63 frames. ], batch size: 43, lr: 2.99e-03, grad_scale: 32.0 2023-04-27 22:59:33,130 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.3437, 3.3163, 2.5019, 3.8568, 3.3624, 3.3117, 1.5148, 3.2781], device='cuda:4'), covar=tensor([0.1723, 0.1479, 0.3328, 0.2302, 0.3579, 0.1998, 0.5614, 0.2724], device='cuda:4'), in_proj_covar=tensor([0.0243, 0.0217, 0.0250, 0.0304, 0.0296, 0.0247, 0.0272, 0.0273], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 23:00:04,052 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.9421, 3.9338, 3.0460, 4.5039, 3.9744, 3.9115, 2.0056, 3.8672], device='cuda:4'), covar=tensor([0.1777, 0.1088, 0.2421, 0.1428, 0.2980, 0.1679, 0.5158, 0.2125], device='cuda:4'), in_proj_covar=tensor([0.0243, 0.0217, 0.0250, 0.0303, 0.0296, 0.0246, 0.0272, 0.0272], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 23:00:24,801 INFO [finetune.py:976] (4/7) Epoch 25, batch 5700, loss[loss=0.1776, simple_loss=0.2391, pruned_loss=0.05801, over 4049.00 frames. ], tot_loss[loss=0.1666, simple_loss=0.2389, pruned_loss=0.04709, over 935012.00 frames. ], batch size: 18, lr: 2.98e-03, grad_scale: 32.0 2023-04-27 23:00:27,757 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.938e+01 1.487e+02 1.759e+02 2.216e+02 4.830e+02, threshold=3.518e+02, percent-clipped=2.0 2023-04-27 23:01:04,606 INFO [finetune.py:976] (4/7) Epoch 26, batch 0, loss[loss=0.1479, simple_loss=0.2279, pruned_loss=0.03394, over 4765.00 frames. ], tot_loss[loss=0.1479, simple_loss=0.2279, pruned_loss=0.03394, over 4765.00 frames. ], batch size: 26, lr: 2.98e-03, grad_scale: 32.0 2023-04-27 23:01:04,607 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 23:01:13,288 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7748, 2.1001, 1.8492, 2.0393, 1.6487, 1.7657, 1.7097, 1.4094], device='cuda:4'), covar=tensor([0.1853, 0.1412, 0.0823, 0.1198, 0.3684, 0.1304, 0.1980, 0.2488], device='cuda:4'), in_proj_covar=tensor([0.0284, 0.0302, 0.0213, 0.0276, 0.0313, 0.0256, 0.0248, 0.0266], device='cuda:4'), out_proj_covar=tensor([1.1349e-04, 1.1920e-04, 8.3963e-05, 1.0865e-04, 1.2629e-04, 1.0072e-04, 1.0005e-04, 1.0505e-04], device='cuda:4') 2023-04-27 23:01:15,352 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4349, 1.2852, 1.6761, 1.6808, 1.3092, 1.2955, 1.3846, 0.8751], device='cuda:4'), covar=tensor([0.0518, 0.0657, 0.0372, 0.0502, 0.0725, 0.1105, 0.0466, 0.0527], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0067, 0.0066, 0.0068, 0.0073, 0.0094, 0.0072, 0.0063], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 23:01:18,325 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2541, 1.4911, 1.7743, 1.8959, 1.8112, 1.9121, 1.7673, 1.8103], device='cuda:4'), covar=tensor([0.3660, 0.5595, 0.4517, 0.4878, 0.5891, 0.7087, 0.5296, 0.4769], device='cuda:4'), in_proj_covar=tensor([0.0341, 0.0375, 0.0328, 0.0341, 0.0349, 0.0395, 0.0360, 0.0332], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 23:01:26,499 INFO [finetune.py:1010] (4/7) Epoch 26, validation: loss=0.1543, simple_loss=0.2237, pruned_loss=0.04251, over 2265189.00 frames. 2023-04-27 23:01:26,499 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6529MB 2023-04-27 23:02:16,869 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=143229.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:02:28,243 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.93 vs. limit=2.0 2023-04-27 23:02:34,931 INFO [finetune.py:976] (4/7) Epoch 26, batch 50, loss[loss=0.1862, simple_loss=0.25, pruned_loss=0.06118, over 4842.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.2418, pruned_loss=0.04622, over 217739.36 frames. ], batch size: 44, lr: 2.98e-03, grad_scale: 32.0 2023-04-27 23:03:09,818 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.418e+01 1.474e+02 1.785e+02 2.280e+02 3.483e+02, threshold=3.571e+02, percent-clipped=0.0 2023-04-27 23:03:19,711 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=143277.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:03:40,345 INFO [finetune.py:976] (4/7) Epoch 26, batch 100, loss[loss=0.1448, simple_loss=0.2216, pruned_loss=0.03401, over 4769.00 frames. ], tot_loss[loss=0.1628, simple_loss=0.235, pruned_loss=0.04523, over 379676.28 frames. ], batch size: 28, lr: 2.98e-03, grad_scale: 32.0 2023-04-27 23:03:54,531 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1766, 2.6752, 0.9625, 1.5657, 1.8829, 1.2217, 3.5319, 1.7693], device='cuda:4'), covar=tensor([0.0641, 0.0630, 0.0767, 0.1199, 0.0534, 0.0989, 0.0199, 0.0577], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0073, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 23:04:19,089 INFO [finetune.py:976] (4/7) Epoch 26, batch 150, loss[loss=0.1543, simple_loss=0.214, pruned_loss=0.04725, over 4719.00 frames. ], tot_loss[loss=0.1599, simple_loss=0.2307, pruned_loss=0.04452, over 508399.80 frames. ], batch size: 23, lr: 2.98e-03, grad_scale: 32.0 2023-04-27 23:04:25,881 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.07 vs. limit=5.0 2023-04-27 23:04:37,629 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.031e+02 1.476e+02 1.692e+02 2.051e+02 5.029e+02, threshold=3.384e+02, percent-clipped=1.0 2023-04-27 23:04:43,891 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=143380.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:04:52,918 INFO [finetune.py:976] (4/7) Epoch 26, batch 200, loss[loss=0.2006, simple_loss=0.2645, pruned_loss=0.06831, over 4827.00 frames. ], tot_loss[loss=0.1593, simple_loss=0.2302, pruned_loss=0.04425, over 606896.26 frames. ], batch size: 33, lr: 2.98e-03, grad_scale: 32.0 2023-04-27 23:04:59,038 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=143400.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 23:05:16,579 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=143428.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:05:16,659 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4752, 2.0280, 2.4057, 3.0718, 2.4062, 1.8622, 1.9105, 2.3606], device='cuda:4'), covar=tensor([0.3109, 0.3189, 0.1585, 0.2026, 0.2650, 0.2631, 0.3547, 0.1986], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0247, 0.0229, 0.0315, 0.0221, 0.0235, 0.0227, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 23:05:31,680 INFO [finetune.py:976] (4/7) Epoch 26, batch 250, loss[loss=0.1774, simple_loss=0.2571, pruned_loss=0.04891, over 4804.00 frames. ], tot_loss[loss=0.1644, simple_loss=0.2358, pruned_loss=0.04652, over 685257.82 frames. ], batch size: 38, lr: 2.98e-03, grad_scale: 32.0 2023-04-27 23:05:44,908 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=143461.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 23:05:47,389 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2820, 3.2958, 2.8671, 3.0882, 3.3518, 3.0666, 4.0986, 2.6558], device='cuda:4'), covar=tensor([0.2736, 0.1545, 0.2945, 0.2430, 0.1301, 0.1821, 0.0997, 0.2725], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0351, 0.0423, 0.0348, 0.0380, 0.0374, 0.0367, 0.0422], device='cuda:4'), out_proj_covar=tensor([9.9711e-05, 1.0461e-04, 1.2805e-04, 1.0466e-04, 1.1269e-04, 1.1139e-04, 1.0755e-04, 1.2697e-04], device='cuda:4') 2023-04-27 23:05:50,285 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.157e+02 1.556e+02 1.824e+02 2.331e+02 6.380e+02, threshold=3.648e+02, percent-clipped=4.0 2023-04-27 23:05:58,103 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=143473.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 23:05:58,133 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1185, 1.4501, 1.9973, 2.4516, 2.1015, 1.5380, 1.2886, 1.8223], device='cuda:4'), covar=tensor([0.3413, 0.3592, 0.1866, 0.2479, 0.2610, 0.2786, 0.4198, 0.2078], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0247, 0.0229, 0.0315, 0.0221, 0.0235, 0.0227, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 23:06:15,192 INFO [finetune.py:976] (4/7) Epoch 26, batch 300, loss[loss=0.1436, simple_loss=0.2194, pruned_loss=0.03385, over 4915.00 frames. ], tot_loss[loss=0.1657, simple_loss=0.2385, pruned_loss=0.04645, over 744428.32 frames. ], batch size: 36, lr: 2.98e-03, grad_scale: 32.0 2023-04-27 23:06:43,065 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=143534.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 23:06:48,240 INFO [finetune.py:976] (4/7) Epoch 26, batch 350, loss[loss=0.1526, simple_loss=0.2285, pruned_loss=0.03838, over 4917.00 frames. ], tot_loss[loss=0.1686, simple_loss=0.2419, pruned_loss=0.04765, over 791472.18 frames. ], batch size: 38, lr: 2.98e-03, grad_scale: 32.0 2023-04-27 23:06:55,951 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8108, 2.4695, 1.9258, 2.3000, 2.5240, 2.1858, 2.9326, 1.8422], device='cuda:4'), covar=tensor([0.3167, 0.1768, 0.4153, 0.3188, 0.1537, 0.2178, 0.1960, 0.4218], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0354, 0.0427, 0.0351, 0.0383, 0.0377, 0.0370, 0.0425], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 23:07:08,173 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.023e+02 1.499e+02 1.727e+02 2.023e+02 3.986e+02, threshold=3.454e+02, percent-clipped=1.0 2023-04-27 23:07:22,105 INFO [finetune.py:976] (4/7) Epoch 26, batch 400, loss[loss=0.1561, simple_loss=0.2261, pruned_loss=0.043, over 4850.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2426, pruned_loss=0.04759, over 826207.92 frames. ], batch size: 44, lr: 2.98e-03, grad_scale: 32.0 2023-04-27 23:07:42,554 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.84 vs. limit=5.0 2023-04-27 23:07:55,515 INFO [finetune.py:976] (4/7) Epoch 26, batch 450, loss[loss=0.1693, simple_loss=0.2384, pruned_loss=0.05007, over 4712.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.2426, pruned_loss=0.0482, over 855850.44 frames. ], batch size: 59, lr: 2.98e-03, grad_scale: 32.0 2023-04-27 23:07:58,555 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4792, 1.5129, 1.8429, 2.8277, 2.8710, 2.2283, 1.9968, 2.4627], device='cuda:4'), covar=tensor([0.0965, 0.2009, 0.1162, 0.0693, 0.0602, 0.1172, 0.0907, 0.0728], device='cuda:4'), in_proj_covar=tensor([0.0184, 0.0203, 0.0185, 0.0171, 0.0177, 0.0177, 0.0150, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 23:08:20,552 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.962e+01 1.532e+02 1.796e+02 2.161e+02 5.029e+02, threshold=3.592e+02, percent-clipped=5.0 2023-04-27 23:08:43,966 INFO [finetune.py:976] (4/7) Epoch 26, batch 500, loss[loss=0.1799, simple_loss=0.2444, pruned_loss=0.05772, over 4832.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.2392, pruned_loss=0.04725, over 877839.87 frames. ], batch size: 40, lr: 2.98e-03, grad_scale: 32.0 2023-04-27 23:09:27,560 INFO [finetune.py:976] (4/7) Epoch 26, batch 550, loss[loss=0.1489, simple_loss=0.2231, pruned_loss=0.03732, over 4912.00 frames. ], tot_loss[loss=0.1631, simple_loss=0.2348, pruned_loss=0.04569, over 894435.62 frames. ], batch size: 36, lr: 2.98e-03, grad_scale: 32.0 2023-04-27 23:09:37,708 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=143756.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 23:09:43,723 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7071, 2.0184, 1.7962, 1.9139, 1.4884, 1.7813, 1.6620, 1.4060], device='cuda:4'), covar=tensor([0.1646, 0.1204, 0.0806, 0.1208, 0.3278, 0.1040, 0.1650, 0.2230], device='cuda:4'), in_proj_covar=tensor([0.0287, 0.0303, 0.0215, 0.0278, 0.0315, 0.0257, 0.0251, 0.0267], device='cuda:4'), out_proj_covar=tensor([1.1464e-04, 1.1964e-04, 8.4491e-05, 1.0936e-04, 1.2713e-04, 1.0095e-04, 1.0102e-04, 1.0537e-04], device='cuda:4') 2023-04-27 23:09:47,112 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.995e+01 1.539e+02 1.829e+02 2.212e+02 3.034e+02, threshold=3.659e+02, percent-clipped=1.0 2023-04-27 23:09:53,948 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4385, 1.3872, 1.5911, 1.6514, 1.3432, 1.1691, 1.4434, 0.9763], device='cuda:4'), covar=tensor([0.0537, 0.0555, 0.0439, 0.0528, 0.0674, 0.0903, 0.0596, 0.0530], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0067, 0.0066, 0.0068, 0.0073, 0.0094, 0.0073, 0.0063], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 23:10:00,518 INFO [finetune.py:976] (4/7) Epoch 26, batch 600, loss[loss=0.1968, simple_loss=0.2722, pruned_loss=0.0607, over 4817.00 frames. ], tot_loss[loss=0.1648, simple_loss=0.2363, pruned_loss=0.04669, over 907753.66 frames. ], batch size: 39, lr: 2.98e-03, grad_scale: 32.0 2023-04-27 23:10:01,249 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5378, 1.3821, 1.7962, 1.8834, 1.4062, 1.2495, 1.5091, 0.9841], device='cuda:4'), covar=tensor([0.0462, 0.0660, 0.0389, 0.0588, 0.0638, 0.1069, 0.0543, 0.0597], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0067, 0.0066, 0.0068, 0.0073, 0.0094, 0.0072, 0.0063], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 23:10:25,720 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=143829.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 23:10:33,616 INFO [finetune.py:976] (4/7) Epoch 26, batch 650, loss[loss=0.1429, simple_loss=0.2299, pruned_loss=0.02795, over 4789.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.2397, pruned_loss=0.04728, over 919307.61 frames. ], batch size: 28, lr: 2.98e-03, grad_scale: 32.0 2023-04-27 23:11:14,499 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.206e+02 1.587e+02 1.916e+02 2.318e+02 7.608e+02, threshold=3.833e+02, percent-clipped=3.0 2023-04-27 23:11:39,522 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5522, 0.9746, 0.3220, 1.2255, 1.0731, 1.4218, 1.3202, 1.2810], device='cuda:4'), covar=tensor([0.0498, 0.0413, 0.0418, 0.0556, 0.0292, 0.0496, 0.0512, 0.0536], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0027, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0046, 0.0038, 0.0053, 0.0039, 0.0050, 0.0050, 0.0052], device='cuda:4') 2023-04-27 23:11:40,023 INFO [finetune.py:976] (4/7) Epoch 26, batch 700, loss[loss=0.1528, simple_loss=0.2211, pruned_loss=0.04227, over 4783.00 frames. ], tot_loss[loss=0.1687, simple_loss=0.2419, pruned_loss=0.04779, over 926224.30 frames. ], batch size: 26, lr: 2.98e-03, grad_scale: 64.0 2023-04-27 23:12:45,799 INFO [finetune.py:976] (4/7) Epoch 26, batch 750, loss[loss=0.2013, simple_loss=0.2788, pruned_loss=0.06186, over 4850.00 frames. ], tot_loss[loss=0.1705, simple_loss=0.2441, pruned_loss=0.04839, over 929165.73 frames. ], batch size: 44, lr: 2.98e-03, grad_scale: 64.0 2023-04-27 23:13:02,612 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0654, 1.7866, 2.3005, 2.5109, 2.1414, 2.0224, 2.1779, 2.1326], device='cuda:4'), covar=tensor([0.4962, 0.6987, 0.6826, 0.5490, 0.6126, 0.8389, 0.9280, 0.9794], device='cuda:4'), in_proj_covar=tensor([0.0443, 0.0424, 0.0518, 0.0511, 0.0471, 0.0507, 0.0509, 0.0521], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 23:13:26,583 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.075e+02 1.623e+02 1.843e+02 2.196e+02 3.710e+02, threshold=3.686e+02, percent-clipped=0.0 2023-04-27 23:13:56,909 INFO [finetune.py:976] (4/7) Epoch 26, batch 800, loss[loss=0.2178, simple_loss=0.2832, pruned_loss=0.07625, over 4799.00 frames. ], tot_loss[loss=0.1696, simple_loss=0.2438, pruned_loss=0.04777, over 935930.74 frames. ], batch size: 45, lr: 2.98e-03, grad_scale: 64.0 2023-04-27 23:13:58,266 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8969, 2.4955, 1.9235, 1.8941, 1.4141, 1.4272, 2.0655, 1.3557], device='cuda:4'), covar=tensor([0.1664, 0.1309, 0.1367, 0.1708, 0.2261, 0.1859, 0.0954, 0.2010], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0209, 0.0168, 0.0204, 0.0201, 0.0185, 0.0157, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 23:14:51,263 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8688, 1.9562, 0.9896, 1.5639, 2.2151, 1.6629, 1.6365, 1.7392], device='cuda:4'), covar=tensor([0.0449, 0.0342, 0.0284, 0.0515, 0.0227, 0.0498, 0.0465, 0.0507], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0027, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0046, 0.0038, 0.0053, 0.0038, 0.0050, 0.0050, 0.0052], device='cuda:4') 2023-04-27 23:15:01,728 INFO [finetune.py:976] (4/7) Epoch 26, batch 850, loss[loss=0.1963, simple_loss=0.267, pruned_loss=0.06277, over 4818.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2428, pruned_loss=0.04798, over 940259.49 frames. ], batch size: 33, lr: 2.98e-03, grad_scale: 64.0 2023-04-27 23:15:15,465 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=144056.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 23:15:35,343 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.021e+02 1.501e+02 1.661e+02 2.180e+02 3.367e+02, threshold=3.322e+02, percent-clipped=0.0 2023-04-27 23:16:05,585 INFO [finetune.py:976] (4/7) Epoch 26, batch 900, loss[loss=0.1389, simple_loss=0.2083, pruned_loss=0.03472, over 4744.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.24, pruned_loss=0.04706, over 945130.59 frames. ], batch size: 27, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:16:06,936 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9558, 2.4238, 1.9459, 1.8431, 1.4631, 1.4539, 2.0198, 1.4208], device='cuda:4'), covar=tensor([0.1404, 0.1251, 0.1203, 0.1517, 0.2126, 0.1726, 0.0852, 0.1862], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0209, 0.0168, 0.0204, 0.0200, 0.0184, 0.0157, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 23:16:12,933 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=144104.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 23:16:16,933 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-04-27 23:16:17,155 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7921, 1.8686, 0.9738, 1.3947, 2.1364, 1.5945, 1.5178, 1.5639], device='cuda:4'), covar=tensor([0.0471, 0.0332, 0.0286, 0.0517, 0.0231, 0.0482, 0.0437, 0.0535], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0027, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0052, 0.0046, 0.0038, 0.0053, 0.0039, 0.0050, 0.0050, 0.0052], device='cuda:4') 2023-04-27 23:16:34,941 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=144129.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 23:16:53,905 INFO [finetune.py:976] (4/7) Epoch 26, batch 950, loss[loss=0.1752, simple_loss=0.2343, pruned_loss=0.05807, over 4002.00 frames. ], tot_loss[loss=0.1659, simple_loss=0.2383, pruned_loss=0.04674, over 946997.37 frames. ], batch size: 65, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:17:28,114 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.085e+02 1.443e+02 1.800e+02 2.192e+02 5.909e+02, threshold=3.600e+02, percent-clipped=3.0 2023-04-27 23:17:37,854 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=144177.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 23:17:59,696 INFO [finetune.py:976] (4/7) Epoch 26, batch 1000, loss[loss=0.1789, simple_loss=0.2697, pruned_loss=0.04399, over 4833.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2415, pruned_loss=0.04811, over 949076.18 frames. ], batch size: 40, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:18:29,345 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.81 vs. limit=2.0 2023-04-27 23:18:32,531 INFO [finetune.py:976] (4/7) Epoch 26, batch 1050, loss[loss=0.1793, simple_loss=0.2481, pruned_loss=0.05521, over 4864.00 frames. ], tot_loss[loss=0.17, simple_loss=0.2435, pruned_loss=0.04828, over 952183.55 frames. ], batch size: 44, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:18:51,251 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.174e+02 1.491e+02 1.893e+02 2.144e+02 7.486e+02, threshold=3.787e+02, percent-clipped=1.0 2023-04-27 23:19:06,557 INFO [finetune.py:976] (4/7) Epoch 26, batch 1100, loss[loss=0.2152, simple_loss=0.2699, pruned_loss=0.08026, over 4179.00 frames. ], tot_loss[loss=0.1701, simple_loss=0.2442, pruned_loss=0.04804, over 953711.45 frames. ], batch size: 65, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:19:32,083 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.44 vs. limit=2.0 2023-04-27 23:19:32,261 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7157, 2.0239, 1.8330, 2.0389, 1.6034, 1.8023, 1.7409, 1.3890], device='cuda:4'), covar=tensor([0.1679, 0.1301, 0.0796, 0.1138, 0.3155, 0.1161, 0.1833, 0.2236], device='cuda:4'), in_proj_covar=tensor([0.0286, 0.0301, 0.0214, 0.0276, 0.0314, 0.0255, 0.0248, 0.0265], device='cuda:4'), out_proj_covar=tensor([1.1428e-04, 1.1875e-04, 8.4096e-05, 1.0887e-04, 1.2673e-04, 1.0029e-04, 1.0011e-04, 1.0475e-04], device='cuda:4') 2023-04-27 23:19:39,804 INFO [finetune.py:976] (4/7) Epoch 26, batch 1150, loss[loss=0.1638, simple_loss=0.2471, pruned_loss=0.04024, over 4817.00 frames. ], tot_loss[loss=0.17, simple_loss=0.2439, pruned_loss=0.048, over 952485.11 frames. ], batch size: 33, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:19:59,727 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.045e+02 1.546e+02 1.832e+02 2.212e+02 3.348e+02, threshold=3.664e+02, percent-clipped=0.0 2023-04-27 23:20:14,236 INFO [finetune.py:976] (4/7) Epoch 26, batch 1200, loss[loss=0.1393, simple_loss=0.2121, pruned_loss=0.03322, over 4909.00 frames. ], tot_loss[loss=0.1681, simple_loss=0.242, pruned_loss=0.04706, over 955029.65 frames. ], batch size: 37, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:20:32,352 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=144411.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:20:50,872 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1825, 3.1329, 2.7622, 2.8107, 3.1164, 2.7974, 4.0919, 2.5698], device='cuda:4'), covar=tensor([0.2833, 0.1832, 0.3129, 0.2657, 0.1621, 0.2236, 0.1047, 0.3214], device='cuda:4'), in_proj_covar=tensor([0.0339, 0.0353, 0.0427, 0.0351, 0.0383, 0.0376, 0.0367, 0.0426], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 23:21:13,635 INFO [finetune.py:976] (4/7) Epoch 26, batch 1250, loss[loss=0.1606, simple_loss=0.2335, pruned_loss=0.04386, over 4816.00 frames. ], tot_loss[loss=0.1669, simple_loss=0.24, pruned_loss=0.04685, over 952693.73 frames. ], batch size: 41, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:21:45,825 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.274e+01 1.443e+02 1.675e+02 2.129e+02 3.494e+02, threshold=3.349e+02, percent-clipped=0.0 2023-04-27 23:21:51,762 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=144472.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:22:16,354 INFO [finetune.py:976] (4/7) Epoch 26, batch 1300, loss[loss=0.1856, simple_loss=0.2549, pruned_loss=0.0582, over 4903.00 frames. ], tot_loss[loss=0.1635, simple_loss=0.2362, pruned_loss=0.04536, over 954479.26 frames. ], batch size: 43, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:22:24,427 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-27 23:23:06,076 INFO [finetune.py:976] (4/7) Epoch 26, batch 1350, loss[loss=0.1788, simple_loss=0.2527, pruned_loss=0.05244, over 4178.00 frames. ], tot_loss[loss=0.1651, simple_loss=0.2378, pruned_loss=0.0462, over 950614.90 frames. ], batch size: 65, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:23:26,216 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.435e+01 1.590e+02 1.903e+02 2.333e+02 4.606e+02, threshold=3.805e+02, percent-clipped=5.0 2023-04-27 23:23:38,292 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8010, 2.4055, 1.8809, 1.7687, 1.3564, 1.3717, 1.9186, 1.2986], device='cuda:4'), covar=tensor([0.1571, 0.1244, 0.1284, 0.1550, 0.2136, 0.1818, 0.0898, 0.1939], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0209, 0.0168, 0.0203, 0.0200, 0.0184, 0.0156, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 23:23:39,995 INFO [finetune.py:976] (4/7) Epoch 26, batch 1400, loss[loss=0.1906, simple_loss=0.2581, pruned_loss=0.06157, over 4931.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.2411, pruned_loss=0.04713, over 953512.32 frames. ], batch size: 33, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:24:06,609 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.84 vs. limit=2.0 2023-04-27 23:24:07,028 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5500, 1.6001, 1.8834, 1.9752, 1.8681, 1.9439, 1.9489, 1.9786], device='cuda:4'), covar=tensor([0.3651, 0.5305, 0.4620, 0.4498, 0.5497, 0.6995, 0.4859, 0.4607], device='cuda:4'), in_proj_covar=tensor([0.0339, 0.0374, 0.0328, 0.0338, 0.0348, 0.0393, 0.0359, 0.0331], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 23:24:12,869 INFO [finetune.py:976] (4/7) Epoch 26, batch 1450, loss[loss=0.1388, simple_loss=0.198, pruned_loss=0.03984, over 4195.00 frames. ], tot_loss[loss=0.1686, simple_loss=0.2426, pruned_loss=0.04733, over 953989.47 frames. ], batch size: 18, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:24:33,311 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.219e+02 1.523e+02 1.914e+02 2.224e+02 4.652e+02, threshold=3.827e+02, percent-clipped=2.0 2023-04-27 23:24:45,858 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.62 vs. limit=2.0 2023-04-27 23:24:46,127 INFO [finetune.py:976] (4/7) Epoch 26, batch 1500, loss[loss=0.1331, simple_loss=0.2182, pruned_loss=0.02395, over 4926.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2436, pruned_loss=0.04711, over 955539.15 frames. ], batch size: 42, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:25:20,118 INFO [finetune.py:976] (4/7) Epoch 26, batch 1550, loss[loss=0.1683, simple_loss=0.2476, pruned_loss=0.04455, over 4891.00 frames. ], tot_loss[loss=0.1682, simple_loss=0.2429, pruned_loss=0.04673, over 955713.58 frames. ], batch size: 35, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:25:24,007 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.32 vs. limit=5.0 2023-04-27 23:25:42,637 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=144767.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:25:45,416 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.907e+01 1.462e+02 1.740e+02 2.100e+02 5.227e+02, threshold=3.480e+02, percent-clipped=2.0 2023-04-27 23:26:04,277 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6462, 2.0492, 1.7977, 1.9860, 1.5469, 1.7048, 1.6961, 1.3571], device='cuda:4'), covar=tensor([0.1875, 0.1143, 0.0752, 0.1160, 0.3115, 0.1155, 0.1707, 0.2141], device='cuda:4'), in_proj_covar=tensor([0.0283, 0.0298, 0.0212, 0.0275, 0.0311, 0.0253, 0.0246, 0.0262], device='cuda:4'), out_proj_covar=tensor([1.1300e-04, 1.1736e-04, 8.3530e-05, 1.0811e-04, 1.2546e-04, 9.9663e-05, 9.9224e-05, 1.0353e-04], device='cuda:4') 2023-04-27 23:26:14,646 INFO [finetune.py:976] (4/7) Epoch 26, batch 1600, loss[loss=0.1206, simple_loss=0.198, pruned_loss=0.02161, over 4804.00 frames. ], tot_loss[loss=0.1674, simple_loss=0.2417, pruned_loss=0.0465, over 955627.46 frames. ], batch size: 25, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:26:31,582 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-27 23:27:00,876 INFO [finetune.py:976] (4/7) Epoch 26, batch 1650, loss[loss=0.1436, simple_loss=0.2163, pruned_loss=0.03539, over 4807.00 frames. ], tot_loss[loss=0.1655, simple_loss=0.2392, pruned_loss=0.04587, over 957680.91 frames. ], batch size: 51, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:27:10,889 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5995, 1.9820, 1.7717, 2.4725, 2.5424, 2.1639, 2.0691, 1.9229], device='cuda:4'), covar=tensor([0.1678, 0.1777, 0.2044, 0.1348, 0.1241, 0.2007, 0.2186, 0.2506], device='cuda:4'), in_proj_covar=tensor([0.0315, 0.0309, 0.0351, 0.0287, 0.0330, 0.0306, 0.0299, 0.0376], device='cuda:4'), out_proj_covar=tensor([6.4575e-05, 6.3420e-05, 7.3831e-05, 5.7565e-05, 6.7625e-05, 6.3842e-05, 6.1934e-05, 7.9733e-05], device='cuda:4') 2023-04-27 23:27:20,926 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.028e+02 1.459e+02 1.734e+02 2.287e+02 5.147e+02, threshold=3.469e+02, percent-clipped=2.0 2023-04-27 23:27:39,744 INFO [finetune.py:976] (4/7) Epoch 26, batch 1700, loss[loss=0.1251, simple_loss=0.2005, pruned_loss=0.02491, over 4381.00 frames. ], tot_loss[loss=0.163, simple_loss=0.2363, pruned_loss=0.04486, over 956883.73 frames. ], batch size: 19, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:28:44,464 INFO [finetune.py:976] (4/7) Epoch 26, batch 1750, loss[loss=0.1815, simple_loss=0.2612, pruned_loss=0.0509, over 4833.00 frames. ], tot_loss[loss=0.164, simple_loss=0.2371, pruned_loss=0.04547, over 955186.43 frames. ], batch size: 33, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:29:25,332 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.053e+02 1.536e+02 1.830e+02 2.200e+02 7.306e+02, threshold=3.661e+02, percent-clipped=1.0 2023-04-27 23:29:26,025 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2163, 2.6131, 1.1122, 1.4962, 2.1296, 1.2974, 3.5304, 1.8801], device='cuda:4'), covar=tensor([0.0657, 0.0548, 0.0797, 0.1308, 0.0489, 0.1042, 0.0202, 0.0619], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0073, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 23:29:50,042 INFO [finetune.py:976] (4/7) Epoch 26, batch 1800, loss[loss=0.1662, simple_loss=0.2505, pruned_loss=0.04095, over 4730.00 frames. ], tot_loss[loss=0.1659, simple_loss=0.2396, pruned_loss=0.04608, over 956663.96 frames. ], batch size: 59, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:29:53,378 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=144996.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:30:15,078 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=145027.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:30:24,557 INFO [finetune.py:976] (4/7) Epoch 26, batch 1850, loss[loss=0.1972, simple_loss=0.2735, pruned_loss=0.06043, over 4841.00 frames. ], tot_loss[loss=0.1678, simple_loss=0.2418, pruned_loss=0.04692, over 957198.22 frames. ], batch size: 47, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:30:34,390 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=145057.0, num_to_drop=1, layers_to_drop={3} 2023-04-27 23:30:41,227 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=145066.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:30:41,819 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=145067.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:30:44,630 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.003e+02 1.513e+02 1.816e+02 2.184e+02 4.128e+02, threshold=3.632e+02, percent-clipped=2.0 2023-04-27 23:30:47,935 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.55 vs. limit=2.0 2023-04-27 23:30:55,919 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=145088.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:30:58,304 INFO [finetune.py:976] (4/7) Epoch 26, batch 1900, loss[loss=0.1517, simple_loss=0.2303, pruned_loss=0.03653, over 4862.00 frames. ], tot_loss[loss=0.1693, simple_loss=0.2435, pruned_loss=0.04755, over 955982.60 frames. ], batch size: 31, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:31:13,800 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=145115.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:31:22,157 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=145127.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:31:30,941 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9999, 2.2865, 1.2176, 1.6439, 2.3610, 1.7588, 1.7351, 1.7735], device='cuda:4'), covar=tensor([0.0457, 0.0339, 0.0277, 0.0528, 0.0217, 0.0485, 0.0458, 0.0539], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0028, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0046, 0.0038, 0.0053, 0.0039, 0.0050, 0.0050, 0.0052], device='cuda:4') 2023-04-27 23:31:31,593 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3909, 1.7253, 1.8411, 1.8845, 1.8889, 1.9313, 1.8664, 1.8882], device='cuda:4'), covar=tensor([0.3712, 0.5311, 0.4713, 0.4890, 0.5726, 0.7102, 0.5254, 0.4744], device='cuda:4'), in_proj_covar=tensor([0.0341, 0.0376, 0.0330, 0.0340, 0.0351, 0.0396, 0.0361, 0.0334], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 23:31:32,060 INFO [finetune.py:976] (4/7) Epoch 26, batch 1950, loss[loss=0.1501, simple_loss=0.2285, pruned_loss=0.03583, over 4755.00 frames. ], tot_loss[loss=0.1691, simple_loss=0.2429, pruned_loss=0.04765, over 956481.34 frames. ], batch size: 26, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:32:02,788 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4614, 1.1319, 0.3593, 1.1771, 1.0665, 1.3376, 1.2783, 1.2360], device='cuda:4'), covar=tensor([0.0466, 0.0366, 0.0421, 0.0523, 0.0311, 0.0458, 0.0433, 0.0534], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0028, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0051, 0.0046, 0.0038, 0.0053, 0.0039, 0.0050, 0.0050, 0.0052], device='cuda:4') 2023-04-27 23:32:03,281 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.496e+01 1.516e+02 2.001e+02 2.404e+02 4.376e+02, threshold=4.002e+02, percent-clipped=4.0 2023-04-27 23:32:23,669 INFO [finetune.py:976] (4/7) Epoch 26, batch 2000, loss[loss=0.1673, simple_loss=0.2334, pruned_loss=0.05064, over 4808.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.2403, pruned_loss=0.04732, over 956959.29 frames. ], batch size: 25, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:32:29,352 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2795, 1.4808, 1.4662, 1.9835, 2.0917, 1.7485, 1.8695, 1.6201], device='cuda:4'), covar=tensor([0.1964, 0.2248, 0.2093, 0.2050, 0.1714, 0.2663, 0.2535, 0.2596], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0306, 0.0348, 0.0284, 0.0327, 0.0303, 0.0295, 0.0372], device='cuda:4'), out_proj_covar=tensor([6.3642e-05, 6.2916e-05, 7.3204e-05, 5.6878e-05, 6.6997e-05, 6.3288e-05, 6.1116e-05, 7.8976e-05], device='cuda:4') 2023-04-27 23:32:34,219 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5920, 1.8186, 1.9416, 2.0343, 1.9054, 2.0161, 2.0084, 1.9785], device='cuda:4'), covar=tensor([0.3402, 0.5144, 0.4517, 0.4038, 0.5701, 0.6933, 0.4758, 0.4570], device='cuda:4'), in_proj_covar=tensor([0.0343, 0.0378, 0.0332, 0.0342, 0.0353, 0.0397, 0.0363, 0.0335], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 23:33:03,263 INFO [finetune.py:976] (4/7) Epoch 26, batch 2050, loss[loss=0.1521, simple_loss=0.2302, pruned_loss=0.03703, over 4934.00 frames. ], tot_loss[loss=0.1657, simple_loss=0.2381, pruned_loss=0.04664, over 956707.63 frames. ], batch size: 33, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:33:43,224 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.085e+02 1.513e+02 1.827e+02 2.273e+02 3.950e+02, threshold=3.653e+02, percent-clipped=0.0 2023-04-27 23:33:58,463 INFO [finetune.py:976] (4/7) Epoch 26, batch 2100, loss[loss=0.1646, simple_loss=0.2346, pruned_loss=0.04729, over 4858.00 frames. ], tot_loss[loss=0.1667, simple_loss=0.2387, pruned_loss=0.04731, over 954956.33 frames. ], batch size: 44, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:34:26,604 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2886, 1.7238, 2.1997, 2.5770, 2.1808, 1.7139, 1.2950, 1.9283], device='cuda:4'), covar=tensor([0.3355, 0.3359, 0.1682, 0.2185, 0.2495, 0.2802, 0.4257, 0.1995], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0245, 0.0227, 0.0313, 0.0220, 0.0235, 0.0227, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 23:34:32,516 INFO [finetune.py:976] (4/7) Epoch 26, batch 2150, loss[loss=0.1524, simple_loss=0.2253, pruned_loss=0.03969, over 4763.00 frames. ], tot_loss[loss=0.1674, simple_loss=0.2402, pruned_loss=0.04735, over 954400.19 frames. ], batch size: 28, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:34:39,185 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=145352.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 23:34:51,080 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.081e+02 1.597e+02 2.010e+02 2.319e+02 4.178e+02, threshold=4.020e+02, percent-clipped=2.0 2023-04-27 23:34:51,215 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=145371.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:34:59,919 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=145383.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:35:10,688 INFO [finetune.py:976] (4/7) Epoch 26, batch 2200, loss[loss=0.1752, simple_loss=0.248, pruned_loss=0.05118, over 4834.00 frames. ], tot_loss[loss=0.1684, simple_loss=0.2415, pruned_loss=0.0477, over 952413.58 frames. ], batch size: 30, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:35:30,066 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=145422.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:35:32,208 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-04-27 23:35:35,614 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-27 23:35:37,114 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=145432.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:35:42,962 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7486, 2.0865, 1.8286, 2.0549, 1.6303, 1.7501, 1.6232, 1.3655], device='cuda:4'), covar=tensor([0.1622, 0.1264, 0.0811, 0.0959, 0.3137, 0.1127, 0.1692, 0.2174], device='cuda:4'), in_proj_covar=tensor([0.0282, 0.0297, 0.0213, 0.0275, 0.0311, 0.0253, 0.0246, 0.0262], device='cuda:4'), out_proj_covar=tensor([1.1263e-04, 1.1697e-04, 8.3824e-05, 1.0810e-04, 1.2523e-04, 9.9629e-05, 9.9228e-05, 1.0333e-04], device='cuda:4') 2023-04-27 23:35:44,058 INFO [finetune.py:976] (4/7) Epoch 26, batch 2250, loss[loss=0.1939, simple_loss=0.2581, pruned_loss=0.06481, over 4925.00 frames. ], tot_loss[loss=0.169, simple_loss=0.2426, pruned_loss=0.04773, over 953215.32 frames. ], batch size: 33, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:36:02,432 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.54 vs. limit=2.0 2023-04-27 23:36:03,180 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.017e+02 1.640e+02 1.880e+02 2.244e+02 3.153e+02, threshold=3.761e+02, percent-clipped=0.0 2023-04-27 23:36:17,798 INFO [finetune.py:976] (4/7) Epoch 26, batch 2300, loss[loss=0.1977, simple_loss=0.2702, pruned_loss=0.06261, over 4813.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2426, pruned_loss=0.04762, over 952564.18 frames. ], batch size: 38, lr: 2.97e-03, grad_scale: 32.0 2023-04-27 23:36:22,032 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5324, 2.3071, 2.4632, 2.8941, 2.8390, 2.2498, 1.9774, 2.5724], device='cuda:4'), covar=tensor([0.0816, 0.0999, 0.0723, 0.0597, 0.0629, 0.1006, 0.0775, 0.0620], device='cuda:4'), in_proj_covar=tensor([0.0186, 0.0204, 0.0186, 0.0172, 0.0177, 0.0179, 0.0151, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 23:36:43,551 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2946, 1.6086, 1.5092, 1.8429, 1.7836, 2.0281, 1.4814, 3.7527], device='cuda:4'), covar=tensor([0.0615, 0.0820, 0.0799, 0.1195, 0.0631, 0.0494, 0.0720, 0.0136], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0042, 0.0039, 0.0037, 0.0037, 0.0054], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0012, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-27 23:36:44,282 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.69 vs. limit=5.0 2023-04-27 23:36:49,319 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5817, 1.9130, 1.9692, 2.0717, 1.9935, 2.0375, 2.0553, 2.0411], device='cuda:4'), covar=tensor([0.3619, 0.5027, 0.4267, 0.4100, 0.5016, 0.6176, 0.4735, 0.4370], device='cuda:4'), in_proj_covar=tensor([0.0341, 0.0375, 0.0329, 0.0339, 0.0350, 0.0394, 0.0361, 0.0332], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 23:36:51,032 INFO [finetune.py:976] (4/7) Epoch 26, batch 2350, loss[loss=0.1067, simple_loss=0.1817, pruned_loss=0.01585, over 4817.00 frames. ], tot_loss[loss=0.1667, simple_loss=0.2405, pruned_loss=0.04639, over 955860.27 frames. ], batch size: 25, lr: 2.96e-03, grad_scale: 32.0 2023-04-27 23:37:10,065 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.053e+02 1.548e+02 1.895e+02 2.111e+02 3.968e+02, threshold=3.791e+02, percent-clipped=2.0 2023-04-27 23:37:40,489 INFO [finetune.py:976] (4/7) Epoch 26, batch 2400, loss[loss=0.1812, simple_loss=0.2459, pruned_loss=0.05832, over 4832.00 frames. ], tot_loss[loss=0.1647, simple_loss=0.2378, pruned_loss=0.04582, over 957226.47 frames. ], batch size: 39, lr: 2.96e-03, grad_scale: 32.0 2023-04-27 23:37:41,825 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=145594.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:38:04,166 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.57 vs. limit=5.0 2023-04-27 23:38:46,676 INFO [finetune.py:976] (4/7) Epoch 26, batch 2450, loss[loss=0.1594, simple_loss=0.2317, pruned_loss=0.04352, over 4815.00 frames. ], tot_loss[loss=0.1635, simple_loss=0.236, pruned_loss=0.04553, over 958012.50 frames. ], batch size: 38, lr: 2.96e-03, grad_scale: 32.0 2023-04-27 23:38:47,939 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=145643.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:39:00,523 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=145652.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 23:39:08,345 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=145655.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 23:39:30,139 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.068e+02 1.495e+02 1.759e+02 2.001e+02 3.629e+02, threshold=3.517e+02, percent-clipped=0.0 2023-04-27 23:39:42,996 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=145683.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:39:48,404 INFO [finetune.py:976] (4/7) Epoch 26, batch 2500, loss[loss=0.1685, simple_loss=0.2298, pruned_loss=0.05359, over 4909.00 frames. ], tot_loss[loss=0.1659, simple_loss=0.2382, pruned_loss=0.04679, over 957156.37 frames. ], batch size: 32, lr: 2.96e-03, grad_scale: 32.0 2023-04-27 23:39:54,789 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=145700.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:39:57,809 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=145704.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:40:09,663 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=145722.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:40:12,707 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=145727.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:40:15,135 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=145731.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:40:20,269 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=5.19 vs. limit=5.0 2023-04-27 23:40:22,234 INFO [finetune.py:976] (4/7) Epoch 26, batch 2550, loss[loss=0.1621, simple_loss=0.2307, pruned_loss=0.04673, over 4779.00 frames. ], tot_loss[loss=0.1683, simple_loss=0.2415, pruned_loss=0.04754, over 957389.41 frames. ], batch size: 26, lr: 2.96e-03, grad_scale: 32.0 2023-04-27 23:40:41,743 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=145770.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:40:42,276 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.201e+01 1.571e+02 1.844e+02 2.174e+02 6.257e+02, threshold=3.689e+02, percent-clipped=2.0 2023-04-27 23:40:56,108 INFO [finetune.py:976] (4/7) Epoch 26, batch 2600, loss[loss=0.1447, simple_loss=0.2277, pruned_loss=0.03084, over 4856.00 frames. ], tot_loss[loss=0.1698, simple_loss=0.2433, pruned_loss=0.04815, over 957516.06 frames. ], batch size: 44, lr: 2.96e-03, grad_scale: 32.0 2023-04-27 23:41:09,799 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.87 vs. limit=5.0 2023-04-27 23:41:14,780 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.26 vs. limit=2.0 2023-04-27 23:41:16,299 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=145821.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 23:41:26,476 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8141, 2.0019, 0.9468, 1.4693, 2.0023, 1.6094, 1.4784, 1.6335], device='cuda:4'), covar=tensor([0.0470, 0.0351, 0.0326, 0.0527, 0.0240, 0.0487, 0.0500, 0.0535], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0028, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0052, 0.0046, 0.0038, 0.0053, 0.0039, 0.0050, 0.0050, 0.0052], device='cuda:4') 2023-04-27 23:41:29,917 INFO [finetune.py:976] (4/7) Epoch 26, batch 2650, loss[loss=0.1562, simple_loss=0.2273, pruned_loss=0.0425, over 4797.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2447, pruned_loss=0.04873, over 955704.93 frames. ], batch size: 29, lr: 2.96e-03, grad_scale: 32.0 2023-04-27 23:41:49,916 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.145e+02 1.525e+02 1.770e+02 2.195e+02 3.362e+02, threshold=3.540e+02, percent-clipped=0.0 2023-04-27 23:41:57,242 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=145882.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 23:42:03,156 INFO [finetune.py:976] (4/7) Epoch 26, batch 2700, loss[loss=0.1354, simple_loss=0.2025, pruned_loss=0.03414, over 4748.00 frames. ], tot_loss[loss=0.1701, simple_loss=0.2433, pruned_loss=0.04846, over 954156.96 frames. ], batch size: 27, lr: 2.96e-03, grad_scale: 32.0 2023-04-27 23:42:36,792 INFO [finetune.py:976] (4/7) Epoch 26, batch 2750, loss[loss=0.1709, simple_loss=0.2285, pruned_loss=0.05663, over 4824.00 frames. ], tot_loss[loss=0.1683, simple_loss=0.2404, pruned_loss=0.04806, over 953511.68 frames. ], batch size: 30, lr: 2.96e-03, grad_scale: 32.0 2023-04-27 23:42:42,209 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=145950.0, num_to_drop=1, layers_to_drop={2} 2023-04-27 23:42:56,787 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.020e+02 1.527e+02 1.797e+02 2.320e+02 4.826e+02, threshold=3.594e+02, percent-clipped=4.0 2023-04-27 23:43:03,039 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0340, 1.5950, 1.8523, 1.8745, 1.8215, 1.5256, 0.9322, 1.5352], device='cuda:4'), covar=tensor([0.3040, 0.2995, 0.1598, 0.1831, 0.2459, 0.2486, 0.3846, 0.1894], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0245, 0.0228, 0.0313, 0.0221, 0.0235, 0.0227, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-27 23:43:10,063 INFO [finetune.py:976] (4/7) Epoch 26, batch 2800, loss[loss=0.1629, simple_loss=0.2414, pruned_loss=0.04224, over 4906.00 frames. ], tot_loss[loss=0.1656, simple_loss=0.2371, pruned_loss=0.04703, over 954103.82 frames. ], batch size: 46, lr: 2.96e-03, grad_scale: 32.0 2023-04-27 23:43:14,401 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=145999.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:43:32,585 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.38 vs. limit=5.0 2023-04-27 23:43:35,186 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=146027.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:43:39,979 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.91 vs. limit=2.0 2023-04-27 23:43:41,159 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.9239, 3.8032, 2.9818, 4.5106, 3.8619, 3.9786, 1.9157, 3.8395], device='cuda:4'), covar=tensor([0.2021, 0.1325, 0.2976, 0.1904, 0.3181, 0.1849, 0.5879, 0.2669], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0218, 0.0252, 0.0305, 0.0299, 0.0246, 0.0272, 0.0273], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 23:43:44,744 INFO [finetune.py:976] (4/7) Epoch 26, batch 2850, loss[loss=0.2198, simple_loss=0.274, pruned_loss=0.08278, over 4937.00 frames. ], tot_loss[loss=0.1654, simple_loss=0.2362, pruned_loss=0.0473, over 951835.61 frames. ], batch size: 33, lr: 2.96e-03, grad_scale: 32.0 2023-04-27 23:44:22,214 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.964e+01 1.598e+02 1.831e+02 2.098e+02 3.575e+02, threshold=3.662e+02, percent-clipped=0.0 2023-04-27 23:44:25,241 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=146075.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:44:42,809 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1436, 2.6141, 1.0805, 1.5782, 2.1107, 1.2980, 3.6190, 1.8922], device='cuda:4'), covar=tensor([0.0689, 0.0757, 0.0824, 0.1241, 0.0516, 0.0990, 0.0213, 0.0583], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0073, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 23:44:43,959 INFO [finetune.py:976] (4/7) Epoch 26, batch 2900, loss[loss=0.1795, simple_loss=0.258, pruned_loss=0.05049, over 4925.00 frames. ], tot_loss[loss=0.1682, simple_loss=0.2396, pruned_loss=0.04846, over 952746.28 frames. ], batch size: 38, lr: 2.96e-03, grad_scale: 64.0 2023-04-27 23:45:49,352 INFO [finetune.py:976] (4/7) Epoch 26, batch 2950, loss[loss=0.1882, simple_loss=0.2617, pruned_loss=0.05731, over 4826.00 frames. ], tot_loss[loss=0.1711, simple_loss=0.2431, pruned_loss=0.04949, over 952567.52 frames. ], batch size: 30, lr: 2.96e-03, grad_scale: 64.0 2023-04-27 23:45:59,511 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=146149.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:45:59,815 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.21 vs. limit=5.0 2023-04-27 23:46:06,693 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8570, 1.5617, 1.4646, 1.6527, 2.0795, 1.6677, 1.4826, 1.4402], device='cuda:4'), covar=tensor([0.1630, 0.1686, 0.1957, 0.1363, 0.0972, 0.1831, 0.2181, 0.2270], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0307, 0.0348, 0.0283, 0.0325, 0.0301, 0.0296, 0.0371], device='cuda:4'), out_proj_covar=tensor([6.3431e-05, 6.3254e-05, 7.3218e-05, 5.6700e-05, 6.6610e-05, 6.2871e-05, 6.1295e-05, 7.8730e-05], device='cuda:4') 2023-04-27 23:46:13,680 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.156e+02 1.553e+02 1.780e+02 2.193e+02 5.128e+02, threshold=3.559e+02, percent-clipped=3.0 2023-04-27 23:46:18,904 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=146177.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 23:46:28,440 INFO [finetune.py:976] (4/7) Epoch 26, batch 3000, loss[loss=0.1918, simple_loss=0.2672, pruned_loss=0.05822, over 4912.00 frames. ], tot_loss[loss=0.1712, simple_loss=0.244, pruned_loss=0.04921, over 953739.44 frames. ], batch size: 36, lr: 2.96e-03, grad_scale: 64.0 2023-04-27 23:46:28,440 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-27 23:46:31,796 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1448, 2.4693, 1.2519, 1.5027, 2.1304, 1.4113, 2.8556, 1.6810], device='cuda:4'), covar=tensor([0.0514, 0.0556, 0.0643, 0.0989, 0.0312, 0.0728, 0.0197, 0.0492], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0073, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 23:46:38,926 INFO [finetune.py:1010] (4/7) Epoch 26, validation: loss=0.1526, simple_loss=0.2216, pruned_loss=0.04183, over 2265189.00 frames. 2023-04-27 23:46:38,927 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6529MB 2023-04-27 23:46:50,695 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=146210.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:47:11,571 INFO [finetune.py:976] (4/7) Epoch 26, batch 3050, loss[loss=0.1732, simple_loss=0.2481, pruned_loss=0.0492, over 4861.00 frames. ], tot_loss[loss=0.1705, simple_loss=0.2442, pruned_loss=0.0484, over 955612.47 frames. ], batch size: 34, lr: 2.96e-03, grad_scale: 64.0 2023-04-27 23:47:17,397 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=146250.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:47:31,022 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.111e+02 1.541e+02 1.745e+02 2.066e+02 6.833e+02, threshold=3.490e+02, percent-clipped=2.0 2023-04-27 23:47:45,139 INFO [finetune.py:976] (4/7) Epoch 26, batch 3100, loss[loss=0.1832, simple_loss=0.2478, pruned_loss=0.05932, over 4838.00 frames. ], tot_loss[loss=0.1692, simple_loss=0.2424, pruned_loss=0.04802, over 954987.51 frames. ], batch size: 49, lr: 2.96e-03, grad_scale: 32.0 2023-04-27 23:47:49,793 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=146298.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:47:50,446 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=146299.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:48:18,962 INFO [finetune.py:976] (4/7) Epoch 26, batch 3150, loss[loss=0.183, simple_loss=0.2378, pruned_loss=0.06409, over 4844.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.2401, pruned_loss=0.04743, over 955568.57 frames. ], batch size: 25, lr: 2.96e-03, grad_scale: 32.0 2023-04-27 23:48:22,014 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5977, 2.4105, 2.4984, 2.9664, 2.8713, 2.3747, 2.2255, 2.7439], device='cuda:4'), covar=tensor([0.0670, 0.0800, 0.0588, 0.0522, 0.0564, 0.0749, 0.0639, 0.0486], device='cuda:4'), in_proj_covar=tensor([0.0184, 0.0202, 0.0184, 0.0170, 0.0176, 0.0177, 0.0149, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 23:48:22,567 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=146347.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:48:26,673 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1768, 2.6865, 1.0489, 1.3985, 2.1429, 1.2311, 3.5124, 1.6512], device='cuda:4'), covar=tensor([0.0703, 0.0716, 0.0891, 0.1491, 0.0533, 0.1121, 0.0321, 0.0748], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0072, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-27 23:48:38,469 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.123e+02 1.566e+02 1.850e+02 2.179e+02 3.570e+02, threshold=3.699e+02, percent-clipped=2.0 2023-04-27 23:48:45,025 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.3676, 3.2175, 2.5585, 3.8552, 3.3330, 3.3183, 1.2836, 3.2552], device='cuda:4'), covar=tensor([0.1908, 0.1634, 0.3308, 0.2520, 0.3814, 0.2169, 0.6219, 0.2736], device='cuda:4'), in_proj_covar=tensor([0.0245, 0.0220, 0.0253, 0.0306, 0.0300, 0.0248, 0.0275, 0.0275], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 23:48:51,062 INFO [finetune.py:976] (4/7) Epoch 26, batch 3200, loss[loss=0.1984, simple_loss=0.2619, pruned_loss=0.06742, over 4838.00 frames. ], tot_loss[loss=0.1645, simple_loss=0.2369, pruned_loss=0.04609, over 955257.89 frames. ], batch size: 39, lr: 2.96e-03, grad_scale: 32.0 2023-04-27 23:49:04,946 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.45 vs. limit=5.0 2023-04-27 23:49:05,569 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-27 23:49:06,116 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5809, 1.9227, 1.9980, 2.0794, 1.9559, 2.0056, 2.0410, 2.0220], device='cuda:4'), covar=tensor([0.3934, 0.5475, 0.4334, 0.4232, 0.5197, 0.6849, 0.5206, 0.4714], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0375, 0.0328, 0.0340, 0.0351, 0.0394, 0.0360, 0.0332], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 23:49:24,698 INFO [finetune.py:976] (4/7) Epoch 26, batch 3250, loss[loss=0.2, simple_loss=0.276, pruned_loss=0.06194, over 4211.00 frames. ], tot_loss[loss=0.1666, simple_loss=0.2387, pruned_loss=0.04727, over 953323.22 frames. ], batch size: 65, lr: 2.96e-03, grad_scale: 32.0 2023-04-27 23:49:45,355 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.653e+01 1.477e+02 1.784e+02 2.151e+02 4.577e+02, threshold=3.568e+02, percent-clipped=2.0 2023-04-27 23:49:46,070 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5591, 3.0105, 1.0610, 1.8197, 1.7820, 2.3758, 1.8376, 1.2529], device='cuda:4'), covar=tensor([0.1374, 0.1188, 0.2052, 0.1300, 0.1133, 0.1066, 0.1531, 0.2218], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0239, 0.0136, 0.0121, 0.0132, 0.0153, 0.0117, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-27 23:49:53,885 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=146477.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 23:49:56,405 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2135, 2.8092, 2.1695, 2.2591, 1.6332, 1.5684, 2.3780, 1.5817], device='cuda:4'), covar=tensor([0.1589, 0.1344, 0.1345, 0.1605, 0.2172, 0.1825, 0.0893, 0.1997], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0210, 0.0170, 0.0204, 0.0200, 0.0186, 0.0157, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 23:49:56,488 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.35 vs. limit=5.0 2023-04-27 23:50:13,001 INFO [finetune.py:976] (4/7) Epoch 26, batch 3300, loss[loss=0.1862, simple_loss=0.254, pruned_loss=0.05916, over 4199.00 frames. ], tot_loss[loss=0.1702, simple_loss=0.2433, pruned_loss=0.04859, over 952443.81 frames. ], batch size: 65, lr: 2.96e-03, grad_scale: 32.0 2023-04-27 23:50:28,267 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=146505.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:50:57,306 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=146525.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 23:51:19,412 INFO [finetune.py:976] (4/7) Epoch 26, batch 3350, loss[loss=0.156, simple_loss=0.2306, pruned_loss=0.04074, over 4903.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2428, pruned_loss=0.04749, over 952543.49 frames. ], batch size: 37, lr: 2.96e-03, grad_scale: 32.0 2023-04-27 23:51:36,965 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2875, 2.2814, 1.8908, 1.9523, 2.4305, 1.9131, 2.9095, 1.7289], device='cuda:4'), covar=tensor([0.3535, 0.2079, 0.4522, 0.3039, 0.1728, 0.2559, 0.1476, 0.4422], device='cuda:4'), in_proj_covar=tensor([0.0334, 0.0348, 0.0418, 0.0345, 0.0377, 0.0370, 0.0362, 0.0418], device='cuda:4'), out_proj_covar=tensor([9.8776e-05, 1.0357e-04, 1.2666e-04, 1.0348e-04, 1.1191e-04, 1.0990e-04, 1.0590e-04, 1.2565e-04], device='cuda:4') 2023-04-27 23:51:45,687 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.686e+01 1.611e+02 1.865e+02 2.186e+02 4.215e+02, threshold=3.729e+02, percent-clipped=3.0 2023-04-27 23:51:57,913 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=146591.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:51:58,398 INFO [finetune.py:976] (4/7) Epoch 26, batch 3400, loss[loss=0.1064, simple_loss=0.1806, pruned_loss=0.01603, over 4685.00 frames. ], tot_loss[loss=0.1702, simple_loss=0.244, pruned_loss=0.04815, over 953921.95 frames. ], batch size: 23, lr: 2.96e-03, grad_scale: 32.0 2023-04-27 23:52:26,078 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0554, 2.5461, 2.0697, 2.4648, 1.9027, 2.0312, 2.0632, 1.5904], device='cuda:4'), covar=tensor([0.1850, 0.1135, 0.0753, 0.1104, 0.3158, 0.1176, 0.1776, 0.2567], device='cuda:4'), in_proj_covar=tensor([0.0289, 0.0303, 0.0218, 0.0280, 0.0317, 0.0256, 0.0251, 0.0267], device='cuda:4'), out_proj_covar=tensor([1.1530e-04, 1.1926e-04, 8.5730e-05, 1.1026e-04, 1.2767e-04, 1.0095e-04, 1.0138e-04, 1.0529e-04], device='cuda:4') 2023-04-27 23:52:27,936 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=146626.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:52:43,201 INFO [finetune.py:976] (4/7) Epoch 26, batch 3450, loss[loss=0.1783, simple_loss=0.2515, pruned_loss=0.05258, over 4858.00 frames. ], tot_loss[loss=0.1687, simple_loss=0.2426, pruned_loss=0.04747, over 954242.23 frames. ], batch size: 34, lr: 2.96e-03, grad_scale: 32.0 2023-04-27 23:52:47,935 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.9580, 3.9069, 2.8051, 4.5774, 4.0404, 4.0236, 1.6266, 3.9602], device='cuda:4'), covar=tensor([0.1482, 0.1225, 0.3123, 0.1481, 0.3087, 0.1528, 0.5759, 0.2145], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0219, 0.0252, 0.0305, 0.0299, 0.0246, 0.0273, 0.0274], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 23:52:49,812 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=146652.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:53:04,800 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.064e+02 1.587e+02 1.818e+02 2.131e+02 4.084e+02, threshold=3.635e+02, percent-clipped=1.0 2023-04-27 23:53:13,322 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.6394, 3.6718, 2.6169, 4.2537, 3.6948, 3.6760, 1.4951, 3.5242], device='cuda:4'), covar=tensor([0.1902, 0.1421, 0.3251, 0.1979, 0.3029, 0.1940, 0.6155, 0.2494], device='cuda:4'), in_proj_covar=tensor([0.0246, 0.0220, 0.0254, 0.0306, 0.0300, 0.0247, 0.0274, 0.0275], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 23:53:13,977 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=146687.0, num_to_drop=1, layers_to_drop={0} 2023-04-27 23:53:16,861 INFO [finetune.py:976] (4/7) Epoch 26, batch 3500, loss[loss=0.1774, simple_loss=0.2511, pruned_loss=0.05182, over 4803.00 frames. ], tot_loss[loss=0.1659, simple_loss=0.239, pruned_loss=0.04642, over 954402.59 frames. ], batch size: 51, lr: 2.96e-03, grad_scale: 32.0 2023-04-27 23:53:50,656 INFO [finetune.py:976] (4/7) Epoch 26, batch 3550, loss[loss=0.1317, simple_loss=0.2022, pruned_loss=0.0306, over 4305.00 frames. ], tot_loss[loss=0.1646, simple_loss=0.2369, pruned_loss=0.04612, over 955397.12 frames. ], batch size: 19, lr: 2.96e-03, grad_scale: 32.0 2023-04-27 23:54:00,400 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0333, 1.9258, 2.4638, 2.6929, 1.7462, 1.6766, 1.9805, 1.0820], device='cuda:4'), covar=tensor([0.0716, 0.0641, 0.0385, 0.0515, 0.0732, 0.1078, 0.0641, 0.0752], device='cuda:4'), in_proj_covar=tensor([0.0069, 0.0067, 0.0066, 0.0068, 0.0074, 0.0094, 0.0072, 0.0063], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-27 23:54:11,379 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.892e+01 1.463e+02 1.726e+02 2.142e+02 5.887e+02, threshold=3.452e+02, percent-clipped=4.0 2023-04-27 23:54:14,460 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9874, 2.6437, 2.0178, 2.1234, 1.4469, 1.4542, 2.1358, 1.4014], device='cuda:4'), covar=tensor([0.1629, 0.1178, 0.1312, 0.1446, 0.2113, 0.1789, 0.0881, 0.2022], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0210, 0.0169, 0.0203, 0.0199, 0.0186, 0.0157, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-27 23:54:24,572 INFO [finetune.py:976] (4/7) Epoch 26, batch 3600, loss[loss=0.1975, simple_loss=0.2752, pruned_loss=0.05993, over 4804.00 frames. ], tot_loss[loss=0.1631, simple_loss=0.2353, pruned_loss=0.04546, over 956427.84 frames. ], batch size: 45, lr: 2.96e-03, grad_scale: 32.0 2023-04-27 23:54:32,793 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=146805.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:54:56,453 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.71 vs. limit=5.0 2023-04-27 23:54:59,555 INFO [finetune.py:976] (4/7) Epoch 26, batch 3650, loss[loss=0.1468, simple_loss=0.1967, pruned_loss=0.04845, over 3910.00 frames. ], tot_loss[loss=0.1669, simple_loss=0.2393, pruned_loss=0.04729, over 953655.25 frames. ], batch size: 17, lr: 2.96e-03, grad_scale: 32.0 2023-04-27 23:55:00,908 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=146844.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:55:11,630 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=146853.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:55:30,101 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.832e+01 1.619e+02 1.915e+02 2.507e+02 7.747e+02, threshold=3.830e+02, percent-clipped=2.0 2023-04-27 23:55:35,092 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-04-27 23:55:48,370 INFO [finetune.py:976] (4/7) Epoch 26, batch 3700, loss[loss=0.1897, simple_loss=0.2681, pruned_loss=0.05568, over 4849.00 frames. ], tot_loss[loss=0.1703, simple_loss=0.2435, pruned_loss=0.04854, over 954627.50 frames. ], batch size: 49, lr: 2.96e-03, grad_scale: 32.0 2023-04-27 23:56:01,869 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=146905.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:56:36,150 INFO [finetune.py:976] (4/7) Epoch 26, batch 3750, loss[loss=0.1536, simple_loss=0.2306, pruned_loss=0.03831, over 4834.00 frames. ], tot_loss[loss=0.1706, simple_loss=0.2441, pruned_loss=0.04854, over 954426.70 frames. ], batch size: 49, lr: 2.96e-03, grad_scale: 32.0 2023-04-27 23:56:44,472 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=146947.0, num_to_drop=0, layers_to_drop=set() 2023-04-27 23:57:06,090 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.923e+01 1.552e+02 1.944e+02 2.307e+02 4.306e+02, threshold=3.888e+02, percent-clipped=2.0 2023-04-27 23:57:09,743 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6883, 1.2748, 1.7744, 2.1831, 1.7783, 1.6852, 1.7513, 1.6577], device='cuda:4'), covar=tensor([0.4298, 0.6937, 0.6475, 0.5484, 0.5726, 0.7603, 0.7754, 0.9682], device='cuda:4'), in_proj_covar=tensor([0.0440, 0.0422, 0.0516, 0.0508, 0.0470, 0.0505, 0.0507, 0.0520], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-27 23:57:13,806 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=146982.0, num_to_drop=1, layers_to_drop={1} 2023-04-27 23:57:20,818 INFO [finetune.py:976] (4/7) Epoch 26, batch 3800, loss[loss=0.145, simple_loss=0.224, pruned_loss=0.03303, over 4772.00 frames. ], tot_loss[loss=0.1713, simple_loss=0.2448, pruned_loss=0.04894, over 955123.07 frames. ], batch size: 27, lr: 2.95e-03, grad_scale: 32.0 2023-04-27 23:58:05,508 INFO [finetune.py:976] (4/7) Epoch 26, batch 3850, loss[loss=0.1664, simple_loss=0.2404, pruned_loss=0.04616, over 4904.00 frames. ], tot_loss[loss=0.1704, simple_loss=0.2436, pruned_loss=0.04863, over 954377.98 frames. ], batch size: 43, lr: 2.95e-03, grad_scale: 32.0 2023-04-27 23:58:08,840 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.34 vs. limit=2.0 2023-04-27 23:58:23,785 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.013e+02 1.597e+02 1.804e+02 2.116e+02 3.825e+02, threshold=3.609e+02, percent-clipped=0.0 2023-04-27 23:58:39,181 INFO [finetune.py:976] (4/7) Epoch 26, batch 3900, loss[loss=0.1881, simple_loss=0.2526, pruned_loss=0.06181, over 4899.00 frames. ], tot_loss[loss=0.1683, simple_loss=0.2406, pruned_loss=0.04801, over 954999.53 frames. ], batch size: 32, lr: 2.95e-03, grad_scale: 32.0 2023-04-27 23:59:08,966 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.14 vs. limit=2.0 2023-04-27 23:59:12,136 INFO [finetune.py:976] (4/7) Epoch 26, batch 3950, loss[loss=0.1756, simple_loss=0.2466, pruned_loss=0.05225, over 4778.00 frames. ], tot_loss[loss=0.1652, simple_loss=0.2372, pruned_loss=0.04662, over 956347.81 frames. ], batch size: 26, lr: 2.95e-03, grad_scale: 32.0 2023-04-27 23:59:17,540 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9375, 1.2139, 1.6365, 1.7176, 1.7000, 1.7465, 1.6582, 1.6123], device='cuda:4'), covar=tensor([0.3711, 0.4591, 0.3883, 0.3668, 0.5160, 0.6580, 0.4280, 0.4054], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0375, 0.0329, 0.0339, 0.0351, 0.0395, 0.0360, 0.0332], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-27 23:59:31,440 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.029e+02 1.596e+02 1.912e+02 2.259e+02 3.879e+02, threshold=3.825e+02, percent-clipped=1.0 2023-04-27 23:59:36,170 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.99 vs. limit=2.0 2023-04-27 23:59:45,575 INFO [finetune.py:976] (4/7) Epoch 26, batch 4000, loss[loss=0.2087, simple_loss=0.2701, pruned_loss=0.07362, over 4811.00 frames. ], tot_loss[loss=0.1661, simple_loss=0.2377, pruned_loss=0.04727, over 955935.95 frames. ], batch size: 40, lr: 2.95e-03, grad_scale: 32.0 2023-04-27 23:59:51,638 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=147200.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:00:02,100 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.42 vs. limit=2.0 2023-04-28 00:00:18,697 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=147241.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:00:19,195 INFO [finetune.py:976] (4/7) Epoch 26, batch 4050, loss[loss=0.1139, simple_loss=0.1853, pruned_loss=0.02131, over 4766.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.2402, pruned_loss=0.04738, over 954801.38 frames. ], batch size: 23, lr: 2.95e-03, grad_scale: 32.0 2023-04-28 00:00:23,234 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=147247.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:00:34,043 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9847, 4.2909, 0.7261, 2.0638, 2.5251, 2.9779, 2.4612, 1.0216], device='cuda:4'), covar=tensor([0.1334, 0.1051, 0.2231, 0.1428, 0.1015, 0.1016, 0.1545, 0.2229], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0240, 0.0137, 0.0122, 0.0133, 0.0154, 0.0118, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 00:00:44,478 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.189e+02 1.656e+02 1.880e+02 2.300e+02 3.890e+02, threshold=3.760e+02, percent-clipped=1.0 2023-04-28 00:00:55,647 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=147282.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:01:10,421 INFO [finetune.py:976] (4/7) Epoch 26, batch 4100, loss[loss=0.185, simple_loss=0.254, pruned_loss=0.05804, over 4830.00 frames. ], tot_loss[loss=0.1701, simple_loss=0.2434, pruned_loss=0.04836, over 955765.90 frames. ], batch size: 33, lr: 2.95e-03, grad_scale: 32.0 2023-04-28 00:01:12,320 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=147295.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:01:23,295 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=147302.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:01:34,375 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=147310.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:01:54,625 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=147330.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:02:13,695 INFO [finetune.py:976] (4/7) Epoch 26, batch 4150, loss[loss=0.1651, simple_loss=0.2342, pruned_loss=0.04801, over 4760.00 frames. ], tot_loss[loss=0.1703, simple_loss=0.2437, pruned_loss=0.04848, over 955275.70 frames. ], batch size: 28, lr: 2.95e-03, grad_scale: 32.0 2023-04-28 00:02:28,411 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.45 vs. limit=2.0 2023-04-28 00:02:56,067 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=147371.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:02:56,541 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.010e+02 1.642e+02 1.885e+02 2.342e+02 4.093e+02, threshold=3.770e+02, percent-clipped=2.0 2023-04-28 00:03:00,401 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6226, 1.1837, 1.3467, 1.2623, 1.7326, 1.4221, 1.1931, 1.3093], device='cuda:4'), covar=tensor([0.1656, 0.1270, 0.1928, 0.1360, 0.0836, 0.1522, 0.1735, 0.2208], device='cuda:4'), in_proj_covar=tensor([0.0315, 0.0310, 0.0353, 0.0287, 0.0327, 0.0305, 0.0298, 0.0375], device='cuda:4'), out_proj_covar=tensor([6.4441e-05, 6.3801e-05, 7.4137e-05, 5.7395e-05, 6.7015e-05, 6.3690e-05, 6.1754e-05, 7.9367e-05], device='cuda:4') 2023-04-28 00:03:09,328 INFO [finetune.py:976] (4/7) Epoch 26, batch 4200, loss[loss=0.161, simple_loss=0.2319, pruned_loss=0.04502, over 4842.00 frames. ], tot_loss[loss=0.1702, simple_loss=0.244, pruned_loss=0.04815, over 956332.37 frames. ], batch size: 44, lr: 2.95e-03, grad_scale: 32.0 2023-04-28 00:03:21,116 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=147407.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:03:37,866 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=147434.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:03:42,686 INFO [finetune.py:976] (4/7) Epoch 26, batch 4250, loss[loss=0.2035, simple_loss=0.2713, pruned_loss=0.06781, over 4812.00 frames. ], tot_loss[loss=0.1682, simple_loss=0.2416, pruned_loss=0.04738, over 956809.53 frames. ], batch size: 41, lr: 2.95e-03, grad_scale: 32.0 2023-04-28 00:03:53,622 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2636, 2.7103, 2.4829, 2.2713, 1.6907, 1.6878, 2.7602, 1.7521], device='cuda:4'), covar=tensor([0.1442, 0.1456, 0.1063, 0.1409, 0.1931, 0.1679, 0.0647, 0.1767], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0211, 0.0170, 0.0204, 0.0201, 0.0187, 0.0157, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 00:04:01,790 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=147468.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:04:04,153 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.178e+01 1.415e+02 1.781e+02 2.171e+02 8.734e+02, threshold=3.562e+02, percent-clipped=6.0 2023-04-28 00:04:16,307 INFO [finetune.py:976] (4/7) Epoch 26, batch 4300, loss[loss=0.139, simple_loss=0.214, pruned_loss=0.03202, over 4760.00 frames. ], tot_loss[loss=0.1657, simple_loss=0.2387, pruned_loss=0.04635, over 958469.53 frames. ], batch size: 28, lr: 2.95e-03, grad_scale: 32.0 2023-04-28 00:04:18,224 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8130, 3.9470, 0.6446, 2.0741, 2.2177, 2.6240, 2.2886, 1.0801], device='cuda:4'), covar=tensor([0.1208, 0.0821, 0.2090, 0.1229, 0.1013, 0.1044, 0.1343, 0.2200], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0237, 0.0135, 0.0121, 0.0131, 0.0152, 0.0117, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 00:04:18,270 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=147495.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:04:18,845 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8881, 3.7814, 0.9089, 1.9027, 2.2199, 2.5945, 1.9948, 1.1250], device='cuda:4'), covar=tensor([0.1229, 0.0855, 0.2137, 0.1296, 0.1019, 0.1047, 0.1653, 0.2038], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0237, 0.0135, 0.0121, 0.0131, 0.0152, 0.0117, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 00:04:22,181 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=147500.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:04:48,271 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-28 00:04:49,577 INFO [finetune.py:976] (4/7) Epoch 26, batch 4350, loss[loss=0.1539, simple_loss=0.226, pruned_loss=0.04092, over 4777.00 frames. ], tot_loss[loss=0.1633, simple_loss=0.2358, pruned_loss=0.04539, over 958874.60 frames. ], batch size: 26, lr: 2.95e-03, grad_scale: 32.0 2023-04-28 00:04:52,797 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0818, 2.4878, 2.1768, 2.4155, 1.7452, 2.1161, 2.0271, 1.6498], device='cuda:4'), covar=tensor([0.1832, 0.1063, 0.0728, 0.1026, 0.3348, 0.1029, 0.2003, 0.2387], device='cuda:4'), in_proj_covar=tensor([0.0285, 0.0300, 0.0216, 0.0278, 0.0315, 0.0254, 0.0250, 0.0264], device='cuda:4'), out_proj_covar=tensor([1.1362e-04, 1.1796e-04, 8.5192e-05, 1.0927e-04, 1.2689e-04, 1.0002e-04, 1.0084e-04, 1.0400e-04], device='cuda:4') 2023-04-28 00:04:53,339 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=147548.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:04:54,252 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.12 vs. limit=2.0 2023-04-28 00:04:59,700 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3629, 1.6116, 1.4932, 1.8411, 1.6939, 1.9667, 1.4441, 3.6018], device='cuda:4'), covar=tensor([0.0553, 0.0762, 0.0759, 0.1158, 0.0604, 0.0442, 0.0708, 0.0135], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0039, 0.0043, 0.0040, 0.0037, 0.0038, 0.0054], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 00:05:10,812 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.116e+01 1.521e+02 1.686e+02 2.044e+02 4.229e+02, threshold=3.372e+02, percent-clipped=1.0 2023-04-28 00:05:18,800 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.2901, 3.3463, 2.5604, 3.8177, 3.3098, 3.2682, 1.4226, 3.2551], device='cuda:4'), covar=tensor([0.2022, 0.1359, 0.3527, 0.2651, 0.4138, 0.2169, 0.5774, 0.2684], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0219, 0.0251, 0.0303, 0.0298, 0.0246, 0.0272, 0.0272], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 00:05:23,019 INFO [finetune.py:976] (4/7) Epoch 26, batch 4400, loss[loss=0.1993, simple_loss=0.2746, pruned_loss=0.06205, over 4871.00 frames. ], tot_loss[loss=0.1645, simple_loss=0.2369, pruned_loss=0.04606, over 957851.41 frames. ], batch size: 34, lr: 2.95e-03, grad_scale: 32.0 2023-04-28 00:05:26,139 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=147597.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:05:36,482 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5750, 3.4660, 0.8676, 1.6491, 1.9461, 2.4485, 1.9233, 1.0933], device='cuda:4'), covar=tensor([0.1406, 0.0855, 0.2071, 0.1353, 0.1126, 0.0991, 0.1661, 0.1969], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0239, 0.0136, 0.0121, 0.0132, 0.0153, 0.0118, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 00:05:56,085 INFO [finetune.py:976] (4/7) Epoch 26, batch 4450, loss[loss=0.1768, simple_loss=0.2472, pruned_loss=0.05315, over 4833.00 frames. ], tot_loss[loss=0.1664, simple_loss=0.2399, pruned_loss=0.04646, over 957546.00 frames. ], batch size: 33, lr: 2.95e-03, grad_scale: 32.0 2023-04-28 00:06:17,560 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=147659.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:06:22,788 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=147666.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:06:31,479 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.062e+02 1.516e+02 1.781e+02 2.208e+02 3.417e+02, threshold=3.563e+02, percent-clipped=1.0 2023-04-28 00:06:35,705 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1460, 1.5054, 1.3815, 1.7076, 1.5571, 1.7413, 1.3672, 3.5318], device='cuda:4'), covar=tensor([0.0668, 0.0898, 0.0879, 0.1305, 0.0709, 0.0551, 0.0827, 0.0152], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0042, 0.0040, 0.0037, 0.0038, 0.0054], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0013, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 00:06:44,121 INFO [finetune.py:976] (4/7) Epoch 26, batch 4500, loss[loss=0.189, simple_loss=0.2595, pruned_loss=0.05919, over 4887.00 frames. ], tot_loss[loss=0.1681, simple_loss=0.2416, pruned_loss=0.04726, over 956405.53 frames. ], batch size: 32, lr: 2.95e-03, grad_scale: 32.0 2023-04-28 00:06:54,512 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-28 00:07:17,732 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=147720.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:07:37,058 INFO [finetune.py:976] (4/7) Epoch 26, batch 4550, loss[loss=0.1764, simple_loss=0.2468, pruned_loss=0.05296, over 4795.00 frames. ], tot_loss[loss=0.1696, simple_loss=0.2433, pruned_loss=0.04794, over 955390.80 frames. ], batch size: 25, lr: 2.95e-03, grad_scale: 32.0 2023-04-28 00:08:01,007 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=147763.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:08:14,011 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.021e+02 1.503e+02 1.815e+02 2.151e+02 5.912e+02, threshold=3.631e+02, percent-clipped=1.0 2023-04-28 00:08:29,100 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8477, 1.4418, 1.8901, 2.3413, 1.9303, 1.8403, 1.8793, 1.8460], device='cuda:4'), covar=tensor([0.4515, 0.6749, 0.6713, 0.5560, 0.6098, 0.7704, 0.8798, 0.8923], device='cuda:4'), in_proj_covar=tensor([0.0444, 0.0424, 0.0518, 0.0509, 0.0471, 0.0507, 0.0509, 0.0521], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 00:08:31,442 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=147790.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:08:32,585 INFO [finetune.py:976] (4/7) Epoch 26, batch 4600, loss[loss=0.1447, simple_loss=0.2131, pruned_loss=0.03815, over 4786.00 frames. ], tot_loss[loss=0.17, simple_loss=0.2434, pruned_loss=0.04834, over 957240.61 frames. ], batch size: 29, lr: 2.95e-03, grad_scale: 32.0 2023-04-28 00:08:46,880 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-28 00:08:47,352 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=147815.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:09:05,602 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3320, 1.6964, 1.5644, 1.9572, 1.7948, 2.0746, 1.5314, 3.7716], device='cuda:4'), covar=tensor([0.0587, 0.0788, 0.0800, 0.1095, 0.0621, 0.0446, 0.0695, 0.0138], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0039, 0.0043, 0.0040, 0.0037, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 00:09:06,110 INFO [finetune.py:976] (4/7) Epoch 26, batch 4650, loss[loss=0.1368, simple_loss=0.2112, pruned_loss=0.03119, over 4750.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.2399, pruned_loss=0.04689, over 956434.07 frames. ], batch size: 28, lr: 2.95e-03, grad_scale: 32.0 2023-04-28 00:09:08,112 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3267, 1.7249, 2.2036, 2.5890, 2.1614, 1.7007, 1.4126, 1.8708], device='cuda:4'), covar=tensor([0.2698, 0.2943, 0.1449, 0.1880, 0.2373, 0.2529, 0.3823, 0.1857], device='cuda:4'), in_proj_covar=tensor([0.0294, 0.0246, 0.0228, 0.0314, 0.0222, 0.0235, 0.0228, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-28 00:09:11,693 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1984, 4.4649, 0.9036, 2.2652, 2.8310, 2.9044, 2.6096, 1.1875], device='cuda:4'), covar=tensor([0.1275, 0.1177, 0.2165, 0.1306, 0.0895, 0.1121, 0.1465, 0.1995], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0237, 0.0135, 0.0120, 0.0131, 0.0152, 0.0117, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 00:09:15,424 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=147857.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:09:24,992 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2264, 2.0507, 2.3629, 2.6785, 2.7264, 2.1964, 1.8392, 2.4432], device='cuda:4'), covar=tensor([0.0807, 0.1089, 0.0639, 0.0598, 0.0623, 0.0807, 0.0874, 0.0614], device='cuda:4'), in_proj_covar=tensor([0.0185, 0.0202, 0.0184, 0.0171, 0.0177, 0.0178, 0.0151, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 00:09:25,966 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 7.854e+01 1.470e+02 1.652e+02 1.982e+02 3.804e+02, threshold=3.304e+02, percent-clipped=1.0 2023-04-28 00:09:29,047 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=147876.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:09:40,051 INFO [finetune.py:976] (4/7) Epoch 26, batch 4700, loss[loss=0.1601, simple_loss=0.2241, pruned_loss=0.04806, over 4744.00 frames. ], tot_loss[loss=0.1639, simple_loss=0.2362, pruned_loss=0.0458, over 956593.96 frames. ], batch size: 59, lr: 2.95e-03, grad_scale: 32.0 2023-04-28 00:09:43,179 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=147897.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:09:51,641 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=147911.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:09:55,899 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=147918.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:10:12,815 INFO [finetune.py:976] (4/7) Epoch 26, batch 4750, loss[loss=0.2059, simple_loss=0.2724, pruned_loss=0.06974, over 4912.00 frames. ], tot_loss[loss=0.1627, simple_loss=0.2345, pruned_loss=0.04544, over 956811.91 frames. ], batch size: 36, lr: 2.95e-03, grad_scale: 32.0 2023-04-28 00:10:15,201 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=147945.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:10:28,067 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=147966.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:10:31,603 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.017e+02 1.547e+02 1.796e+02 2.219e+02 3.890e+02, threshold=3.592e+02, percent-clipped=1.0 2023-04-28 00:10:31,745 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=147972.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:10:45,443 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7634, 1.5785, 4.6473, 4.3935, 3.9899, 4.4634, 4.1965, 4.0787], device='cuda:4'), covar=tensor([0.6714, 0.5462, 0.0922, 0.1399, 0.1095, 0.1790, 0.1511, 0.1491], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0308, 0.0407, 0.0409, 0.0347, 0.0413, 0.0318, 0.0364], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-28 00:10:46,578 INFO [finetune.py:976] (4/7) Epoch 26, batch 4800, loss[loss=0.1468, simple_loss=0.2354, pruned_loss=0.02909, over 4768.00 frames. ], tot_loss[loss=0.1652, simple_loss=0.238, pruned_loss=0.04621, over 954724.31 frames. ], batch size: 28, lr: 2.95e-03, grad_scale: 32.0 2023-04-28 00:10:51,470 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5136, 1.0721, 1.2481, 1.2200, 1.6502, 1.3057, 1.1546, 1.1897], device='cuda:4'), covar=tensor([0.1709, 0.1535, 0.1954, 0.1491, 0.0840, 0.1749, 0.2051, 0.2405], device='cuda:4'), in_proj_covar=tensor([0.0315, 0.0310, 0.0353, 0.0287, 0.0328, 0.0305, 0.0299, 0.0376], device='cuda:4'), out_proj_covar=tensor([6.4306e-05, 6.3714e-05, 7.4087e-05, 5.7488e-05, 6.7026e-05, 6.3779e-05, 6.1836e-05, 7.9760e-05], device='cuda:4') 2023-04-28 00:11:00,207 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2070, 1.4890, 1.9967, 2.5971, 2.1111, 1.5959, 1.4541, 1.8079], device='cuda:4'), covar=tensor([0.3621, 0.3898, 0.1963, 0.2487, 0.2947, 0.2970, 0.4287, 0.2251], device='cuda:4'), in_proj_covar=tensor([0.0295, 0.0247, 0.0229, 0.0315, 0.0223, 0.0236, 0.0229, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-28 00:11:01,953 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=148014.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:11:02,584 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=148015.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:11:21,744 INFO [finetune.py:976] (4/7) Epoch 26, batch 4850, loss[loss=0.2123, simple_loss=0.2817, pruned_loss=0.07151, over 4839.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2427, pruned_loss=0.04743, over 955687.40 frames. ], batch size: 49, lr: 2.95e-03, grad_scale: 32.0 2023-04-28 00:11:22,496 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7814, 1.7207, 2.1447, 2.2063, 1.5484, 1.4891, 1.7509, 0.8785], device='cuda:4'), covar=tensor([0.0561, 0.0679, 0.0465, 0.0722, 0.0694, 0.0974, 0.0665, 0.0729], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0067, 0.0066, 0.0069, 0.0074, 0.0094, 0.0072, 0.0063], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-28 00:11:24,854 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-28 00:11:32,722 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1435, 1.8847, 2.0880, 2.3856, 2.5976, 1.9880, 1.8587, 2.2273], device='cuda:4'), covar=tensor([0.0803, 0.1190, 0.0708, 0.0596, 0.0562, 0.0874, 0.0802, 0.0596], device='cuda:4'), in_proj_covar=tensor([0.0185, 0.0202, 0.0184, 0.0171, 0.0177, 0.0177, 0.0152, 0.0178], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 00:11:44,624 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=148063.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:11:55,271 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.111e+02 1.622e+02 1.940e+02 2.340e+02 4.108e+02, threshold=3.881e+02, percent-clipped=3.0 2023-04-28 00:12:18,046 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=148090.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:12:24,897 INFO [finetune.py:976] (4/7) Epoch 26, batch 4900, loss[loss=0.1459, simple_loss=0.2273, pruned_loss=0.03227, over 4796.00 frames. ], tot_loss[loss=0.1692, simple_loss=0.2433, pruned_loss=0.04759, over 954127.52 frames. ], batch size: 29, lr: 2.95e-03, grad_scale: 32.0 2023-04-28 00:12:37,090 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5861, 1.4595, 4.2246, 3.9529, 3.7149, 4.0361, 4.0470, 3.7461], device='cuda:4'), covar=tensor([0.6679, 0.5502, 0.0979, 0.1637, 0.1228, 0.1691, 0.1094, 0.1349], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0309, 0.0407, 0.0409, 0.0347, 0.0412, 0.0318, 0.0364], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-28 00:12:47,901 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=148111.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:13:22,193 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=148138.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:13:23,454 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6471, 1.7729, 0.7411, 1.3276, 1.7154, 1.4918, 1.4189, 1.4621], device='cuda:4'), covar=tensor([0.0519, 0.0356, 0.0356, 0.0558, 0.0277, 0.0540, 0.0529, 0.0602], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0028, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0052, 0.0046, 0.0038, 0.0053, 0.0039, 0.0051, 0.0050, 0.0052], device='cuda:4') 2023-04-28 00:13:30,217 INFO [finetune.py:976] (4/7) Epoch 26, batch 4950, loss[loss=0.1671, simple_loss=0.2369, pruned_loss=0.04871, over 4881.00 frames. ], tot_loss[loss=0.1687, simple_loss=0.2432, pruned_loss=0.04706, over 954449.31 frames. ], batch size: 32, lr: 2.95e-03, grad_scale: 32.0 2023-04-28 00:13:52,206 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=148171.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:13:52,740 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.839e+01 1.452e+02 1.801e+02 2.197e+02 4.892e+02, threshold=3.601e+02, percent-clipped=1.0 2023-04-28 00:13:53,434 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6080, 1.3086, 4.3792, 4.0659, 3.7951, 4.2012, 4.0753, 3.8333], device='cuda:4'), covar=tensor([0.7504, 0.6209, 0.1128, 0.1933, 0.1213, 0.1935, 0.1733, 0.1702], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0310, 0.0408, 0.0410, 0.0347, 0.0413, 0.0319, 0.0365], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-28 00:14:06,561 INFO [finetune.py:976] (4/7) Epoch 26, batch 5000, loss[loss=0.1582, simple_loss=0.2216, pruned_loss=0.04739, over 4905.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.2415, pruned_loss=0.04695, over 954674.03 frames. ], batch size: 32, lr: 2.95e-03, grad_scale: 32.0 2023-04-28 00:14:21,042 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=148213.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:14:39,766 INFO [finetune.py:976] (4/7) Epoch 26, batch 5050, loss[loss=0.1314, simple_loss=0.1939, pruned_loss=0.03448, over 4098.00 frames. ], tot_loss[loss=0.1661, simple_loss=0.2392, pruned_loss=0.0465, over 955139.24 frames. ], batch size: 17, lr: 2.95e-03, grad_scale: 32.0 2023-04-28 00:14:56,597 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=148267.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:14:59,524 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.057e+01 1.422e+02 1.750e+02 2.059e+02 5.482e+02, threshold=3.501e+02, percent-clipped=2.0 2023-04-28 00:15:08,809 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=148287.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:15:11,733 INFO [finetune.py:976] (4/7) Epoch 26, batch 5100, loss[loss=0.1137, simple_loss=0.1826, pruned_loss=0.02246, over 4719.00 frames. ], tot_loss[loss=0.1645, simple_loss=0.2371, pruned_loss=0.04597, over 952872.51 frames. ], batch size: 23, lr: 2.95e-03, grad_scale: 64.0 2023-04-28 00:15:26,473 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5346, 1.8287, 1.9363, 2.0173, 1.9460, 2.0095, 2.0332, 1.9794], device='cuda:4'), covar=tensor([0.3771, 0.5453, 0.4297, 0.4233, 0.5127, 0.6720, 0.4800, 0.4514], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0375, 0.0330, 0.0340, 0.0351, 0.0395, 0.0361, 0.0333], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 00:15:28,681 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=148315.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:15:45,039 INFO [finetune.py:976] (4/7) Epoch 26, batch 5150, loss[loss=0.1434, simple_loss=0.2157, pruned_loss=0.03557, over 4709.00 frames. ], tot_loss[loss=0.1644, simple_loss=0.2366, pruned_loss=0.04608, over 952137.03 frames. ], batch size: 23, lr: 2.95e-03, grad_scale: 64.0 2023-04-28 00:15:49,310 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=148348.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:16:00,255 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=148363.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:16:06,188 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.267e+01 1.646e+02 1.916e+02 2.227e+02 4.429e+02, threshold=3.832e+02, percent-clipped=2.0 2023-04-28 00:16:18,411 INFO [finetune.py:976] (4/7) Epoch 26, batch 5200, loss[loss=0.2229, simple_loss=0.2926, pruned_loss=0.07659, over 4741.00 frames. ], tot_loss[loss=0.1686, simple_loss=0.2413, pruned_loss=0.04792, over 951838.05 frames. ], batch size: 54, lr: 2.95e-03, grad_scale: 64.0 2023-04-28 00:16:38,166 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.16 vs. limit=5.0 2023-04-28 00:16:44,049 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=148429.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:16:44,079 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.9324, 2.2023, 2.1625, 2.2941, 2.0994, 2.2184, 2.2517, 2.1300], device='cuda:4'), covar=tensor([0.4159, 0.6245, 0.4723, 0.4251, 0.5841, 0.7142, 0.5727, 0.5804], device='cuda:4'), in_proj_covar=tensor([0.0338, 0.0374, 0.0329, 0.0339, 0.0350, 0.0394, 0.0360, 0.0332], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 00:16:51,852 INFO [finetune.py:976] (4/7) Epoch 26, batch 5250, loss[loss=0.2354, simple_loss=0.3036, pruned_loss=0.08358, over 4190.00 frames. ], tot_loss[loss=0.1691, simple_loss=0.2421, pruned_loss=0.04804, over 951756.75 frames. ], batch size: 66, lr: 2.94e-03, grad_scale: 64.0 2023-04-28 00:17:12,094 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=148471.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:17:12,617 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.069e+02 1.667e+02 1.923e+02 2.369e+02 3.667e+02, threshold=3.847e+02, percent-clipped=0.0 2023-04-28 00:17:23,758 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=148490.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:17:24,853 INFO [finetune.py:976] (4/7) Epoch 26, batch 5300, loss[loss=0.1885, simple_loss=0.2558, pruned_loss=0.06054, over 4699.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2428, pruned_loss=0.04752, over 952957.90 frames. ], batch size: 59, lr: 2.94e-03, grad_scale: 64.0 2023-04-28 00:17:28,536 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=148498.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:17:52,366 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=148513.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:17:55,920 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-28 00:17:56,056 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=148519.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:18:24,822 INFO [finetune.py:976] (4/7) Epoch 26, batch 5350, loss[loss=0.1534, simple_loss=0.2201, pruned_loss=0.04337, over 4845.00 frames. ], tot_loss[loss=0.169, simple_loss=0.2431, pruned_loss=0.04741, over 954303.29 frames. ], batch size: 49, lr: 2.94e-03, grad_scale: 64.0 2023-04-28 00:18:45,987 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=148559.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:18:47,667 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=148561.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:18:55,819 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-28 00:18:56,338 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=148567.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:18:59,258 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.079e+02 1.519e+02 1.727e+02 2.044e+02 3.466e+02, threshold=3.453e+02, percent-clipped=0.0 2023-04-28 00:19:22,274 INFO [finetune.py:976] (4/7) Epoch 26, batch 5400, loss[loss=0.1418, simple_loss=0.2034, pruned_loss=0.04008, over 4811.00 frames. ], tot_loss[loss=0.1672, simple_loss=0.2406, pruned_loss=0.0469, over 953784.48 frames. ], batch size: 25, lr: 2.94e-03, grad_scale: 64.0 2023-04-28 00:19:51,579 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-28 00:19:54,123 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=148615.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:20:04,245 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.72 vs. limit=2.0 2023-04-28 00:20:11,714 INFO [finetune.py:976] (4/7) Epoch 26, batch 5450, loss[loss=0.1541, simple_loss=0.2357, pruned_loss=0.03622, over 4855.00 frames. ], tot_loss[loss=0.1658, simple_loss=0.2384, pruned_loss=0.04665, over 955094.99 frames. ], batch size: 44, lr: 2.94e-03, grad_scale: 64.0 2023-04-28 00:20:12,389 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=148643.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:20:17,269 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=148651.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:20:31,290 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.056e+02 1.574e+02 1.926e+02 2.349e+02 4.613e+02, threshold=3.853e+02, percent-clipped=6.0 2023-04-28 00:20:45,353 INFO [finetune.py:976] (4/7) Epoch 26, batch 5500, loss[loss=0.1371, simple_loss=0.2217, pruned_loss=0.02626, over 4738.00 frames. ], tot_loss[loss=0.165, simple_loss=0.2367, pruned_loss=0.04666, over 955624.48 frames. ], batch size: 54, lr: 2.94e-03, grad_scale: 64.0 2023-04-28 00:20:57,704 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=148712.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:21:00,116 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=148716.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:21:06,349 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.90 vs. limit=2.0 2023-04-28 00:21:18,757 INFO [finetune.py:976] (4/7) Epoch 26, batch 5550, loss[loss=0.163, simple_loss=0.2429, pruned_loss=0.04157, over 4867.00 frames. ], tot_loss[loss=0.1656, simple_loss=0.2377, pruned_loss=0.04675, over 954690.79 frames. ], batch size: 31, lr: 2.94e-03, grad_scale: 64.0 2023-04-28 00:21:38,041 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.670e+01 1.487e+02 1.767e+02 2.141e+02 3.989e+02, threshold=3.535e+02, percent-clipped=1.0 2023-04-28 00:21:41,039 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=148777.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:21:42,149 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.6159, 4.6097, 3.0392, 5.2007, 4.6089, 4.5620, 1.8143, 4.4330], device='cuda:4'), covar=tensor([0.1380, 0.0905, 0.3208, 0.1007, 0.3423, 0.1525, 0.5834, 0.2242], device='cuda:4'), in_proj_covar=tensor([0.0247, 0.0223, 0.0255, 0.0306, 0.0302, 0.0250, 0.0277, 0.0278], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 00:21:45,568 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=148785.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:21:49,252 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-28 00:21:49,595 INFO [finetune.py:976] (4/7) Epoch 26, batch 5600, loss[loss=0.1971, simple_loss=0.2714, pruned_loss=0.06141, over 4917.00 frames. ], tot_loss[loss=0.1681, simple_loss=0.2412, pruned_loss=0.04751, over 956198.15 frames. ], batch size: 36, lr: 2.94e-03, grad_scale: 64.0 2023-04-28 00:22:20,665 INFO [finetune.py:976] (4/7) Epoch 26, batch 5650, loss[loss=0.1417, simple_loss=0.208, pruned_loss=0.03766, over 4670.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.2428, pruned_loss=0.04811, over 955404.95 frames. ], batch size: 23, lr: 2.94e-03, grad_scale: 32.0 2023-04-28 00:22:27,776 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=148854.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:22:39,028 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.125e+02 1.517e+02 1.815e+02 2.254e+02 4.751e+02, threshold=3.630e+02, percent-clipped=7.0 2023-04-28 00:22:50,263 INFO [finetune.py:976] (4/7) Epoch 26, batch 5700, loss[loss=0.1932, simple_loss=0.2385, pruned_loss=0.07392, over 4268.00 frames. ], tot_loss[loss=0.1663, simple_loss=0.2384, pruned_loss=0.04713, over 935447.88 frames. ], batch size: 18, lr: 2.94e-03, grad_scale: 32.0 2023-04-28 00:22:55,760 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.8569, 2.1170, 1.9985, 2.8069, 2.8180, 2.3960, 2.5037, 1.9662], device='cuda:4'), covar=tensor([0.1813, 0.1837, 0.1856, 0.1492, 0.1168, 0.1814, 0.2046, 0.2576], device='cuda:4'), in_proj_covar=tensor([0.0317, 0.0311, 0.0354, 0.0289, 0.0329, 0.0308, 0.0302, 0.0377], device='cuda:4'), out_proj_covar=tensor([6.4791e-05, 6.3888e-05, 7.4393e-05, 5.7959e-05, 6.7213e-05, 6.4240e-05, 6.2484e-05, 7.9881e-05], device='cuda:4') 2023-04-28 00:23:20,151 INFO [finetune.py:976] (4/7) Epoch 27, batch 0, loss[loss=0.1781, simple_loss=0.2625, pruned_loss=0.04684, over 4888.00 frames. ], tot_loss[loss=0.1781, simple_loss=0.2625, pruned_loss=0.04684, over 4888.00 frames. ], batch size: 46, lr: 2.94e-03, grad_scale: 32.0 2023-04-28 00:23:20,151 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-28 00:23:31,756 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1490, 2.5058, 1.0702, 1.4701, 2.0708, 1.2863, 3.0381, 1.7496], device='cuda:4'), covar=tensor([0.0642, 0.0617, 0.0674, 0.1156, 0.0374, 0.0892, 0.0212, 0.0542], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0064, 0.0047, 0.0046, 0.0050, 0.0051, 0.0073, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-28 00:23:32,229 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4616, 1.3726, 1.6541, 1.6997, 1.3772, 1.2938, 1.4671, 0.8791], device='cuda:4'), covar=tensor([0.0494, 0.0666, 0.0515, 0.0412, 0.0677, 0.1086, 0.0501, 0.0608], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0067, 0.0066, 0.0068, 0.0074, 0.0094, 0.0072, 0.0063], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-28 00:23:41,724 INFO [finetune.py:1010] (4/7) Epoch 27, validation: loss=0.1548, simple_loss=0.2237, pruned_loss=0.04298, over 2265189.00 frames. 2023-04-28 00:23:41,724 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6529MB 2023-04-28 00:24:11,139 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=148943.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:24:41,615 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-28 00:24:43,145 INFO [finetune.py:976] (4/7) Epoch 27, batch 50, loss[loss=0.2074, simple_loss=0.2879, pruned_loss=0.06346, over 4811.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2432, pruned_loss=0.04728, over 216128.59 frames. ], batch size: 38, lr: 2.94e-03, grad_scale: 32.0 2023-04-28 00:24:45,491 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.880e+01 1.482e+02 1.782e+02 2.132e+02 4.593e+02, threshold=3.564e+02, percent-clipped=2.0 2023-04-28 00:24:53,727 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5974, 1.2211, 4.3021, 4.0717, 3.6961, 4.0513, 3.9696, 3.7522], device='cuda:4'), covar=tensor([0.7385, 0.5920, 0.1080, 0.1625, 0.1277, 0.1392, 0.1831, 0.1604], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0310, 0.0410, 0.0411, 0.0349, 0.0415, 0.0320, 0.0367], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-28 00:25:07,487 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=148991.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:25:28,985 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=149007.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:25:37,831 INFO [finetune.py:976] (4/7) Epoch 27, batch 100, loss[loss=0.1453, simple_loss=0.2115, pruned_loss=0.03952, over 4700.00 frames. ], tot_loss[loss=0.1639, simple_loss=0.2363, pruned_loss=0.04579, over 379742.40 frames. ], batch size: 23, lr: 2.94e-03, grad_scale: 32.0 2023-04-28 00:25:48,516 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8332, 1.6842, 1.9428, 2.1933, 2.1959, 1.7556, 1.6069, 2.1239], device='cuda:4'), covar=tensor([0.0768, 0.1050, 0.0687, 0.0558, 0.0599, 0.0786, 0.0690, 0.0497], device='cuda:4'), in_proj_covar=tensor([0.0184, 0.0200, 0.0182, 0.0169, 0.0176, 0.0175, 0.0150, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 00:25:58,869 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.4487, 4.3503, 3.0346, 5.0934, 4.3979, 4.3529, 1.8012, 4.3453], device='cuda:4'), covar=tensor([0.1522, 0.1051, 0.3549, 0.0993, 0.2949, 0.1791, 0.5929, 0.2027], device='cuda:4'), in_proj_covar=tensor([0.0246, 0.0222, 0.0255, 0.0306, 0.0301, 0.0250, 0.0276, 0.0276], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 00:26:03,059 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8761, 2.3482, 1.9359, 1.7606, 1.3870, 1.4357, 2.0435, 1.3663], device='cuda:4'), covar=tensor([0.1736, 0.1415, 0.1460, 0.1848, 0.2407, 0.1954, 0.1001, 0.2167], device='cuda:4'), in_proj_covar=tensor([0.0196, 0.0209, 0.0169, 0.0203, 0.0199, 0.0185, 0.0156, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 00:26:11,901 INFO [finetune.py:976] (4/7) Epoch 27, batch 150, loss[loss=0.1265, simple_loss=0.2025, pruned_loss=0.02523, over 4862.00 frames. ], tot_loss[loss=0.158, simple_loss=0.2304, pruned_loss=0.04279, over 508136.08 frames. ], batch size: 49, lr: 2.94e-03, grad_scale: 32.0 2023-04-28 00:26:13,183 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=149072.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:26:13,727 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.057e+02 1.439e+02 1.708e+02 1.990e+02 3.271e+02, threshold=3.417e+02, percent-clipped=0.0 2023-04-28 00:26:22,669 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=149085.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:26:45,182 INFO [finetune.py:976] (4/7) Epoch 27, batch 200, loss[loss=0.1752, simple_loss=0.249, pruned_loss=0.0507, over 4872.00 frames. ], tot_loss[loss=0.1597, simple_loss=0.2314, pruned_loss=0.04398, over 607252.10 frames. ], batch size: 31, lr: 2.94e-03, grad_scale: 32.0 2023-04-28 00:26:55,611 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=149133.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:27:08,548 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=149154.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:27:18,795 INFO [finetune.py:976] (4/7) Epoch 27, batch 250, loss[loss=0.1651, simple_loss=0.2321, pruned_loss=0.04905, over 4868.00 frames. ], tot_loss[loss=0.1631, simple_loss=0.2354, pruned_loss=0.04534, over 685663.95 frames. ], batch size: 34, lr: 2.94e-03, grad_scale: 32.0 2023-04-28 00:27:21,650 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.056e+02 1.484e+02 1.804e+02 2.162e+02 4.404e+02, threshold=3.609e+02, percent-clipped=2.0 2023-04-28 00:27:29,602 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7319, 1.5639, 1.9525, 2.1943, 1.5473, 1.3799, 1.7464, 1.0550], device='cuda:4'), covar=tensor([0.0408, 0.0570, 0.0381, 0.0431, 0.0633, 0.1028, 0.0470, 0.0587], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0067, 0.0066, 0.0069, 0.0075, 0.0094, 0.0072, 0.0063], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-28 00:27:34,369 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-04-28 00:27:35,860 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.53 vs. limit=5.0 2023-04-28 00:27:41,049 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=149202.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:27:50,699 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-28 00:27:52,341 INFO [finetune.py:976] (4/7) Epoch 27, batch 300, loss[loss=0.1482, simple_loss=0.2354, pruned_loss=0.03053, over 4823.00 frames. ], tot_loss[loss=0.1666, simple_loss=0.24, pruned_loss=0.04653, over 746503.36 frames. ], batch size: 40, lr: 2.94e-03, grad_scale: 32.0 2023-04-28 00:27:55,210 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-28 00:28:22,270 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.8221, 2.1681, 5.6130, 5.2876, 4.9102, 5.2469, 4.9170, 5.0412], device='cuda:4'), covar=tensor([0.5738, 0.4758, 0.0825, 0.1637, 0.1049, 0.1076, 0.1170, 0.1339], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0312, 0.0413, 0.0413, 0.0350, 0.0416, 0.0321, 0.0367], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-28 00:28:25,830 INFO [finetune.py:976] (4/7) Epoch 27, batch 350, loss[loss=0.1267, simple_loss=0.2055, pruned_loss=0.02399, over 4792.00 frames. ], tot_loss[loss=0.168, simple_loss=0.2417, pruned_loss=0.0472, over 791940.01 frames. ], batch size: 29, lr: 2.94e-03, grad_scale: 32.0 2023-04-28 00:28:28,123 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.225e+02 1.612e+02 1.887e+02 2.251e+02 5.949e+02, threshold=3.774e+02, percent-clipped=2.0 2023-04-28 00:28:51,973 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=149307.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:28:59,852 INFO [finetune.py:976] (4/7) Epoch 27, batch 400, loss[loss=0.1865, simple_loss=0.2539, pruned_loss=0.05958, over 4888.00 frames. ], tot_loss[loss=0.1685, simple_loss=0.2427, pruned_loss=0.04713, over 828333.68 frames. ], batch size: 35, lr: 2.94e-03, grad_scale: 32.0 2023-04-28 00:29:36,641 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.75 vs. limit=5.0 2023-04-28 00:29:38,827 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.96 vs. limit=2.0 2023-04-28 00:29:39,547 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=149355.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:29:40,818 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0620, 2.5403, 1.0642, 1.3672, 2.2408, 1.1941, 3.5680, 1.7455], device='cuda:4'), covar=tensor([0.0711, 0.0593, 0.0775, 0.1258, 0.0474, 0.1074, 0.0258, 0.0648], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0063, 0.0047, 0.0045, 0.0049, 0.0051, 0.0072, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0007, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-28 00:29:48,686 INFO [finetune.py:976] (4/7) Epoch 27, batch 450, loss[loss=0.1536, simple_loss=0.2126, pruned_loss=0.04729, over 4902.00 frames. ], tot_loss[loss=0.1672, simple_loss=0.2412, pruned_loss=0.04659, over 857200.24 frames. ], batch size: 32, lr: 2.94e-03, grad_scale: 32.0 2023-04-28 00:29:50,018 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=149372.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:29:50,985 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.462e+01 1.496e+02 1.752e+02 2.050e+02 4.938e+02, threshold=3.505e+02, percent-clipped=1.0 2023-04-28 00:29:56,468 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.8616, 4.9025, 3.2948, 5.5533, 4.8620, 4.8757, 2.0923, 4.7750], device='cuda:4'), covar=tensor([0.1584, 0.0892, 0.2588, 0.0902, 0.3334, 0.1529, 0.5588, 0.2057], device='cuda:4'), in_proj_covar=tensor([0.0247, 0.0222, 0.0254, 0.0305, 0.0302, 0.0250, 0.0276, 0.0276], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 00:30:41,614 INFO [finetune.py:976] (4/7) Epoch 27, batch 500, loss[loss=0.1604, simple_loss=0.2258, pruned_loss=0.04746, over 4874.00 frames. ], tot_loss[loss=0.1658, simple_loss=0.2389, pruned_loss=0.04631, over 877588.40 frames. ], batch size: 34, lr: 2.93e-03, grad_scale: 32.0 2023-04-28 00:30:41,677 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=149420.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:31:43,468 INFO [finetune.py:976] (4/7) Epoch 27, batch 550, loss[loss=0.1589, simple_loss=0.2314, pruned_loss=0.04324, over 4759.00 frames. ], tot_loss[loss=0.1642, simple_loss=0.2364, pruned_loss=0.046, over 895035.19 frames. ], batch size: 59, lr: 2.93e-03, grad_scale: 32.0 2023-04-28 00:31:45,299 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.135e+02 1.482e+02 1.685e+02 2.089e+02 3.058e+02, threshold=3.371e+02, percent-clipped=0.0 2023-04-28 00:32:06,929 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9320, 1.1854, 3.2475, 3.0013, 2.9292, 3.1406, 3.1382, 2.8797], device='cuda:4'), covar=tensor([0.7568, 0.5290, 0.1468, 0.2107, 0.1428, 0.2142, 0.2320, 0.1742], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0310, 0.0410, 0.0411, 0.0348, 0.0415, 0.0319, 0.0365], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-28 00:32:08,163 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1946, 1.5308, 1.4300, 1.7872, 1.5992, 1.9342, 1.4923, 3.5388], device='cuda:4'), covar=tensor([0.0627, 0.0833, 0.0824, 0.1210, 0.0657, 0.0482, 0.0754, 0.0126], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 00:32:13,099 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=149514.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:32:16,579 INFO [finetune.py:976] (4/7) Epoch 27, batch 600, loss[loss=0.1654, simple_loss=0.2341, pruned_loss=0.04838, over 4907.00 frames. ], tot_loss[loss=0.165, simple_loss=0.2371, pruned_loss=0.04643, over 909575.27 frames. ], batch size: 37, lr: 2.93e-03, grad_scale: 32.0 2023-04-28 00:32:47,510 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=149565.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:32:50,410 INFO [finetune.py:976] (4/7) Epoch 27, batch 650, loss[loss=0.2018, simple_loss=0.2691, pruned_loss=0.06723, over 4845.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.2398, pruned_loss=0.04689, over 917231.85 frames. ], batch size: 49, lr: 2.93e-03, grad_scale: 32.0 2023-04-28 00:32:52,256 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.202e+02 1.546e+02 1.916e+02 2.340e+02 4.465e+02, threshold=3.832e+02, percent-clipped=5.0 2023-04-28 00:32:53,595 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=149575.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:33:23,678 INFO [finetune.py:976] (4/7) Epoch 27, batch 700, loss[loss=0.2135, simple_loss=0.29, pruned_loss=0.06851, over 4809.00 frames. ], tot_loss[loss=0.1679, simple_loss=0.2415, pruned_loss=0.04712, over 927428.17 frames. ], batch size: 40, lr: 2.93e-03, grad_scale: 32.0 2023-04-28 00:33:27,487 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=149626.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 00:33:35,933 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.68 vs. limit=5.0 2023-04-28 00:33:53,504 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.0499, 2.3814, 2.2879, 2.3641, 2.1957, 2.2897, 2.3149, 2.2367], device='cuda:4'), covar=tensor([0.3711, 0.5703, 0.4848, 0.4754, 0.5854, 0.6859, 0.5943, 0.5953], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0375, 0.0329, 0.0339, 0.0350, 0.0394, 0.0360, 0.0332], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 00:33:56,993 INFO [finetune.py:976] (4/7) Epoch 27, batch 750, loss[loss=0.1897, simple_loss=0.2547, pruned_loss=0.06234, over 4895.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.2418, pruned_loss=0.0468, over 935463.83 frames. ], batch size: 37, lr: 2.93e-03, grad_scale: 32.0 2023-04-28 00:33:58,788 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.029e+02 1.603e+02 1.862e+02 2.208e+02 4.099e+02, threshold=3.723e+02, percent-clipped=2.0 2023-04-28 00:34:42,994 INFO [finetune.py:976] (4/7) Epoch 27, batch 800, loss[loss=0.1497, simple_loss=0.2218, pruned_loss=0.03882, over 4852.00 frames. ], tot_loss[loss=0.1684, simple_loss=0.2422, pruned_loss=0.04727, over 939373.36 frames. ], batch size: 49, lr: 2.93e-03, grad_scale: 32.0 2023-04-28 00:34:44,962 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0623, 1.7787, 2.2933, 2.4610, 2.1019, 1.9994, 2.1230, 2.1284], device='cuda:4'), covar=tensor([0.5160, 0.7397, 0.7483, 0.6447, 0.6591, 0.9783, 0.9426, 0.9596], device='cuda:4'), in_proj_covar=tensor([0.0444, 0.0423, 0.0518, 0.0508, 0.0471, 0.0508, 0.0510, 0.0523], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 00:34:54,716 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6489, 1.6527, 0.8084, 1.3289, 1.6995, 1.5467, 1.4207, 1.4942], device='cuda:4'), covar=tensor([0.0479, 0.0370, 0.0350, 0.0552, 0.0266, 0.0483, 0.0499, 0.0555], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0027, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0052, 0.0046, 0.0038, 0.0053, 0.0039, 0.0051, 0.0050, 0.0052], device='cuda:4') 2023-04-28 00:35:14,876 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=149746.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 00:35:48,631 INFO [finetune.py:976] (4/7) Epoch 27, batch 850, loss[loss=0.1432, simple_loss=0.2216, pruned_loss=0.03244, over 4747.00 frames. ], tot_loss[loss=0.1672, simple_loss=0.2404, pruned_loss=0.04696, over 943200.13 frames. ], batch size: 54, lr: 2.93e-03, grad_scale: 32.0 2023-04-28 00:35:55,546 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.068e+02 1.475e+02 1.749e+02 2.076e+02 4.110e+02, threshold=3.498e+02, percent-clipped=2.0 2023-04-28 00:36:06,648 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3454, 1.7118, 1.5801, 2.2145, 2.3212, 1.9229, 1.9252, 1.7323], device='cuda:4'), covar=tensor([0.1553, 0.1864, 0.1775, 0.1485, 0.1467, 0.1785, 0.2065, 0.2098], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0303, 0.0346, 0.0282, 0.0322, 0.0301, 0.0294, 0.0367], device='cuda:4'), out_proj_covar=tensor([6.3025e-05, 6.2319e-05, 7.2552e-05, 5.6427e-05, 6.5854e-05, 6.2767e-05, 6.0826e-05, 7.7732e-05], device='cuda:4') 2023-04-28 00:36:09,625 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6225, 1.5899, 1.5356, 1.9287, 1.7283, 2.2276, 1.5065, 3.8508], device='cuda:4'), covar=tensor([0.0494, 0.0801, 0.0777, 0.1102, 0.0621, 0.0421, 0.0698, 0.0118], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 00:36:24,416 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=149807.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 00:36:33,260 INFO [finetune.py:976] (4/7) Epoch 27, batch 900, loss[loss=0.117, simple_loss=0.1848, pruned_loss=0.02462, over 4745.00 frames. ], tot_loss[loss=0.1649, simple_loss=0.2373, pruned_loss=0.04621, over 945122.78 frames. ], batch size: 54, lr: 2.93e-03, grad_scale: 32.0 2023-04-28 00:36:51,939 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5722, 2.0128, 1.9985, 2.2672, 2.1416, 2.2938, 1.9817, 4.8002], device='cuda:4'), covar=tensor([0.0509, 0.0723, 0.0690, 0.1050, 0.0594, 0.0454, 0.0623, 0.0106], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0039, 0.0042, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 00:37:06,210 INFO [finetune.py:976] (4/7) Epoch 27, batch 950, loss[loss=0.2037, simple_loss=0.2705, pruned_loss=0.06851, over 4827.00 frames. ], tot_loss[loss=0.1643, simple_loss=0.2363, pruned_loss=0.04617, over 946743.66 frames. ], batch size: 41, lr: 2.93e-03, grad_scale: 32.0 2023-04-28 00:37:06,276 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=149870.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:37:08,538 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.083e+02 1.514e+02 1.776e+02 2.110e+02 4.018e+02, threshold=3.552e+02, percent-clipped=2.0 2023-04-28 00:37:40,039 INFO [finetune.py:976] (4/7) Epoch 27, batch 1000, loss[loss=0.1521, simple_loss=0.2308, pruned_loss=0.03672, over 4752.00 frames. ], tot_loss[loss=0.1654, simple_loss=0.2379, pruned_loss=0.04642, over 950051.32 frames. ], batch size: 27, lr: 2.93e-03, grad_scale: 32.0 2023-04-28 00:37:40,686 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=149921.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 00:37:43,619 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4132, 3.3515, 0.8109, 1.8355, 1.6897, 2.3408, 1.8626, 0.9671], device='cuda:4'), covar=tensor([0.1444, 0.0780, 0.2015, 0.1193, 0.1149, 0.1058, 0.1558, 0.2166], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0237, 0.0135, 0.0121, 0.0131, 0.0151, 0.0117, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 00:38:13,689 INFO [finetune.py:976] (4/7) Epoch 27, batch 1050, loss[loss=0.2009, simple_loss=0.2754, pruned_loss=0.06317, over 4868.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.2406, pruned_loss=0.04679, over 952191.00 frames. ], batch size: 31, lr: 2.93e-03, grad_scale: 32.0 2023-04-28 00:38:15,492 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.212e+02 1.681e+02 2.029e+02 2.486e+02 4.683e+02, threshold=4.057e+02, percent-clipped=3.0 2023-04-28 00:38:48,335 INFO [finetune.py:976] (4/7) Epoch 27, batch 1100, loss[loss=0.179, simple_loss=0.2543, pruned_loss=0.05184, over 4852.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.242, pruned_loss=0.04649, over 953637.47 frames. ], batch size: 31, lr: 2.93e-03, grad_scale: 32.0 2023-04-28 00:38:50,726 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.9210, 3.8347, 2.7880, 4.5191, 3.9514, 3.8798, 1.7027, 3.8784], device='cuda:4'), covar=tensor([0.1817, 0.1115, 0.3242, 0.1422, 0.2562, 0.1837, 0.5594, 0.2318], device='cuda:4'), in_proj_covar=tensor([0.0248, 0.0222, 0.0255, 0.0305, 0.0302, 0.0250, 0.0277, 0.0276], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 00:38:53,560 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.46 vs. limit=5.0 2023-04-28 00:39:22,074 INFO [finetune.py:976] (4/7) Epoch 27, batch 1150, loss[loss=0.1895, simple_loss=0.2699, pruned_loss=0.05452, over 4822.00 frames. ], tot_loss[loss=0.1693, simple_loss=0.2436, pruned_loss=0.04743, over 953270.80 frames. ], batch size: 33, lr: 2.93e-03, grad_scale: 32.0 2023-04-28 00:39:23,883 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.145e+02 1.471e+02 1.817e+02 2.070e+02 3.709e+02, threshold=3.635e+02, percent-clipped=0.0 2023-04-28 00:39:28,114 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=150079.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:39:43,680 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=150102.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 00:39:56,125 INFO [finetune.py:976] (4/7) Epoch 27, batch 1200, loss[loss=0.1764, simple_loss=0.2508, pruned_loss=0.05105, over 4901.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2427, pruned_loss=0.04744, over 954388.35 frames. ], batch size: 35, lr: 2.93e-03, grad_scale: 32.0 2023-04-28 00:40:10,326 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=150140.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:40:45,077 INFO [finetune.py:976] (4/7) Epoch 27, batch 1250, loss[loss=0.1416, simple_loss=0.2195, pruned_loss=0.03184, over 4799.00 frames. ], tot_loss[loss=0.1661, simple_loss=0.2395, pruned_loss=0.04633, over 954168.51 frames. ], batch size: 29, lr: 2.93e-03, grad_scale: 32.0 2023-04-28 00:40:45,704 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=150170.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:40:47,937 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.413e+01 1.487e+02 1.765e+02 2.046e+02 3.394e+02, threshold=3.529e+02, percent-clipped=0.0 2023-04-28 00:41:41,402 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3363, 1.6995, 1.7728, 1.8667, 1.7941, 1.8299, 1.8317, 1.7788], device='cuda:4'), covar=tensor([0.3658, 0.4889, 0.4046, 0.4011, 0.4976, 0.6463, 0.4707, 0.4507], device='cuda:4'), in_proj_covar=tensor([0.0339, 0.0373, 0.0329, 0.0339, 0.0349, 0.0392, 0.0358, 0.0330], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 00:41:49,586 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=150218.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:41:50,799 INFO [finetune.py:976] (4/7) Epoch 27, batch 1300, loss[loss=0.1354, simple_loss=0.209, pruned_loss=0.03091, over 4795.00 frames. ], tot_loss[loss=0.1638, simple_loss=0.2367, pruned_loss=0.04545, over 955624.61 frames. ], batch size: 25, lr: 2.93e-03, grad_scale: 32.0 2023-04-28 00:41:51,515 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=150221.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:42:56,733 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=150269.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:42:57,288 INFO [finetune.py:976] (4/7) Epoch 27, batch 1350, loss[loss=0.1808, simple_loss=0.2564, pruned_loss=0.05262, over 4925.00 frames. ], tot_loss[loss=0.1638, simple_loss=0.2367, pruned_loss=0.04543, over 956314.87 frames. ], batch size: 36, lr: 2.93e-03, grad_scale: 32.0 2023-04-28 00:42:59,131 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 1.474e+02 1.801e+02 2.070e+02 3.416e+02, threshold=3.602e+02, percent-clipped=0.0 2023-04-28 00:43:05,181 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=150274.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:43:19,079 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8086, 1.7120, 2.2530, 2.2913, 1.5956, 1.4834, 1.7121, 0.9892], device='cuda:4'), covar=tensor([0.0616, 0.0628, 0.0411, 0.0721, 0.0767, 0.1120, 0.0626, 0.0758], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0067, 0.0066, 0.0069, 0.0074, 0.0094, 0.0072, 0.0062], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-28 00:44:01,100 INFO [finetune.py:976] (4/7) Epoch 27, batch 1400, loss[loss=0.1819, simple_loss=0.2576, pruned_loss=0.05312, over 4819.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.2401, pruned_loss=0.04672, over 956707.36 frames. ], batch size: 39, lr: 2.93e-03, grad_scale: 16.0 2023-04-28 00:44:22,625 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=150335.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:45:00,623 INFO [finetune.py:976] (4/7) Epoch 27, batch 1450, loss[loss=0.1806, simple_loss=0.2554, pruned_loss=0.05287, over 4915.00 frames. ], tot_loss[loss=0.1669, simple_loss=0.241, pruned_loss=0.04641, over 953922.18 frames. ], batch size: 36, lr: 2.93e-03, grad_scale: 16.0 2023-04-28 00:45:03,058 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.061e+02 1.594e+02 1.851e+02 2.331e+02 4.105e+02, threshold=3.701e+02, percent-clipped=2.0 2023-04-28 00:45:22,884 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=150402.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 00:45:33,783 INFO [finetune.py:976] (4/7) Epoch 27, batch 1500, loss[loss=0.1376, simple_loss=0.2149, pruned_loss=0.03014, over 4781.00 frames. ], tot_loss[loss=0.1676, simple_loss=0.2416, pruned_loss=0.04677, over 953324.07 frames. ], batch size: 29, lr: 2.93e-03, grad_scale: 16.0 2023-04-28 00:45:44,392 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=150435.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:45:54,543 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=150450.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 00:46:06,955 INFO [finetune.py:976] (4/7) Epoch 27, batch 1550, loss[loss=0.1722, simple_loss=0.2429, pruned_loss=0.05081, over 4687.00 frames. ], tot_loss[loss=0.168, simple_loss=0.2418, pruned_loss=0.04707, over 952379.17 frames. ], batch size: 23, lr: 2.93e-03, grad_scale: 16.0 2023-04-28 00:46:09,361 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.042e+02 1.486e+02 1.780e+02 2.149e+02 3.630e+02, threshold=3.561e+02, percent-clipped=0.0 2023-04-28 00:46:19,683 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-28 00:46:22,360 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2071, 1.3841, 1.3126, 1.7002, 1.4737, 1.6022, 1.2826, 2.9790], device='cuda:4'), covar=tensor([0.0735, 0.1048, 0.0948, 0.1352, 0.0814, 0.0686, 0.0985, 0.0255], device='cuda:4'), in_proj_covar=tensor([0.0038, 0.0039, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0014, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 00:46:40,764 INFO [finetune.py:976] (4/7) Epoch 27, batch 1600, loss[loss=0.1864, simple_loss=0.2488, pruned_loss=0.06202, over 4851.00 frames. ], tot_loss[loss=0.1658, simple_loss=0.2392, pruned_loss=0.04623, over 952396.17 frames. ], batch size: 47, lr: 2.93e-03, grad_scale: 16.0 2023-04-28 00:46:59,332 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5910, 1.8750, 1.7244, 2.4276, 2.5444, 2.0673, 2.0453, 1.8106], device='cuda:4'), covar=tensor([0.1743, 0.1763, 0.1798, 0.1726, 0.1120, 0.2026, 0.2465, 0.2495], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0306, 0.0348, 0.0285, 0.0325, 0.0305, 0.0297, 0.0372], device='cuda:4'), out_proj_covar=tensor([6.3604e-05, 6.2921e-05, 7.3116e-05, 5.6946e-05, 6.6419e-05, 6.3567e-05, 6.1442e-05, 7.8719e-05], device='cuda:4') 2023-04-28 00:47:01,609 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8307, 1.9255, 1.2161, 1.5830, 2.0358, 1.6979, 1.6471, 1.7163], device='cuda:4'), covar=tensor([0.0407, 0.0284, 0.0295, 0.0428, 0.0278, 0.0380, 0.0378, 0.0450], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0027, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0052, 0.0046, 0.0039, 0.0053, 0.0039, 0.0051, 0.0051, 0.0053], device='cuda:4') 2023-04-28 00:47:08,131 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=150537.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:47:41,251 INFO [finetune.py:976] (4/7) Epoch 27, batch 1650, loss[loss=0.1388, simple_loss=0.2126, pruned_loss=0.0325, over 4773.00 frames. ], tot_loss[loss=0.1639, simple_loss=0.2369, pruned_loss=0.04544, over 953259.60 frames. ], batch size: 28, lr: 2.93e-03, grad_scale: 16.0 2023-04-28 00:47:41,387 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4954, 3.2230, 2.8270, 3.0307, 2.3131, 2.6974, 2.8626, 2.2709], device='cuda:4'), covar=tensor([0.2040, 0.1057, 0.0685, 0.1228, 0.2950, 0.1133, 0.1718, 0.2514], device='cuda:4'), in_proj_covar=tensor([0.0286, 0.0302, 0.0217, 0.0279, 0.0316, 0.0255, 0.0251, 0.0266], device='cuda:4'), out_proj_covar=tensor([1.1432e-04, 1.1903e-04, 8.5539e-05, 1.0992e-04, 1.2739e-04, 1.0027e-04, 1.0115e-04, 1.0506e-04], device='cuda:4') 2023-04-28 00:47:43,702 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.015e+02 1.594e+02 1.886e+02 2.369e+02 4.479e+02, threshold=3.773e+02, percent-clipped=3.0 2023-04-28 00:48:00,517 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=150598.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:48:14,795 INFO [finetune.py:976] (4/7) Epoch 27, batch 1700, loss[loss=0.1285, simple_loss=0.2051, pruned_loss=0.02589, over 4768.00 frames. ], tot_loss[loss=0.1623, simple_loss=0.235, pruned_loss=0.04483, over 954959.63 frames. ], batch size: 28, lr: 2.93e-03, grad_scale: 16.0 2023-04-28 00:48:20,945 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=150630.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:48:48,162 INFO [finetune.py:976] (4/7) Epoch 27, batch 1750, loss[loss=0.1214, simple_loss=0.1866, pruned_loss=0.02813, over 4768.00 frames. ], tot_loss[loss=0.164, simple_loss=0.237, pruned_loss=0.04547, over 953793.28 frames. ], batch size: 26, lr: 2.93e-03, grad_scale: 16.0 2023-04-28 00:48:50,603 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.161e+02 1.571e+02 1.900e+02 2.319e+02 4.181e+02, threshold=3.799e+02, percent-clipped=4.0 2023-04-28 00:49:11,792 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.7334, 3.6416, 2.7910, 4.2969, 3.6954, 3.6338, 1.5209, 3.7070], device='cuda:4'), covar=tensor([0.1736, 0.1158, 0.3831, 0.1610, 0.3453, 0.1880, 0.5791, 0.2288], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0219, 0.0251, 0.0301, 0.0297, 0.0247, 0.0274, 0.0272], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 00:49:21,909 INFO [finetune.py:976] (4/7) Epoch 27, batch 1800, loss[loss=0.2139, simple_loss=0.2853, pruned_loss=0.07121, over 4722.00 frames. ], tot_loss[loss=0.1667, simple_loss=0.2407, pruned_loss=0.04641, over 954617.83 frames. ], batch size: 59, lr: 2.93e-03, grad_scale: 16.0 2023-04-28 00:49:22,684 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3180, 1.8431, 2.2415, 2.6522, 2.2561, 1.7551, 1.5547, 2.1741], device='cuda:4'), covar=tensor([0.3435, 0.3095, 0.1698, 0.2181, 0.2473, 0.2782, 0.3851, 0.1856], device='cuda:4'), in_proj_covar=tensor([0.0294, 0.0247, 0.0229, 0.0315, 0.0221, 0.0235, 0.0228, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-28 00:49:29,202 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=150732.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:49:31,009 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=150735.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:49:54,545 INFO [finetune.py:976] (4/7) Epoch 27, batch 1850, loss[loss=0.1262, simple_loss=0.2073, pruned_loss=0.02257, over 4736.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2427, pruned_loss=0.04755, over 954665.61 frames. ], batch size: 27, lr: 2.93e-03, grad_scale: 16.0 2023-04-28 00:50:02,278 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.087e+02 1.541e+02 1.934e+02 2.290e+02 4.009e+02, threshold=3.867e+02, percent-clipped=2.0 2023-04-28 00:50:13,366 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=150783.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:50:25,745 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=150793.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 00:50:56,512 INFO [finetune.py:976] (4/7) Epoch 27, batch 1900, loss[loss=0.1517, simple_loss=0.2243, pruned_loss=0.03951, over 4718.00 frames. ], tot_loss[loss=0.1697, simple_loss=0.2437, pruned_loss=0.04785, over 955692.89 frames. ], batch size: 59, lr: 2.93e-03, grad_scale: 16.0 2023-04-28 00:51:29,797 INFO [finetune.py:976] (4/7) Epoch 27, batch 1950, loss[loss=0.1659, simple_loss=0.2313, pruned_loss=0.05024, over 4892.00 frames. ], tot_loss[loss=0.17, simple_loss=0.2439, pruned_loss=0.048, over 957232.56 frames. ], batch size: 32, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 00:51:32,199 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.093e+02 1.535e+02 1.747e+02 2.093e+02 5.445e+02, threshold=3.494e+02, percent-clipped=3.0 2023-04-28 00:51:44,420 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=150893.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:51:49,219 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=150900.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:51:50,411 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8063, 2.0180, 1.3368, 1.5904, 2.0106, 1.6617, 1.6195, 1.6988], device='cuda:4'), covar=tensor([0.0392, 0.0276, 0.0271, 0.0426, 0.0253, 0.0415, 0.0382, 0.0440], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0028, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0052, 0.0046, 0.0038, 0.0053, 0.0039, 0.0051, 0.0051, 0.0052], device='cuda:4') 2023-04-28 00:52:03,059 INFO [finetune.py:976] (4/7) Epoch 27, batch 2000, loss[loss=0.1725, simple_loss=0.241, pruned_loss=0.05198, over 4930.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2419, pruned_loss=0.04789, over 955598.40 frames. ], batch size: 33, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 00:52:09,182 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=150930.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:52:20,123 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0754, 1.7894, 2.2612, 2.4162, 2.1302, 1.9814, 2.1120, 2.0171], device='cuda:4'), covar=tensor([0.4505, 0.7213, 0.6663, 0.5192, 0.5639, 0.8456, 0.8292, 1.0480], device='cuda:4'), in_proj_covar=tensor([0.0443, 0.0422, 0.0519, 0.0507, 0.0471, 0.0506, 0.0508, 0.0522], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 00:52:32,008 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7123, 2.6536, 2.1436, 2.3733, 2.7478, 2.2355, 3.4883, 2.1270], device='cuda:4'), covar=tensor([0.3602, 0.2067, 0.4213, 0.3360, 0.1860, 0.2632, 0.1553, 0.3978], device='cuda:4'), in_proj_covar=tensor([0.0338, 0.0354, 0.0425, 0.0350, 0.0379, 0.0375, 0.0365, 0.0420], device='cuda:4'), out_proj_covar=tensor([9.9898e-05, 1.0531e-04, 1.2866e-04, 1.0464e-04, 1.1227e-04, 1.1143e-04, 1.0683e-04, 1.2635e-04], device='cuda:4') 2023-04-28 00:52:51,915 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=150961.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:52:54,940 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1765, 1.3731, 1.2942, 1.6957, 1.5454, 1.5976, 1.4077, 2.4232], device='cuda:4'), covar=tensor([0.0617, 0.0785, 0.0746, 0.1123, 0.0598, 0.0450, 0.0678, 0.0207], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0039, 0.0042, 0.0040, 0.0037, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 00:53:02,103 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.74 vs. limit=5.0 2023-04-28 00:53:02,955 INFO [finetune.py:976] (4/7) Epoch 27, batch 2050, loss[loss=0.1095, simple_loss=0.1856, pruned_loss=0.01669, over 4818.00 frames. ], tot_loss[loss=0.1652, simple_loss=0.238, pruned_loss=0.04622, over 957367.65 frames. ], batch size: 25, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 00:53:04,683 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.52 vs. limit=5.0 2023-04-28 00:53:05,397 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.068e+02 1.536e+02 1.853e+02 2.161e+02 4.480e+02, threshold=3.705e+02, percent-clipped=3.0 2023-04-28 00:53:12,614 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=150978.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:53:48,013 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.75 vs. limit=2.0 2023-04-28 00:54:07,738 INFO [finetune.py:976] (4/7) Epoch 27, batch 2100, loss[loss=0.1826, simple_loss=0.2458, pruned_loss=0.0597, over 4828.00 frames. ], tot_loss[loss=0.1651, simple_loss=0.2375, pruned_loss=0.04636, over 956331.41 frames. ], batch size: 33, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 00:55:00,545 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4414, 1.8287, 1.8092, 1.9601, 1.7791, 1.7747, 1.9022, 1.8730], device='cuda:4'), covar=tensor([0.3945, 0.5328, 0.4593, 0.4401, 0.5662, 0.7279, 0.4847, 0.4867], device='cuda:4'), in_proj_covar=tensor([0.0343, 0.0378, 0.0332, 0.0343, 0.0352, 0.0397, 0.0362, 0.0335], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 00:55:11,731 INFO [finetune.py:976] (4/7) Epoch 27, batch 2150, loss[loss=0.2077, simple_loss=0.2774, pruned_loss=0.06902, over 4898.00 frames. ], tot_loss[loss=0.1686, simple_loss=0.2412, pruned_loss=0.04801, over 956193.25 frames. ], batch size: 37, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 00:55:19,340 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.036e+02 1.480e+02 1.779e+02 2.139e+02 3.586e+02, threshold=3.558e+02, percent-clipped=0.0 2023-04-28 00:55:33,248 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=151088.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 00:56:16,163 INFO [finetune.py:976] (4/7) Epoch 27, batch 2200, loss[loss=0.1555, simple_loss=0.234, pruned_loss=0.03849, over 4833.00 frames. ], tot_loss[loss=0.1702, simple_loss=0.2436, pruned_loss=0.04843, over 956663.37 frames. ], batch size: 49, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 00:56:34,895 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7390, 1.2901, 1.8492, 2.2019, 1.8526, 1.7404, 1.8140, 1.7518], device='cuda:4'), covar=tensor([0.4466, 0.6738, 0.6183, 0.5272, 0.5593, 0.7522, 0.7412, 0.9065], device='cuda:4'), in_proj_covar=tensor([0.0443, 0.0421, 0.0518, 0.0507, 0.0470, 0.0506, 0.0506, 0.0522], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 00:57:19,654 INFO [finetune.py:976] (4/7) Epoch 27, batch 2250, loss[loss=0.1748, simple_loss=0.2595, pruned_loss=0.04502, over 4862.00 frames. ], tot_loss[loss=0.1704, simple_loss=0.244, pruned_loss=0.04843, over 957252.86 frames. ], batch size: 31, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 00:57:27,225 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.948e+01 1.668e+02 1.911e+02 2.463e+02 4.700e+02, threshold=3.822e+02, percent-clipped=3.0 2023-04-28 00:57:36,748 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6890, 1.8533, 0.9263, 1.4377, 2.1382, 1.5194, 1.4556, 1.6070], device='cuda:4'), covar=tensor([0.0473, 0.0363, 0.0289, 0.0515, 0.0240, 0.0470, 0.0475, 0.0544], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0028, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0052, 0.0046, 0.0038, 0.0053, 0.0039, 0.0051, 0.0051, 0.0052], device='cuda:4') 2023-04-28 00:57:38,566 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=151184.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:57:49,167 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=151193.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:58:01,674 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5242, 2.9758, 0.9571, 1.7504, 2.2159, 1.6104, 4.2303, 2.2583], device='cuda:4'), covar=tensor([0.0649, 0.0732, 0.0861, 0.1190, 0.0564, 0.0982, 0.0198, 0.0581], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0072, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-28 00:58:22,084 INFO [finetune.py:976] (4/7) Epoch 27, batch 2300, loss[loss=0.1419, simple_loss=0.2028, pruned_loss=0.04047, over 4751.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2436, pruned_loss=0.04764, over 954962.49 frames. ], batch size: 28, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 00:58:42,720 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=151241.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:58:45,169 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=151245.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:58:52,342 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=151256.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:59:01,205 INFO [finetune.py:976] (4/7) Epoch 27, batch 2350, loss[loss=0.1789, simple_loss=0.2516, pruned_loss=0.05308, over 4736.00 frames. ], tot_loss[loss=0.1686, simple_loss=0.2425, pruned_loss=0.04738, over 955017.04 frames. ], batch size: 54, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 00:59:04,084 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.014e+02 1.580e+02 1.872e+02 2.240e+02 4.120e+02, threshold=3.744e+02, percent-clipped=1.0 2023-04-28 00:59:08,916 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6914, 3.6022, 0.9400, 2.0855, 1.9528, 2.5281, 2.0441, 0.9815], device='cuda:4'), covar=tensor([0.1355, 0.0851, 0.2024, 0.1134, 0.1077, 0.0965, 0.1458, 0.2088], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0237, 0.0134, 0.0120, 0.0131, 0.0152, 0.0117, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 00:59:10,174 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=151282.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:59:12,264 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-28 00:59:18,171 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-28 00:59:20,486 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=151298.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:59:34,701 INFO [finetune.py:976] (4/7) Epoch 27, batch 2400, loss[loss=0.137, simple_loss=0.2108, pruned_loss=0.03157, over 4766.00 frames. ], tot_loss[loss=0.1666, simple_loss=0.2397, pruned_loss=0.04673, over 954957.95 frames. ], batch size: 23, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 00:59:50,903 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=151343.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 00:59:51,572 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.52 vs. limit=2.0 2023-04-28 00:59:54,551 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1723, 2.7372, 1.0351, 1.4769, 2.0237, 1.2586, 3.6320, 2.0715], device='cuda:4'), covar=tensor([0.0671, 0.0536, 0.0756, 0.1234, 0.0537, 0.1037, 0.0241, 0.0550], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0072, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-28 01:00:01,139 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=151359.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:00:07,713 INFO [finetune.py:976] (4/7) Epoch 27, batch 2450, loss[loss=0.1801, simple_loss=0.2614, pruned_loss=0.04942, over 4781.00 frames. ], tot_loss[loss=0.1648, simple_loss=0.2374, pruned_loss=0.04611, over 955328.67 frames. ], batch size: 59, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 01:00:10,600 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.893e+01 1.571e+02 1.897e+02 2.275e+02 8.002e+02, threshold=3.795e+02, percent-clipped=4.0 2023-04-28 01:00:15,524 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6465, 2.0436, 2.1670, 2.2219, 2.1579, 2.2154, 2.2400, 2.1915], device='cuda:4'), covar=tensor([0.3511, 0.5050, 0.4409, 0.4553, 0.4739, 0.6358, 0.4741, 0.4233], device='cuda:4'), in_proj_covar=tensor([0.0343, 0.0377, 0.0331, 0.0341, 0.0351, 0.0395, 0.0361, 0.0334], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 01:00:21,187 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=151388.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:00:41,811 INFO [finetune.py:976] (4/7) Epoch 27, batch 2500, loss[loss=0.1669, simple_loss=0.2384, pruned_loss=0.04775, over 4875.00 frames. ], tot_loss[loss=0.1655, simple_loss=0.238, pruned_loss=0.04649, over 956442.08 frames. ], batch size: 34, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 01:00:53,956 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=151436.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:01:02,960 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5048, 1.5117, 4.4831, 4.2121, 3.9176, 4.3024, 4.1955, 3.9317], device='cuda:4'), covar=tensor([0.7769, 0.5952, 0.1057, 0.1880, 0.1186, 0.2055, 0.1372, 0.1522], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0309, 0.0410, 0.0410, 0.0349, 0.0415, 0.0319, 0.0364], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-28 01:01:07,906 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.76 vs. limit=2.0 2023-04-28 01:01:14,722 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=5.21 vs. limit=5.0 2023-04-28 01:01:15,577 INFO [finetune.py:976] (4/7) Epoch 27, batch 2550, loss[loss=0.1594, simple_loss=0.2351, pruned_loss=0.04188, over 4817.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.2404, pruned_loss=0.0469, over 955950.35 frames. ], batch size: 40, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 01:01:17,943 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.064e+02 1.630e+02 1.877e+02 2.163e+02 4.851e+02, threshold=3.753e+02, percent-clipped=1.0 2023-04-28 01:01:20,036 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-04-28 01:01:29,230 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.79 vs. limit=5.0 2023-04-28 01:01:48,926 INFO [finetune.py:976] (4/7) Epoch 27, batch 2600, loss[loss=0.1893, simple_loss=0.2644, pruned_loss=0.0571, over 4814.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.243, pruned_loss=0.04801, over 957410.62 frames. ], batch size: 33, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 01:02:00,468 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-28 01:02:02,559 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=151540.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:02:19,374 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=151556.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:02:33,622 INFO [finetune.py:976] (4/7) Epoch 27, batch 2650, loss[loss=0.1505, simple_loss=0.2255, pruned_loss=0.03779, over 4921.00 frames. ], tot_loss[loss=0.1687, simple_loss=0.2424, pruned_loss=0.0475, over 953866.84 frames. ], batch size: 33, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 01:02:41,497 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.080e+02 1.560e+02 1.768e+02 2.112e+02 4.460e+02, threshold=3.536e+02, percent-clipped=1.0 2023-04-28 01:02:53,944 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.37 vs. limit=2.0 2023-04-28 01:03:12,949 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8130, 1.2241, 1.8407, 2.2660, 1.9274, 1.7570, 1.8022, 1.7366], device='cuda:4'), covar=tensor([0.4486, 0.6626, 0.5890, 0.5452, 0.5715, 0.7341, 0.7611, 0.8741], device='cuda:4'), in_proj_covar=tensor([0.0441, 0.0420, 0.0516, 0.0505, 0.0468, 0.0504, 0.0505, 0.0519], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 01:03:23,448 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=151604.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:03:39,057 INFO [finetune.py:976] (4/7) Epoch 27, batch 2700, loss[loss=0.1179, simple_loss=0.1992, pruned_loss=0.01824, over 4769.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.2417, pruned_loss=0.04668, over 954694.99 frames. ], batch size: 28, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 01:04:05,033 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=151638.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:04:21,252 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=151654.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:04:40,565 INFO [finetune.py:976] (4/7) Epoch 27, batch 2750, loss[loss=0.1803, simple_loss=0.2474, pruned_loss=0.05656, over 4932.00 frames. ], tot_loss[loss=0.1672, simple_loss=0.2403, pruned_loss=0.04702, over 954155.19 frames. ], batch size: 38, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 01:04:48,521 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.458e+01 1.381e+02 1.737e+02 2.155e+02 3.699e+02, threshold=3.473e+02, percent-clipped=1.0 2023-04-28 01:05:07,214 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=151688.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:05:31,029 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6939, 1.4178, 1.4047, 1.5394, 1.9422, 1.5492, 1.3409, 1.3504], device='cuda:4'), covar=tensor([0.1830, 0.1495, 0.1705, 0.1280, 0.0848, 0.1933, 0.2178, 0.2285], device='cuda:4'), in_proj_covar=tensor([0.0317, 0.0309, 0.0352, 0.0288, 0.0329, 0.0308, 0.0301, 0.0376], device='cuda:4'), out_proj_covar=tensor([6.4644e-05, 6.3502e-05, 7.3952e-05, 5.7580e-05, 6.7232e-05, 6.4222e-05, 6.2311e-05, 7.9531e-05], device='cuda:4') 2023-04-28 01:05:44,740 INFO [finetune.py:976] (4/7) Epoch 27, batch 2800, loss[loss=0.1389, simple_loss=0.2099, pruned_loss=0.03399, over 4834.00 frames. ], tot_loss[loss=0.1648, simple_loss=0.2372, pruned_loss=0.04627, over 953632.68 frames. ], batch size: 38, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 01:05:54,673 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2884, 1.7334, 2.1397, 2.4202, 2.0609, 1.7046, 1.2898, 1.9403], device='cuda:4'), covar=tensor([0.3033, 0.2915, 0.1573, 0.2008, 0.2566, 0.2569, 0.3991, 0.1810], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0246, 0.0227, 0.0313, 0.0220, 0.0233, 0.0227, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-28 01:06:24,905 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=151749.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:06:53,981 INFO [finetune.py:976] (4/7) Epoch 27, batch 2850, loss[loss=0.1549, simple_loss=0.2377, pruned_loss=0.03608, over 4820.00 frames. ], tot_loss[loss=0.1631, simple_loss=0.235, pruned_loss=0.04558, over 954818.12 frames. ], batch size: 51, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 01:06:56,484 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.135e+02 1.506e+02 1.770e+02 2.075e+02 3.794e+02, threshold=3.540e+02, percent-clipped=1.0 2023-04-28 01:07:04,862 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1236, 4.2276, 0.7006, 2.3244, 2.3844, 2.8146, 2.6366, 1.0023], device='cuda:4'), covar=tensor([0.1252, 0.0754, 0.2137, 0.1141, 0.1000, 0.1054, 0.1375, 0.2263], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0237, 0.0134, 0.0120, 0.0131, 0.0151, 0.0117, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 01:07:29,951 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=151801.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:07:58,282 INFO [finetune.py:976] (4/7) Epoch 27, batch 2900, loss[loss=0.1774, simple_loss=0.2559, pruned_loss=0.04945, over 4860.00 frames. ], tot_loss[loss=0.1653, simple_loss=0.2375, pruned_loss=0.0465, over 955900.94 frames. ], batch size: 44, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 01:08:10,970 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-28 01:08:11,762 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=151840.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:08:32,191 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=151862.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:08:32,839 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2506, 2.1444, 1.7951, 1.8897, 2.3033, 1.8594, 2.8054, 1.6597], device='cuda:4'), covar=tensor([0.4048, 0.2039, 0.5412, 0.2884, 0.1733, 0.2567, 0.1471, 0.4754], device='cuda:4'), in_proj_covar=tensor([0.0338, 0.0354, 0.0425, 0.0349, 0.0380, 0.0374, 0.0365, 0.0421], device='cuda:4'), out_proj_covar=tensor([9.9783e-05, 1.0546e-04, 1.2852e-04, 1.0460e-04, 1.1266e-04, 1.1115e-04, 1.0678e-04, 1.2657e-04], device='cuda:4') 2023-04-28 01:08:42,906 INFO [finetune.py:976] (4/7) Epoch 27, batch 2950, loss[loss=0.16, simple_loss=0.2419, pruned_loss=0.03901, over 4884.00 frames. ], tot_loss[loss=0.1683, simple_loss=0.2413, pruned_loss=0.04767, over 956819.98 frames. ], batch size: 35, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 01:08:50,554 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.751e+01 1.623e+02 1.878e+02 2.443e+02 4.733e+02, threshold=3.756e+02, percent-clipped=2.0 2023-04-28 01:09:04,992 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=151888.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:09:24,791 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.63 vs. limit=2.0 2023-04-28 01:09:47,998 INFO [finetune.py:976] (4/7) Epoch 27, batch 3000, loss[loss=0.1812, simple_loss=0.25, pruned_loss=0.05617, over 4870.00 frames. ], tot_loss[loss=0.1708, simple_loss=0.2439, pruned_loss=0.04883, over 954396.02 frames. ], batch size: 34, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 01:09:47,999 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-28 01:10:08,675 INFO [finetune.py:1010] (4/7) Epoch 27, validation: loss=0.1539, simple_loss=0.2224, pruned_loss=0.04268, over 2265189.00 frames. 2023-04-28 01:10:08,675 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6529MB 2023-04-28 01:10:24,801 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=151938.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:10:35,098 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=151954.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:10:43,247 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-28 01:10:44,120 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5816, 1.2917, 4.4743, 4.2112, 3.8627, 4.2479, 4.1491, 3.9239], device='cuda:4'), covar=tensor([0.7539, 0.6325, 0.1123, 0.1630, 0.1268, 0.2138, 0.1465, 0.1591], device='cuda:4'), in_proj_covar=tensor([0.0315, 0.0311, 0.0414, 0.0411, 0.0352, 0.0419, 0.0322, 0.0367], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-28 01:10:45,719 INFO [finetune.py:976] (4/7) Epoch 27, batch 3050, loss[loss=0.1527, simple_loss=0.2259, pruned_loss=0.03975, over 4800.00 frames. ], tot_loss[loss=0.1709, simple_loss=0.2444, pruned_loss=0.04867, over 953954.32 frames. ], batch size: 25, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 01:10:48,112 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.167e+02 1.615e+02 1.912e+02 2.194e+02 5.604e+02, threshold=3.825e+02, percent-clipped=1.0 2023-04-28 01:10:57,066 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=151986.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:11:08,925 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=152002.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:11:20,460 INFO [finetune.py:976] (4/7) Epoch 27, batch 3100, loss[loss=0.1392, simple_loss=0.216, pruned_loss=0.03121, over 4908.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2426, pruned_loss=0.04757, over 954538.41 frames. ], batch size: 36, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 01:11:21,665 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8554, 2.5025, 1.9539, 1.8881, 1.3521, 1.3681, 1.9997, 1.2907], device='cuda:4'), covar=tensor([0.1672, 0.1323, 0.1331, 0.1629, 0.2314, 0.1947, 0.0938, 0.2069], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0209, 0.0170, 0.0204, 0.0201, 0.0187, 0.0156, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 01:11:37,594 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=152044.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:11:54,338 INFO [finetune.py:976] (4/7) Epoch 27, batch 3150, loss[loss=0.1584, simple_loss=0.2278, pruned_loss=0.04446, over 4819.00 frames. ], tot_loss[loss=0.167, simple_loss=0.2401, pruned_loss=0.04697, over 954769.89 frames. ], batch size: 38, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 01:11:56,742 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.740e+01 1.539e+02 1.884e+02 2.286e+02 5.253e+02, threshold=3.767e+02, percent-clipped=2.0 2023-04-28 01:12:27,068 INFO [finetune.py:976] (4/7) Epoch 27, batch 3200, loss[loss=0.1654, simple_loss=0.248, pruned_loss=0.04143, over 4933.00 frames. ], tot_loss[loss=0.164, simple_loss=0.2366, pruned_loss=0.04571, over 953345.99 frames. ], batch size: 38, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 01:12:52,030 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=152157.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:13:00,433 INFO [finetune.py:976] (4/7) Epoch 27, batch 3250, loss[loss=0.2022, simple_loss=0.2649, pruned_loss=0.06979, over 4760.00 frames. ], tot_loss[loss=0.1656, simple_loss=0.2382, pruned_loss=0.04655, over 954726.59 frames. ], batch size: 54, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 01:13:02,805 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.216e+01 1.479e+02 1.800e+02 2.164e+02 4.753e+02, threshold=3.600e+02, percent-clipped=3.0 2023-04-28 01:13:07,362 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-04-28 01:13:21,044 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-28 01:13:22,611 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4170, 2.8483, 1.0727, 1.7725, 1.6236, 2.2887, 1.7530, 1.0782], device='cuda:4'), covar=tensor([0.1226, 0.0836, 0.1632, 0.1141, 0.1105, 0.0788, 0.1364, 0.2090], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0239, 0.0135, 0.0121, 0.0131, 0.0153, 0.0117, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 01:13:33,550 INFO [finetune.py:976] (4/7) Epoch 27, batch 3300, loss[loss=0.1759, simple_loss=0.2559, pruned_loss=0.04795, over 4805.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2421, pruned_loss=0.04785, over 951289.55 frames. ], batch size: 41, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 01:13:36,719 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=152225.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:14:13,059 INFO [finetune.py:976] (4/7) Epoch 27, batch 3350, loss[loss=0.152, simple_loss=0.238, pruned_loss=0.033, over 4887.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2432, pruned_loss=0.04778, over 953906.33 frames. ], batch size: 32, lr: 2.92e-03, grad_scale: 16.0 2023-04-28 01:14:14,901 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=152273.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:14:15,405 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.182e+02 1.492e+02 1.749e+02 2.149e+02 5.486e+02, threshold=3.498e+02, percent-clipped=3.0 2023-04-28 01:14:35,129 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=152286.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:14:54,601 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2820, 1.5744, 1.4099, 1.8478, 1.6992, 1.8097, 1.4266, 3.4279], device='cuda:4'), covar=tensor([0.0615, 0.0793, 0.0796, 0.1177, 0.0618, 0.0509, 0.0735, 0.0140], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 01:15:05,999 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6714, 1.4667, 1.6589, 1.9385, 1.9546, 1.6268, 1.3346, 1.7895], device='cuda:4'), covar=tensor([0.0653, 0.1064, 0.0690, 0.0464, 0.0524, 0.0694, 0.0702, 0.0511], device='cuda:4'), in_proj_covar=tensor([0.0183, 0.0199, 0.0182, 0.0168, 0.0176, 0.0176, 0.0150, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 01:15:09,162 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.8803, 2.2528, 2.1206, 2.8069, 2.9983, 2.5685, 2.5392, 2.0775], device='cuda:4'), covar=tensor([0.1591, 0.1695, 0.1925, 0.1436, 0.0996, 0.1695, 0.2313, 0.2396], device='cuda:4'), in_proj_covar=tensor([0.0316, 0.0310, 0.0353, 0.0288, 0.0329, 0.0307, 0.0302, 0.0377], device='cuda:4'), out_proj_covar=tensor([6.4534e-05, 6.3562e-05, 7.4220e-05, 5.7560e-05, 6.7221e-05, 6.4123e-05, 6.2448e-05, 7.9742e-05], device='cuda:4') 2023-04-28 01:15:17,723 INFO [finetune.py:976] (4/7) Epoch 27, batch 3400, loss[loss=0.2097, simple_loss=0.2821, pruned_loss=0.06858, over 4737.00 frames. ], tot_loss[loss=0.1714, simple_loss=0.2448, pruned_loss=0.04899, over 952590.86 frames. ], batch size: 54, lr: 2.92e-03, grad_scale: 32.0 2023-04-28 01:15:36,884 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=152334.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:15:50,221 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=152344.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:16:09,159 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-28 01:16:22,884 INFO [finetune.py:976] (4/7) Epoch 27, batch 3450, loss[loss=0.1342, simple_loss=0.2113, pruned_loss=0.02858, over 4741.00 frames. ], tot_loss[loss=0.1704, simple_loss=0.2442, pruned_loss=0.04829, over 952587.18 frames. ], batch size: 54, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:16:31,451 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.088e+02 1.589e+02 1.871e+02 2.255e+02 4.038e+02, threshold=3.742e+02, percent-clipped=2.0 2023-04-28 01:16:33,495 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.00 vs. limit=5.0 2023-04-28 01:16:53,822 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=152392.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:17:28,478 INFO [finetune.py:976] (4/7) Epoch 27, batch 3500, loss[loss=0.168, simple_loss=0.2453, pruned_loss=0.04532, over 4776.00 frames. ], tot_loss[loss=0.1679, simple_loss=0.2413, pruned_loss=0.04726, over 953462.80 frames. ], batch size: 26, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:17:59,814 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.38 vs. limit=5.0 2023-04-28 01:18:00,021 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=152445.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:18:18,958 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=152457.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:18:32,263 INFO [finetune.py:976] (4/7) Epoch 27, batch 3550, loss[loss=0.1453, simple_loss=0.2158, pruned_loss=0.03744, over 4923.00 frames. ], tot_loss[loss=0.1654, simple_loss=0.238, pruned_loss=0.04639, over 953478.93 frames. ], batch size: 38, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:18:34,193 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.5763, 4.5035, 3.1788, 5.2601, 4.6195, 4.5900, 2.1316, 4.4727], device='cuda:4'), covar=tensor([0.1436, 0.1115, 0.3104, 0.0988, 0.2920, 0.1550, 0.5431, 0.2028], device='cuda:4'), in_proj_covar=tensor([0.0245, 0.0218, 0.0251, 0.0303, 0.0298, 0.0248, 0.0274, 0.0273], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 01:18:34,714 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.620e+01 1.515e+02 1.760e+02 2.193e+02 3.921e+02, threshold=3.521e+02, percent-clipped=1.0 2023-04-28 01:18:52,612 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.70 vs. limit=2.0 2023-04-28 01:19:25,050 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=152505.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:19:25,751 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=152506.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:19:39,436 INFO [finetune.py:976] (4/7) Epoch 27, batch 3600, loss[loss=0.1649, simple_loss=0.2381, pruned_loss=0.04586, over 4819.00 frames. ], tot_loss[loss=0.1641, simple_loss=0.2363, pruned_loss=0.04598, over 954588.49 frames. ], batch size: 40, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:20:18,887 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.23 vs. limit=2.0 2023-04-28 01:20:32,645 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3699, 1.6174, 1.4587, 1.7590, 1.7092, 1.7553, 1.5190, 2.8188], device='cuda:4'), covar=tensor([0.0587, 0.0629, 0.0655, 0.0967, 0.0518, 0.0610, 0.0637, 0.0202], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 01:20:43,946 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.84 vs. limit=2.0 2023-04-28 01:20:44,729 INFO [finetune.py:976] (4/7) Epoch 27, batch 3650, loss[loss=0.1976, simple_loss=0.2745, pruned_loss=0.06034, over 4809.00 frames. ], tot_loss[loss=0.1664, simple_loss=0.2387, pruned_loss=0.04701, over 953516.46 frames. ], batch size: 41, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:20:46,056 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2313, 2.0782, 1.6923, 1.7752, 2.2526, 1.6991, 2.6463, 1.5419], device='cuda:4'), covar=tensor([0.3851, 0.2280, 0.5178, 0.3144, 0.1695, 0.2604, 0.1481, 0.4619], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0354, 0.0421, 0.0350, 0.0379, 0.0374, 0.0366, 0.0420], device='cuda:4'), out_proj_covar=tensor([9.9562e-05, 1.0528e-04, 1.2750e-04, 1.0480e-04, 1.1208e-04, 1.1116e-04, 1.0706e-04, 1.2633e-04], device='cuda:4') 2023-04-28 01:20:51,634 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.035e+02 1.532e+02 1.875e+02 2.206e+02 4.612e+02, threshold=3.749e+02, percent-clipped=4.0 2023-04-28 01:20:56,047 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=152581.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:21:02,382 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=152582.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:21:03,548 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2047, 1.5573, 2.0371, 2.2853, 2.0438, 1.5947, 1.2671, 1.8501], device='cuda:4'), covar=tensor([0.2665, 0.2990, 0.1489, 0.2171, 0.2377, 0.2459, 0.4232, 0.1824], device='cuda:4'), in_proj_covar=tensor([0.0295, 0.0248, 0.0230, 0.0317, 0.0223, 0.0235, 0.0230, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-28 01:21:12,788 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=152589.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:21:49,538 INFO [finetune.py:976] (4/7) Epoch 27, batch 3700, loss[loss=0.197, simple_loss=0.2711, pruned_loss=0.06147, over 4730.00 frames. ], tot_loss[loss=0.1669, simple_loss=0.2405, pruned_loss=0.04661, over 952080.12 frames. ], batch size: 54, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:22:00,774 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=152629.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:22:21,245 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=152643.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:22:31,086 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=152650.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:22:56,763 INFO [finetune.py:976] (4/7) Epoch 27, batch 3750, loss[loss=0.2828, simple_loss=0.3342, pruned_loss=0.1157, over 4839.00 frames. ], tot_loss[loss=0.1674, simple_loss=0.2418, pruned_loss=0.04652, over 952663.25 frames. ], batch size: 44, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:23:03,932 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6080, 1.7459, 1.5630, 2.1123, 2.0734, 2.2473, 1.7552, 4.5729], device='cuda:4'), covar=tensor([0.0520, 0.0798, 0.0784, 0.1146, 0.0597, 0.0458, 0.0695, 0.0093], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 01:23:04,991 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.032e+02 1.529e+02 1.747e+02 2.097e+02 4.287e+02, threshold=3.495e+02, percent-clipped=2.0 2023-04-28 01:23:18,957 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.90 vs. limit=5.0 2023-04-28 01:24:01,817 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.53 vs. limit=2.0 2023-04-28 01:24:08,008 INFO [finetune.py:976] (4/7) Epoch 27, batch 3800, loss[loss=0.1583, simple_loss=0.2304, pruned_loss=0.04311, over 4910.00 frames. ], tot_loss[loss=0.1686, simple_loss=0.2428, pruned_loss=0.04725, over 953174.51 frames. ], batch size: 36, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:24:11,805 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=152726.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:25:01,635 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5130, 1.9445, 1.6401, 2.3862, 2.4975, 2.0105, 2.1226, 1.7197], device='cuda:4'), covar=tensor([0.1765, 0.1536, 0.2178, 0.1540, 0.1099, 0.1998, 0.1924, 0.2527], device='cuda:4'), in_proj_covar=tensor([0.0316, 0.0308, 0.0353, 0.0286, 0.0328, 0.0306, 0.0300, 0.0376], device='cuda:4'), out_proj_covar=tensor([6.4347e-05, 6.3335e-05, 7.4085e-05, 5.7247e-05, 6.7027e-05, 6.3791e-05, 6.2155e-05, 7.9582e-05], device='cuda:4') 2023-04-28 01:25:13,176 INFO [finetune.py:976] (4/7) Epoch 27, batch 3850, loss[loss=0.1686, simple_loss=0.2417, pruned_loss=0.04779, over 4734.00 frames. ], tot_loss[loss=0.1679, simple_loss=0.2418, pruned_loss=0.04699, over 951817.77 frames. ], batch size: 54, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:25:15,589 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.111e+02 1.420e+02 1.784e+02 2.045e+02 3.500e+02, threshold=3.567e+02, percent-clipped=1.0 2023-04-28 01:25:33,576 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=152787.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:25:47,903 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=152801.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:26:10,605 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.6353, 2.1185, 2.4802, 3.0035, 2.4934, 1.9427, 2.0894, 2.4822], device='cuda:4'), covar=tensor([0.2737, 0.2811, 0.1436, 0.2080, 0.2514, 0.2386, 0.3428, 0.1926], device='cuda:4'), in_proj_covar=tensor([0.0294, 0.0248, 0.0229, 0.0315, 0.0222, 0.0235, 0.0229, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-28 01:26:17,504 INFO [finetune.py:976] (4/7) Epoch 27, batch 3900, loss[loss=0.1199, simple_loss=0.1918, pruned_loss=0.02402, over 4312.00 frames. ], tot_loss[loss=0.1657, simple_loss=0.2389, pruned_loss=0.04623, over 951778.48 frames. ], batch size: 65, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:26:51,242 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1238, 2.4033, 0.7846, 1.3091, 1.3471, 1.8020, 1.5536, 0.8530], device='cuda:4'), covar=tensor([0.2035, 0.1845, 0.2247, 0.2079, 0.1593, 0.1390, 0.1894, 0.1928], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0239, 0.0134, 0.0120, 0.0131, 0.0152, 0.0117, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 01:27:22,907 INFO [finetune.py:976] (4/7) Epoch 27, batch 3950, loss[loss=0.168, simple_loss=0.2319, pruned_loss=0.0521, over 4761.00 frames. ], tot_loss[loss=0.1636, simple_loss=0.2359, pruned_loss=0.04567, over 953945.77 frames. ], batch size: 23, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:27:26,329 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.030e+02 1.447e+02 1.748e+02 2.157e+02 5.230e+02, threshold=3.496e+02, percent-clipped=2.0 2023-04-28 01:27:36,690 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=152881.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:27:43,435 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6308, 3.3447, 0.9094, 1.8730, 1.6952, 2.3330, 1.8678, 0.9747], device='cuda:4'), covar=tensor([0.1315, 0.0911, 0.2002, 0.1200, 0.1183, 0.1038, 0.1544, 0.1947], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0238, 0.0134, 0.0120, 0.0131, 0.0152, 0.0117, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 01:27:47,544 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.27 vs. limit=2.0 2023-04-28 01:28:08,107 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-28 01:28:27,785 INFO [finetune.py:976] (4/7) Epoch 27, batch 4000, loss[loss=0.1352, simple_loss=0.2162, pruned_loss=0.02708, over 4785.00 frames. ], tot_loss[loss=0.1626, simple_loss=0.2348, pruned_loss=0.04519, over 954930.28 frames. ], batch size: 29, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:28:29,023 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.8419, 2.0514, 2.0832, 2.2313, 1.8957, 1.9862, 2.1883, 2.1098], device='cuda:4'), covar=tensor([0.3691, 0.6327, 0.5181, 0.4213, 0.6001, 0.6683, 0.6117, 0.5412], device='cuda:4'), in_proj_covar=tensor([0.0342, 0.0374, 0.0330, 0.0340, 0.0349, 0.0394, 0.0360, 0.0334], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 01:28:39,701 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=152929.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:28:39,727 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=152929.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:28:50,571 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=152938.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:29:00,138 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=152945.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:29:32,517 INFO [finetune.py:976] (4/7) Epoch 27, batch 4050, loss[loss=0.1868, simple_loss=0.2679, pruned_loss=0.05289, over 4821.00 frames. ], tot_loss[loss=0.1657, simple_loss=0.2385, pruned_loss=0.04644, over 951894.88 frames. ], batch size: 39, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:29:35,451 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.138e+02 1.640e+02 1.918e+02 2.261e+02 4.754e+02, threshold=3.835e+02, percent-clipped=2.0 2023-04-28 01:29:43,788 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=152977.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:29:57,222 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-04-28 01:30:03,171 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9211, 2.3435, 2.0949, 2.2858, 1.8038, 2.0281, 1.9756, 1.5151], device='cuda:4'), covar=tensor([0.1640, 0.1224, 0.0667, 0.0962, 0.2988, 0.1100, 0.1651, 0.2309], device='cuda:4'), in_proj_covar=tensor([0.0282, 0.0300, 0.0215, 0.0276, 0.0314, 0.0253, 0.0246, 0.0263], device='cuda:4'), out_proj_covar=tensor([1.1271e-04, 1.1800e-04, 8.4621e-05, 1.0851e-04, 1.2633e-04, 9.9420e-05, 9.9249e-05, 1.0373e-04], device='cuda:4') 2023-04-28 01:30:36,408 INFO [finetune.py:976] (4/7) Epoch 27, batch 4100, loss[loss=0.2325, simple_loss=0.2958, pruned_loss=0.0846, over 4927.00 frames. ], tot_loss[loss=0.1676, simple_loss=0.2408, pruned_loss=0.04716, over 953468.42 frames. ], batch size: 38, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:31:30,334 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8316, 2.2907, 1.9158, 1.8074, 1.3338, 1.4336, 2.0313, 1.3106], device='cuda:4'), covar=tensor([0.1689, 0.1334, 0.1350, 0.1538, 0.2351, 0.1869, 0.0928, 0.2036], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0210, 0.0170, 0.0204, 0.0201, 0.0186, 0.0156, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 01:31:40,934 INFO [finetune.py:976] (4/7) Epoch 27, batch 4150, loss[loss=0.1835, simple_loss=0.2646, pruned_loss=0.05116, over 4860.00 frames. ], tot_loss[loss=0.1664, simple_loss=0.2402, pruned_loss=0.04627, over 953192.34 frames. ], batch size: 31, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:31:43,390 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.012e+02 1.659e+02 1.933e+02 2.318e+02 5.641e+02, threshold=3.866e+02, percent-clipped=2.0 2023-04-28 01:31:59,395 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=153082.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:32:14,445 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.45 vs. limit=2.0 2023-04-28 01:32:22,878 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=153101.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:32:24,773 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5139, 1.6878, 1.8752, 1.9717, 1.7958, 1.8577, 1.9482, 2.0090], device='cuda:4'), covar=tensor([0.3808, 0.5420, 0.4684, 0.4764, 0.5509, 0.6855, 0.5347, 0.4576], device='cuda:4'), in_proj_covar=tensor([0.0341, 0.0373, 0.0329, 0.0340, 0.0349, 0.0392, 0.0359, 0.0333], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 01:32:34,881 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1686, 1.4141, 1.2627, 1.6159, 1.5178, 1.7187, 1.3141, 3.0381], device='cuda:4'), covar=tensor([0.0602, 0.0826, 0.0798, 0.1258, 0.0630, 0.0465, 0.0749, 0.0161], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0039, 0.0043, 0.0040, 0.0037, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 01:32:36,420 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.09 vs. limit=2.0 2023-04-28 01:32:43,325 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=153117.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:32:45,076 INFO [finetune.py:976] (4/7) Epoch 27, batch 4200, loss[loss=0.1553, simple_loss=0.2362, pruned_loss=0.03721, over 4905.00 frames. ], tot_loss[loss=0.1656, simple_loss=0.24, pruned_loss=0.04561, over 952827.72 frames. ], batch size: 37, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:33:26,254 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=153149.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:33:49,141 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.9384, 3.9233, 2.9224, 4.4853, 3.9292, 3.8980, 2.1098, 3.8803], device='cuda:4'), covar=tensor([0.1597, 0.1023, 0.3086, 0.1673, 0.3028, 0.1797, 0.4963, 0.2229], device='cuda:4'), in_proj_covar=tensor([0.0245, 0.0218, 0.0251, 0.0302, 0.0299, 0.0248, 0.0273, 0.0273], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 01:33:49,693 INFO [finetune.py:976] (4/7) Epoch 27, batch 4250, loss[loss=0.1549, simple_loss=0.2269, pruned_loss=0.04143, over 4114.00 frames. ], tot_loss[loss=0.1655, simple_loss=0.2393, pruned_loss=0.04586, over 953215.71 frames. ], batch size: 18, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:33:57,256 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.090e+02 1.449e+02 1.716e+02 2.053e+02 3.736e+02, threshold=3.432e+02, percent-clipped=0.0 2023-04-28 01:33:59,820 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=153178.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:34:12,175 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6385, 2.6463, 2.0909, 2.2783, 2.5860, 2.2724, 3.4230, 1.9325], device='cuda:4'), covar=tensor([0.3693, 0.2078, 0.4570, 0.3139, 0.1933, 0.2370, 0.1677, 0.4624], device='cuda:4'), in_proj_covar=tensor([0.0337, 0.0352, 0.0421, 0.0349, 0.0379, 0.0372, 0.0366, 0.0420], device='cuda:4'), out_proj_covar=tensor([9.9465e-05, 1.0479e-04, 1.2737e-04, 1.0451e-04, 1.1229e-04, 1.1047e-04, 1.0687e-04, 1.2633e-04], device='cuda:4') 2023-04-28 01:34:51,799 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2613, 1.3381, 3.7493, 3.5232, 3.2815, 3.5597, 3.5238, 3.3153], device='cuda:4'), covar=tensor([0.7137, 0.5583, 0.1202, 0.1672, 0.1224, 0.2006, 0.2113, 0.1524], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0308, 0.0408, 0.0407, 0.0348, 0.0417, 0.0318, 0.0364], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-28 01:34:54,746 INFO [finetune.py:976] (4/7) Epoch 27, batch 4300, loss[loss=0.15, simple_loss=0.2256, pruned_loss=0.03718, over 4790.00 frames. ], tot_loss[loss=0.1641, simple_loss=0.237, pruned_loss=0.04559, over 954489.31 frames. ], batch size: 29, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:35:17,351 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=153238.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:35:23,569 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=153245.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:35:25,998 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8202, 3.5633, 1.0904, 1.9221, 2.0000, 2.5915, 2.0380, 1.1463], device='cuda:4'), covar=tensor([0.1222, 0.0730, 0.1800, 0.1218, 0.1035, 0.0949, 0.1605, 0.1782], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0240, 0.0136, 0.0122, 0.0133, 0.0153, 0.0118, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 01:35:33,114 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.57 vs. limit=2.0 2023-04-28 01:35:34,540 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8120, 2.1093, 1.7265, 1.4943, 1.3575, 1.3849, 1.7710, 1.2940], device='cuda:4'), covar=tensor([0.1579, 0.1208, 0.1280, 0.1622, 0.2192, 0.1818, 0.0974, 0.2026], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0210, 0.0170, 0.0205, 0.0201, 0.0187, 0.0156, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 01:35:38,705 INFO [finetune.py:976] (4/7) Epoch 27, batch 4350, loss[loss=0.1162, simple_loss=0.1899, pruned_loss=0.02129, over 4786.00 frames. ], tot_loss[loss=0.1614, simple_loss=0.234, pruned_loss=0.04443, over 952547.43 frames. ], batch size: 26, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:35:41,101 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.135e+02 1.585e+02 1.848e+02 2.377e+02 5.336e+02, threshold=3.696e+02, percent-clipped=3.0 2023-04-28 01:35:44,137 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=153278.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:35:48,972 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=153286.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:35:54,194 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=153293.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:36:12,552 INFO [finetune.py:976] (4/7) Epoch 27, batch 4400, loss[loss=0.2005, simple_loss=0.2681, pruned_loss=0.06646, over 4821.00 frames. ], tot_loss[loss=0.1629, simple_loss=0.2353, pruned_loss=0.04524, over 952318.55 frames. ], batch size: 39, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:36:33,309 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=153339.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:37:16,506 INFO [finetune.py:976] (4/7) Epoch 27, batch 4450, loss[loss=0.1567, simple_loss=0.2331, pruned_loss=0.04019, over 4836.00 frames. ], tot_loss[loss=0.1648, simple_loss=0.238, pruned_loss=0.04581, over 949908.95 frames. ], batch size: 30, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:37:18,877 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.171e+02 1.598e+02 1.820e+02 2.307e+02 3.312e+02, threshold=3.640e+02, percent-clipped=0.0 2023-04-28 01:37:18,984 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1326, 2.4763, 1.0393, 1.3285, 1.9750, 1.2561, 3.3639, 1.8293], device='cuda:4'), covar=tensor([0.0750, 0.0692, 0.0852, 0.1396, 0.0542, 0.1053, 0.0316, 0.0652], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0064, 0.0047, 0.0046, 0.0050, 0.0051, 0.0073, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-28 01:37:30,038 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=153382.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:38:10,736 INFO [finetune.py:976] (4/7) Epoch 27, batch 4500, loss[loss=0.2067, simple_loss=0.282, pruned_loss=0.0657, over 4752.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.2405, pruned_loss=0.04683, over 951499.16 frames. ], batch size: 54, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:38:11,440 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=153421.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:38:16,910 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=153430.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:38:18,873 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.24 vs. limit=2.0 2023-04-28 01:38:43,733 INFO [finetune.py:976] (4/7) Epoch 27, batch 4550, loss[loss=0.181, simple_loss=0.2533, pruned_loss=0.05432, over 4824.00 frames. ], tot_loss[loss=0.1682, simple_loss=0.242, pruned_loss=0.04725, over 950642.74 frames. ], batch size: 30, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:38:45,570 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=153473.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:38:46,107 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.179e+01 1.577e+02 1.859e+02 2.196e+02 3.775e+02, threshold=3.717e+02, percent-clipped=3.0 2023-04-28 01:38:51,086 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=153482.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:39:00,169 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7020, 1.0495, 1.7216, 2.1594, 1.7731, 1.6472, 1.7140, 1.6560], device='cuda:4'), covar=tensor([0.4186, 0.6595, 0.5806, 0.5195, 0.5505, 0.7127, 0.7297, 0.8455], device='cuda:4'), in_proj_covar=tensor([0.0444, 0.0425, 0.0520, 0.0509, 0.0471, 0.0509, 0.0510, 0.0526], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 01:39:15,749 INFO [finetune.py:976] (4/7) Epoch 27, batch 4600, loss[loss=0.1987, simple_loss=0.2766, pruned_loss=0.06045, over 4718.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.2415, pruned_loss=0.04694, over 950687.77 frames. ], batch size: 23, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:39:48,981 INFO [finetune.py:976] (4/7) Epoch 27, batch 4650, loss[loss=0.164, simple_loss=0.2371, pruned_loss=0.04545, over 4769.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.2402, pruned_loss=0.04706, over 951259.90 frames. ], batch size: 26, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:39:51,381 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.905e+01 1.489e+02 1.829e+02 2.275e+02 4.563e+02, threshold=3.657e+02, percent-clipped=2.0 2023-04-28 01:40:08,526 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.85 vs. limit=2.0 2023-04-28 01:40:27,424 INFO [finetune.py:976] (4/7) Epoch 27, batch 4700, loss[loss=0.1494, simple_loss=0.2149, pruned_loss=0.04196, over 4817.00 frames. ], tot_loss[loss=0.1644, simple_loss=0.2372, pruned_loss=0.04583, over 952891.12 frames. ], batch size: 25, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:40:41,152 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=153634.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:41:06,296 INFO [finetune.py:976] (4/7) Epoch 27, batch 4750, loss[loss=0.1241, simple_loss=0.2011, pruned_loss=0.02352, over 4718.00 frames. ], tot_loss[loss=0.1627, simple_loss=0.235, pruned_loss=0.04522, over 952753.07 frames. ], batch size: 23, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:41:08,702 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.046e+02 1.498e+02 1.801e+02 2.144e+02 5.776e+02, threshold=3.603e+02, percent-clipped=1.0 2023-04-28 01:41:39,700 INFO [finetune.py:976] (4/7) Epoch 27, batch 4800, loss[loss=0.1639, simple_loss=0.2422, pruned_loss=0.0428, over 4826.00 frames. ], tot_loss[loss=0.1647, simple_loss=0.2374, pruned_loss=0.04602, over 953465.50 frames. ], batch size: 30, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:42:03,566 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0013, 2.6329, 1.0998, 1.3827, 1.9658, 1.2614, 3.4325, 1.8535], device='cuda:4'), covar=tensor([0.0711, 0.0615, 0.0774, 0.1163, 0.0526, 0.1001, 0.0199, 0.0601], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0064, 0.0047, 0.0046, 0.0050, 0.0051, 0.0072, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-28 01:42:13,211 INFO [finetune.py:976] (4/7) Epoch 27, batch 4850, loss[loss=0.193, simple_loss=0.2707, pruned_loss=0.05763, over 4921.00 frames. ], tot_loss[loss=0.1678, simple_loss=0.2415, pruned_loss=0.04705, over 954191.33 frames. ], batch size: 38, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:42:15,634 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=153773.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:42:16,122 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.154e+02 1.558e+02 1.785e+02 2.146e+02 3.572e+02, threshold=3.570e+02, percent-clipped=0.0 2023-04-28 01:42:17,996 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=153777.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:42:44,727 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9473, 1.9127, 1.7213, 1.4975, 2.0201, 1.6344, 2.5426, 1.5371], device='cuda:4'), covar=tensor([0.3514, 0.1862, 0.5149, 0.3063, 0.1627, 0.2396, 0.1472, 0.4671], device='cuda:4'), in_proj_covar=tensor([0.0338, 0.0355, 0.0424, 0.0351, 0.0380, 0.0374, 0.0369, 0.0422], device='cuda:4'), out_proj_covar=tensor([9.9878e-05, 1.0563e-04, 1.2844e-04, 1.0511e-04, 1.1236e-04, 1.1121e-04, 1.0798e-04, 1.2682e-04], device='cuda:4') 2023-04-28 01:43:04,189 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2355, 2.7977, 2.4432, 2.6747, 1.9364, 2.4754, 2.2777, 1.8756], device='cuda:4'), covar=tensor([0.1964, 0.1179, 0.0742, 0.1109, 0.3191, 0.1008, 0.1894, 0.2489], device='cuda:4'), in_proj_covar=tensor([0.0283, 0.0300, 0.0215, 0.0276, 0.0315, 0.0253, 0.0247, 0.0263], device='cuda:4'), out_proj_covar=tensor([1.1275e-04, 1.1801e-04, 8.4550e-05, 1.0840e-04, 1.2692e-04, 9.9387e-05, 9.9598e-05, 1.0362e-04], device='cuda:4') 2023-04-28 01:43:05,707 INFO [finetune.py:976] (4/7) Epoch 27, batch 4900, loss[loss=0.223, simple_loss=0.291, pruned_loss=0.07752, over 4905.00 frames. ], tot_loss[loss=0.1695, simple_loss=0.2427, pruned_loss=0.04809, over 954001.28 frames. ], batch size: 36, lr: 2.91e-03, grad_scale: 32.0 2023-04-28 01:43:05,828 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0795, 1.9939, 1.7815, 1.5674, 2.0767, 1.6644, 2.6441, 1.6255], device='cuda:4'), covar=tensor([0.3575, 0.1916, 0.4693, 0.3247, 0.1661, 0.2425, 0.1277, 0.4108], device='cuda:4'), in_proj_covar=tensor([0.0338, 0.0355, 0.0424, 0.0351, 0.0379, 0.0374, 0.0369, 0.0422], device='cuda:4'), out_proj_covar=tensor([9.9797e-05, 1.0564e-04, 1.2840e-04, 1.0506e-04, 1.1227e-04, 1.1111e-04, 1.0791e-04, 1.2680e-04], device='cuda:4') 2023-04-28 01:43:06,909 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=153821.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:44:09,303 INFO [finetune.py:976] (4/7) Epoch 27, batch 4950, loss[loss=0.1891, simple_loss=0.2691, pruned_loss=0.0545, over 4815.00 frames. ], tot_loss[loss=0.1699, simple_loss=0.2439, pruned_loss=0.04793, over 956207.00 frames. ], batch size: 38, lr: 2.90e-03, grad_scale: 32.0 2023-04-28 01:44:17,535 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0802, 2.3557, 2.2644, 2.6815, 2.5374, 2.6896, 2.3143, 4.9046], device='cuda:4'), covar=tensor([0.0460, 0.0663, 0.0675, 0.1000, 0.0533, 0.0420, 0.0573, 0.0117], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0039, 0.0042, 0.0040, 0.0037, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 01:44:18,026 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.075e+02 1.601e+02 1.869e+02 2.254e+02 5.628e+02, threshold=3.738e+02, percent-clipped=5.0 2023-04-28 01:44:51,821 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6175, 1.5868, 0.8873, 1.3329, 1.8944, 1.4839, 1.3993, 1.5065], device='cuda:4'), covar=tensor([0.0479, 0.0376, 0.0322, 0.0542, 0.0267, 0.0484, 0.0477, 0.0561], device='cuda:4'), in_proj_covar=tensor([0.0027, 0.0023, 0.0022, 0.0028, 0.0019, 0.0027, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0052, 0.0046, 0.0038, 0.0053, 0.0039, 0.0051, 0.0051, 0.0052], device='cuda:4') 2023-04-28 01:45:13,176 INFO [finetune.py:976] (4/7) Epoch 27, batch 5000, loss[loss=0.1627, simple_loss=0.2353, pruned_loss=0.04507, over 4901.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.2414, pruned_loss=0.04678, over 956284.52 frames. ], batch size: 37, lr: 2.90e-03, grad_scale: 32.0 2023-04-28 01:45:21,159 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=153923.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:45:34,547 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=153934.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:46:15,671 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8873, 1.6671, 2.0109, 2.2067, 2.2975, 1.6955, 1.5654, 1.9922], device='cuda:4'), covar=tensor([0.0817, 0.1211, 0.0684, 0.0690, 0.0603, 0.0866, 0.0722, 0.0591], device='cuda:4'), in_proj_covar=tensor([0.0186, 0.0203, 0.0185, 0.0173, 0.0179, 0.0180, 0.0153, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 01:46:17,425 INFO [finetune.py:976] (4/7) Epoch 27, batch 5050, loss[loss=0.1574, simple_loss=0.2253, pruned_loss=0.0447, over 4830.00 frames. ], tot_loss[loss=0.1651, simple_loss=0.2381, pruned_loss=0.04603, over 955949.16 frames. ], batch size: 30, lr: 2.90e-03, grad_scale: 32.0 2023-04-28 01:46:25,380 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.912e+01 1.615e+02 1.861e+02 2.291e+02 3.934e+02, threshold=3.722e+02, percent-clipped=1.0 2023-04-28 01:46:32,333 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=153982.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:46:34,111 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=153984.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:46:57,619 INFO [finetune.py:976] (4/7) Epoch 27, batch 5100, loss[loss=0.1542, simple_loss=0.2262, pruned_loss=0.04106, over 4821.00 frames. ], tot_loss[loss=0.161, simple_loss=0.234, pruned_loss=0.04401, over 956854.94 frames. ], batch size: 41, lr: 2.90e-03, grad_scale: 32.0 2023-04-28 01:47:18,517 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=154049.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 01:47:28,931 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.5031, 1.3714, 1.4446, 1.0503, 1.4037, 1.1919, 1.7465, 1.3472], device='cuda:4'), covar=tensor([0.3419, 0.1756, 0.4554, 0.2513, 0.1460, 0.2048, 0.1499, 0.4448], device='cuda:4'), in_proj_covar=tensor([0.0338, 0.0353, 0.0422, 0.0350, 0.0379, 0.0374, 0.0368, 0.0420], device='cuda:4'), out_proj_covar=tensor([9.9649e-05, 1.0522e-04, 1.2782e-04, 1.0480e-04, 1.1214e-04, 1.1119e-04, 1.0749e-04, 1.2628e-04], device='cuda:4') 2023-04-28 01:47:31,250 INFO [finetune.py:976] (4/7) Epoch 27, batch 5150, loss[loss=0.2004, simple_loss=0.2749, pruned_loss=0.06295, over 4817.00 frames. ], tot_loss[loss=0.1609, simple_loss=0.234, pruned_loss=0.04388, over 954214.23 frames. ], batch size: 38, lr: 2.90e-03, grad_scale: 32.0 2023-04-28 01:47:33,691 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.527e+01 1.424e+02 1.690e+02 2.200e+02 4.419e+02, threshold=3.381e+02, percent-clipped=1.0 2023-04-28 01:47:35,042 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=154076.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:47:35,621 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=154077.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:47:41,332 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=154084.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:47:41,348 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3537, 1.8087, 1.5252, 2.1586, 2.3397, 1.9261, 1.8874, 1.5762], device='cuda:4'), covar=tensor([0.1840, 0.1492, 0.1828, 0.1395, 0.1016, 0.1748, 0.1932, 0.2423], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0305, 0.0349, 0.0284, 0.0325, 0.0303, 0.0299, 0.0374], device='cuda:4'), out_proj_covar=tensor([6.3812e-05, 6.2691e-05, 7.3277e-05, 5.6744e-05, 6.6439e-05, 6.3204e-05, 6.1810e-05, 7.9120e-05], device='cuda:4') 2023-04-28 01:48:06,059 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.48 vs. limit=2.0 2023-04-28 01:48:15,436 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=154110.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 01:48:26,896 INFO [finetune.py:976] (4/7) Epoch 27, batch 5200, loss[loss=0.1991, simple_loss=0.2732, pruned_loss=0.06253, over 4913.00 frames. ], tot_loss[loss=0.1654, simple_loss=0.2388, pruned_loss=0.04605, over 952308.76 frames. ], batch size: 36, lr: 2.90e-03, grad_scale: 32.0 2023-04-28 01:48:35,436 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=154125.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:48:49,925 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=154137.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 01:49:00,255 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=154145.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:49:09,708 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.85 vs. limit=2.0 2023-04-28 01:49:31,139 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.8737, 2.8303, 2.1346, 3.2663, 2.8500, 2.8543, 1.2074, 2.8183], device='cuda:4'), covar=tensor([0.2086, 0.1669, 0.3505, 0.3008, 0.3310, 0.2064, 0.5486, 0.2875], device='cuda:4'), in_proj_covar=tensor([0.0244, 0.0218, 0.0250, 0.0303, 0.0298, 0.0247, 0.0273, 0.0272], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 01:49:31,678 INFO [finetune.py:976] (4/7) Epoch 27, batch 5250, loss[loss=0.188, simple_loss=0.2622, pruned_loss=0.05689, over 4832.00 frames. ], tot_loss[loss=0.1674, simple_loss=0.2413, pruned_loss=0.04673, over 953137.86 frames. ], batch size: 49, lr: 2.90e-03, grad_scale: 32.0 2023-04-28 01:49:34,128 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.088e+02 1.655e+02 1.963e+02 2.223e+02 5.480e+02, threshold=3.927e+02, percent-clipped=5.0 2023-04-28 01:50:24,936 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9131, 1.7801, 2.3080, 2.3310, 1.7732, 1.5415, 1.8959, 1.0130], device='cuda:4'), covar=tensor([0.0592, 0.0708, 0.0351, 0.0708, 0.0672, 0.1039, 0.0588, 0.0656], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0067, 0.0066, 0.0069, 0.0075, 0.0095, 0.0073, 0.0063], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-28 01:50:25,960 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.45 vs. limit=2.0 2023-04-28 01:50:35,791 INFO [finetune.py:976] (4/7) Epoch 27, batch 5300, loss[loss=0.1917, simple_loss=0.2678, pruned_loss=0.05782, over 4907.00 frames. ], tot_loss[loss=0.1703, simple_loss=0.2446, pruned_loss=0.04802, over 955462.04 frames. ], batch size: 36, lr: 2.90e-03, grad_scale: 32.0 2023-04-28 01:50:36,542 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=154221.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:51:41,570 INFO [finetune.py:976] (4/7) Epoch 27, batch 5350, loss[loss=0.1498, simple_loss=0.2149, pruned_loss=0.04235, over 4811.00 frames. ], tot_loss[loss=0.1691, simple_loss=0.2439, pruned_loss=0.04717, over 956162.65 frames. ], batch size: 25, lr: 2.90e-03, grad_scale: 32.0 2023-04-28 01:51:49,111 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.757e+01 1.510e+02 1.847e+02 2.221e+02 4.452e+02, threshold=3.694e+02, percent-clipped=2.0 2023-04-28 01:51:52,262 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=154279.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:51:54,122 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=154282.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 01:51:56,594 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=154286.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:52:00,123 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=154291.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:52:06,615 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.58 vs. limit=2.0 2023-04-28 01:52:19,625 INFO [finetune.py:976] (4/7) Epoch 27, batch 5400, loss[loss=0.1198, simple_loss=0.2042, pruned_loss=0.01771, over 4781.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.2411, pruned_loss=0.04659, over 955408.89 frames. ], batch size: 29, lr: 2.90e-03, grad_scale: 64.0 2023-04-28 01:52:30,057 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=154337.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 01:52:37,689 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=154347.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:52:41,219 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=154352.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:52:52,200 INFO [finetune.py:976] (4/7) Epoch 27, batch 5450, loss[loss=0.1426, simple_loss=0.2123, pruned_loss=0.03652, over 4839.00 frames. ], tot_loss[loss=0.1651, simple_loss=0.2383, pruned_loss=0.0459, over 957690.04 frames. ], batch size: 40, lr: 2.90e-03, grad_scale: 64.0 2023-04-28 01:52:54,629 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.131e+02 1.557e+02 1.812e+02 2.063e+02 3.462e+02, threshold=3.624e+02, percent-clipped=0.0 2023-04-28 01:53:09,446 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=154398.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 01:53:16,764 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=154405.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 01:53:20,466 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1574, 2.6756, 2.2726, 2.5400, 1.8648, 2.3577, 2.2667, 1.8008], device='cuda:4'), covar=tensor([0.1726, 0.1012, 0.0712, 0.1049, 0.3125, 0.0915, 0.1860, 0.2378], device='cuda:4'), in_proj_covar=tensor([0.0283, 0.0302, 0.0216, 0.0277, 0.0314, 0.0254, 0.0248, 0.0262], device='cuda:4'), out_proj_covar=tensor([1.1299e-04, 1.1871e-04, 8.4936e-05, 1.0889e-04, 1.2676e-04, 1.0001e-04, 9.9976e-05, 1.0340e-04], device='cuda:4') 2023-04-28 01:53:31,462 INFO [finetune.py:976] (4/7) Epoch 27, batch 5500, loss[loss=0.171, simple_loss=0.2459, pruned_loss=0.04806, over 4851.00 frames. ], tot_loss[loss=0.163, simple_loss=0.2358, pruned_loss=0.04508, over 958321.32 frames. ], batch size: 44, lr: 2.90e-03, grad_scale: 64.0 2023-04-28 01:53:44,350 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=154432.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 01:53:53,051 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.11 vs. limit=2.0 2023-04-28 01:53:54,046 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=154440.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:54:35,187 INFO [finetune.py:976] (4/7) Epoch 27, batch 5550, loss[loss=0.2136, simple_loss=0.2861, pruned_loss=0.07056, over 4764.00 frames. ], tot_loss[loss=0.1648, simple_loss=0.2379, pruned_loss=0.04587, over 959593.23 frames. ], batch size: 54, lr: 2.90e-03, grad_scale: 64.0 2023-04-28 01:54:37,625 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.014e+02 1.508e+02 1.736e+02 2.106e+02 4.174e+02, threshold=3.472e+02, percent-clipped=1.0 2023-04-28 01:54:37,743 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=154474.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:55:07,209 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9493, 1.0095, 1.2358, 1.1806, 0.9637, 0.8998, 0.9965, 0.5798], device='cuda:4'), covar=tensor([0.0552, 0.0651, 0.0466, 0.0500, 0.0686, 0.1225, 0.0450, 0.0613], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0067, 0.0065, 0.0069, 0.0075, 0.0094, 0.0072, 0.0062], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-28 01:55:31,959 INFO [finetune.py:976] (4/7) Epoch 27, batch 5600, loss[loss=0.181, simple_loss=0.2633, pruned_loss=0.04939, over 4841.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.241, pruned_loss=0.04629, over 955723.60 frames. ], batch size: 49, lr: 2.90e-03, grad_scale: 64.0 2023-04-28 01:55:51,026 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=154535.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:56:01,155 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.64 vs. limit=2.0 2023-04-28 01:56:32,738 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4839, 1.0684, 0.4191, 1.1356, 1.0248, 1.3420, 1.2343, 1.2260], device='cuda:4'), covar=tensor([0.0514, 0.0410, 0.0401, 0.0580, 0.0315, 0.0531, 0.0535, 0.0592], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0028, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0052, 0.0046, 0.0039, 0.0053, 0.0039, 0.0051, 0.0051, 0.0053], device='cuda:4') 2023-04-28 01:56:35,595 INFO [finetune.py:976] (4/7) Epoch 27, batch 5650, loss[loss=0.1378, simple_loss=0.2066, pruned_loss=0.03449, over 3972.00 frames. ], tot_loss[loss=0.1676, simple_loss=0.2426, pruned_loss=0.04628, over 956113.75 frames. ], batch size: 17, lr: 2.90e-03, grad_scale: 64.0 2023-04-28 01:56:42,720 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.767e+01 1.509e+02 1.920e+02 2.295e+02 5.018e+02, threshold=3.840e+02, percent-clipped=4.0 2023-04-28 01:56:44,491 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=154577.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 01:56:45,665 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=154579.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:57:17,097 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6700, 2.1088, 1.7923, 2.0492, 1.7063, 1.7401, 1.7870, 1.4516], device='cuda:4'), covar=tensor([0.1577, 0.1266, 0.0826, 0.1076, 0.3084, 0.1275, 0.1708, 0.2129], device='cuda:4'), in_proj_covar=tensor([0.0284, 0.0304, 0.0217, 0.0277, 0.0315, 0.0256, 0.0248, 0.0263], device='cuda:4'), out_proj_covar=tensor([1.1351e-04, 1.1955e-04, 8.5337e-05, 1.0896e-04, 1.2716e-04, 1.0041e-04, 1.0022e-04, 1.0381e-04], device='cuda:4') 2023-04-28 01:57:36,098 INFO [finetune.py:976] (4/7) Epoch 27, batch 5700, loss[loss=0.1404, simple_loss=0.2035, pruned_loss=0.03867, over 3984.00 frames. ], tot_loss[loss=0.1663, simple_loss=0.2394, pruned_loss=0.04664, over 935840.54 frames. ], batch size: 17, lr: 2.90e-03, grad_scale: 32.0 2023-04-28 01:57:36,181 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2378, 3.2324, 2.8518, 3.0459, 3.3416, 2.8119, 4.0974, 2.7785], device='cuda:4'), covar=tensor([0.2821, 0.1506, 0.2412, 0.2052, 0.1216, 0.2080, 0.1039, 0.2485], device='cuda:4'), in_proj_covar=tensor([0.0338, 0.0354, 0.0425, 0.0351, 0.0380, 0.0377, 0.0370, 0.0423], device='cuda:4'), out_proj_covar=tensor([9.9833e-05, 1.0529e-04, 1.2857e-04, 1.0501e-04, 1.1237e-04, 1.1208e-04, 1.0813e-04, 1.2723e-04], device='cuda:4') 2023-04-28 01:57:45,550 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=154627.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:57:46,206 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=154628.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:57:59,106 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=154642.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:58:12,012 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=154647.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:58:12,561 INFO [finetune.py:976] (4/7) Epoch 28, batch 0, loss[loss=0.1988, simple_loss=0.2761, pruned_loss=0.06077, over 4920.00 frames. ], tot_loss[loss=0.1988, simple_loss=0.2761, pruned_loss=0.06077, over 4920.00 frames. ], batch size: 42, lr: 2.90e-03, grad_scale: 32.0 2023-04-28 01:58:12,561 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-28 01:58:17,803 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6321, 1.2954, 1.4074, 1.3844, 1.7926, 1.4983, 1.2569, 1.3789], device='cuda:4'), covar=tensor([0.2094, 0.1325, 0.2116, 0.1431, 0.0969, 0.1608, 0.2028, 0.2892], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0306, 0.0350, 0.0286, 0.0325, 0.0304, 0.0300, 0.0375], device='cuda:4'), out_proj_covar=tensor([6.3781e-05, 6.2941e-05, 7.3412e-05, 5.7246e-05, 6.6436e-05, 6.3401e-05, 6.2023e-05, 7.9490e-05], device='cuda:4') 2023-04-28 01:58:29,430 INFO [finetune.py:1010] (4/7) Epoch 28, validation: loss=0.1549, simple_loss=0.224, pruned_loss=0.04297, over 2265189.00 frames. 2023-04-28 01:58:29,430 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6529MB 2023-04-28 01:58:33,234 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0506, 2.5496, 1.0550, 1.3878, 1.9172, 1.1662, 3.3198, 1.7416], device='cuda:4'), covar=tensor([0.0685, 0.0689, 0.0832, 0.1146, 0.0513, 0.1021, 0.0199, 0.0618], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0064, 0.0047, 0.0046, 0.0050, 0.0051, 0.0073, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-28 01:58:43,442 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0973, 4.0726, 0.7777, 2.3748, 2.4699, 2.8697, 2.3662, 1.0366], device='cuda:4'), covar=tensor([0.1355, 0.1249, 0.2333, 0.1247, 0.1024, 0.1092, 0.1653, 0.2261], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0240, 0.0136, 0.0121, 0.0132, 0.0152, 0.0118, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 01:58:55,769 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.710e+01 1.461e+02 1.801e+02 2.323e+02 7.600e+02, threshold=3.601e+02, percent-clipped=3.0 2023-04-28 01:59:04,494 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=154689.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:59:06,870 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=154693.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 01:59:09,850 INFO [finetune.py:976] (4/7) Epoch 28, batch 50, loss[loss=0.1765, simple_loss=0.2471, pruned_loss=0.05301, over 4883.00 frames. ], tot_loss[loss=0.1703, simple_loss=0.2443, pruned_loss=0.04821, over 215647.03 frames. ], batch size: 35, lr: 2.90e-03, grad_scale: 32.0 2023-04-28 01:59:17,153 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=154705.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 01:59:26,317 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=154720.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:59:30,384 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.54 vs. limit=2.0 2023-04-28 01:59:33,621 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=154732.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:59:38,569 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=154740.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 01:59:43,316 INFO [finetune.py:976] (4/7) Epoch 28, batch 100, loss[loss=0.1591, simple_loss=0.2313, pruned_loss=0.04352, over 4825.00 frames. ], tot_loss[loss=0.1628, simple_loss=0.2355, pruned_loss=0.04501, over 380737.60 frames. ], batch size: 33, lr: 2.90e-03, grad_scale: 32.0 2023-04-28 01:59:48,389 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=154753.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 02:00:02,206 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.686e+01 1.538e+02 1.845e+02 2.159e+02 3.671e+02, threshold=3.690e+02, percent-clipped=1.0 2023-04-28 02:00:05,349 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=154780.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:00:06,015 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=154781.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:00:10,233 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=154788.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:00:16,273 INFO [finetune.py:976] (4/7) Epoch 28, batch 150, loss[loss=0.1275, simple_loss=0.2097, pruned_loss=0.02264, over 4828.00 frames. ], tot_loss[loss=0.16, simple_loss=0.2324, pruned_loss=0.04382, over 509919.01 frames. ], batch size: 40, lr: 2.90e-03, grad_scale: 32.0 2023-04-28 02:00:38,346 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=154830.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:00:49,254 INFO [finetune.py:976] (4/7) Epoch 28, batch 200, loss[loss=0.1762, simple_loss=0.245, pruned_loss=0.05374, over 4926.00 frames. ], tot_loss[loss=0.1592, simple_loss=0.2315, pruned_loss=0.04345, over 608758.51 frames. ], batch size: 38, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:01:08,233 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.084e+02 1.505e+02 1.793e+02 2.296e+02 3.844e+02, threshold=3.586e+02, percent-clipped=2.0 2023-04-28 02:01:09,541 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=154877.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 02:01:22,151 INFO [finetune.py:976] (4/7) Epoch 28, batch 250, loss[loss=0.1609, simple_loss=0.2367, pruned_loss=0.04255, over 4816.00 frames. ], tot_loss[loss=0.1624, simple_loss=0.2353, pruned_loss=0.04476, over 687730.81 frames. ], batch size: 39, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:01:34,802 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0453, 1.2807, 4.8319, 4.5260, 4.2064, 4.6334, 4.3519, 4.2647], device='cuda:4'), covar=tensor([0.7111, 0.6295, 0.1073, 0.1862, 0.1251, 0.1466, 0.1727, 0.1712], device='cuda:4'), in_proj_covar=tensor([0.0315, 0.0313, 0.0411, 0.0412, 0.0352, 0.0422, 0.0322, 0.0370], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-28 02:01:41,470 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=154925.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:01:51,947 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=154942.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:02:00,124 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=154947.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:02:00,648 INFO [finetune.py:976] (4/7) Epoch 28, batch 300, loss[loss=0.1785, simple_loss=0.2647, pruned_loss=0.04612, over 4862.00 frames. ], tot_loss[loss=0.167, simple_loss=0.2406, pruned_loss=0.04675, over 747759.06 frames. ], batch size: 34, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:02:36,612 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.011e+02 1.547e+02 1.784e+02 2.160e+02 3.859e+02, threshold=3.568e+02, percent-clipped=1.0 2023-04-28 02:02:37,340 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=154976.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:02:47,491 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=154984.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:02:56,119 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=154990.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:02:56,763 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3087, 1.6083, 1.4927, 1.7563, 1.8009, 1.9296, 1.4871, 3.7361], device='cuda:4'), covar=tensor([0.0600, 0.0800, 0.0774, 0.1263, 0.0595, 0.0508, 0.0752, 0.0148], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 02:02:57,984 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=154993.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 02:02:59,161 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=154995.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:03:06,551 INFO [finetune.py:976] (4/7) Epoch 28, batch 350, loss[loss=0.1568, simple_loss=0.2421, pruned_loss=0.03576, over 4886.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.2403, pruned_loss=0.04697, over 791340.09 frames. ], batch size: 35, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:03:45,296 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=155037.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:03:47,613 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=155041.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 02:03:51,702 INFO [finetune.py:976] (4/7) Epoch 28, batch 400, loss[loss=0.2098, simple_loss=0.2741, pruned_loss=0.07279, over 4818.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.2417, pruned_loss=0.04685, over 828636.54 frames. ], batch size: 38, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:04:06,219 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=155069.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:04:11,281 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.089e+02 1.524e+02 1.961e+02 2.438e+02 3.962e+02, threshold=3.922e+02, percent-clipped=3.0 2023-04-28 02:04:11,989 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=155076.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:04:25,383 INFO [finetune.py:976] (4/7) Epoch 28, batch 450, loss[loss=0.1166, simple_loss=0.1977, pruned_loss=0.01778, over 4018.00 frames. ], tot_loss[loss=0.1663, simple_loss=0.2402, pruned_loss=0.04622, over 857227.67 frames. ], batch size: 17, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:04:31,704 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.10 vs. limit=5.0 2023-04-28 02:04:40,293 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9250, 4.2042, 0.7599, 2.2281, 2.4771, 2.8511, 2.4170, 0.9246], device='cuda:4'), covar=tensor([0.1287, 0.0894, 0.2100, 0.1183, 0.0931, 0.0956, 0.1344, 0.2055], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0240, 0.0136, 0.0121, 0.0132, 0.0152, 0.0117, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 02:04:48,049 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=155130.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:04:48,092 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=155130.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:04:58,930 INFO [finetune.py:976] (4/7) Epoch 28, batch 500, loss[loss=0.163, simple_loss=0.2229, pruned_loss=0.05159, over 4861.00 frames. ], tot_loss[loss=0.1652, simple_loss=0.2386, pruned_loss=0.04589, over 881502.41 frames. ], batch size: 44, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:05:16,309 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1964, 2.1001, 1.9840, 1.7996, 2.3137, 1.8658, 2.7524, 1.7847], device='cuda:4'), covar=tensor([0.3506, 0.1831, 0.4803, 0.2826, 0.1471, 0.2559, 0.1376, 0.4046], device='cuda:4'), in_proj_covar=tensor([0.0342, 0.0357, 0.0424, 0.0352, 0.0381, 0.0377, 0.0372, 0.0423], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 02:05:17,836 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.043e+02 1.583e+02 1.761e+02 2.203e+02 6.801e+02, threshold=3.523e+02, percent-clipped=1.0 2023-04-28 02:05:20,242 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=155178.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:05:21,494 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3188, 1.6419, 1.5001, 1.7795, 1.8196, 2.0107, 1.4744, 3.5341], device='cuda:4'), covar=tensor([0.0604, 0.0758, 0.0738, 0.1163, 0.0590, 0.0527, 0.0718, 0.0153], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0042, 0.0040, 0.0037, 0.0038, 0.0054], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 02:05:32,277 INFO [finetune.py:976] (4/7) Epoch 28, batch 550, loss[loss=0.145, simple_loss=0.2241, pruned_loss=0.03292, over 4818.00 frames. ], tot_loss[loss=0.1634, simple_loss=0.2358, pruned_loss=0.04545, over 897676.77 frames. ], batch size: 25, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:05:44,715 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4510, 1.3268, 1.6604, 1.7177, 1.2739, 1.1574, 1.4205, 0.8537], device='cuda:4'), covar=tensor([0.0504, 0.0659, 0.0393, 0.0530, 0.0672, 0.1114, 0.0627, 0.0562], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0068, 0.0066, 0.0069, 0.0075, 0.0095, 0.0073, 0.0062], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-28 02:06:05,934 INFO [finetune.py:976] (4/7) Epoch 28, batch 600, loss[loss=0.2303, simple_loss=0.3003, pruned_loss=0.08017, over 4811.00 frames. ], tot_loss[loss=0.1648, simple_loss=0.2373, pruned_loss=0.04613, over 910178.63 frames. ], batch size: 45, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:06:23,749 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.531e+01 1.578e+02 2.066e+02 2.357e+02 4.358e+02, threshold=4.132e+02, percent-clipped=3.0 2023-04-28 02:06:30,723 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=155284.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:06:39,097 INFO [finetune.py:976] (4/7) Epoch 28, batch 650, loss[loss=0.1409, simple_loss=0.222, pruned_loss=0.02988, over 4802.00 frames. ], tot_loss[loss=0.1681, simple_loss=0.2413, pruned_loss=0.04746, over 920458.51 frames. ], batch size: 45, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:06:41,051 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1945, 1.4847, 1.7607, 2.5169, 2.5890, 2.0128, 1.8400, 2.3370], device='cuda:4'), covar=tensor([0.0938, 0.1870, 0.1175, 0.0681, 0.0617, 0.1075, 0.0846, 0.0595], device='cuda:4'), in_proj_covar=tensor([0.0184, 0.0201, 0.0183, 0.0170, 0.0177, 0.0176, 0.0150, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 02:07:02,743 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=155332.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:07:02,753 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=155332.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:07:12,881 INFO [finetune.py:976] (4/7) Epoch 28, batch 700, loss[loss=0.1716, simple_loss=0.2562, pruned_loss=0.04347, over 4917.00 frames. ], tot_loss[loss=0.1692, simple_loss=0.2425, pruned_loss=0.04794, over 927721.14 frames. ], batch size: 38, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:07:18,231 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.56 vs. limit=2.0 2023-04-28 02:07:34,633 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.574e+02 1.852e+02 2.166e+02 4.284e+02, threshold=3.703e+02, percent-clipped=1.0 2023-04-28 02:07:40,201 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=155376.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:08:09,452 INFO [finetune.py:976] (4/7) Epoch 28, batch 750, loss[loss=0.1876, simple_loss=0.2503, pruned_loss=0.06246, over 4767.00 frames. ], tot_loss[loss=0.1692, simple_loss=0.2426, pruned_loss=0.04791, over 933751.72 frames. ], batch size: 28, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:08:41,162 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=155424.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:08:41,842 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=155425.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:09:12,225 INFO [finetune.py:976] (4/7) Epoch 28, batch 800, loss[loss=0.1989, simple_loss=0.2656, pruned_loss=0.06614, over 4902.00 frames. ], tot_loss[loss=0.1682, simple_loss=0.2419, pruned_loss=0.04728, over 939676.95 frames. ], batch size: 37, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:09:22,984 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=155457.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:09:42,702 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.932e+01 1.490e+02 1.749e+02 2.216e+02 3.602e+02, threshold=3.499e+02, percent-clipped=0.0 2023-04-28 02:09:59,774 INFO [finetune.py:976] (4/7) Epoch 28, batch 850, loss[loss=0.1418, simple_loss=0.2167, pruned_loss=0.03339, over 4773.00 frames. ], tot_loss[loss=0.1662, simple_loss=0.2396, pruned_loss=0.0464, over 939701.39 frames. ], batch size: 28, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:10:11,983 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=155518.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:10:32,814 INFO [finetune.py:976] (4/7) Epoch 28, batch 900, loss[loss=0.1747, simple_loss=0.2336, pruned_loss=0.05797, over 4876.00 frames. ], tot_loss[loss=0.1645, simple_loss=0.2373, pruned_loss=0.04587, over 943633.39 frames. ], batch size: 34, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:10:49,543 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.500e+01 1.455e+02 1.735e+02 2.204e+02 5.182e+02, threshold=3.469e+02, percent-clipped=3.0 2023-04-28 02:10:55,783 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2218, 1.6966, 2.0570, 2.4972, 1.9510, 1.6425, 1.2352, 1.7770], device='cuda:4'), covar=tensor([0.2875, 0.3098, 0.1585, 0.2010, 0.2478, 0.2643, 0.4269, 0.2076], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0246, 0.0229, 0.0315, 0.0222, 0.0235, 0.0228, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-28 02:11:05,999 INFO [finetune.py:976] (4/7) Epoch 28, batch 950, loss[loss=0.161, simple_loss=0.2267, pruned_loss=0.04762, over 4829.00 frames. ], tot_loss[loss=0.1626, simple_loss=0.2351, pruned_loss=0.04503, over 947881.57 frames. ], batch size: 25, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:11:10,958 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=155605.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:11:19,474 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2993, 1.5040, 1.4151, 1.7138, 1.6426, 1.7601, 1.4268, 3.2849], device='cuda:4'), covar=tensor([0.0630, 0.0848, 0.0826, 0.1271, 0.0656, 0.0521, 0.0756, 0.0159], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0039, 0.0042, 0.0040, 0.0037, 0.0038, 0.0054], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 02:11:27,369 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=155632.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:11:39,449 INFO [finetune.py:976] (4/7) Epoch 28, batch 1000, loss[loss=0.1589, simple_loss=0.2217, pruned_loss=0.04798, over 4887.00 frames. ], tot_loss[loss=0.1655, simple_loss=0.2383, pruned_loss=0.04633, over 950595.08 frames. ], batch size: 32, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:11:45,561 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=155657.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:11:51,071 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=155666.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:11:56,470 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.942e+01 1.547e+02 1.789e+02 2.038e+02 3.472e+02, threshold=3.578e+02, percent-clipped=1.0 2023-04-28 02:11:57,813 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0875, 0.8439, 0.8783, 0.8063, 1.2367, 1.0173, 0.8821, 0.9453], device='cuda:4'), covar=tensor([0.1832, 0.1401, 0.2093, 0.1701, 0.1071, 0.1405, 0.1679, 0.2435], device='cuda:4'), in_proj_covar=tensor([0.0316, 0.0308, 0.0352, 0.0286, 0.0326, 0.0306, 0.0301, 0.0377], device='cuda:4'), out_proj_covar=tensor([6.4443e-05, 6.3347e-05, 7.3907e-05, 5.7337e-05, 6.6549e-05, 6.3708e-05, 6.2246e-05, 7.9734e-05], device='cuda:4') 2023-04-28 02:11:59,554 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=155680.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:12:12,449 INFO [finetune.py:976] (4/7) Epoch 28, batch 1050, loss[loss=0.1698, simple_loss=0.2474, pruned_loss=0.04608, over 4827.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.2412, pruned_loss=0.04684, over 952865.08 frames. ], batch size: 30, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:12:25,718 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=155718.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:12:29,951 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=155725.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:12:45,201 INFO [finetune.py:976] (4/7) Epoch 28, batch 1100, loss[loss=0.1823, simple_loss=0.2538, pruned_loss=0.05541, over 4701.00 frames. ], tot_loss[loss=0.1687, simple_loss=0.2426, pruned_loss=0.04741, over 954307.63 frames. ], batch size: 59, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:12:48,444 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6646, 1.4535, 4.2555, 3.9890, 3.6617, 4.0632, 3.9177, 3.8058], device='cuda:4'), covar=tensor([0.7196, 0.5470, 0.1048, 0.1780, 0.1312, 0.1369, 0.2185, 0.1483], device='cuda:4'), in_proj_covar=tensor([0.0314, 0.0312, 0.0410, 0.0411, 0.0352, 0.0420, 0.0321, 0.0367], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-28 02:13:13,291 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=155773.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:13:13,353 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4468, 1.3797, 1.7770, 1.7284, 1.2838, 1.1167, 1.4820, 1.0366], device='cuda:4'), covar=tensor([0.0525, 0.0612, 0.0392, 0.0636, 0.0703, 0.1055, 0.0595, 0.0587], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0068, 0.0066, 0.0069, 0.0075, 0.0094, 0.0073, 0.0062], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-28 02:13:19,855 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.115e+02 1.609e+02 1.883e+02 2.305e+02 3.828e+02, threshold=3.767e+02, percent-clipped=2.0 2023-04-28 02:13:30,964 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=155784.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:13:46,059 INFO [finetune.py:976] (4/7) Epoch 28, batch 1150, loss[loss=0.1418, simple_loss=0.2143, pruned_loss=0.03468, over 4766.00 frames. ], tot_loss[loss=0.1684, simple_loss=0.2426, pruned_loss=0.04712, over 955773.05 frames. ], batch size: 26, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:13:58,499 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=155813.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:14:18,111 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=155845.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 02:14:19,787 INFO [finetune.py:976] (4/7) Epoch 28, batch 1200, loss[loss=0.1665, simple_loss=0.2425, pruned_loss=0.04526, over 4929.00 frames. ], tot_loss[loss=0.1664, simple_loss=0.2407, pruned_loss=0.04609, over 955784.96 frames. ], batch size: 38, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:14:54,755 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.082e+02 1.563e+02 1.837e+02 2.158e+02 4.161e+02, threshold=3.673e+02, percent-clipped=1.0 2023-04-28 02:15:25,828 INFO [finetune.py:976] (4/7) Epoch 28, batch 1250, loss[loss=0.1866, simple_loss=0.2609, pruned_loss=0.05618, over 4910.00 frames. ], tot_loss[loss=0.1655, simple_loss=0.2387, pruned_loss=0.04609, over 956515.34 frames. ], batch size: 32, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:15:57,630 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.14 vs. limit=2.0 2023-04-28 02:16:21,913 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.99 vs. limit=5.0 2023-04-28 02:16:24,805 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6253, 1.7064, 1.5105, 1.0829, 1.1843, 1.1624, 1.5054, 1.1370], device='cuda:4'), covar=tensor([0.1772, 0.1277, 0.1651, 0.1905, 0.2460, 0.2200, 0.1095, 0.2236], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0209, 0.0170, 0.0204, 0.0200, 0.0186, 0.0156, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 02:16:33,390 INFO [finetune.py:976] (4/7) Epoch 28, batch 1300, loss[loss=0.119, simple_loss=0.1931, pruned_loss=0.02241, over 4761.00 frames. ], tot_loss[loss=0.1625, simple_loss=0.2352, pruned_loss=0.0449, over 957447.26 frames. ], batch size: 27, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:16:53,061 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=155961.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:17:07,931 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.397e+01 1.571e+02 1.820e+02 2.327e+02 4.182e+02, threshold=3.640e+02, percent-clipped=2.0 2023-04-28 02:17:37,319 INFO [finetune.py:976] (4/7) Epoch 28, batch 1350, loss[loss=0.1656, simple_loss=0.2408, pruned_loss=0.04519, over 4825.00 frames. ], tot_loss[loss=0.1632, simple_loss=0.2363, pruned_loss=0.04508, over 958556.63 frames. ], batch size: 51, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:17:56,778 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=156013.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:18:40,869 INFO [finetune.py:976] (4/7) Epoch 28, batch 1400, loss[loss=0.2109, simple_loss=0.2767, pruned_loss=0.07258, over 4886.00 frames. ], tot_loss[loss=0.1648, simple_loss=0.2389, pruned_loss=0.04538, over 958074.70 frames. ], batch size: 32, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:19:09,176 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8593, 2.5853, 1.9565, 1.8820, 1.3468, 1.3792, 2.0177, 1.3073], device='cuda:4'), covar=tensor([0.1577, 0.1204, 0.1281, 0.1628, 0.2323, 0.1858, 0.0943, 0.2008], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0209, 0.0170, 0.0205, 0.0200, 0.0186, 0.0156, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 02:19:12,160 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=156070.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:19:20,864 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.099e+02 1.674e+02 1.966e+02 2.354e+02 4.293e+02, threshold=3.933e+02, percent-clipped=2.0 2023-04-28 02:19:32,255 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9245, 2.3081, 1.0286, 1.3165, 1.6560, 1.3140, 2.4835, 1.5317], device='cuda:4'), covar=tensor([0.0711, 0.0509, 0.0623, 0.1209, 0.0451, 0.0943, 0.0343, 0.0654], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0073, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-28 02:19:36,476 INFO [finetune.py:976] (4/7) Epoch 28, batch 1450, loss[loss=0.146, simple_loss=0.2262, pruned_loss=0.03286, over 4790.00 frames. ], tot_loss[loss=0.1657, simple_loss=0.2401, pruned_loss=0.04562, over 956549.34 frames. ], batch size: 25, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:19:46,155 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=156113.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:19:50,680 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.7884, 2.1663, 1.9963, 2.1683, 1.9734, 2.0120, 2.0490, 2.0550], device='cuda:4'), covar=tensor([0.3878, 0.5421, 0.5001, 0.4286, 0.5681, 0.6770, 0.6412, 0.5406], device='cuda:4'), in_proj_covar=tensor([0.0344, 0.0377, 0.0333, 0.0344, 0.0352, 0.0394, 0.0363, 0.0336], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 02:19:59,745 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=156131.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 02:20:05,160 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=156140.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 02:20:09,930 INFO [finetune.py:976] (4/7) Epoch 28, batch 1500, loss[loss=0.1649, simple_loss=0.2343, pruned_loss=0.0477, over 4820.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.2417, pruned_loss=0.0468, over 957822.29 frames. ], batch size: 30, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:20:23,249 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=156161.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:20:43,693 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.067e+02 1.591e+02 1.859e+02 2.303e+02 4.450e+02, threshold=3.717e+02, percent-clipped=1.0 2023-04-28 02:21:08,784 INFO [finetune.py:976] (4/7) Epoch 28, batch 1550, loss[loss=0.1522, simple_loss=0.2274, pruned_loss=0.03851, over 4902.00 frames. ], tot_loss[loss=0.1667, simple_loss=0.2403, pruned_loss=0.04653, over 956835.66 frames. ], batch size: 43, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:21:31,143 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=156216.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:22:01,927 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0115, 2.4342, 2.1954, 2.3789, 1.6512, 2.1793, 2.1520, 1.6741], device='cuda:4'), covar=tensor([0.2123, 0.1122, 0.0799, 0.1254, 0.3398, 0.1023, 0.1907, 0.2469], device='cuda:4'), in_proj_covar=tensor([0.0283, 0.0304, 0.0218, 0.0277, 0.0316, 0.0255, 0.0249, 0.0264], device='cuda:4'), out_proj_covar=tensor([1.1302e-04, 1.1944e-04, 8.5574e-05, 1.0893e-04, 1.2726e-04, 1.0030e-04, 1.0030e-04, 1.0421e-04], device='cuda:4') 2023-04-28 02:22:04,808 INFO [finetune.py:976] (4/7) Epoch 28, batch 1600, loss[loss=0.136, simple_loss=0.2068, pruned_loss=0.03262, over 4764.00 frames. ], tot_loss[loss=0.1639, simple_loss=0.2372, pruned_loss=0.04536, over 958046.38 frames. ], batch size: 26, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:22:12,823 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=156261.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:22:24,227 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.128e+02 1.624e+02 1.857e+02 2.215e+02 3.679e+02, threshold=3.713e+02, percent-clipped=0.0 2023-04-28 02:22:25,632 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=156277.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:22:34,044 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=156290.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:22:38,824 INFO [finetune.py:976] (4/7) Epoch 28, batch 1650, loss[loss=0.1565, simple_loss=0.2281, pruned_loss=0.04252, over 4753.00 frames. ], tot_loss[loss=0.1604, simple_loss=0.2335, pruned_loss=0.04367, over 957235.46 frames. ], batch size: 54, lr: 2.89e-03, grad_scale: 32.0 2023-04-28 02:22:45,576 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=156309.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:22:48,048 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=156313.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:23:12,278 INFO [finetune.py:976] (4/7) Epoch 28, batch 1700, loss[loss=0.176, simple_loss=0.2541, pruned_loss=0.04891, over 4920.00 frames. ], tot_loss[loss=0.1602, simple_loss=0.233, pruned_loss=0.04374, over 958041.39 frames. ], batch size: 37, lr: 2.88e-03, grad_scale: 32.0 2023-04-28 02:23:14,233 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=156351.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:23:19,690 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1952, 2.4977, 1.1021, 1.4482, 2.0471, 1.3231, 3.4117, 1.9485], device='cuda:4'), covar=tensor([0.0695, 0.0633, 0.0770, 0.1302, 0.0493, 0.1014, 0.0235, 0.0605], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0073, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-28 02:23:20,261 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=156361.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:23:28,664 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.471e+01 1.503e+02 1.857e+02 2.212e+02 3.931e+02, threshold=3.714e+02, percent-clipped=1.0 2023-04-28 02:23:45,145 INFO [finetune.py:976] (4/7) Epoch 28, batch 1750, loss[loss=0.1531, simple_loss=0.2373, pruned_loss=0.03447, over 4845.00 frames. ], tot_loss[loss=0.163, simple_loss=0.236, pruned_loss=0.04495, over 957692.25 frames. ], batch size: 47, lr: 2.88e-03, grad_scale: 32.0 2023-04-28 02:24:08,544 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=156426.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 02:24:29,101 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=156440.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 02:24:39,114 INFO [finetune.py:976] (4/7) Epoch 28, batch 1800, loss[loss=0.1926, simple_loss=0.232, pruned_loss=0.07661, over 4295.00 frames. ], tot_loss[loss=0.1656, simple_loss=0.239, pruned_loss=0.0461, over 957370.69 frames. ], batch size: 18, lr: 2.88e-03, grad_scale: 32.0 2023-04-28 02:25:02,479 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-04-28 02:25:11,111 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.353e+01 1.610e+02 1.909e+02 2.260e+02 4.171e+02, threshold=3.817e+02, percent-clipped=3.0 2023-04-28 02:25:31,042 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=156488.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:25:42,454 INFO [finetune.py:976] (4/7) Epoch 28, batch 1850, loss[loss=0.1467, simple_loss=0.2196, pruned_loss=0.0369, over 4782.00 frames. ], tot_loss[loss=0.1669, simple_loss=0.2405, pruned_loss=0.04666, over 955018.80 frames. ], batch size: 26, lr: 2.88e-03, grad_scale: 32.0 2023-04-28 02:26:38,731 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6003, 2.5481, 1.9924, 2.2366, 2.5769, 2.0379, 3.3852, 1.9488], device='cuda:4'), covar=tensor([0.3748, 0.2158, 0.4555, 0.3278, 0.1735, 0.2789, 0.1317, 0.4476], device='cuda:4'), in_proj_covar=tensor([0.0341, 0.0357, 0.0427, 0.0352, 0.0383, 0.0379, 0.0372, 0.0425], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 02:26:48,594 INFO [finetune.py:976] (4/7) Epoch 28, batch 1900, loss[loss=0.1489, simple_loss=0.2212, pruned_loss=0.03831, over 4780.00 frames. ], tot_loss[loss=0.1686, simple_loss=0.2426, pruned_loss=0.04733, over 955802.98 frames. ], batch size: 26, lr: 2.88e-03, grad_scale: 32.0 2023-04-28 02:27:01,796 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.19 vs. limit=5.0 2023-04-28 02:27:20,597 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=156572.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:27:22,350 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.009e+02 1.575e+02 1.976e+02 2.372e+02 4.648e+02, threshold=3.953e+02, percent-clipped=2.0 2023-04-28 02:27:54,785 INFO [finetune.py:976] (4/7) Epoch 28, batch 1950, loss[loss=0.2202, simple_loss=0.2675, pruned_loss=0.08648, over 4845.00 frames. ], tot_loss[loss=0.1684, simple_loss=0.2418, pruned_loss=0.04753, over 955208.77 frames. ], batch size: 49, lr: 2.88e-03, grad_scale: 32.0 2023-04-28 02:28:07,451 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.89 vs. limit=2.0 2023-04-28 02:29:00,506 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=156646.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:29:01,648 INFO [finetune.py:976] (4/7) Epoch 28, batch 2000, loss[loss=0.1921, simple_loss=0.2462, pruned_loss=0.06905, over 4872.00 frames. ], tot_loss[loss=0.1663, simple_loss=0.239, pruned_loss=0.0468, over 954877.10 frames. ], batch size: 31, lr: 2.88e-03, grad_scale: 64.0 2023-04-28 02:29:34,745 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.047e+02 1.509e+02 1.833e+02 2.149e+02 4.980e+02, threshold=3.665e+02, percent-clipped=1.0 2023-04-28 02:29:51,509 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-04-28 02:30:06,061 INFO [finetune.py:976] (4/7) Epoch 28, batch 2050, loss[loss=0.1505, simple_loss=0.2209, pruned_loss=0.04004, over 4912.00 frames. ], tot_loss[loss=0.1628, simple_loss=0.2352, pruned_loss=0.04522, over 955811.21 frames. ], batch size: 36, lr: 2.88e-03, grad_scale: 64.0 2023-04-28 02:30:06,128 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.8290, 3.7476, 2.7014, 4.4324, 3.8766, 3.8413, 1.4766, 3.7433], device='cuda:4'), covar=tensor([0.1749, 0.1398, 0.3355, 0.1538, 0.4014, 0.1821, 0.6614, 0.2421], device='cuda:4'), in_proj_covar=tensor([0.0249, 0.0224, 0.0255, 0.0307, 0.0303, 0.0253, 0.0278, 0.0277], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 02:30:12,964 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9994, 2.5631, 1.9173, 1.9143, 1.4157, 1.5093, 1.9252, 1.3646], device='cuda:4'), covar=tensor([0.1545, 0.1273, 0.1420, 0.1614, 0.2160, 0.1838, 0.0952, 0.1959], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0209, 0.0169, 0.0204, 0.0200, 0.0185, 0.0155, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 02:30:43,840 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=156726.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 02:31:05,849 INFO [finetune.py:976] (4/7) Epoch 28, batch 2100, loss[loss=0.1791, simple_loss=0.2569, pruned_loss=0.0506, over 4818.00 frames. ], tot_loss[loss=0.1649, simple_loss=0.2369, pruned_loss=0.04639, over 954228.50 frames. ], batch size: 45, lr: 2.88e-03, grad_scale: 64.0 2023-04-28 02:31:07,080 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=156749.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:31:39,407 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=156774.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:31:39,961 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.904e+01 1.566e+02 1.862e+02 2.081e+02 4.387e+02, threshold=3.725e+02, percent-clipped=1.0 2023-04-28 02:31:57,092 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.30 vs. limit=5.0 2023-04-28 02:32:10,052 INFO [finetune.py:976] (4/7) Epoch 28, batch 2150, loss[loss=0.189, simple_loss=0.2676, pruned_loss=0.05517, over 4835.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.24, pruned_loss=0.0471, over 954984.66 frames. ], batch size: 47, lr: 2.88e-03, grad_scale: 32.0 2023-04-28 02:32:27,874 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=156810.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:32:31,197 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.64 vs. limit=2.0 2023-04-28 02:32:57,260 INFO [finetune.py:976] (4/7) Epoch 28, batch 2200, loss[loss=0.1853, simple_loss=0.2624, pruned_loss=0.05405, over 4803.00 frames. ], tot_loss[loss=0.1684, simple_loss=0.2421, pruned_loss=0.04738, over 954066.63 frames. ], batch size: 41, lr: 2.88e-03, grad_scale: 32.0 2023-04-28 02:33:15,058 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=156872.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:33:17,408 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4637, 1.4247, 1.8309, 1.8030, 1.3789, 1.2745, 1.5311, 0.8661], device='cuda:4'), covar=tensor([0.0484, 0.0526, 0.0321, 0.0503, 0.0647, 0.0932, 0.0552, 0.0617], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0067, 0.0065, 0.0069, 0.0075, 0.0094, 0.0072, 0.0062], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-28 02:33:17,876 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.069e+02 1.564e+02 1.773e+02 2.100e+02 3.301e+02, threshold=3.547e+02, percent-clipped=0.0 2023-04-28 02:33:23,528 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=156885.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:33:31,445 INFO [finetune.py:976] (4/7) Epoch 28, batch 2250, loss[loss=0.165, simple_loss=0.2446, pruned_loss=0.04269, over 4866.00 frames. ], tot_loss[loss=0.1703, simple_loss=0.2443, pruned_loss=0.04816, over 954263.12 frames. ], batch size: 34, lr: 2.88e-03, grad_scale: 32.0 2023-04-28 02:33:47,976 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=156920.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:34:04,147 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=156946.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:34:04,165 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=156946.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:34:05,298 INFO [finetune.py:976] (4/7) Epoch 28, batch 2300, loss[loss=0.145, simple_loss=0.2407, pruned_loss=0.02462, over 4845.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2439, pruned_loss=0.0474, over 955389.01 frames. ], batch size: 49, lr: 2.88e-03, grad_scale: 32.0 2023-04-28 02:34:25,204 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.469e+02 1.728e+02 2.151e+02 3.571e+02, threshold=3.456e+02, percent-clipped=1.0 2023-04-28 02:34:36,694 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=156994.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:34:39,088 INFO [finetune.py:976] (4/7) Epoch 28, batch 2350, loss[loss=0.1487, simple_loss=0.2129, pruned_loss=0.04229, over 4780.00 frames. ], tot_loss[loss=0.1664, simple_loss=0.2404, pruned_loss=0.04619, over 954117.21 frames. ], batch size: 29, lr: 2.88e-03, grad_scale: 32.0 2023-04-28 02:34:40,527 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5077, 1.0282, 0.5498, 1.2488, 1.0878, 1.4195, 1.3333, 1.3090], device='cuda:4'), covar=tensor([0.0478, 0.0403, 0.0360, 0.0554, 0.0299, 0.0505, 0.0468, 0.0558], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0028, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0052, 0.0046, 0.0039, 0.0053, 0.0039, 0.0051, 0.0051, 0.0053], device='cuda:4') 2023-04-28 02:34:48,515 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=157010.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:35:23,596 INFO [finetune.py:976] (4/7) Epoch 28, batch 2400, loss[loss=0.16, simple_loss=0.2325, pruned_loss=0.04374, over 4906.00 frames. ], tot_loss[loss=0.1644, simple_loss=0.2378, pruned_loss=0.04553, over 954871.72 frames. ], batch size: 36, lr: 2.88e-03, grad_scale: 32.0 2023-04-28 02:35:25,513 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4010, 1.3133, 3.7215, 3.2396, 3.3093, 3.5310, 3.5621, 3.1938], device='cuda:4'), covar=tensor([0.9144, 0.7482, 0.1935, 0.3680, 0.2373, 0.3638, 0.2332, 0.3235], device='cuda:4'), in_proj_covar=tensor([0.0310, 0.0308, 0.0407, 0.0407, 0.0347, 0.0415, 0.0318, 0.0364], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-28 02:35:50,223 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=157071.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:35:53,121 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.080e+02 1.553e+02 1.893e+02 2.191e+02 4.408e+02, threshold=3.786e+02, percent-clipped=4.0 2023-04-28 02:36:07,063 INFO [finetune.py:976] (4/7) Epoch 28, batch 2450, loss[loss=0.1664, simple_loss=0.2423, pruned_loss=0.04521, over 4910.00 frames. ], tot_loss[loss=0.1625, simple_loss=0.2353, pruned_loss=0.04482, over 956313.88 frames. ], batch size: 35, lr: 2.88e-03, grad_scale: 32.0 2023-04-28 02:36:11,413 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=157105.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:36:22,726 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6579, 2.2350, 1.8776, 2.0050, 2.2839, 2.0230, 2.5325, 1.7140], device='cuda:4'), covar=tensor([0.2760, 0.1603, 0.3625, 0.2381, 0.1523, 0.1929, 0.1535, 0.3765], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0355, 0.0424, 0.0351, 0.0382, 0.0377, 0.0371, 0.0423], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 02:36:58,748 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=157142.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:37:07,639 INFO [finetune.py:976] (4/7) Epoch 28, batch 2500, loss[loss=0.1646, simple_loss=0.2426, pruned_loss=0.04325, over 4826.00 frames. ], tot_loss[loss=0.1636, simple_loss=0.2368, pruned_loss=0.04519, over 954098.59 frames. ], batch size: 39, lr: 2.88e-03, grad_scale: 32.0 2023-04-28 02:37:44,331 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.070e+02 1.608e+02 1.980e+02 2.407e+02 6.066e+02, threshold=3.960e+02, percent-clipped=5.0 2023-04-28 02:38:00,486 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-04-28 02:38:14,044 INFO [finetune.py:976] (4/7) Epoch 28, batch 2550, loss[loss=0.134, simple_loss=0.2279, pruned_loss=0.01999, over 4778.00 frames. ], tot_loss[loss=0.1651, simple_loss=0.2392, pruned_loss=0.04544, over 954697.95 frames. ], batch size: 28, lr: 2.88e-03, grad_scale: 32.0 2023-04-28 02:38:22,905 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=157203.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:39:11,689 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=157241.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:39:16,445 INFO [finetune.py:976] (4/7) Epoch 28, batch 2600, loss[loss=0.177, simple_loss=0.25, pruned_loss=0.05201, over 4824.00 frames. ], tot_loss[loss=0.1676, simple_loss=0.2416, pruned_loss=0.04678, over 956416.10 frames. ], batch size: 39, lr: 2.88e-03, grad_scale: 32.0 2023-04-28 02:39:50,660 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.059e+02 1.630e+02 1.851e+02 2.303e+02 4.332e+02, threshold=3.701e+02, percent-clipped=1.0 2023-04-28 02:40:05,059 INFO [finetune.py:976] (4/7) Epoch 28, batch 2650, loss[loss=0.1687, simple_loss=0.2505, pruned_loss=0.0434, over 4728.00 frames. ], tot_loss[loss=0.1686, simple_loss=0.2427, pruned_loss=0.0473, over 957142.07 frames. ], batch size: 59, lr: 2.88e-03, grad_scale: 32.0 2023-04-28 02:40:53,196 INFO [finetune.py:976] (4/7) Epoch 28, batch 2700, loss[loss=0.1345, simple_loss=0.2018, pruned_loss=0.0336, over 4227.00 frames. ], tot_loss[loss=0.1669, simple_loss=0.2412, pruned_loss=0.04632, over 956022.97 frames. ], batch size: 18, lr: 2.88e-03, grad_scale: 32.0 2023-04-28 02:40:56,702 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.48 vs. limit=2.0 2023-04-28 02:40:58,718 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=157356.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:41:04,717 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=157366.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:41:10,645 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.027e+02 1.425e+02 1.670e+02 2.111e+02 3.242e+02, threshold=3.340e+02, percent-clipped=0.0 2023-04-28 02:41:15,913 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=157382.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:41:26,096 INFO [finetune.py:976] (4/7) Epoch 28, batch 2750, loss[loss=0.2257, simple_loss=0.2816, pruned_loss=0.08492, over 4806.00 frames. ], tot_loss[loss=0.1662, simple_loss=0.2397, pruned_loss=0.04638, over 954313.06 frames. ], batch size: 45, lr: 2.88e-03, grad_scale: 32.0 2023-04-28 02:41:30,009 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=157404.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 02:41:31,079 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=157405.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:41:36,421 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.5534, 3.4824, 2.6514, 4.1672, 3.6606, 3.5936, 1.7265, 3.4818], device='cuda:4'), covar=tensor([0.1993, 0.1457, 0.3289, 0.1942, 0.3386, 0.2065, 0.5872, 0.2776], device='cuda:4'), in_proj_covar=tensor([0.0247, 0.0221, 0.0252, 0.0305, 0.0300, 0.0249, 0.0275, 0.0272], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 02:41:38,333 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=157417.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:41:46,153 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.97 vs. limit=2.0 2023-04-28 02:41:50,742 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2607, 3.0055, 0.9452, 1.6307, 1.6517, 2.2892, 1.7938, 1.1072], device='cuda:4'), covar=tensor([0.1771, 0.1352, 0.2138, 0.1677, 0.1288, 0.1179, 0.1607, 0.2006], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0238, 0.0134, 0.0120, 0.0131, 0.0152, 0.0116, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 02:41:53,597 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.61 vs. limit=2.0 2023-04-28 02:42:02,281 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=157443.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:42:04,685 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6832, 1.6155, 0.7937, 1.4457, 1.6538, 1.5678, 1.4510, 1.5472], device='cuda:4'), covar=tensor([0.0478, 0.0345, 0.0350, 0.0517, 0.0290, 0.0487, 0.0478, 0.0519], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0028, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0052, 0.0047, 0.0039, 0.0054, 0.0039, 0.0051, 0.0051, 0.0053], device='cuda:4') 2023-04-28 02:42:05,176 INFO [finetune.py:976] (4/7) Epoch 28, batch 2800, loss[loss=0.144, simple_loss=0.2163, pruned_loss=0.0359, over 4934.00 frames. ], tot_loss[loss=0.1635, simple_loss=0.2364, pruned_loss=0.0453, over 953972.53 frames. ], batch size: 38, lr: 2.88e-03, grad_scale: 32.0 2023-04-28 02:42:13,772 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=157453.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:42:26,587 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=157465.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 02:42:38,119 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.163e+02 1.553e+02 1.775e+02 2.169e+02 3.521e+02, threshold=3.549e+02, percent-clipped=1.0 2023-04-28 02:43:07,287 INFO [finetune.py:976] (4/7) Epoch 28, batch 2850, loss[loss=0.1682, simple_loss=0.2428, pruned_loss=0.04677, over 4778.00 frames. ], tot_loss[loss=0.1624, simple_loss=0.2347, pruned_loss=0.04502, over 954130.66 frames. ], batch size: 29, lr: 2.88e-03, grad_scale: 32.0 2023-04-28 02:43:07,364 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=157498.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:43:17,205 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=157506.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 02:44:00,027 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=157541.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:44:09,939 INFO [finetune.py:976] (4/7) Epoch 28, batch 2900, loss[loss=0.1821, simple_loss=0.2619, pruned_loss=0.05112, over 4840.00 frames. ], tot_loss[loss=0.1653, simple_loss=0.2384, pruned_loss=0.04609, over 953914.80 frames. ], batch size: 47, lr: 2.88e-03, grad_scale: 32.0 2023-04-28 02:44:33,699 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=157567.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 02:44:34,292 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=157568.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:44:43,298 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4187, 1.6908, 1.8680, 1.9516, 1.8606, 1.9587, 1.9311, 1.8667], device='cuda:4'), covar=tensor([0.3959, 0.5207, 0.4519, 0.4647, 0.5428, 0.7153, 0.4530, 0.4562], device='cuda:4'), in_proj_covar=tensor([0.0343, 0.0377, 0.0330, 0.0343, 0.0353, 0.0394, 0.0360, 0.0335], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 02:44:43,842 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8109, 3.6912, 1.0056, 1.9959, 2.1628, 2.5753, 2.0892, 1.0077], device='cuda:4'), covar=tensor([0.1314, 0.0858, 0.1998, 0.1248, 0.1014, 0.1021, 0.1497, 0.2029], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0239, 0.0135, 0.0120, 0.0132, 0.0152, 0.0116, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 02:44:44,346 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.234e+02 1.564e+02 1.889e+02 2.208e+02 3.535e+02, threshold=3.777e+02, percent-clipped=0.0 2023-04-28 02:45:03,636 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=157589.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:45:14,538 INFO [finetune.py:976] (4/7) Epoch 28, batch 2950, loss[loss=0.1767, simple_loss=0.2505, pruned_loss=0.05145, over 4817.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.2418, pruned_loss=0.04683, over 955619.22 frames. ], batch size: 33, lr: 2.88e-03, grad_scale: 32.0 2023-04-28 02:45:50,085 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=157629.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:45:58,262 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3387, 2.2034, 1.8182, 1.9188, 2.2899, 1.9338, 2.7205, 1.6798], device='cuda:4'), covar=tensor([0.3756, 0.2100, 0.4846, 0.3227, 0.1741, 0.2418, 0.1542, 0.4425], device='cuda:4'), in_proj_covar=tensor([0.0341, 0.0357, 0.0427, 0.0352, 0.0385, 0.0380, 0.0374, 0.0426], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 02:46:10,744 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1609, 2.0734, 1.8016, 1.7166, 2.1402, 1.8284, 2.6737, 1.6593], device='cuda:4'), covar=tensor([0.3846, 0.1986, 0.4675, 0.3060, 0.1672, 0.2395, 0.1288, 0.4319], device='cuda:4'), in_proj_covar=tensor([0.0341, 0.0357, 0.0427, 0.0352, 0.0385, 0.0380, 0.0374, 0.0426], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 02:46:18,330 INFO [finetune.py:976] (4/7) Epoch 28, batch 3000, loss[loss=0.1978, simple_loss=0.2592, pruned_loss=0.06822, over 4275.00 frames. ], tot_loss[loss=0.1687, simple_loss=0.2424, pruned_loss=0.04752, over 956228.42 frames. ], batch size: 65, lr: 2.88e-03, grad_scale: 32.0 2023-04-28 02:46:18,331 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-28 02:46:37,393 INFO [finetune.py:1010] (4/7) Epoch 28, validation: loss=0.153, simple_loss=0.2217, pruned_loss=0.04213, over 2265189.00 frames. 2023-04-28 02:46:37,393 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6529MB 2023-04-28 02:46:41,903 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.66 vs. limit=2.0 2023-04-28 02:46:49,120 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=157666.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:46:55,060 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.120e+02 1.557e+02 1.829e+02 2.170e+02 4.324e+02, threshold=3.658e+02, percent-clipped=1.0 2023-04-28 02:47:00,706 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.25 vs. limit=2.0 2023-04-28 02:47:00,860 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.63 vs. limit=5.0 2023-04-28 02:47:09,010 INFO [finetune.py:976] (4/7) Epoch 28, batch 3050, loss[loss=0.1812, simple_loss=0.2482, pruned_loss=0.05709, over 4907.00 frames. ], tot_loss[loss=0.1691, simple_loss=0.2431, pruned_loss=0.04752, over 957183.11 frames. ], batch size: 36, lr: 2.88e-03, grad_scale: 32.0 2023-04-28 02:47:27,242 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=157712.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:47:28,441 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=157714.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:47:58,456 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=157738.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:48:09,325 INFO [finetune.py:976] (4/7) Epoch 28, batch 3100, loss[loss=0.1165, simple_loss=0.194, pruned_loss=0.01953, over 4769.00 frames. ], tot_loss[loss=0.1674, simple_loss=0.241, pruned_loss=0.04687, over 955449.59 frames. ], batch size: 28, lr: 2.88e-03, grad_scale: 32.0 2023-04-28 02:48:22,593 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=157760.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 02:48:32,274 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.835e+01 1.560e+02 1.859e+02 2.154e+02 3.848e+02, threshold=3.719e+02, percent-clipped=2.0 2023-04-28 02:48:46,288 INFO [finetune.py:976] (4/7) Epoch 28, batch 3150, loss[loss=0.1688, simple_loss=0.2364, pruned_loss=0.05063, over 4747.00 frames. ], tot_loss[loss=0.1648, simple_loss=0.2382, pruned_loss=0.04571, over 956179.54 frames. ], batch size: 27, lr: 2.88e-03, grad_scale: 32.0 2023-04-28 02:48:46,382 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=157798.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:49:18,161 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=157846.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:49:19,358 INFO [finetune.py:976] (4/7) Epoch 28, batch 3200, loss[loss=0.1451, simple_loss=0.2146, pruned_loss=0.03786, over 4860.00 frames. ], tot_loss[loss=0.1627, simple_loss=0.2352, pruned_loss=0.04516, over 956949.14 frames. ], batch size: 31, lr: 2.88e-03, grad_scale: 32.0 2023-04-28 02:49:22,526 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=157853.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:49:27,985 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=157862.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 02:49:38,913 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.062e+02 1.560e+02 1.811e+02 2.192e+02 5.644e+02, threshold=3.622e+02, percent-clipped=2.0 2023-04-28 02:50:07,745 INFO [finetune.py:976] (4/7) Epoch 28, batch 3250, loss[loss=0.1827, simple_loss=0.2465, pruned_loss=0.05943, over 4823.00 frames. ], tot_loss[loss=0.1639, simple_loss=0.2364, pruned_loss=0.04576, over 956578.05 frames. ], batch size: 33, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 02:50:29,710 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=157914.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:50:43,190 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=157924.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:51:14,742 INFO [finetune.py:976] (4/7) Epoch 28, batch 3300, loss[loss=0.176, simple_loss=0.2543, pruned_loss=0.04883, over 4175.00 frames. ], tot_loss[loss=0.1674, simple_loss=0.2405, pruned_loss=0.04712, over 955827.12 frames. ], batch size: 65, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 02:51:51,135 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.051e+02 1.721e+02 1.967e+02 2.277e+02 5.584e+02, threshold=3.933e+02, percent-clipped=3.0 2023-04-28 02:52:10,304 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.3580, 3.2951, 2.4374, 3.8491, 3.3005, 3.3700, 1.5226, 3.3325], device='cuda:4'), covar=tensor([0.1841, 0.1363, 0.3468, 0.2276, 0.3157, 0.1873, 0.5168, 0.2457], device='cuda:4'), in_proj_covar=tensor([0.0243, 0.0217, 0.0250, 0.0301, 0.0296, 0.0246, 0.0272, 0.0269], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 02:52:20,398 INFO [finetune.py:976] (4/7) Epoch 28, batch 3350, loss[loss=0.1459, simple_loss=0.218, pruned_loss=0.03684, over 4853.00 frames. ], tot_loss[loss=0.1696, simple_loss=0.2431, pruned_loss=0.04808, over 957271.93 frames. ], batch size: 31, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 02:52:41,469 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=158012.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:53:03,023 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5582, 1.0988, 4.1412, 3.8559, 3.6087, 3.9051, 3.8560, 3.6744], device='cuda:4'), covar=tensor([0.6936, 0.6051, 0.0973, 0.1733, 0.1175, 0.1332, 0.1730, 0.1391], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0314, 0.0410, 0.0411, 0.0354, 0.0421, 0.0322, 0.0368], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-28 02:53:12,140 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=158038.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:53:23,378 INFO [finetune.py:976] (4/7) Epoch 28, batch 3400, loss[loss=0.1819, simple_loss=0.2532, pruned_loss=0.05531, over 4902.00 frames. ], tot_loss[loss=0.17, simple_loss=0.2439, pruned_loss=0.04804, over 957039.99 frames. ], batch size: 36, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 02:53:41,206 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=158060.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:53:41,252 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=158060.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 02:53:56,198 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.582e+02 1.900e+02 2.246e+02 3.787e+02, threshold=3.800e+02, percent-clipped=0.0 2023-04-28 02:54:10,637 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=158086.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:54:23,032 INFO [finetune.py:976] (4/7) Epoch 28, batch 3450, loss[loss=0.1765, simple_loss=0.2459, pruned_loss=0.05357, over 4925.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2436, pruned_loss=0.0476, over 956800.97 frames. ], batch size: 33, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 02:54:34,533 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=158108.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 02:54:54,848 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=158122.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:55:29,823 INFO [finetune.py:976] (4/7) Epoch 28, batch 3500, loss[loss=0.1519, simple_loss=0.2192, pruned_loss=0.04229, over 4754.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.2407, pruned_loss=0.0465, over 956045.37 frames. ], batch size: 23, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 02:55:40,236 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-28 02:55:48,567 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=158162.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 02:56:07,655 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.013e+02 1.474e+02 1.724e+02 2.129e+02 5.829e+02, threshold=3.447e+02, percent-clipped=1.0 2023-04-28 02:56:12,614 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=158183.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:56:18,393 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8478, 2.4589, 0.8984, 1.2051, 1.5227, 1.2054, 2.4918, 1.4626], device='cuda:4'), covar=tensor([0.0843, 0.0503, 0.0762, 0.1557, 0.0577, 0.1271, 0.0385, 0.0824], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0073, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-28 02:56:23,135 INFO [finetune.py:976] (4/7) Epoch 28, batch 3550, loss[loss=0.1409, simple_loss=0.209, pruned_loss=0.03639, over 4749.00 frames. ], tot_loss[loss=0.1654, simple_loss=0.2384, pruned_loss=0.04617, over 955134.00 frames. ], batch size: 54, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 02:56:30,181 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=158209.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:56:30,778 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=158210.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 02:56:39,842 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=158224.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:56:56,635 INFO [finetune.py:976] (4/7) Epoch 28, batch 3600, loss[loss=0.129, simple_loss=0.2089, pruned_loss=0.02458, over 4765.00 frames. ], tot_loss[loss=0.1626, simple_loss=0.2354, pruned_loss=0.04494, over 955609.41 frames. ], batch size: 28, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 02:57:11,761 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=158272.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:57:14,137 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.946e+01 1.583e+02 1.935e+02 2.237e+02 3.659e+02, threshold=3.870e+02, percent-clipped=1.0 2023-04-28 02:57:29,422 INFO [finetune.py:976] (4/7) Epoch 28, batch 3650, loss[loss=0.1955, simple_loss=0.2662, pruned_loss=0.06235, over 4804.00 frames. ], tot_loss[loss=0.1651, simple_loss=0.2378, pruned_loss=0.04625, over 953893.86 frames. ], batch size: 41, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 02:58:02,903 INFO [finetune.py:976] (4/7) Epoch 28, batch 3700, loss[loss=0.1193, simple_loss=0.2038, pruned_loss=0.01741, over 4784.00 frames. ], tot_loss[loss=0.1674, simple_loss=0.2407, pruned_loss=0.047, over 951289.83 frames. ], batch size: 25, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 02:58:15,787 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.7054, 2.1047, 2.1446, 2.2216, 2.1491, 2.1641, 2.2290, 2.1448], device='cuda:4'), covar=tensor([0.3729, 0.5369, 0.4546, 0.4363, 0.5685, 0.6571, 0.5086, 0.5047], device='cuda:4'), in_proj_covar=tensor([0.0343, 0.0377, 0.0331, 0.0343, 0.0351, 0.0393, 0.0362, 0.0335], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 02:58:19,866 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.152e+02 1.537e+02 1.915e+02 2.211e+02 4.327e+02, threshold=3.830e+02, percent-clipped=1.0 2023-04-28 02:58:25,396 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=158384.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:58:35,213 INFO [finetune.py:976] (4/7) Epoch 28, batch 3750, loss[loss=0.1761, simple_loss=0.2452, pruned_loss=0.05348, over 4772.00 frames. ], tot_loss[loss=0.168, simple_loss=0.2414, pruned_loss=0.04733, over 949968.57 frames. ], batch size: 28, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 02:59:05,532 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=158445.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 02:59:07,664 INFO [finetune.py:976] (4/7) Epoch 28, batch 3800, loss[loss=0.1642, simple_loss=0.2489, pruned_loss=0.03974, over 4818.00 frames. ], tot_loss[loss=0.1678, simple_loss=0.2416, pruned_loss=0.04701, over 952530.55 frames. ], batch size: 39, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 02:59:42,092 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.322e+01 1.536e+02 1.803e+02 2.140e+02 3.917e+02, threshold=3.606e+02, percent-clipped=1.0 2023-04-28 02:59:43,432 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=158478.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:00:07,899 INFO [finetune.py:976] (4/7) Epoch 28, batch 3850, loss[loss=0.1809, simple_loss=0.269, pruned_loss=0.04643, over 4908.00 frames. ], tot_loss[loss=0.1665, simple_loss=0.2407, pruned_loss=0.04618, over 954926.91 frames. ], batch size: 43, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 03:00:26,842 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=158509.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:01:09,879 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.8854, 2.2821, 2.1357, 2.2565, 2.1061, 2.2235, 2.2326, 2.1412], device='cuda:4'), covar=tensor([0.3831, 0.5931, 0.5147, 0.4703, 0.5580, 0.6842, 0.6148, 0.5636], device='cuda:4'), in_proj_covar=tensor([0.0343, 0.0377, 0.0331, 0.0343, 0.0351, 0.0394, 0.0361, 0.0336], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 03:01:10,982 INFO [finetune.py:976] (4/7) Epoch 28, batch 3900, loss[loss=0.1471, simple_loss=0.228, pruned_loss=0.03308, over 4831.00 frames. ], tot_loss[loss=0.1653, simple_loss=0.2384, pruned_loss=0.04607, over 955904.45 frames. ], batch size: 33, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 03:01:28,077 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=158557.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:01:51,819 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.216e+01 1.537e+02 1.820e+02 2.187e+02 3.711e+02, threshold=3.639e+02, percent-clipped=1.0 2023-04-28 03:02:23,131 INFO [finetune.py:976] (4/7) Epoch 28, batch 3950, loss[loss=0.1641, simple_loss=0.2267, pruned_loss=0.05079, over 4204.00 frames. ], tot_loss[loss=0.1633, simple_loss=0.2362, pruned_loss=0.04526, over 953627.80 frames. ], batch size: 65, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 03:03:31,311 INFO [finetune.py:976] (4/7) Epoch 28, batch 4000, loss[loss=0.1307, simple_loss=0.1936, pruned_loss=0.03384, over 4234.00 frames. ], tot_loss[loss=0.164, simple_loss=0.2362, pruned_loss=0.04584, over 953923.40 frames. ], batch size: 18, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 03:04:13,586 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.149e+01 1.505e+02 1.815e+02 2.168e+02 4.885e+02, threshold=3.630e+02, percent-clipped=2.0 2023-04-28 03:04:38,151 INFO [finetune.py:976] (4/7) Epoch 28, batch 4050, loss[loss=0.1779, simple_loss=0.2565, pruned_loss=0.04962, over 4816.00 frames. ], tot_loss[loss=0.1679, simple_loss=0.2403, pruned_loss=0.04774, over 952120.38 frames. ], batch size: 51, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 03:05:33,924 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=158740.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:05:43,860 INFO [finetune.py:976] (4/7) Epoch 28, batch 4100, loss[loss=0.1439, simple_loss=0.2226, pruned_loss=0.0326, over 4874.00 frames. ], tot_loss[loss=0.1688, simple_loss=0.2426, pruned_loss=0.0475, over 952792.76 frames. ], batch size: 34, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 03:05:52,641 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.79 vs. limit=2.0 2023-04-28 03:06:25,665 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.124e+02 1.629e+02 1.870e+02 2.384e+02 6.257e+02, threshold=3.741e+02, percent-clipped=1.0 2023-04-28 03:06:26,403 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=158777.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:06:27,018 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=158778.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:06:35,230 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=158783.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:06:49,757 INFO [finetune.py:976] (4/7) Epoch 28, batch 4150, loss[loss=0.207, simple_loss=0.2693, pruned_loss=0.0724, over 4817.00 frames. ], tot_loss[loss=0.17, simple_loss=0.244, pruned_loss=0.04804, over 953513.02 frames. ], batch size: 30, lr: 2.87e-03, grad_scale: 64.0 2023-04-28 03:07:30,855 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=158826.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:07:41,993 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=158835.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 03:07:43,884 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=158838.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:07:52,647 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=158844.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:07:53,500 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.73 vs. limit=5.0 2023-04-28 03:07:54,973 INFO [finetune.py:976] (4/7) Epoch 28, batch 4200, loss[loss=0.1816, simple_loss=0.2605, pruned_loss=0.0514, over 4924.00 frames. ], tot_loss[loss=0.1684, simple_loss=0.243, pruned_loss=0.0469, over 952381.08 frames. ], batch size: 33, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 03:08:35,720 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4852, 1.9734, 2.3954, 2.7266, 2.3987, 1.9075, 1.6806, 2.1766], device='cuda:4'), covar=tensor([0.2949, 0.2657, 0.1465, 0.2068, 0.2283, 0.2358, 0.3674, 0.1817], device='cuda:4'), in_proj_covar=tensor([0.0291, 0.0245, 0.0228, 0.0314, 0.0222, 0.0235, 0.0228, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-28 03:08:37,325 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.042e+02 1.525e+02 1.688e+02 1.998e+02 3.338e+02, threshold=3.377e+02, percent-clipped=0.0 2023-04-28 03:09:04,990 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=158896.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 03:09:06,084 INFO [finetune.py:976] (4/7) Epoch 28, batch 4250, loss[loss=0.1661, simple_loss=0.2486, pruned_loss=0.04179, over 4903.00 frames. ], tot_loss[loss=0.1664, simple_loss=0.2406, pruned_loss=0.04612, over 952027.56 frames. ], batch size: 36, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 03:09:17,706 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5660, 2.2785, 1.9544, 1.9645, 2.3300, 1.9039, 2.5546, 1.7695], device='cuda:4'), covar=tensor([0.3286, 0.1779, 0.3532, 0.2675, 0.1440, 0.2370, 0.1883, 0.4044], device='cuda:4'), in_proj_covar=tensor([0.0339, 0.0356, 0.0424, 0.0352, 0.0384, 0.0378, 0.0373, 0.0424], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 03:10:04,960 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5005, 2.1282, 2.2280, 3.0127, 2.5099, 1.9125, 2.1437, 2.3671], device='cuda:4'), covar=tensor([0.2343, 0.2564, 0.1415, 0.1954, 0.2222, 0.2226, 0.3330, 0.1897], device='cuda:4'), in_proj_covar=tensor([0.0291, 0.0245, 0.0228, 0.0314, 0.0223, 0.0235, 0.0228, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-28 03:10:12,870 INFO [finetune.py:976] (4/7) Epoch 28, batch 4300, loss[loss=0.1632, simple_loss=0.2307, pruned_loss=0.04783, over 4856.00 frames. ], tot_loss[loss=0.1638, simple_loss=0.2376, pruned_loss=0.04502, over 952899.82 frames. ], batch size: 49, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 03:10:48,762 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.467e+01 1.425e+02 1.689e+02 1.969e+02 4.397e+02, threshold=3.377e+02, percent-clipped=1.0 2023-04-28 03:11:17,424 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.89 vs. limit=2.0 2023-04-28 03:11:18,967 INFO [finetune.py:976] (4/7) Epoch 28, batch 4350, loss[loss=0.1359, simple_loss=0.2076, pruned_loss=0.03212, over 4849.00 frames. ], tot_loss[loss=0.1609, simple_loss=0.234, pruned_loss=0.04385, over 952732.30 frames. ], batch size: 44, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 03:12:15,671 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=159040.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:12:26,357 INFO [finetune.py:976] (4/7) Epoch 28, batch 4400, loss[loss=0.1622, simple_loss=0.216, pruned_loss=0.05422, over 4151.00 frames. ], tot_loss[loss=0.1612, simple_loss=0.2341, pruned_loss=0.04415, over 949474.04 frames. ], batch size: 18, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 03:12:27,696 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2747, 2.9613, 0.8368, 1.6377, 1.4784, 2.0523, 1.7267, 1.0173], device='cuda:4'), covar=tensor([0.1482, 0.0858, 0.2065, 0.1248, 0.1236, 0.1068, 0.1612, 0.1870], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0240, 0.0137, 0.0121, 0.0133, 0.0154, 0.0117, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 03:12:54,454 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=159069.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:13:05,081 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.784e+01 1.599e+02 1.852e+02 2.200e+02 4.625e+02, threshold=3.703e+02, percent-clipped=1.0 2023-04-28 03:13:19,038 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=159088.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:13:29,732 INFO [finetune.py:976] (4/7) Epoch 28, batch 4450, loss[loss=0.1696, simple_loss=0.238, pruned_loss=0.0506, over 4934.00 frames. ], tot_loss[loss=0.1653, simple_loss=0.2393, pruned_loss=0.04565, over 951667.89 frames. ], batch size: 33, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 03:13:47,988 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4928, 3.4966, 0.8813, 1.8933, 1.7072, 2.4166, 1.9366, 1.1056], device='cuda:4'), covar=tensor([0.1366, 0.1075, 0.2021, 0.1197, 0.1143, 0.0996, 0.1466, 0.1913], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0240, 0.0136, 0.0121, 0.0132, 0.0154, 0.0117, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 03:13:57,410 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0048, 2.1775, 1.3270, 1.7671, 2.2211, 1.8665, 1.8259, 1.8531], device='cuda:4'), covar=tensor([0.0456, 0.0336, 0.0299, 0.0500, 0.0251, 0.0462, 0.0456, 0.0552], device='cuda:4'), in_proj_covar=tensor([0.0027, 0.0023, 0.0022, 0.0028, 0.0019, 0.0027, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0052, 0.0046, 0.0039, 0.0053, 0.0039, 0.0051, 0.0051, 0.0053], device='cuda:4') 2023-04-28 03:14:09,982 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=159130.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:14:12,209 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=159133.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:14:21,840 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=159139.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:14:38,435 INFO [finetune.py:976] (4/7) Epoch 28, batch 4500, loss[loss=0.2233, simple_loss=0.2906, pruned_loss=0.07798, over 4810.00 frames. ], tot_loss[loss=0.1657, simple_loss=0.2401, pruned_loss=0.04568, over 952013.19 frames. ], batch size: 45, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 03:14:49,769 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=159159.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:15:11,106 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.154e+02 1.553e+02 1.824e+02 2.169e+02 4.146e+02, threshold=3.648e+02, percent-clipped=1.0 2023-04-28 03:15:31,034 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=159191.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 03:15:41,718 INFO [finetune.py:976] (4/7) Epoch 28, batch 4550, loss[loss=0.1715, simple_loss=0.2508, pruned_loss=0.04615, over 4869.00 frames. ], tot_loss[loss=0.1667, simple_loss=0.2414, pruned_loss=0.04603, over 952763.13 frames. ], batch size: 34, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 03:16:05,597 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=159220.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:16:24,883 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9598, 1.7458, 1.9619, 2.2963, 2.3252, 1.9226, 1.6309, 1.9736], device='cuda:4'), covar=tensor([0.0872, 0.1189, 0.0779, 0.0538, 0.0630, 0.0811, 0.0751, 0.0588], device='cuda:4'), in_proj_covar=tensor([0.0182, 0.0202, 0.0182, 0.0169, 0.0177, 0.0176, 0.0149, 0.0174], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 03:16:45,688 INFO [finetune.py:976] (4/7) Epoch 28, batch 4600, loss[loss=0.147, simple_loss=0.2233, pruned_loss=0.03535, over 4828.00 frames. ], tot_loss[loss=0.1683, simple_loss=0.2427, pruned_loss=0.04696, over 954903.30 frames. ], batch size: 39, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 03:17:18,839 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.827e+01 1.517e+02 1.769e+02 2.322e+02 3.935e+02, threshold=3.539e+02, percent-clipped=1.0 2023-04-28 03:17:18,982 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2965, 2.6520, 2.2530, 2.2169, 1.5773, 1.6386, 2.3545, 1.6560], device='cuda:4'), covar=tensor([0.1584, 0.1581, 0.1303, 0.1682, 0.2297, 0.1808, 0.0948, 0.1986], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0210, 0.0171, 0.0205, 0.0201, 0.0187, 0.0157, 0.0189], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 03:17:38,887 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=159290.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:17:49,022 INFO [finetune.py:976] (4/7) Epoch 28, batch 4650, loss[loss=0.1532, simple_loss=0.2324, pruned_loss=0.03697, over 4822.00 frames. ], tot_loss[loss=0.1669, simple_loss=0.2403, pruned_loss=0.04669, over 953954.08 frames. ], batch size: 38, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 03:18:53,430 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4179, 1.7151, 1.6115, 1.8770, 1.8177, 2.0105, 1.5484, 3.6895], device='cuda:4'), covar=tensor([0.0574, 0.0801, 0.0776, 0.1194, 0.0633, 0.0475, 0.0731, 0.0137], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0039, 0.0043, 0.0040, 0.0038, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0014, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 03:18:54,507 INFO [finetune.py:976] (4/7) Epoch 28, batch 4700, loss[loss=0.1318, simple_loss=0.2037, pruned_loss=0.02994, over 4828.00 frames. ], tot_loss[loss=0.1648, simple_loss=0.2378, pruned_loss=0.04593, over 954819.27 frames. ], batch size: 40, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 03:19:03,238 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=159351.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:19:36,243 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.039e+02 1.471e+02 1.707e+02 1.992e+02 4.348e+02, threshold=3.414e+02, percent-clipped=1.0 2023-04-28 03:19:56,806 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.65 vs. limit=2.0 2023-04-28 03:20:00,614 INFO [finetune.py:976] (4/7) Epoch 28, batch 4750, loss[loss=0.181, simple_loss=0.2493, pruned_loss=0.0564, over 4867.00 frames. ], tot_loss[loss=0.1639, simple_loss=0.2361, pruned_loss=0.04591, over 953054.60 frames. ], batch size: 34, lr: 2.87e-03, grad_scale: 32.0 2023-04-28 03:20:41,282 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=159425.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:20:51,605 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=159433.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:20:55,246 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=159439.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:21:06,689 INFO [finetune.py:976] (4/7) Epoch 28, batch 4800, loss[loss=0.2117, simple_loss=0.2898, pruned_loss=0.06681, over 4833.00 frames. ], tot_loss[loss=0.1662, simple_loss=0.2387, pruned_loss=0.04686, over 953244.37 frames. ], batch size: 49, lr: 2.86e-03, grad_scale: 32.0 2023-04-28 03:21:48,625 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.050e+02 1.679e+02 2.036e+02 2.333e+02 4.105e+02, threshold=4.072e+02, percent-clipped=2.0 2023-04-28 03:21:51,196 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=159481.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:21:59,479 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=159487.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:22:07,512 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=159491.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 03:22:12,300 INFO [finetune.py:976] (4/7) Epoch 28, batch 4850, loss[loss=0.1998, simple_loss=0.2831, pruned_loss=0.0582, over 4831.00 frames. ], tot_loss[loss=0.1678, simple_loss=0.2411, pruned_loss=0.04728, over 953803.64 frames. ], batch size: 30, lr: 2.86e-03, grad_scale: 32.0 2023-04-28 03:22:40,592 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=159515.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:22:51,368 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3211, 2.0909, 2.3633, 2.8766, 2.7714, 2.4661, 2.0271, 2.5082], device='cuda:4'), covar=tensor([0.0800, 0.1019, 0.0605, 0.0479, 0.0555, 0.0683, 0.0667, 0.0472], device='cuda:4'), in_proj_covar=tensor([0.0183, 0.0202, 0.0182, 0.0169, 0.0177, 0.0176, 0.0149, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 03:23:05,855 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=159539.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 03:23:16,525 INFO [finetune.py:976] (4/7) Epoch 28, batch 4900, loss[loss=0.1634, simple_loss=0.2409, pruned_loss=0.04296, over 4907.00 frames. ], tot_loss[loss=0.1694, simple_loss=0.2431, pruned_loss=0.04785, over 955791.83 frames. ], batch size: 37, lr: 2.86e-03, grad_scale: 32.0 2023-04-28 03:23:56,947 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.158e+02 1.586e+02 1.930e+02 2.270e+02 5.838e+02, threshold=3.860e+02, percent-clipped=1.0 2023-04-28 03:24:20,243 INFO [finetune.py:976] (4/7) Epoch 28, batch 4950, loss[loss=0.1536, simple_loss=0.2289, pruned_loss=0.03919, over 4726.00 frames. ], tot_loss[loss=0.1704, simple_loss=0.2439, pruned_loss=0.04842, over 954955.62 frames. ], batch size: 54, lr: 2.86e-03, grad_scale: 32.0 2023-04-28 03:25:08,555 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=159631.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:25:23,010 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=159646.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:25:23,663 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7810, 2.9040, 2.2185, 2.5116, 2.8760, 2.6664, 3.7309, 2.2090], device='cuda:4'), covar=tensor([0.3842, 0.2362, 0.4790, 0.3251, 0.1844, 0.2421, 0.1413, 0.4055], device='cuda:4'), in_proj_covar=tensor([0.0336, 0.0353, 0.0421, 0.0350, 0.0381, 0.0373, 0.0370, 0.0421], device='cuda:4'), out_proj_covar=tensor([9.9286e-05, 1.0501e-04, 1.2725e-04, 1.0457e-04, 1.1254e-04, 1.1065e-04, 1.0808e-04, 1.2658e-04], device='cuda:4') 2023-04-28 03:25:24,155 INFO [finetune.py:976] (4/7) Epoch 28, batch 5000, loss[loss=0.1254, simple_loss=0.2039, pruned_loss=0.02347, over 4754.00 frames. ], tot_loss[loss=0.1683, simple_loss=0.2417, pruned_loss=0.04749, over 955838.88 frames. ], batch size: 27, lr: 2.86e-03, grad_scale: 32.0 2023-04-28 03:26:05,241 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.105e+02 1.569e+02 1.814e+02 2.278e+02 4.708e+02, threshold=3.628e+02, percent-clipped=1.0 2023-04-28 03:26:13,635 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9981, 2.3707, 0.9669, 1.2778, 1.9186, 1.1296, 3.1410, 1.4201], device='cuda:4'), covar=tensor([0.0679, 0.0756, 0.0812, 0.1132, 0.0463, 0.0971, 0.0201, 0.0631], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0073, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-28 03:26:24,450 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=159692.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:26:32,514 INFO [finetune.py:976] (4/7) Epoch 28, batch 5050, loss[loss=0.1375, simple_loss=0.2177, pruned_loss=0.02865, over 4826.00 frames. ], tot_loss[loss=0.1666, simple_loss=0.2392, pruned_loss=0.04703, over 955874.13 frames. ], batch size: 25, lr: 2.86e-03, grad_scale: 32.0 2023-04-28 03:27:06,800 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=159725.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:27:16,890 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8703, 2.2984, 1.9780, 2.2720, 1.7928, 1.9793, 1.9416, 1.5678], device='cuda:4'), covar=tensor([0.1923, 0.1294, 0.0857, 0.1152, 0.3221, 0.1174, 0.2040, 0.2590], device='cuda:4'), in_proj_covar=tensor([0.0283, 0.0301, 0.0217, 0.0275, 0.0313, 0.0253, 0.0247, 0.0263], device='cuda:4'), out_proj_covar=tensor([1.1256e-04, 1.1839e-04, 8.5374e-05, 1.0810e-04, 1.2622e-04, 9.9212e-05, 9.9589e-05, 1.0350e-04], device='cuda:4') 2023-04-28 03:27:35,600 INFO [finetune.py:976] (4/7) Epoch 28, batch 5100, loss[loss=0.1894, simple_loss=0.2501, pruned_loss=0.06434, over 4866.00 frames. ], tot_loss[loss=0.1629, simple_loss=0.2352, pruned_loss=0.04533, over 955939.80 frames. ], batch size: 31, lr: 2.86e-03, grad_scale: 32.0 2023-04-28 03:28:09,664 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=159773.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:28:13,003 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.052e+01 1.474e+02 1.729e+02 2.168e+02 4.974e+02, threshold=3.458e+02, percent-clipped=1.0 2023-04-28 03:28:41,782 INFO [finetune.py:976] (4/7) Epoch 28, batch 5150, loss[loss=0.1621, simple_loss=0.2527, pruned_loss=0.03569, over 4820.00 frames. ], tot_loss[loss=0.1644, simple_loss=0.2368, pruned_loss=0.04598, over 954712.84 frames. ], batch size: 39, lr: 2.86e-03, grad_scale: 32.0 2023-04-28 03:29:03,530 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=159815.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:29:45,773 INFO [finetune.py:976] (4/7) Epoch 28, batch 5200, loss[loss=0.1591, simple_loss=0.2371, pruned_loss=0.04056, over 4775.00 frames. ], tot_loss[loss=0.1664, simple_loss=0.2395, pruned_loss=0.04662, over 953084.05 frames. ], batch size: 26, lr: 2.86e-03, grad_scale: 32.0 2023-04-28 03:29:55,498 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=159855.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:30:05,392 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=159863.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:30:26,723 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.215e+02 1.682e+02 1.955e+02 2.433e+02 4.346e+02, threshold=3.909e+02, percent-clipped=3.0 2023-04-28 03:30:51,957 INFO [finetune.py:976] (4/7) Epoch 28, batch 5250, loss[loss=0.1824, simple_loss=0.2596, pruned_loss=0.05258, over 4865.00 frames. ], tot_loss[loss=0.1661, simple_loss=0.2401, pruned_loss=0.04602, over 954351.18 frames. ], batch size: 34, lr: 2.86e-03, grad_scale: 32.0 2023-04-28 03:30:58,135 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6052, 2.0569, 1.7247, 2.0329, 1.6082, 1.6970, 1.7039, 1.4099], device='cuda:4'), covar=tensor([0.1841, 0.1238, 0.0814, 0.1109, 0.3102, 0.1136, 0.1796, 0.2232], device='cuda:4'), in_proj_covar=tensor([0.0282, 0.0300, 0.0217, 0.0275, 0.0312, 0.0252, 0.0247, 0.0263], device='cuda:4'), out_proj_covar=tensor([1.1247e-04, 1.1806e-04, 8.5226e-05, 1.0812e-04, 1.2593e-04, 9.9101e-05, 9.9498e-05, 1.0353e-04], device='cuda:4') 2023-04-28 03:31:00,014 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.03 vs. limit=5.0 2023-04-28 03:31:09,940 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=159910.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:31:19,015 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=159916.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:31:55,426 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=159946.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:32:02,070 INFO [finetune.py:976] (4/7) Epoch 28, batch 5300, loss[loss=0.1977, simple_loss=0.2767, pruned_loss=0.05932, over 4928.00 frames. ], tot_loss[loss=0.168, simple_loss=0.2425, pruned_loss=0.04676, over 955802.96 frames. ], batch size: 42, lr: 2.86e-03, grad_scale: 32.0 2023-04-28 03:32:26,771 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=159971.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:32:26,804 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2184, 1.8403, 2.1146, 2.1548, 2.0737, 1.7252, 1.2576, 1.7992], device='cuda:4'), covar=tensor([0.2975, 0.2748, 0.1481, 0.2115, 0.2367, 0.2352, 0.3765, 0.1715], device='cuda:4'), in_proj_covar=tensor([0.0291, 0.0245, 0.0227, 0.0313, 0.0222, 0.0234, 0.0228, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-28 03:32:35,450 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.404e+01 1.544e+02 1.829e+02 2.287e+02 5.054e+02, threshold=3.659e+02, percent-clipped=1.0 2023-04-28 03:32:48,735 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=159987.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:32:58,791 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=159994.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:33:06,691 INFO [finetune.py:976] (4/7) Epoch 28, batch 5350, loss[loss=0.1692, simple_loss=0.2461, pruned_loss=0.04619, over 4864.00 frames. ], tot_loss[loss=0.1674, simple_loss=0.242, pruned_loss=0.04646, over 952809.82 frames. ], batch size: 34, lr: 2.86e-03, grad_scale: 32.0 2023-04-28 03:33:17,040 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3427, 1.5767, 1.8039, 1.9266, 1.8223, 1.8616, 1.8034, 1.8243], device='cuda:4'), covar=tensor([0.3348, 0.5160, 0.4261, 0.3924, 0.5142, 0.6584, 0.4812, 0.4372], device='cuda:4'), in_proj_covar=tensor([0.0343, 0.0377, 0.0331, 0.0344, 0.0353, 0.0394, 0.0364, 0.0335], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 03:33:32,228 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2762, 1.8649, 5.1784, 4.9098, 4.4703, 4.9490, 4.5943, 4.6097], device='cuda:4'), covar=tensor([0.6378, 0.5048, 0.1101, 0.1631, 0.1074, 0.1697, 0.1416, 0.1627], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0310, 0.0405, 0.0408, 0.0351, 0.0416, 0.0319, 0.0365], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-28 03:34:13,400 INFO [finetune.py:976] (4/7) Epoch 28, batch 5400, loss[loss=0.157, simple_loss=0.2355, pruned_loss=0.03924, over 4770.00 frames. ], tot_loss[loss=0.1662, simple_loss=0.2399, pruned_loss=0.04626, over 953745.84 frames. ], batch size: 27, lr: 2.86e-03, grad_scale: 32.0 2023-04-28 03:34:46,552 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.246e+02 1.579e+02 1.912e+02 2.220e+02 4.270e+02, threshold=3.825e+02, percent-clipped=1.0 2023-04-28 03:35:03,331 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9424, 1.5859, 2.0351, 2.4276, 2.0695, 1.8990, 2.0307, 1.9535], device='cuda:4'), covar=tensor([0.4192, 0.6324, 0.6238, 0.5338, 0.5607, 0.7346, 0.7426, 0.9132], device='cuda:4'), in_proj_covar=tensor([0.0445, 0.0425, 0.0520, 0.0509, 0.0474, 0.0511, 0.0513, 0.0527], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 03:35:18,041 INFO [finetune.py:976] (4/7) Epoch 28, batch 5450, loss[loss=0.1387, simple_loss=0.2135, pruned_loss=0.03197, over 4855.00 frames. ], tot_loss[loss=0.165, simple_loss=0.2377, pruned_loss=0.04618, over 955284.30 frames. ], batch size: 44, lr: 2.86e-03, grad_scale: 32.0 2023-04-28 03:35:27,494 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3544, 1.7616, 2.1455, 2.9598, 2.9427, 2.3088, 2.1390, 2.6792], device='cuda:4'), covar=tensor([0.1043, 0.1630, 0.0973, 0.0676, 0.0580, 0.1097, 0.0796, 0.0607], device='cuda:4'), in_proj_covar=tensor([0.0184, 0.0203, 0.0183, 0.0170, 0.0177, 0.0177, 0.0150, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 03:35:27,502 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5006, 3.0305, 2.4600, 2.8446, 2.0353, 2.5122, 2.7753, 2.1163], device='cuda:4'), covar=tensor([0.1957, 0.1327, 0.0886, 0.1212, 0.3415, 0.1226, 0.1855, 0.2680], device='cuda:4'), in_proj_covar=tensor([0.0282, 0.0299, 0.0216, 0.0275, 0.0312, 0.0252, 0.0247, 0.0263], device='cuda:4'), out_proj_covar=tensor([1.1218e-04, 1.1773e-04, 8.4813e-05, 1.0793e-04, 1.2571e-04, 9.8819e-05, 9.9400e-05, 1.0336e-04], device='cuda:4') 2023-04-28 03:36:22,771 INFO [finetune.py:976] (4/7) Epoch 28, batch 5500, loss[loss=0.1566, simple_loss=0.2341, pruned_loss=0.03954, over 4787.00 frames. ], tot_loss[loss=0.1625, simple_loss=0.2351, pruned_loss=0.04493, over 955164.73 frames. ], batch size: 29, lr: 2.86e-03, grad_scale: 32.0 2023-04-28 03:37:00,972 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 7.993e+01 1.507e+02 1.787e+02 2.170e+02 4.329e+02, threshold=3.573e+02, percent-clipped=2.0 2023-04-28 03:37:27,651 INFO [finetune.py:976] (4/7) Epoch 28, batch 5550, loss[loss=0.1721, simple_loss=0.2575, pruned_loss=0.04333, over 4931.00 frames. ], tot_loss[loss=0.1638, simple_loss=0.2363, pruned_loss=0.04563, over 953338.20 frames. ], batch size: 38, lr: 2.86e-03, grad_scale: 32.0 2023-04-28 03:37:41,112 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=160211.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:38:03,241 INFO [finetune.py:976] (4/7) Epoch 28, batch 5600, loss[loss=0.1844, simple_loss=0.2636, pruned_loss=0.05262, over 4871.00 frames. ], tot_loss[loss=0.1654, simple_loss=0.2394, pruned_loss=0.04573, over 953961.31 frames. ], batch size: 34, lr: 2.86e-03, grad_scale: 32.0 2023-04-28 03:38:13,712 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=160266.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:38:20,149 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.023e+02 1.566e+02 2.005e+02 2.303e+02 3.818e+02, threshold=4.010e+02, percent-clipped=2.0 2023-04-28 03:38:26,529 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=160287.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:38:32,860 INFO [finetune.py:976] (4/7) Epoch 28, batch 5650, loss[loss=0.1502, simple_loss=0.228, pruned_loss=0.03615, over 4856.00 frames. ], tot_loss[loss=0.1662, simple_loss=0.241, pruned_loss=0.04574, over 952117.55 frames. ], batch size: 49, lr: 2.86e-03, grad_scale: 32.0 2023-04-28 03:38:38,250 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=160307.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 03:38:49,409 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5397, 2.3828, 2.2327, 2.2006, 2.6028, 2.2274, 3.0837, 2.0312], device='cuda:4'), covar=tensor([0.3017, 0.1920, 0.3915, 0.2656, 0.1398, 0.2242, 0.1156, 0.3913], device='cuda:4'), in_proj_covar=tensor([0.0339, 0.0355, 0.0422, 0.0353, 0.0384, 0.0377, 0.0372, 0.0426], device='cuda:4'), out_proj_covar=tensor([9.9887e-05, 1.0553e-04, 1.2756e-04, 1.0555e-04, 1.1363e-04, 1.1181e-04, 1.0847e-04, 1.2786e-04], device='cuda:4') 2023-04-28 03:38:54,999 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=160335.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:39:02,716 INFO [finetune.py:976] (4/7) Epoch 28, batch 5700, loss[loss=0.126, simple_loss=0.2003, pruned_loss=0.02585, over 4200.00 frames. ], tot_loss[loss=0.1635, simple_loss=0.2368, pruned_loss=0.04514, over 933308.17 frames. ], batch size: 18, lr: 2.86e-03, grad_scale: 32.0 2023-04-28 03:39:14,603 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=160368.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 03:39:31,531 INFO [finetune.py:976] (4/7) Epoch 29, batch 0, loss[loss=0.1986, simple_loss=0.2643, pruned_loss=0.06642, over 4919.00 frames. ], tot_loss[loss=0.1986, simple_loss=0.2643, pruned_loss=0.06642, over 4919.00 frames. ], batch size: 33, lr: 2.86e-03, grad_scale: 32.0 2023-04-28 03:39:31,532 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-28 03:39:42,780 INFO [finetune.py:1010] (4/7) Epoch 29, validation: loss=0.1546, simple_loss=0.2236, pruned_loss=0.04278, over 2265189.00 frames. 2023-04-28 03:39:42,780 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6529MB 2023-04-28 03:39:44,401 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.451e+01 1.469e+02 1.726e+02 2.023e+02 3.272e+02, threshold=3.452e+02, percent-clipped=0.0 2023-04-28 03:39:45,705 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9497, 4.2262, 0.9127, 2.3553, 2.5504, 3.0229, 2.6317, 1.1957], device='cuda:4'), covar=tensor([0.1316, 0.1006, 0.1995, 0.1156, 0.0910, 0.0951, 0.1317, 0.2001], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0238, 0.0136, 0.0120, 0.0131, 0.0152, 0.0117, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 03:40:16,114 INFO [finetune.py:976] (4/7) Epoch 29, batch 50, loss[loss=0.1351, simple_loss=0.2108, pruned_loss=0.02968, over 4880.00 frames. ], tot_loss[loss=0.1698, simple_loss=0.2452, pruned_loss=0.04723, over 217000.93 frames. ], batch size: 35, lr: 2.85e-03, grad_scale: 32.0 2023-04-28 03:40:21,814 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-28 03:40:33,634 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.40 vs. limit=5.0 2023-04-28 03:40:45,248 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-04-28 03:41:16,377 INFO [finetune.py:976] (4/7) Epoch 29, batch 100, loss[loss=0.1533, simple_loss=0.2262, pruned_loss=0.04017, over 4911.00 frames. ], tot_loss[loss=0.1629, simple_loss=0.236, pruned_loss=0.04493, over 380418.14 frames. ], batch size: 46, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 03:41:24,290 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.006e+02 1.575e+02 1.942e+02 2.371e+02 3.324e+02, threshold=3.883e+02, percent-clipped=0.0 2023-04-28 03:42:07,348 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=160511.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:42:22,987 INFO [finetune.py:976] (4/7) Epoch 29, batch 150, loss[loss=0.1859, simple_loss=0.2532, pruned_loss=0.05929, over 4926.00 frames. ], tot_loss[loss=0.1605, simple_loss=0.2324, pruned_loss=0.04426, over 506524.48 frames. ], batch size: 33, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 03:42:43,469 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5741, 2.2009, 2.4923, 3.0310, 2.4233, 1.9254, 1.9390, 2.4202], device='cuda:4'), covar=tensor([0.3166, 0.2839, 0.1539, 0.2292, 0.2604, 0.2609, 0.3489, 0.1859], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0247, 0.0228, 0.0315, 0.0223, 0.0236, 0.0229, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-28 03:43:05,314 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=160559.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:43:14,856 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=160566.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:43:27,106 INFO [finetune.py:976] (4/7) Epoch 29, batch 200, loss[loss=0.1859, simple_loss=0.2534, pruned_loss=0.05922, over 4801.00 frames. ], tot_loss[loss=0.1591, simple_loss=0.2307, pruned_loss=0.04382, over 607926.10 frames. ], batch size: 51, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 03:43:34,236 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.002e+02 1.562e+02 1.851e+02 2.230e+02 3.985e+02, threshold=3.702e+02, percent-clipped=1.0 2023-04-28 03:43:37,299 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8127, 4.2157, 0.7800, 2.2469, 2.3796, 2.7338, 2.3632, 0.9735], device='cuda:4'), covar=tensor([0.1463, 0.0788, 0.2173, 0.1217, 0.1056, 0.1155, 0.1415, 0.2280], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0238, 0.0135, 0.0119, 0.0131, 0.0152, 0.0117, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 03:44:17,861 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=160614.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:44:30,032 INFO [finetune.py:976] (4/7) Epoch 29, batch 250, loss[loss=0.1645, simple_loss=0.2376, pruned_loss=0.04568, over 4826.00 frames. ], tot_loss[loss=0.1603, simple_loss=0.232, pruned_loss=0.0443, over 686691.54 frames. ], batch size: 33, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 03:45:20,093 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=160663.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 03:45:32,745 INFO [finetune.py:976] (4/7) Epoch 29, batch 300, loss[loss=0.1494, simple_loss=0.2231, pruned_loss=0.03788, over 4801.00 frames. ], tot_loss[loss=0.1635, simple_loss=0.2362, pruned_loss=0.04538, over 747436.99 frames. ], batch size: 25, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 03:45:38,925 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.030e+02 1.518e+02 1.888e+02 2.303e+02 4.692e+02, threshold=3.776e+02, percent-clipped=1.0 2023-04-28 03:46:01,457 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-28 03:46:15,477 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-28 03:46:37,396 INFO [finetune.py:976] (4/7) Epoch 29, batch 350, loss[loss=0.1284, simple_loss=0.2006, pruned_loss=0.02816, over 4724.00 frames. ], tot_loss[loss=0.1651, simple_loss=0.2382, pruned_loss=0.04599, over 794349.12 frames. ], batch size: 23, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 03:47:08,178 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9569, 1.6187, 1.8928, 2.2887, 2.2552, 1.8648, 1.6159, 2.1136], device='cuda:4'), covar=tensor([0.0741, 0.1225, 0.0730, 0.0523, 0.0571, 0.0829, 0.0665, 0.0488], device='cuda:4'), in_proj_covar=tensor([0.0184, 0.0205, 0.0184, 0.0171, 0.0178, 0.0177, 0.0151, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 03:47:35,777 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-28 03:47:45,909 INFO [finetune.py:976] (4/7) Epoch 29, batch 400, loss[loss=0.1925, simple_loss=0.2675, pruned_loss=0.05879, over 4878.00 frames. ], tot_loss[loss=0.1669, simple_loss=0.2404, pruned_loss=0.04668, over 828098.91 frames. ], batch size: 35, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 03:47:47,780 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.194e+02 1.618e+02 1.920e+02 2.400e+02 5.071e+02, threshold=3.839e+02, percent-clipped=2.0 2023-04-28 03:48:21,597 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.0932, 2.3607, 2.3221, 2.8592, 2.9963, 2.6546, 2.5993, 2.2716], device='cuda:4'), covar=tensor([0.1079, 0.1423, 0.1367, 0.1495, 0.0930, 0.1148, 0.1382, 0.1900], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0305, 0.0346, 0.0284, 0.0322, 0.0302, 0.0297, 0.0372], device='cuda:4'), out_proj_covar=tensor([6.3550e-05, 6.2497e-05, 7.2385e-05, 5.6854e-05, 6.5639e-05, 6.2742e-05, 6.1197e-05, 7.8507e-05], device='cuda:4') 2023-04-28 03:48:36,948 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6543, 1.7449, 0.7969, 1.3914, 1.7871, 1.5502, 1.4501, 1.5114], device='cuda:4'), covar=tensor([0.0496, 0.0355, 0.0332, 0.0528, 0.0259, 0.0500, 0.0481, 0.0550], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0028, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0052, 0.0046, 0.0039, 0.0053, 0.0039, 0.0051, 0.0051, 0.0053], device='cuda:4') 2023-04-28 03:48:48,105 INFO [finetune.py:976] (4/7) Epoch 29, batch 450, loss[loss=0.2051, simple_loss=0.2657, pruned_loss=0.0723, over 4686.00 frames. ], tot_loss[loss=0.1666, simple_loss=0.2403, pruned_loss=0.04648, over 853774.37 frames. ], batch size: 23, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 03:49:19,876 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.3989, 3.2849, 2.5595, 3.8617, 3.3241, 3.4456, 1.7402, 3.3352], device='cuda:4'), covar=tensor([0.1551, 0.1228, 0.2748, 0.1621, 0.2661, 0.1583, 0.4619, 0.2063], device='cuda:4'), in_proj_covar=tensor([0.0250, 0.0223, 0.0254, 0.0308, 0.0303, 0.0254, 0.0280, 0.0278], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 03:49:51,959 INFO [finetune.py:976] (4/7) Epoch 29, batch 500, loss[loss=0.1704, simple_loss=0.2346, pruned_loss=0.05316, over 4837.00 frames. ], tot_loss[loss=0.1639, simple_loss=0.2368, pruned_loss=0.04548, over 874948.78 frames. ], batch size: 33, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 03:49:53,843 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.077e+02 1.560e+02 1.871e+02 2.250e+02 5.218e+02, threshold=3.742e+02, percent-clipped=1.0 2023-04-28 03:50:02,890 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.30 vs. limit=2.0 2023-04-28 03:50:25,306 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=160901.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:50:56,032 INFO [finetune.py:976] (4/7) Epoch 29, batch 550, loss[loss=0.1401, simple_loss=0.2171, pruned_loss=0.03154, over 4817.00 frames. ], tot_loss[loss=0.1614, simple_loss=0.234, pruned_loss=0.04435, over 894816.69 frames. ], batch size: 40, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 03:51:48,580 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=160962.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:51:49,199 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=160963.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 03:51:50,432 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=160965.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:51:51,656 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4604, 1.3925, 1.7624, 1.7509, 1.3097, 1.2767, 1.4648, 0.9813], device='cuda:4'), covar=tensor([0.0494, 0.0598, 0.0341, 0.0577, 0.0750, 0.1052, 0.0502, 0.0512], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0066, 0.0065, 0.0068, 0.0074, 0.0093, 0.0072, 0.0062], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-28 03:52:01,358 INFO [finetune.py:976] (4/7) Epoch 29, batch 600, loss[loss=0.1913, simple_loss=0.2715, pruned_loss=0.05554, over 4819.00 frames. ], tot_loss[loss=0.162, simple_loss=0.2346, pruned_loss=0.04475, over 907963.99 frames. ], batch size: 40, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 03:52:03,146 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.040e+02 1.581e+02 1.979e+02 2.385e+02 4.353e+02, threshold=3.958e+02, percent-clipped=2.0 2023-04-28 03:52:35,724 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=161011.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 03:52:36,357 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8795, 2.4425, 2.0732, 2.3036, 1.6669, 2.0679, 1.9431, 1.5843], device='cuda:4'), covar=tensor([0.2105, 0.1103, 0.0714, 0.1280, 0.3528, 0.1065, 0.2067, 0.2664], device='cuda:4'), in_proj_covar=tensor([0.0282, 0.0299, 0.0216, 0.0275, 0.0313, 0.0253, 0.0248, 0.0263], device='cuda:4'), out_proj_covar=tensor([1.1216e-04, 1.1769e-04, 8.4773e-05, 1.0791e-04, 1.2607e-04, 9.9356e-05, 9.9928e-05, 1.0375e-04], device='cuda:4') 2023-04-28 03:52:38,799 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4127, 1.3426, 1.4228, 1.6337, 1.6433, 1.4021, 1.0533, 1.5246], device='cuda:4'), covar=tensor([0.0808, 0.1419, 0.0964, 0.0655, 0.0729, 0.0877, 0.0814, 0.0579], device='cuda:4'), in_proj_covar=tensor([0.0184, 0.0203, 0.0183, 0.0171, 0.0177, 0.0177, 0.0150, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 03:52:44,112 INFO [finetune.py:976] (4/7) Epoch 29, batch 650, loss[loss=0.1744, simple_loss=0.2502, pruned_loss=0.04931, over 4905.00 frames. ], tot_loss[loss=0.1634, simple_loss=0.2367, pruned_loss=0.04509, over 919978.70 frames. ], batch size: 36, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 03:52:44,878 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=161026.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:52:51,962 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.31 vs. limit=2.0 2023-04-28 03:53:17,108 INFO [finetune.py:976] (4/7) Epoch 29, batch 700, loss[loss=0.1908, simple_loss=0.2685, pruned_loss=0.05653, over 4929.00 frames. ], tot_loss[loss=0.1655, simple_loss=0.2388, pruned_loss=0.04603, over 926155.16 frames. ], batch size: 38, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 03:53:18,920 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.470e+01 1.599e+02 1.884e+02 2.263e+02 4.345e+02, threshold=3.768e+02, percent-clipped=2.0 2023-04-28 03:53:50,554 INFO [finetune.py:976] (4/7) Epoch 29, batch 750, loss[loss=0.1586, simple_loss=0.2349, pruned_loss=0.04121, over 4733.00 frames. ], tot_loss[loss=0.1673, simple_loss=0.2412, pruned_loss=0.0467, over 932176.78 frames. ], batch size: 54, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 03:53:55,699 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3547, 1.6893, 1.5620, 2.2227, 2.3656, 1.9377, 1.8589, 1.6113], device='cuda:4'), covar=tensor([0.1683, 0.2031, 0.2035, 0.1460, 0.1161, 0.2061, 0.2519, 0.2506], device='cuda:4'), in_proj_covar=tensor([0.0314, 0.0308, 0.0350, 0.0287, 0.0325, 0.0305, 0.0300, 0.0376], device='cuda:4'), out_proj_covar=tensor([6.3982e-05, 6.3123e-05, 7.3327e-05, 5.7366e-05, 6.6170e-05, 6.3397e-05, 6.1804e-05, 7.9502e-05], device='cuda:4') 2023-04-28 03:53:56,327 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9321, 2.4371, 1.9617, 1.7436, 1.3472, 1.4553, 1.9877, 1.3560], device='cuda:4'), covar=tensor([0.1629, 0.1258, 0.1317, 0.1696, 0.2395, 0.2009, 0.1007, 0.2081], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0210, 0.0170, 0.0204, 0.0202, 0.0186, 0.0156, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 03:54:24,404 INFO [finetune.py:976] (4/7) Epoch 29, batch 800, loss[loss=0.1491, simple_loss=0.2128, pruned_loss=0.04271, over 4835.00 frames. ], tot_loss[loss=0.1664, simple_loss=0.2407, pruned_loss=0.04608, over 936444.93 frames. ], batch size: 49, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 03:54:25,699 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4533, 1.5026, 4.1500, 3.9177, 3.6503, 3.9912, 3.9189, 3.6246], device='cuda:4'), covar=tensor([0.7066, 0.5225, 0.1055, 0.1606, 0.1184, 0.1632, 0.1456, 0.1536], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0307, 0.0403, 0.0406, 0.0348, 0.0413, 0.0316, 0.0361], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-28 03:54:26,206 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.992e+01 1.674e+02 1.972e+02 2.388e+02 4.465e+02, threshold=3.944e+02, percent-clipped=4.0 2023-04-28 03:54:52,596 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=161216.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:54:56,244 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7230, 1.3369, 1.8606, 2.0197, 1.7694, 1.6915, 1.7633, 1.7800], device='cuda:4'), covar=tensor([0.5097, 0.7369, 0.6664, 0.6861, 0.6429, 0.8581, 0.8205, 0.9072], device='cuda:4'), in_proj_covar=tensor([0.0448, 0.0427, 0.0523, 0.0513, 0.0477, 0.0513, 0.0516, 0.0531], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 03:54:57,905 INFO [finetune.py:976] (4/7) Epoch 29, batch 850, loss[loss=0.146, simple_loss=0.2231, pruned_loss=0.03447, over 4856.00 frames. ], tot_loss[loss=0.1645, simple_loss=0.2383, pruned_loss=0.0454, over 941225.27 frames. ], batch size: 49, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 03:55:01,671 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=161231.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:55:29,546 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=161257.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:55:49,080 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9560, 2.4354, 2.1631, 1.8557, 1.4140, 1.5051, 2.1443, 1.4143], device='cuda:4'), covar=tensor([0.1631, 0.1266, 0.1166, 0.1563, 0.2224, 0.1797, 0.0871, 0.1968], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0210, 0.0171, 0.0205, 0.0202, 0.0187, 0.0157, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 03:55:49,584 INFO [finetune.py:976] (4/7) Epoch 29, batch 900, loss[loss=0.1433, simple_loss=0.2241, pruned_loss=0.03124, over 4790.00 frames. ], tot_loss[loss=0.1626, simple_loss=0.2362, pruned_loss=0.04454, over 945628.48 frames. ], batch size: 29, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 03:55:50,930 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=161277.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:55:51,421 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.567e+01 1.487e+02 1.850e+02 2.194e+02 4.508e+02, threshold=3.700e+02, percent-clipped=1.0 2023-04-28 03:56:00,172 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=161292.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:56:09,234 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5855, 1.7516, 1.6300, 2.0540, 1.8995, 2.2730, 1.6897, 4.3871], device='cuda:4'), covar=tensor([0.0536, 0.0762, 0.0788, 0.1194, 0.0604, 0.0495, 0.0713, 0.0106], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0040, 0.0037, 0.0038, 0.0054], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0014, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 03:56:21,349 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=161321.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:56:22,008 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0242, 1.0363, 1.2022, 1.1453, 0.9929, 0.9054, 0.9715, 0.5371], device='cuda:4'), covar=tensor([0.0524, 0.0548, 0.0419, 0.0498, 0.0660, 0.1020, 0.0431, 0.0564], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0067, 0.0065, 0.0069, 0.0075, 0.0094, 0.0072, 0.0062], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-28 03:56:23,751 INFO [finetune.py:976] (4/7) Epoch 29, batch 950, loss[loss=0.1822, simple_loss=0.2495, pruned_loss=0.05742, over 4870.00 frames. ], tot_loss[loss=0.1616, simple_loss=0.235, pruned_loss=0.04411, over 949190.35 frames. ], batch size: 34, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 03:56:28,735 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3717, 3.3072, 0.9554, 1.7751, 1.8276, 2.4070, 1.9288, 1.0008], device='cuda:4'), covar=tensor([0.1484, 0.0949, 0.1946, 0.1253, 0.1123, 0.0985, 0.1483, 0.2104], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0237, 0.0135, 0.0119, 0.0131, 0.0152, 0.0117, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 03:57:07,675 INFO [finetune.py:976] (4/7) Epoch 29, batch 1000, loss[loss=0.1603, simple_loss=0.2467, pruned_loss=0.03697, over 4900.00 frames. ], tot_loss[loss=0.1641, simple_loss=0.2376, pruned_loss=0.04532, over 950455.57 frames. ], batch size: 35, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 03:57:09,497 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.605e+02 1.983e+02 2.337e+02 3.564e+02, threshold=3.965e+02, percent-clipped=0.0 2023-04-28 03:57:38,280 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8722, 1.4138, 1.5312, 1.6964, 1.9887, 1.6742, 1.3995, 1.4622], device='cuda:4'), covar=tensor([0.1507, 0.1643, 0.1809, 0.1246, 0.0950, 0.1534, 0.1870, 0.2212], device='cuda:4'), in_proj_covar=tensor([0.0316, 0.0309, 0.0351, 0.0288, 0.0328, 0.0307, 0.0301, 0.0377], device='cuda:4'), out_proj_covar=tensor([6.4496e-05, 6.3351e-05, 7.3511e-05, 5.7688e-05, 6.6889e-05, 6.3702e-05, 6.2145e-05, 7.9699e-05], device='cuda:4') 2023-04-28 03:57:39,574 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=161401.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:58:12,304 INFO [finetune.py:976] (4/7) Epoch 29, batch 1050, loss[loss=0.1747, simple_loss=0.2527, pruned_loss=0.0484, over 4814.00 frames. ], tot_loss[loss=0.1662, simple_loss=0.2407, pruned_loss=0.0459, over 952221.31 frames. ], batch size: 39, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 03:58:54,279 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5309, 1.4758, 1.8870, 1.8162, 1.3988, 1.3084, 1.5079, 0.9007], device='cuda:4'), covar=tensor([0.0520, 0.0603, 0.0302, 0.0634, 0.0715, 0.0981, 0.0579, 0.0551], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0067, 0.0065, 0.0069, 0.0075, 0.0094, 0.0072, 0.0062], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-28 03:58:55,464 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=161460.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:58:56,752 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=161462.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 03:59:05,917 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9156, 1.7285, 1.9238, 2.3477, 2.3722, 1.9567, 1.6316, 2.1059], device='cuda:4'), covar=tensor([0.0854, 0.1231, 0.0884, 0.0602, 0.0603, 0.0860, 0.0745, 0.0578], device='cuda:4'), in_proj_covar=tensor([0.0184, 0.0204, 0.0184, 0.0170, 0.0178, 0.0177, 0.0151, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 03:59:17,656 INFO [finetune.py:976] (4/7) Epoch 29, batch 1100, loss[loss=0.1509, simple_loss=0.2324, pruned_loss=0.03469, over 4743.00 frames. ], tot_loss[loss=0.1673, simple_loss=0.242, pruned_loss=0.04634, over 953490.59 frames. ], batch size: 27, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 03:59:20,410 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.123e+02 1.507e+02 1.843e+02 2.398e+02 4.910e+02, threshold=3.687e+02, percent-clipped=4.0 2023-04-28 04:00:02,497 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-28 04:00:20,258 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=161521.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:00:23,058 INFO [finetune.py:976] (4/7) Epoch 29, batch 1150, loss[loss=0.1532, simple_loss=0.2297, pruned_loss=0.03836, over 4897.00 frames. ], tot_loss[loss=0.1664, simple_loss=0.2415, pruned_loss=0.04568, over 954155.82 frames. ], batch size: 46, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 04:00:53,905 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3913, 1.3243, 1.7002, 1.7419, 1.3183, 1.1759, 1.3955, 0.8366], device='cuda:4'), covar=tensor([0.0539, 0.0603, 0.0363, 0.0534, 0.0706, 0.0925, 0.0652, 0.0551], device='cuda:4'), in_proj_covar=tensor([0.0071, 0.0067, 0.0066, 0.0069, 0.0076, 0.0095, 0.0073, 0.0062], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-28 04:01:04,587 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=161557.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:01:07,694 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0260, 2.4058, 2.1531, 2.3656, 1.7363, 2.0490, 2.0969, 1.5857], device='cuda:4'), covar=tensor([0.1665, 0.1061, 0.0647, 0.1122, 0.3132, 0.0989, 0.1801, 0.2450], device='cuda:4'), in_proj_covar=tensor([0.0281, 0.0299, 0.0216, 0.0275, 0.0312, 0.0253, 0.0247, 0.0263], device='cuda:4'), out_proj_covar=tensor([1.1181e-04, 1.1755e-04, 8.4583e-05, 1.0808e-04, 1.2574e-04, 9.9298e-05, 9.9385e-05, 1.0364e-04], device='cuda:4') 2023-04-28 04:01:25,592 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=161572.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:01:27,862 INFO [finetune.py:976] (4/7) Epoch 29, batch 1200, loss[loss=0.1634, simple_loss=0.2371, pruned_loss=0.04487, over 4802.00 frames. ], tot_loss[loss=0.1664, simple_loss=0.2411, pruned_loss=0.0459, over 953093.08 frames. ], batch size: 29, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 04:01:29,681 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.012e+02 1.552e+02 1.827e+02 2.258e+02 5.032e+02, threshold=3.654e+02, percent-clipped=3.0 2023-04-28 04:01:47,188 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=161587.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:02:08,741 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1333, 2.6029, 2.1859, 2.4943, 1.6808, 2.0836, 2.2427, 1.6825], device='cuda:4'), covar=tensor([0.1988, 0.1293, 0.0853, 0.1202, 0.3685, 0.1081, 0.1838, 0.2732], device='cuda:4'), in_proj_covar=tensor([0.0279, 0.0297, 0.0214, 0.0274, 0.0310, 0.0251, 0.0246, 0.0262], device='cuda:4'), out_proj_covar=tensor([1.1130e-04, 1.1686e-04, 8.4064e-05, 1.0759e-04, 1.2498e-04, 9.8828e-05, 9.8893e-05, 1.0320e-04], device='cuda:4') 2023-04-28 04:02:09,298 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=161605.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:02:14,216 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6324, 1.6184, 0.8842, 1.3766, 1.8775, 1.5021, 1.3905, 1.5157], device='cuda:4'), covar=tensor([0.0476, 0.0370, 0.0321, 0.0543, 0.0252, 0.0484, 0.0462, 0.0569], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0027, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0052, 0.0046, 0.0039, 0.0053, 0.0039, 0.0051, 0.0051, 0.0053], device='cuda:4') 2023-04-28 04:02:19,284 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6152, 2.6269, 2.0394, 2.2947, 2.6833, 2.3011, 3.3510, 1.9401], device='cuda:4'), covar=tensor([0.3627, 0.2601, 0.4471, 0.3252, 0.1809, 0.2290, 0.1776, 0.4330], device='cuda:4'), in_proj_covar=tensor([0.0339, 0.0356, 0.0424, 0.0353, 0.0386, 0.0376, 0.0373, 0.0424], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 04:02:23,980 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=161621.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:02:33,466 INFO [finetune.py:976] (4/7) Epoch 29, batch 1250, loss[loss=0.1555, simple_loss=0.2202, pruned_loss=0.04542, over 4818.00 frames. ], tot_loss[loss=0.1647, simple_loss=0.2385, pruned_loss=0.04545, over 952677.18 frames. ], batch size: 25, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 04:02:41,597 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.95 vs. limit=2.0 2023-04-28 04:03:28,499 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=161669.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:03:35,474 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2551, 2.1749, 1.9141, 1.8138, 2.3349, 1.9068, 2.8467, 1.7241], device='cuda:4'), covar=tensor([0.3606, 0.2031, 0.4389, 0.3068, 0.1687, 0.2498, 0.1223, 0.4121], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0356, 0.0424, 0.0353, 0.0386, 0.0377, 0.0373, 0.0424], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 04:03:38,166 INFO [finetune.py:976] (4/7) Epoch 29, batch 1300, loss[loss=0.1496, simple_loss=0.2205, pruned_loss=0.03931, over 4895.00 frames. ], tot_loss[loss=0.1618, simple_loss=0.2352, pruned_loss=0.04423, over 953669.77 frames. ], batch size: 32, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 04:03:46,269 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.807e+01 1.476e+02 1.722e+02 2.175e+02 4.011e+02, threshold=3.444e+02, percent-clipped=1.0 2023-04-28 04:04:09,213 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6124, 1.4989, 1.8952, 1.9842, 1.4681, 1.3410, 1.5628, 0.9333], device='cuda:4'), covar=tensor([0.0633, 0.0656, 0.0423, 0.0528, 0.0832, 0.1256, 0.0652, 0.0643], device='cuda:4'), in_proj_covar=tensor([0.0071, 0.0067, 0.0065, 0.0069, 0.0075, 0.0094, 0.0072, 0.0062], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-28 04:04:42,649 INFO [finetune.py:976] (4/7) Epoch 29, batch 1350, loss[loss=0.1672, simple_loss=0.2405, pruned_loss=0.04698, over 4866.00 frames. ], tot_loss[loss=0.1621, simple_loss=0.2356, pruned_loss=0.04433, over 955578.88 frames. ], batch size: 34, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 04:04:54,108 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7255, 2.2092, 1.8996, 2.1466, 1.6123, 1.8235, 1.7267, 1.3194], device='cuda:4'), covar=tensor([0.2034, 0.1393, 0.0819, 0.1163, 0.3506, 0.1088, 0.1997, 0.2735], device='cuda:4'), in_proj_covar=tensor([0.0280, 0.0298, 0.0215, 0.0274, 0.0311, 0.0252, 0.0246, 0.0262], device='cuda:4'), out_proj_covar=tensor([1.1164e-04, 1.1718e-04, 8.4384e-05, 1.0784e-04, 1.2517e-04, 9.8873e-05, 9.9056e-05, 1.0340e-04], device='cuda:4') 2023-04-28 04:05:27,904 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=161757.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:05:50,110 INFO [finetune.py:976] (4/7) Epoch 29, batch 1400, loss[loss=0.1195, simple_loss=0.1932, pruned_loss=0.0229, over 4770.00 frames. ], tot_loss[loss=0.1647, simple_loss=0.2387, pruned_loss=0.04534, over 956362.24 frames. ], batch size: 26, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 04:05:58,004 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.100e+02 1.598e+02 1.895e+02 2.318e+02 6.343e+02, threshold=3.789e+02, percent-clipped=7.0 2023-04-28 04:06:51,804 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=161815.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:06:52,364 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=161816.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:07:03,299 INFO [finetune.py:976] (4/7) Epoch 29, batch 1450, loss[loss=0.1574, simple_loss=0.2458, pruned_loss=0.03447, over 4768.00 frames. ], tot_loss[loss=0.1651, simple_loss=0.2394, pruned_loss=0.04535, over 955421.34 frames. ], batch size: 28, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 04:07:18,043 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.93 vs. limit=5.0 2023-04-28 04:07:28,293 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8195, 2.2722, 1.8237, 1.6952, 1.3305, 1.3676, 1.9106, 1.3037], device='cuda:4'), covar=tensor([0.1790, 0.1337, 0.1419, 0.1712, 0.2362, 0.1943, 0.0994, 0.2115], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0210, 0.0171, 0.0205, 0.0202, 0.0187, 0.0157, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 04:07:36,446 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1055, 2.4960, 2.1593, 2.4126, 1.7176, 2.0855, 1.9904, 1.7780], device='cuda:4'), covar=tensor([0.1799, 0.1271, 0.0759, 0.1028, 0.3457, 0.0986, 0.1899, 0.2432], device='cuda:4'), in_proj_covar=tensor([0.0280, 0.0298, 0.0215, 0.0274, 0.0311, 0.0251, 0.0246, 0.0262], device='cuda:4'), out_proj_covar=tensor([1.1162e-04, 1.1712e-04, 8.4266e-05, 1.0779e-04, 1.2530e-04, 9.8668e-05, 9.9089e-05, 1.0318e-04], device='cuda:4') 2023-04-28 04:08:00,373 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=161872.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:08:02,915 INFO [finetune.py:976] (4/7) Epoch 29, batch 1500, loss[loss=0.1881, simple_loss=0.2552, pruned_loss=0.06055, over 4900.00 frames. ], tot_loss[loss=0.1676, simple_loss=0.242, pruned_loss=0.04664, over 957003.21 frames. ], batch size: 36, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 04:08:03,653 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=161876.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:08:04,722 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.141e+02 1.641e+02 1.910e+02 2.350e+02 4.691e+02, threshold=3.820e+02, percent-clipped=1.0 2023-04-28 04:08:16,476 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=161887.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:08:38,564 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.68 vs. limit=2.0 2023-04-28 04:09:01,027 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=161920.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:09:09,620 INFO [finetune.py:976] (4/7) Epoch 29, batch 1550, loss[loss=0.1852, simple_loss=0.249, pruned_loss=0.06069, over 4899.00 frames. ], tot_loss[loss=0.1689, simple_loss=0.2434, pruned_loss=0.04718, over 955898.69 frames. ], batch size: 43, lr: 2.85e-03, grad_scale: 16.0 2023-04-28 04:09:21,319 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=161935.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:10:14,713 INFO [finetune.py:976] (4/7) Epoch 29, batch 1600, loss[loss=0.1502, simple_loss=0.2232, pruned_loss=0.03856, over 4888.00 frames. ], tot_loss[loss=0.168, simple_loss=0.2416, pruned_loss=0.04715, over 957289.96 frames. ], batch size: 32, lr: 2.84e-03, grad_scale: 16.0 2023-04-28 04:10:16,469 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 7.098e+01 1.531e+02 1.823e+02 2.197e+02 4.043e+02, threshold=3.646e+02, percent-clipped=1.0 2023-04-28 04:11:10,810 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.16 vs. limit=2.0 2023-04-28 04:11:20,428 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9664, 1.5054, 1.7744, 1.7977, 1.7215, 1.4557, 0.9299, 1.4707], device='cuda:4'), covar=tensor([0.2749, 0.2766, 0.1479, 0.1875, 0.2261, 0.2295, 0.3763, 0.1776], device='cuda:4'), in_proj_covar=tensor([0.0289, 0.0244, 0.0226, 0.0311, 0.0221, 0.0233, 0.0226, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-28 04:11:22,118 INFO [finetune.py:976] (4/7) Epoch 29, batch 1650, loss[loss=0.1635, simple_loss=0.233, pruned_loss=0.04694, over 4904.00 frames. ], tot_loss[loss=0.1653, simple_loss=0.2382, pruned_loss=0.04619, over 955615.12 frames. ], batch size: 32, lr: 2.84e-03, grad_scale: 16.0 2023-04-28 04:12:07,061 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=162057.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:12:29,332 INFO [finetune.py:976] (4/7) Epoch 29, batch 1700, loss[loss=0.1308, simple_loss=0.213, pruned_loss=0.0243, over 4748.00 frames. ], tot_loss[loss=0.1633, simple_loss=0.2355, pruned_loss=0.04554, over 953141.07 frames. ], batch size: 59, lr: 2.84e-03, grad_scale: 16.0 2023-04-28 04:12:36,096 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=162077.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:12:36,610 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.915e+01 1.582e+02 1.975e+02 2.280e+02 6.731e+02, threshold=3.951e+02, percent-clipped=4.0 2023-04-28 04:13:11,491 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=162105.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:13:24,528 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=162116.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:13:29,897 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9062, 2.0029, 1.1014, 1.6803, 2.3386, 1.7891, 1.6505, 1.8953], device='cuda:4'), covar=tensor([0.0470, 0.0338, 0.0273, 0.0510, 0.0227, 0.0454, 0.0445, 0.0514], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0028, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0052, 0.0047, 0.0039, 0.0054, 0.0039, 0.0051, 0.0051, 0.0053], device='cuda:4') 2023-04-28 04:13:34,110 INFO [finetune.py:976] (4/7) Epoch 29, batch 1750, loss[loss=0.1449, simple_loss=0.2171, pruned_loss=0.03638, over 4757.00 frames. ], tot_loss[loss=0.1641, simple_loss=0.2367, pruned_loss=0.04576, over 953110.14 frames. ], batch size: 23, lr: 2.84e-03, grad_scale: 16.0 2023-04-28 04:13:40,744 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0693, 2.6278, 1.0891, 1.3678, 2.0955, 1.2459, 3.5979, 1.8511], device='cuda:4'), covar=tensor([0.0710, 0.0601, 0.0827, 0.1337, 0.0489, 0.1057, 0.0250, 0.0628], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0073, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-28 04:13:51,950 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=162138.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:14:27,313 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=162164.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:14:37,342 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=162171.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:14:45,211 INFO [finetune.py:976] (4/7) Epoch 29, batch 1800, loss[loss=0.2235, simple_loss=0.2997, pruned_loss=0.07371, over 4793.00 frames. ], tot_loss[loss=0.1663, simple_loss=0.2402, pruned_loss=0.0462, over 954562.21 frames. ], batch size: 45, lr: 2.84e-03, grad_scale: 16.0 2023-04-28 04:14:47,021 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.098e+02 1.596e+02 1.871e+02 2.266e+02 6.327e+02, threshold=3.743e+02, percent-clipped=2.0 2023-04-28 04:14:56,440 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2570, 1.5101, 1.4452, 1.7081, 1.6163, 1.8058, 1.4213, 3.3410], device='cuda:4'), covar=tensor([0.0633, 0.0825, 0.0790, 0.1255, 0.0630, 0.0529, 0.0765, 0.0136], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0042, 0.0040, 0.0037, 0.0038, 0.0054], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0014, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 04:15:20,870 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9409, 2.2606, 0.8872, 1.1968, 1.5514, 1.1099, 2.3935, 1.4305], device='cuda:4'), covar=tensor([0.0607, 0.0451, 0.0582, 0.1187, 0.0418, 0.0979, 0.0385, 0.0612], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0073, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-28 04:15:40,573 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2500, 1.5440, 1.3467, 1.5261, 1.3656, 1.2985, 1.2725, 1.0664], device='cuda:4'), covar=tensor([0.1755, 0.1398, 0.1020, 0.1315, 0.3632, 0.1259, 0.1996, 0.2357], device='cuda:4'), in_proj_covar=tensor([0.0282, 0.0300, 0.0216, 0.0276, 0.0314, 0.0252, 0.0248, 0.0263], device='cuda:4'), out_proj_covar=tensor([1.1244e-04, 1.1792e-04, 8.4774e-05, 1.0846e-04, 1.2629e-04, 9.9113e-05, 9.9742e-05, 1.0366e-04], device='cuda:4') 2023-04-28 04:15:49,784 INFO [finetune.py:976] (4/7) Epoch 29, batch 1850, loss[loss=0.184, simple_loss=0.2677, pruned_loss=0.05018, over 4911.00 frames. ], tot_loss[loss=0.1666, simple_loss=0.2409, pruned_loss=0.04619, over 954653.25 frames. ], batch size: 36, lr: 2.84e-03, grad_scale: 16.0 2023-04-28 04:15:53,545 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=162231.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:16:53,819 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-04-28 04:16:53,968 INFO [finetune.py:976] (4/7) Epoch 29, batch 1900, loss[loss=0.1578, simple_loss=0.2323, pruned_loss=0.04169, over 4872.00 frames. ], tot_loss[loss=0.1685, simple_loss=0.2425, pruned_loss=0.04726, over 954770.52 frames. ], batch size: 35, lr: 2.84e-03, grad_scale: 16.0 2023-04-28 04:16:55,788 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.192e+01 1.518e+02 1.776e+02 2.128e+02 3.542e+02, threshold=3.553e+02, percent-clipped=0.0 2023-04-28 04:17:14,635 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=162292.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:17:35,914 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-04-28 04:17:58,316 INFO [finetune.py:976] (4/7) Epoch 29, batch 1950, loss[loss=0.1558, simple_loss=0.2288, pruned_loss=0.04133, over 4889.00 frames. ], tot_loss[loss=0.1678, simple_loss=0.2418, pruned_loss=0.04689, over 956096.41 frames. ], batch size: 32, lr: 2.84e-03, grad_scale: 16.0 2023-04-28 04:18:51,450 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6354, 1.5568, 1.7708, 2.0409, 2.0369, 1.6080, 1.4279, 1.7391], device='cuda:4'), covar=tensor([0.0753, 0.1091, 0.0723, 0.0471, 0.0551, 0.0747, 0.0664, 0.0551], device='cuda:4'), in_proj_covar=tensor([0.0184, 0.0203, 0.0184, 0.0170, 0.0177, 0.0177, 0.0151, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 04:19:03,740 INFO [finetune.py:976] (4/7) Epoch 29, batch 2000, loss[loss=0.1642, simple_loss=0.2302, pruned_loss=0.04913, over 4826.00 frames. ], tot_loss[loss=0.1653, simple_loss=0.2388, pruned_loss=0.04588, over 954932.99 frames. ], batch size: 30, lr: 2.84e-03, grad_scale: 16.0 2023-04-28 04:19:05,552 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.225e+01 1.505e+02 1.783e+02 2.109e+02 5.594e+02, threshold=3.566e+02, percent-clipped=2.0 2023-04-28 04:19:30,710 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.90 vs. limit=5.0 2023-04-28 04:19:43,205 INFO [finetune.py:976] (4/7) Epoch 29, batch 2050, loss[loss=0.1573, simple_loss=0.2278, pruned_loss=0.04334, over 4836.00 frames. ], tot_loss[loss=0.1616, simple_loss=0.2346, pruned_loss=0.04426, over 957176.25 frames. ], batch size: 30, lr: 2.84e-03, grad_scale: 16.0 2023-04-28 04:19:48,143 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=162433.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:20:13,307 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=162471.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:20:16,687 INFO [finetune.py:976] (4/7) Epoch 29, batch 2100, loss[loss=0.1612, simple_loss=0.2399, pruned_loss=0.0413, over 4718.00 frames. ], tot_loss[loss=0.1622, simple_loss=0.2352, pruned_loss=0.04463, over 954844.48 frames. ], batch size: 23, lr: 2.84e-03, grad_scale: 32.0 2023-04-28 04:20:19,022 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.161e+01 1.530e+02 1.785e+02 2.249e+02 3.474e+02, threshold=3.570e+02, percent-clipped=1.0 2023-04-28 04:20:39,348 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=162511.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:20:44,606 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.61 vs. limit=2.0 2023-04-28 04:20:45,707 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=162519.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:20:50,372 INFO [finetune.py:976] (4/7) Epoch 29, batch 2150, loss[loss=0.1598, simple_loss=0.2543, pruned_loss=0.0327, over 4785.00 frames. ], tot_loss[loss=0.1642, simple_loss=0.238, pruned_loss=0.04516, over 952502.72 frames. ], batch size: 29, lr: 2.84e-03, grad_scale: 32.0 2023-04-28 04:20:52,009 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4149, 1.7876, 1.6003, 2.1418, 1.9867, 1.9614, 1.7408, 4.6230], device='cuda:4'), covar=tensor([0.0575, 0.0843, 0.0845, 0.1229, 0.0649, 0.0529, 0.0741, 0.0083], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0042, 0.0040, 0.0037, 0.0038, 0.0054], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 04:21:21,583 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=162572.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 04:21:23,767 INFO [finetune.py:976] (4/7) Epoch 29, batch 2200, loss[loss=0.1833, simple_loss=0.2563, pruned_loss=0.05518, over 4899.00 frames. ], tot_loss[loss=0.1658, simple_loss=0.2398, pruned_loss=0.04591, over 953897.47 frames. ], batch size: 36, lr: 2.84e-03, grad_scale: 32.0 2023-04-28 04:21:26,030 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.398e+01 1.632e+02 1.891e+02 2.223e+02 3.490e+02, threshold=3.782e+02, percent-clipped=0.0 2023-04-28 04:21:37,946 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=162587.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:21:39,916 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-28 04:22:23,022 INFO [finetune.py:976] (4/7) Epoch 29, batch 2250, loss[loss=0.1369, simple_loss=0.2112, pruned_loss=0.03132, over 4682.00 frames. ], tot_loss[loss=0.1661, simple_loss=0.2408, pruned_loss=0.04576, over 954619.93 frames. ], batch size: 23, lr: 2.84e-03, grad_scale: 32.0 2023-04-28 04:23:28,884 INFO [finetune.py:976] (4/7) Epoch 29, batch 2300, loss[loss=0.1972, simple_loss=0.258, pruned_loss=0.06814, over 4907.00 frames. ], tot_loss[loss=0.1665, simple_loss=0.2411, pruned_loss=0.04593, over 953829.09 frames. ], batch size: 37, lr: 2.84e-03, grad_scale: 32.0 2023-04-28 04:23:36,410 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.802e+01 1.609e+02 1.813e+02 2.047e+02 3.617e+02, threshold=3.626e+02, percent-clipped=0.0 2023-04-28 04:23:36,554 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8732, 1.8044, 2.4934, 2.4790, 1.7828, 1.6441, 1.9362, 1.0704], device='cuda:4'), covar=tensor([0.0580, 0.0658, 0.0289, 0.0669, 0.0718, 0.0960, 0.0566, 0.0668], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0067, 0.0065, 0.0069, 0.0075, 0.0094, 0.0072, 0.0062], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-28 04:23:47,282 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1750, 1.5825, 1.3858, 1.7940, 1.6901, 1.7554, 1.4493, 3.4407], device='cuda:4'), covar=tensor([0.0613, 0.0766, 0.0784, 0.1153, 0.0611, 0.0529, 0.0718, 0.0152], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0042, 0.0040, 0.0037, 0.0037, 0.0054], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 04:23:58,745 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9582, 1.8137, 2.0293, 2.2731, 2.3668, 1.8791, 1.7020, 2.0669], device='cuda:4'), covar=tensor([0.0791, 0.1045, 0.0658, 0.0561, 0.0588, 0.0852, 0.0716, 0.0537], device='cuda:4'), in_proj_covar=tensor([0.0184, 0.0202, 0.0183, 0.0170, 0.0177, 0.0177, 0.0151, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 04:24:29,165 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.21 vs. limit=2.0 2023-04-28 04:24:31,808 INFO [finetune.py:976] (4/7) Epoch 29, batch 2350, loss[loss=0.1318, simple_loss=0.2005, pruned_loss=0.03154, over 4913.00 frames. ], tot_loss[loss=0.1658, simple_loss=0.2396, pruned_loss=0.04604, over 955934.92 frames. ], batch size: 46, lr: 2.84e-03, grad_scale: 32.0 2023-04-28 04:24:32,262 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.35 vs. limit=2.0 2023-04-28 04:24:42,013 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=162733.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:25:30,907 INFO [finetune.py:976] (4/7) Epoch 29, batch 2400, loss[loss=0.1344, simple_loss=0.2027, pruned_loss=0.03306, over 4861.00 frames. ], tot_loss[loss=0.1644, simple_loss=0.2375, pruned_loss=0.04567, over 957524.65 frames. ], batch size: 31, lr: 2.84e-03, grad_scale: 32.0 2023-04-28 04:25:33,145 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.191e+01 1.527e+02 1.810e+02 2.223e+02 4.938e+02, threshold=3.619e+02, percent-clipped=3.0 2023-04-28 04:25:36,110 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=162781.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:25:37,386 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3709, 1.7976, 1.6540, 2.3165, 2.4651, 2.0560, 1.9568, 1.6924], device='cuda:4'), covar=tensor([0.1699, 0.1752, 0.1708, 0.1310, 0.1004, 0.1585, 0.1979, 0.2058], device='cuda:4'), in_proj_covar=tensor([0.0316, 0.0309, 0.0351, 0.0287, 0.0326, 0.0306, 0.0302, 0.0377], device='cuda:4'), out_proj_covar=tensor([6.4484e-05, 6.3309e-05, 7.3328e-05, 5.7334e-05, 6.6403e-05, 6.3478e-05, 6.2210e-05, 7.9725e-05], device='cuda:4') 2023-04-28 04:26:04,588 INFO [finetune.py:976] (4/7) Epoch 29, batch 2450, loss[loss=0.1672, simple_loss=0.2306, pruned_loss=0.05189, over 4750.00 frames. ], tot_loss[loss=0.1615, simple_loss=0.234, pruned_loss=0.04447, over 957937.37 frames. ], batch size: 28, lr: 2.84e-03, grad_scale: 32.0 2023-04-28 04:26:05,961 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=162827.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:26:33,770 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=162867.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 04:26:38,577 INFO [finetune.py:976] (4/7) Epoch 29, batch 2500, loss[loss=0.1275, simple_loss=0.2069, pruned_loss=0.02401, over 4761.00 frames. ], tot_loss[loss=0.1624, simple_loss=0.2351, pruned_loss=0.04482, over 957851.16 frames. ], batch size: 27, lr: 2.84e-03, grad_scale: 32.0 2023-04-28 04:26:40,361 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.423e+01 1.461e+02 1.777e+02 2.062e+02 3.489e+02, threshold=3.555e+02, percent-clipped=0.0 2023-04-28 04:26:47,957 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=162887.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:26:48,582 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=162888.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:27:12,361 INFO [finetune.py:976] (4/7) Epoch 29, batch 2550, loss[loss=0.1475, simple_loss=0.2253, pruned_loss=0.03479, over 4779.00 frames. ], tot_loss[loss=0.165, simple_loss=0.2389, pruned_loss=0.04554, over 955984.58 frames. ], batch size: 29, lr: 2.84e-03, grad_scale: 32.0 2023-04-28 04:27:18,508 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2676, 1.4336, 1.6623, 1.7584, 1.6771, 1.7739, 1.7209, 1.6969], device='cuda:4'), covar=tensor([0.3563, 0.4557, 0.4191, 0.4130, 0.4828, 0.6310, 0.4463, 0.4230], device='cuda:4'), in_proj_covar=tensor([0.0344, 0.0375, 0.0332, 0.0343, 0.0352, 0.0396, 0.0363, 0.0335], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 04:27:19,031 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=162935.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:27:46,026 INFO [finetune.py:976] (4/7) Epoch 29, batch 2600, loss[loss=0.1482, simple_loss=0.2201, pruned_loss=0.03815, over 4871.00 frames. ], tot_loss[loss=0.1674, simple_loss=0.2414, pruned_loss=0.04674, over 955711.34 frames. ], batch size: 34, lr: 2.84e-03, grad_scale: 32.0 2023-04-28 04:27:47,802 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.782e+01 1.602e+02 1.933e+02 2.328e+02 3.675e+02, threshold=3.867e+02, percent-clipped=1.0 2023-04-28 04:28:12,096 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.81 vs. limit=2.0 2023-04-28 04:28:19,511 INFO [finetune.py:976] (4/7) Epoch 29, batch 2650, loss[loss=0.159, simple_loss=0.2466, pruned_loss=0.03568, over 4800.00 frames. ], tot_loss[loss=0.1679, simple_loss=0.2423, pruned_loss=0.04679, over 956422.88 frames. ], batch size: 41, lr: 2.84e-03, grad_scale: 32.0 2023-04-28 04:28:57,568 INFO [finetune.py:976] (4/7) Epoch 29, batch 2700, loss[loss=0.1579, simple_loss=0.2173, pruned_loss=0.04927, over 4776.00 frames. ], tot_loss[loss=0.167, simple_loss=0.2412, pruned_loss=0.04638, over 956164.82 frames. ], batch size: 23, lr: 2.84e-03, grad_scale: 32.0 2023-04-28 04:29:04,329 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.042e+02 1.516e+02 1.799e+02 2.193e+02 4.304e+02, threshold=3.599e+02, percent-clipped=1.0 2023-04-28 04:29:15,565 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8193, 2.3051, 1.8606, 1.7383, 1.3455, 1.3932, 1.8509, 1.3534], device='cuda:4'), covar=tensor([0.1594, 0.1191, 0.1315, 0.1544, 0.2237, 0.1918, 0.0913, 0.1927], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0209, 0.0170, 0.0203, 0.0200, 0.0186, 0.0155, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 04:29:17,297 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=163090.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:30:02,210 INFO [finetune.py:976] (4/7) Epoch 29, batch 2750, loss[loss=0.1572, simple_loss=0.2195, pruned_loss=0.04743, over 4064.00 frames. ], tot_loss[loss=0.1656, simple_loss=0.2393, pruned_loss=0.04591, over 955017.81 frames. ], batch size: 17, lr: 2.84e-03, grad_scale: 32.0 2023-04-28 04:30:11,869 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1532, 2.7613, 2.1723, 2.2621, 1.5748, 1.6285, 2.2353, 1.5826], device='cuda:4'), covar=tensor([0.1615, 0.1441, 0.1257, 0.1515, 0.2230, 0.1807, 0.0897, 0.1912], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0209, 0.0169, 0.0203, 0.0200, 0.0186, 0.0155, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 04:30:38,998 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=163151.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:30:50,661 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=163167.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 04:31:01,088 INFO [finetune.py:976] (4/7) Epoch 29, batch 2800, loss[loss=0.1549, simple_loss=0.2298, pruned_loss=0.04003, over 4923.00 frames. ], tot_loss[loss=0.1632, simple_loss=0.2365, pruned_loss=0.04494, over 955045.32 frames. ], batch size: 28, lr: 2.84e-03, grad_scale: 32.0 2023-04-28 04:31:08,097 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.000e+02 1.442e+02 1.723e+02 2.053e+02 5.173e+02, threshold=3.446e+02, percent-clipped=1.0 2023-04-28 04:31:11,254 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=163183.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:31:52,469 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=163215.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:31:59,019 INFO [finetune.py:976] (4/7) Epoch 29, batch 2850, loss[loss=0.1295, simple_loss=0.2021, pruned_loss=0.02841, over 4780.00 frames. ], tot_loss[loss=0.1603, simple_loss=0.2337, pruned_loss=0.04344, over 955366.67 frames. ], batch size: 26, lr: 2.84e-03, grad_scale: 32.0 2023-04-28 04:32:08,286 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=163240.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:32:31,405 INFO [finetune.py:976] (4/7) Epoch 29, batch 2900, loss[loss=0.159, simple_loss=0.2336, pruned_loss=0.04219, over 4762.00 frames. ], tot_loss[loss=0.1635, simple_loss=0.2368, pruned_loss=0.04509, over 956569.17 frames. ], batch size: 27, lr: 2.84e-03, grad_scale: 32.0 2023-04-28 04:32:33,721 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.059e+02 1.435e+02 1.823e+02 2.184e+02 4.986e+02, threshold=3.647e+02, percent-clipped=1.0 2023-04-28 04:32:47,862 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=163301.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:33:04,178 INFO [finetune.py:976] (4/7) Epoch 29, batch 2950, loss[loss=0.1695, simple_loss=0.2384, pruned_loss=0.05024, over 4819.00 frames. ], tot_loss[loss=0.1656, simple_loss=0.2392, pruned_loss=0.046, over 954988.10 frames. ], batch size: 39, lr: 2.84e-03, grad_scale: 32.0 2023-04-28 04:33:27,669 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6343, 1.6609, 0.7406, 1.3511, 1.7755, 1.5276, 1.4156, 1.4927], device='cuda:4'), covar=tensor([0.0488, 0.0387, 0.0352, 0.0556, 0.0269, 0.0524, 0.0483, 0.0571], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0028, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0052, 0.0047, 0.0039, 0.0054, 0.0039, 0.0051, 0.0051, 0.0053], device='cuda:4') 2023-04-28 04:33:37,006 INFO [finetune.py:976] (4/7) Epoch 29, batch 3000, loss[loss=0.1459, simple_loss=0.2358, pruned_loss=0.02802, over 4819.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.2408, pruned_loss=0.04638, over 956073.45 frames. ], batch size: 38, lr: 2.84e-03, grad_scale: 32.0 2023-04-28 04:33:37,006 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-28 04:33:43,981 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0819, 2.5492, 1.0712, 1.3579, 1.9386, 1.3106, 3.0161, 1.6558], device='cuda:4'), covar=tensor([0.0611, 0.0508, 0.0674, 0.1152, 0.0419, 0.0872, 0.0235, 0.0564], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0063, 0.0047, 0.0046, 0.0049, 0.0051, 0.0072, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-28 04:33:47,842 INFO [finetune.py:1010] (4/7) Epoch 29, validation: loss=0.1535, simple_loss=0.222, pruned_loss=0.04252, over 2265189.00 frames. 2023-04-28 04:33:47,842 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6529MB 2023-04-28 04:33:49,649 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.160e+02 1.593e+02 1.913e+02 2.268e+02 4.606e+02, threshold=3.825e+02, percent-clipped=1.0 2023-04-28 04:34:17,716 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.17 vs. limit=2.0 2023-04-28 04:34:19,351 INFO [finetune.py:976] (4/7) Epoch 29, batch 3050, loss[loss=0.1511, simple_loss=0.2349, pruned_loss=0.03367, over 4820.00 frames. ], tot_loss[loss=0.1669, simple_loss=0.2418, pruned_loss=0.04603, over 956162.08 frames. ], batch size: 41, lr: 2.84e-03, grad_scale: 32.0 2023-04-28 04:34:34,896 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=163446.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:34:53,042 INFO [finetune.py:976] (4/7) Epoch 29, batch 3100, loss[loss=0.1233, simple_loss=0.1988, pruned_loss=0.02389, over 4743.00 frames. ], tot_loss[loss=0.1653, simple_loss=0.2395, pruned_loss=0.04549, over 954891.48 frames. ], batch size: 23, lr: 2.84e-03, grad_scale: 32.0 2023-04-28 04:34:55,820 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.898e+01 1.464e+02 1.707e+02 2.164e+02 5.622e+02, threshold=3.413e+02, percent-clipped=1.0 2023-04-28 04:34:59,439 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=163483.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:35:12,450 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0404, 1.4447, 1.3333, 1.7371, 1.5100, 1.7917, 1.2835, 3.4027], device='cuda:4'), covar=tensor([0.0748, 0.1049, 0.1018, 0.1375, 0.0805, 0.0650, 0.0974, 0.0219], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0042, 0.0040, 0.0037, 0.0038, 0.0054], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 04:35:27,055 INFO [finetune.py:976] (4/7) Epoch 29, batch 3150, loss[loss=0.1362, simple_loss=0.21, pruned_loss=0.03125, over 4888.00 frames. ], tot_loss[loss=0.1622, simple_loss=0.2357, pruned_loss=0.04436, over 955413.81 frames. ], batch size: 32, lr: 2.84e-03, grad_scale: 32.0 2023-04-28 04:35:31,735 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=163531.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:36:03,515 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.38 vs. limit=5.0 2023-04-28 04:36:10,391 INFO [finetune.py:976] (4/7) Epoch 29, batch 3200, loss[loss=0.1716, simple_loss=0.2413, pruned_loss=0.05094, over 4931.00 frames. ], tot_loss[loss=0.1607, simple_loss=0.2331, pruned_loss=0.0441, over 957686.54 frames. ], batch size: 33, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 04:36:12,741 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.620e+01 1.506e+02 1.753e+02 2.068e+02 8.624e+02, threshold=3.506e+02, percent-clipped=5.0 2023-04-28 04:36:42,516 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=163596.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:37:17,128 INFO [finetune.py:976] (4/7) Epoch 29, batch 3250, loss[loss=0.1501, simple_loss=0.2308, pruned_loss=0.03475, over 4803.00 frames. ], tot_loss[loss=0.1615, simple_loss=0.2339, pruned_loss=0.04455, over 956603.12 frames. ], batch size: 45, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 04:37:39,298 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3325, 1.8500, 2.2455, 2.5558, 2.2482, 1.8048, 1.4036, 1.9889], device='cuda:4'), covar=tensor([0.2851, 0.2769, 0.1476, 0.1961, 0.2376, 0.2460, 0.3894, 0.1822], device='cuda:4'), in_proj_covar=tensor([0.0290, 0.0244, 0.0226, 0.0311, 0.0220, 0.0234, 0.0227, 0.0184], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-28 04:37:57,201 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.6965, 3.6126, 2.8503, 4.2928, 3.6469, 3.6992, 1.4565, 3.6220], device='cuda:4'), covar=tensor([0.1846, 0.1309, 0.3255, 0.1886, 0.3977, 0.1889, 0.6533, 0.2529], device='cuda:4'), in_proj_covar=tensor([0.0249, 0.0222, 0.0254, 0.0308, 0.0304, 0.0254, 0.0279, 0.0278], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 04:38:20,408 INFO [finetune.py:976] (4/7) Epoch 29, batch 3300, loss[loss=0.1833, simple_loss=0.2679, pruned_loss=0.04934, over 4908.00 frames. ], tot_loss[loss=0.1643, simple_loss=0.2375, pruned_loss=0.04558, over 958378.35 frames. ], batch size: 35, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 04:38:22,206 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.182e+02 1.605e+02 1.841e+02 2.298e+02 3.971e+02, threshold=3.681e+02, percent-clipped=3.0 2023-04-28 04:39:13,617 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8788, 2.1882, 1.2224, 1.5735, 2.4532, 1.7950, 1.6556, 1.8123], device='cuda:4'), covar=tensor([0.0466, 0.0327, 0.0266, 0.0506, 0.0207, 0.0470, 0.0445, 0.0529], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0027, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0052, 0.0047, 0.0039, 0.0053, 0.0039, 0.0051, 0.0051, 0.0053], device='cuda:4') 2023-04-28 04:39:14,452 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.61 vs. limit=2.0 2023-04-28 04:39:22,576 INFO [finetune.py:976] (4/7) Epoch 29, batch 3350, loss[loss=0.154, simple_loss=0.2344, pruned_loss=0.03687, over 4728.00 frames. ], tot_loss[loss=0.1652, simple_loss=0.2394, pruned_loss=0.04546, over 958380.80 frames. ], batch size: 54, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 04:39:46,996 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=163746.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:39:48,624 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.39 vs. limit=2.0 2023-04-28 04:40:27,095 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8092, 1.5065, 1.4149, 1.5919, 2.0326, 1.6339, 1.4333, 1.3664], device='cuda:4'), covar=tensor([0.2135, 0.1568, 0.2388, 0.1493, 0.0875, 0.1935, 0.2138, 0.2532], device='cuda:4'), in_proj_covar=tensor([0.0314, 0.0308, 0.0349, 0.0286, 0.0325, 0.0305, 0.0300, 0.0375], device='cuda:4'), out_proj_covar=tensor([6.3896e-05, 6.3087e-05, 7.3067e-05, 5.7157e-05, 6.6349e-05, 6.3360e-05, 6.1873e-05, 7.9099e-05], device='cuda:4') 2023-04-28 04:40:27,594 INFO [finetune.py:976] (4/7) Epoch 29, batch 3400, loss[loss=0.1573, simple_loss=0.2406, pruned_loss=0.03694, over 4800.00 frames. ], tot_loss[loss=0.1667, simple_loss=0.2411, pruned_loss=0.04616, over 958514.58 frames. ], batch size: 29, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 04:40:29,472 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.152e+02 1.589e+02 1.829e+02 2.322e+02 5.159e+02, threshold=3.657e+02, percent-clipped=5.0 2023-04-28 04:40:50,563 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=163794.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:41:20,998 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-28 04:41:32,560 INFO [finetune.py:976] (4/7) Epoch 29, batch 3450, loss[loss=0.179, simple_loss=0.2603, pruned_loss=0.04886, over 4799.00 frames. ], tot_loss[loss=0.1663, simple_loss=0.2411, pruned_loss=0.04579, over 956159.10 frames. ], batch size: 29, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 04:41:34,580 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4230, 1.9374, 2.2780, 2.9756, 2.3162, 1.9343, 1.9184, 2.1528], device='cuda:4'), covar=tensor([0.2782, 0.2936, 0.1479, 0.1933, 0.2532, 0.2339, 0.3468, 0.1996], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0246, 0.0229, 0.0314, 0.0223, 0.0237, 0.0229, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-28 04:42:36,323 INFO [finetune.py:976] (4/7) Epoch 29, batch 3500, loss[loss=0.1588, simple_loss=0.2312, pruned_loss=0.04325, over 4930.00 frames. ], tot_loss[loss=0.1649, simple_loss=0.2387, pruned_loss=0.04561, over 954685.99 frames. ], batch size: 38, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 04:42:37,683 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0705, 0.7180, 0.8945, 0.7752, 1.1928, 0.9534, 0.8948, 0.9118], device='cuda:4'), covar=tensor([0.1562, 0.1604, 0.1837, 0.1601, 0.1030, 0.1441, 0.1476, 0.2090], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0307, 0.0347, 0.0285, 0.0324, 0.0304, 0.0298, 0.0373], device='cuda:4'), out_proj_covar=tensor([6.3694e-05, 6.2839e-05, 7.2565e-05, 5.7053e-05, 6.6138e-05, 6.3091e-05, 6.1513e-05, 7.8719e-05], device='cuda:4') 2023-04-28 04:42:38,148 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.042e+02 1.447e+02 1.753e+02 2.110e+02 3.235e+02, threshold=3.507e+02, percent-clipped=0.0 2023-04-28 04:42:59,214 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=163896.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:43:28,395 INFO [finetune.py:976] (4/7) Epoch 29, batch 3550, loss[loss=0.1546, simple_loss=0.2348, pruned_loss=0.03718, over 4829.00 frames. ], tot_loss[loss=0.1626, simple_loss=0.2358, pruned_loss=0.04469, over 956058.94 frames. ], batch size: 33, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 04:43:39,968 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=163944.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:43:55,255 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0548, 1.3378, 1.2602, 1.6186, 1.4690, 1.4626, 1.2600, 2.3567], device='cuda:4'), covar=tensor([0.0562, 0.0703, 0.0748, 0.1095, 0.0596, 0.0428, 0.0675, 0.0202], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0042, 0.0040, 0.0037, 0.0037, 0.0054], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 04:44:01,287 INFO [finetune.py:976] (4/7) Epoch 29, batch 3600, loss[loss=0.1923, simple_loss=0.2602, pruned_loss=0.06226, over 4755.00 frames. ], tot_loss[loss=0.1598, simple_loss=0.2327, pruned_loss=0.04347, over 956235.27 frames. ], batch size: 54, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 04:44:03,074 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5462, 2.5352, 1.9168, 2.2181, 2.5564, 2.1591, 3.1825, 1.7918], device='cuda:4'), covar=tensor([0.3863, 0.2344, 0.4868, 0.3127, 0.1883, 0.2516, 0.1713, 0.4817], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0356, 0.0424, 0.0354, 0.0386, 0.0377, 0.0372, 0.0427], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 04:44:03,546 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.301e+01 1.508e+02 1.790e+02 2.030e+02 3.807e+02, threshold=3.580e+02, percent-clipped=2.0 2023-04-28 04:44:16,913 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-04-28 04:44:28,508 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0818, 1.3912, 5.1511, 4.8604, 4.4627, 4.9676, 4.6083, 4.5134], device='cuda:4'), covar=tensor([0.7258, 0.6491, 0.0968, 0.1794, 0.1192, 0.1664, 0.1275, 0.1703], device='cuda:4'), in_proj_covar=tensor([0.0307, 0.0307, 0.0405, 0.0407, 0.0348, 0.0412, 0.0317, 0.0360], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-28 04:44:34,415 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1740, 2.5111, 0.8706, 1.5435, 1.5163, 1.8575, 1.6375, 0.9163], device='cuda:4'), covar=tensor([0.1378, 0.1135, 0.1662, 0.1187, 0.1073, 0.0920, 0.1414, 0.1635], device='cuda:4'), in_proj_covar=tensor([0.0118, 0.0238, 0.0136, 0.0121, 0.0132, 0.0153, 0.0117, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 04:44:36,739 INFO [finetune.py:976] (4/7) Epoch 29, batch 3650, loss[loss=0.204, simple_loss=0.2765, pruned_loss=0.0658, over 4802.00 frames. ], tot_loss[loss=0.1634, simple_loss=0.236, pruned_loss=0.0454, over 953933.43 frames. ], batch size: 51, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 04:45:10,039 INFO [finetune.py:976] (4/7) Epoch 29, batch 3700, loss[loss=0.1581, simple_loss=0.2342, pruned_loss=0.04101, over 4829.00 frames. ], tot_loss[loss=0.1654, simple_loss=0.2389, pruned_loss=0.04592, over 954608.87 frames. ], batch size: 33, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 04:45:11,858 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.127e+02 1.572e+02 1.919e+02 2.476e+02 4.831e+02, threshold=3.838e+02, percent-clipped=5.0 2023-04-28 04:45:19,687 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=164090.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:45:43,377 INFO [finetune.py:976] (4/7) Epoch 29, batch 3750, loss[loss=0.1766, simple_loss=0.2507, pruned_loss=0.05126, over 4821.00 frames. ], tot_loss[loss=0.1651, simple_loss=0.239, pruned_loss=0.0456, over 955017.53 frames. ], batch size: 39, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 04:46:11,049 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=164151.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:46:18,654 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=164155.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:46:43,098 INFO [finetune.py:976] (4/7) Epoch 29, batch 3800, loss[loss=0.1835, simple_loss=0.2533, pruned_loss=0.05684, over 4786.00 frames. ], tot_loss[loss=0.1676, simple_loss=0.2416, pruned_loss=0.04682, over 952869.19 frames. ], batch size: 25, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 04:46:47,406 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.309e+01 1.497e+02 1.761e+02 2.099e+02 4.096e+02, threshold=3.523e+02, percent-clipped=1.0 2023-04-28 04:47:05,292 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4935, 1.4631, 1.8347, 1.8332, 1.3531, 1.2913, 1.4891, 0.9582], device='cuda:4'), covar=tensor([0.0550, 0.0535, 0.0365, 0.0505, 0.0671, 0.0932, 0.0583, 0.0550], device='cuda:4'), in_proj_covar=tensor([0.0071, 0.0067, 0.0065, 0.0069, 0.0076, 0.0095, 0.0073, 0.0062], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-28 04:47:23,339 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=164216.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:47:34,752 INFO [finetune.py:976] (4/7) Epoch 29, batch 3850, loss[loss=0.1387, simple_loss=0.2109, pruned_loss=0.03326, over 4760.00 frames. ], tot_loss[loss=0.1651, simple_loss=0.2392, pruned_loss=0.04554, over 954007.15 frames. ], batch size: 28, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 04:48:09,467 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6707, 1.6943, 0.9880, 1.3651, 1.9586, 1.5427, 1.4224, 1.5421], device='cuda:4'), covar=tensor([0.0485, 0.0344, 0.0306, 0.0522, 0.0250, 0.0476, 0.0450, 0.0542], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0028, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0052, 0.0046, 0.0039, 0.0053, 0.0039, 0.0051, 0.0051, 0.0053], device='cuda:4') 2023-04-28 04:48:38,394 INFO [finetune.py:976] (4/7) Epoch 29, batch 3900, loss[loss=0.1561, simple_loss=0.2117, pruned_loss=0.0502, over 4110.00 frames. ], tot_loss[loss=0.1637, simple_loss=0.237, pruned_loss=0.0452, over 953418.90 frames. ], batch size: 18, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 04:48:40,203 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.003e+02 1.498e+02 1.753e+02 2.124e+02 6.066e+02, threshold=3.506e+02, percent-clipped=2.0 2023-04-28 04:49:39,856 INFO [finetune.py:976] (4/7) Epoch 29, batch 3950, loss[loss=0.1287, simple_loss=0.2089, pruned_loss=0.02426, over 4812.00 frames. ], tot_loss[loss=0.1614, simple_loss=0.2338, pruned_loss=0.04448, over 951375.03 frames. ], batch size: 25, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 04:50:22,552 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9062, 1.3703, 1.9691, 2.3500, 1.9475, 1.8444, 1.9174, 1.9099], device='cuda:4'), covar=tensor([0.4911, 0.7105, 0.6277, 0.5591, 0.6113, 0.8047, 0.7880, 0.8273], device='cuda:4'), in_proj_covar=tensor([0.0448, 0.0425, 0.0521, 0.0509, 0.0475, 0.0514, 0.0515, 0.0530], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 04:50:24,381 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=164360.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:50:45,028 INFO [finetune.py:976] (4/7) Epoch 29, batch 4000, loss[loss=0.1568, simple_loss=0.2393, pruned_loss=0.03715, over 4930.00 frames. ], tot_loss[loss=0.1615, simple_loss=0.2337, pruned_loss=0.0446, over 952467.75 frames. ], batch size: 38, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 04:50:47,325 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.986e+01 1.466e+02 1.722e+02 2.043e+02 3.565e+02, threshold=3.444e+02, percent-clipped=1.0 2023-04-28 04:51:18,764 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.41 vs. limit=5.0 2023-04-28 04:51:41,797 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=164421.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:51:48,793 INFO [finetune.py:976] (4/7) Epoch 29, batch 4050, loss[loss=0.1749, simple_loss=0.25, pruned_loss=0.04989, over 4861.00 frames. ], tot_loss[loss=0.1635, simple_loss=0.2363, pruned_loss=0.04535, over 953143.53 frames. ], batch size: 31, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 04:52:20,692 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=164446.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:52:54,275 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7079, 2.1407, 1.6829, 1.5429, 1.3038, 1.2755, 1.6735, 1.2047], device='cuda:4'), covar=tensor([0.1540, 0.1211, 0.1362, 0.1634, 0.2341, 0.1971, 0.0974, 0.2089], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0210, 0.0171, 0.0204, 0.0201, 0.0187, 0.0157, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 04:52:55,387 INFO [finetune.py:976] (4/7) Epoch 29, batch 4100, loss[loss=0.1778, simple_loss=0.2444, pruned_loss=0.05559, over 4896.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.2402, pruned_loss=0.04669, over 952119.46 frames. ], batch size: 37, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 04:53:02,652 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.013e+02 1.616e+02 1.841e+02 2.144e+02 4.180e+02, threshold=3.683e+02, percent-clipped=3.0 2023-04-28 04:53:41,168 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=164511.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:53:54,831 INFO [finetune.py:976] (4/7) Epoch 29, batch 4150, loss[loss=0.1576, simple_loss=0.23, pruned_loss=0.04254, over 4787.00 frames. ], tot_loss[loss=0.1669, simple_loss=0.2405, pruned_loss=0.04668, over 950239.92 frames. ], batch size: 29, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 04:54:35,909 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=164568.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:54:40,036 INFO [finetune.py:976] (4/7) Epoch 29, batch 4200, loss[loss=0.1641, simple_loss=0.2361, pruned_loss=0.04608, over 4838.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.2416, pruned_loss=0.04695, over 949574.73 frames. ], batch size: 49, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 04:54:41,956 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=164578.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:54:42,444 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.830e+01 1.587e+02 1.797e+02 2.330e+02 9.173e+02, threshold=3.593e+02, percent-clipped=2.0 2023-04-28 04:54:43,836 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7278, 0.9921, 1.6696, 2.2328, 1.8047, 1.6223, 1.6655, 1.6580], device='cuda:4'), covar=tensor([0.4244, 0.6742, 0.5928, 0.5062, 0.5490, 0.7236, 0.7392, 0.8660], device='cuda:4'), in_proj_covar=tensor([0.0448, 0.0426, 0.0520, 0.0510, 0.0476, 0.0514, 0.0516, 0.0530], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 04:55:08,339 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6492, 1.7534, 1.5227, 1.1443, 1.2900, 1.2323, 1.4967, 1.2091], device='cuda:4'), covar=tensor([0.1820, 0.1289, 0.1521, 0.1766, 0.2444, 0.2116, 0.1071, 0.2155], device='cuda:4'), in_proj_covar=tensor([0.0198, 0.0210, 0.0171, 0.0204, 0.0201, 0.0187, 0.0157, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 04:55:14,304 INFO [finetune.py:976] (4/7) Epoch 29, batch 4250, loss[loss=0.1653, simple_loss=0.2316, pruned_loss=0.04945, over 4932.00 frames. ], tot_loss[loss=0.1667, simple_loss=0.2403, pruned_loss=0.04653, over 951454.16 frames. ], batch size: 33, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 04:55:15,061 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7722, 1.7126, 1.9621, 2.1670, 2.2578, 1.8240, 1.5429, 1.9236], device='cuda:4'), covar=tensor([0.0824, 0.1093, 0.0690, 0.0586, 0.0595, 0.0784, 0.0693, 0.0590], device='cuda:4'), in_proj_covar=tensor([0.0183, 0.0200, 0.0183, 0.0170, 0.0176, 0.0175, 0.0150, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 04:55:16,912 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=164629.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:55:24,095 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=164639.0, num_to_drop=1, layers_to_drop={3} 2023-04-28 04:55:35,803 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6952, 1.6701, 0.9898, 1.3648, 1.9102, 1.5747, 1.3830, 1.5313], device='cuda:4'), covar=tensor([0.0490, 0.0375, 0.0306, 0.0548, 0.0246, 0.0498, 0.0476, 0.0561], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0022, 0.0028, 0.0019, 0.0028, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0052, 0.0047, 0.0039, 0.0054, 0.0039, 0.0051, 0.0051, 0.0053], device='cuda:4') 2023-04-28 04:55:37,640 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=164658.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:55:48,427 INFO [finetune.py:976] (4/7) Epoch 29, batch 4300, loss[loss=0.1701, simple_loss=0.2511, pruned_loss=0.04455, over 4851.00 frames. ], tot_loss[loss=0.1647, simple_loss=0.2376, pruned_loss=0.04587, over 955060.09 frames. ], batch size: 44, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 04:55:50,845 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.176e+01 1.533e+02 1.704e+02 2.198e+02 4.636e+02, threshold=3.409e+02, percent-clipped=3.0 2023-04-28 04:55:57,662 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.58 vs. limit=2.0 2023-04-28 04:56:09,248 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2829, 1.7830, 2.1240, 2.2686, 2.1368, 1.7322, 1.2003, 1.8387], device='cuda:4'), covar=tensor([0.3352, 0.3094, 0.1775, 0.2230, 0.2734, 0.2721, 0.4124, 0.2042], device='cuda:4'), in_proj_covar=tensor([0.0297, 0.0249, 0.0231, 0.0318, 0.0225, 0.0239, 0.0232, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-28 04:56:16,862 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=164716.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:56:18,773 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=164719.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:56:22,290 INFO [finetune.py:976] (4/7) Epoch 29, batch 4350, loss[loss=0.188, simple_loss=0.2434, pruned_loss=0.06624, over 4798.00 frames. ], tot_loss[loss=0.1624, simple_loss=0.2346, pruned_loss=0.04512, over 953429.18 frames. ], batch size: 45, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 04:56:33,998 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7898, 1.3254, 1.9056, 2.3044, 1.9064, 1.8004, 1.8236, 1.7934], device='cuda:4'), covar=tensor([0.4226, 0.6444, 0.6020, 0.5301, 0.5386, 0.7227, 0.7546, 0.8717], device='cuda:4'), in_proj_covar=tensor([0.0448, 0.0425, 0.0522, 0.0511, 0.0476, 0.0514, 0.0517, 0.0530], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 04:56:36,302 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=164746.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:56:36,317 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=164746.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:57:06,725 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.8725, 4.7547, 3.2835, 5.5562, 4.9007, 4.7966, 1.9805, 4.8196], device='cuda:4'), covar=tensor([0.1403, 0.0944, 0.2717, 0.0926, 0.4223, 0.1592, 0.6100, 0.1882], device='cuda:4'), in_proj_covar=tensor([0.0247, 0.0221, 0.0254, 0.0306, 0.0303, 0.0252, 0.0276, 0.0278], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 04:57:17,722 INFO [finetune.py:976] (4/7) Epoch 29, batch 4400, loss[loss=0.2028, simple_loss=0.277, pruned_loss=0.06427, over 4751.00 frames. ], tot_loss[loss=0.1618, simple_loss=0.2341, pruned_loss=0.04479, over 949834.04 frames. ], batch size: 54, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 04:57:20,180 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.037e+02 1.542e+02 1.919e+02 2.288e+02 3.219e+02, threshold=3.837e+02, percent-clipped=0.0 2023-04-28 04:57:41,276 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=164794.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:58:01,574 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=164807.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:58:04,446 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=164811.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:58:11,125 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1059, 2.7066, 2.2944, 2.5428, 1.8913, 2.3166, 2.5108, 1.8285], device='cuda:4'), covar=tensor([0.2182, 0.1358, 0.0907, 0.1316, 0.3610, 0.1224, 0.1823, 0.2748], device='cuda:4'), in_proj_covar=tensor([0.0282, 0.0301, 0.0217, 0.0276, 0.0313, 0.0254, 0.0247, 0.0263], device='cuda:4'), out_proj_covar=tensor([1.1239e-04, 1.1823e-04, 8.4924e-05, 1.0841e-04, 1.2617e-04, 9.9546e-05, 9.9507e-05, 1.0368e-04], device='cuda:4') 2023-04-28 04:58:24,166 INFO [finetune.py:976] (4/7) Epoch 29, batch 4450, loss[loss=0.1317, simple_loss=0.2071, pruned_loss=0.02818, over 4748.00 frames. ], tot_loss[loss=0.1626, simple_loss=0.2361, pruned_loss=0.04457, over 950968.06 frames. ], batch size: 23, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 04:58:55,984 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4061, 1.2610, 4.1766, 3.9199, 3.6824, 4.0238, 3.9994, 3.6682], device='cuda:4'), covar=tensor([0.8040, 0.6382, 0.1315, 0.2041, 0.1289, 0.1662, 0.1394, 0.1643], device='cuda:4'), in_proj_covar=tensor([0.0311, 0.0311, 0.0409, 0.0412, 0.0352, 0.0416, 0.0320, 0.0365], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-28 04:58:59,954 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=164859.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 04:59:04,746 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5523, 3.1859, 0.9151, 1.6027, 2.2673, 1.5179, 4.2357, 2.0701], device='cuda:4'), covar=tensor([0.0613, 0.0640, 0.0904, 0.1260, 0.0544, 0.1009, 0.0196, 0.0590], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0072, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-28 04:59:10,503 INFO [finetune.py:976] (4/7) Epoch 29, batch 4500, loss[loss=0.1601, simple_loss=0.2424, pruned_loss=0.03888, over 4825.00 frames. ], tot_loss[loss=0.164, simple_loss=0.2384, pruned_loss=0.0448, over 952936.76 frames. ], batch size: 33, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 04:59:18,301 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.040e+02 1.480e+02 1.804e+02 2.196e+02 5.471e+02, threshold=3.609e+02, percent-clipped=1.0 2023-04-28 04:59:27,506 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.66 vs. limit=2.0 2023-04-28 05:00:20,418 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=164924.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:00:20,956 INFO [finetune.py:976] (4/7) Epoch 29, batch 4550, loss[loss=0.1674, simple_loss=0.244, pruned_loss=0.04541, over 4883.00 frames. ], tot_loss[loss=0.1654, simple_loss=0.2401, pruned_loss=0.04537, over 954820.51 frames. ], batch size: 32, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 05:00:23,730 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.46 vs. limit=2.0 2023-04-28 05:00:32,586 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=164934.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 05:01:17,425 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1880, 2.5721, 0.7666, 1.5064, 1.5597, 1.8373, 1.6178, 0.8264], device='cuda:4'), covar=tensor([0.1407, 0.1017, 0.1789, 0.1242, 0.1047, 0.0948, 0.1498, 0.1629], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0235, 0.0135, 0.0119, 0.0130, 0.0151, 0.0116, 0.0117], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 05:01:26,332 INFO [finetune.py:976] (4/7) Epoch 29, batch 4600, loss[loss=0.1372, simple_loss=0.2072, pruned_loss=0.03356, over 4795.00 frames. ], tot_loss[loss=0.1642, simple_loss=0.239, pruned_loss=0.04472, over 955235.15 frames. ], batch size: 29, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 05:01:29,218 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.013e+02 1.461e+02 1.731e+02 1.968e+02 2.942e+02, threshold=3.463e+02, percent-clipped=1.0 2023-04-28 05:02:09,393 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.76 vs. limit=5.0 2023-04-28 05:02:19,694 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=165014.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:02:20,899 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=165016.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:02:32,276 INFO [finetune.py:976] (4/7) Epoch 29, batch 4650, loss[loss=0.1277, simple_loss=0.2041, pruned_loss=0.02568, over 4690.00 frames. ], tot_loss[loss=0.1623, simple_loss=0.2363, pruned_loss=0.04421, over 953551.93 frames. ], batch size: 23, lr: 2.83e-03, grad_scale: 32.0 2023-04-28 05:03:18,781 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=165064.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:03:36,022 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1740, 2.7980, 2.2074, 2.2079, 1.5935, 1.5550, 2.3771, 1.6400], device='cuda:4'), covar=tensor([0.1635, 0.1375, 0.1364, 0.1593, 0.2149, 0.1907, 0.0879, 0.2026], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0210, 0.0171, 0.0204, 0.0202, 0.0188, 0.0157, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 05:03:37,007 INFO [finetune.py:976] (4/7) Epoch 29, batch 4700, loss[loss=0.1415, simple_loss=0.2084, pruned_loss=0.03725, over 4819.00 frames. ], tot_loss[loss=0.1604, simple_loss=0.2337, pruned_loss=0.04357, over 955507.98 frames. ], batch size: 38, lr: 2.83e-03, grad_scale: 16.0 2023-04-28 05:03:40,046 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.099e+02 1.538e+02 1.838e+02 2.257e+02 4.770e+02, threshold=3.676e+02, percent-clipped=2.0 2023-04-28 05:03:58,088 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1988, 3.1485, 0.9313, 1.6698, 1.6244, 2.3259, 1.8133, 1.0304], device='cuda:4'), covar=tensor([0.1784, 0.1404, 0.2214, 0.1599, 0.1361, 0.1261, 0.1685, 0.2084], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0236, 0.0135, 0.0120, 0.0131, 0.0152, 0.0116, 0.0117], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 05:04:09,168 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=165102.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:04:09,393 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.20 vs. limit=2.0 2023-04-28 05:04:18,496 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-28 05:04:41,336 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=165123.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:04:42,454 INFO [finetune.py:976] (4/7) Epoch 29, batch 4750, loss[loss=0.189, simple_loss=0.2554, pruned_loss=0.06132, over 4896.00 frames. ], tot_loss[loss=0.1597, simple_loss=0.2324, pruned_loss=0.04353, over 952971.98 frames. ], batch size: 32, lr: 2.83e-03, grad_scale: 16.0 2023-04-28 05:04:43,244 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4401, 1.6639, 1.9364, 2.0222, 1.9196, 1.9177, 1.9479, 1.9305], device='cuda:4'), covar=tensor([0.3848, 0.5396, 0.4281, 0.4071, 0.5196, 0.7067, 0.5046, 0.4824], device='cuda:4'), in_proj_covar=tensor([0.0344, 0.0374, 0.0331, 0.0343, 0.0352, 0.0394, 0.0363, 0.0334], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 05:05:25,306 INFO [finetune.py:976] (4/7) Epoch 29, batch 4800, loss[loss=0.1784, simple_loss=0.2647, pruned_loss=0.04603, over 4832.00 frames. ], tot_loss[loss=0.1639, simple_loss=0.2366, pruned_loss=0.0456, over 952008.94 frames. ], batch size: 47, lr: 2.82e-03, grad_scale: 16.0 2023-04-28 05:05:28,761 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.123e+02 1.552e+02 1.836e+02 2.145e+02 4.672e+02, threshold=3.672e+02, percent-clipped=1.0 2023-04-28 05:05:31,322 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=165184.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:05:48,808 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9854, 2.4609, 1.0292, 1.3168, 1.6988, 1.2132, 3.0710, 1.6111], device='cuda:4'), covar=tensor([0.0660, 0.0521, 0.0705, 0.1212, 0.0488, 0.0993, 0.0253, 0.0585], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0073, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-28 05:05:57,699 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=165224.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:05:58,222 INFO [finetune.py:976] (4/7) Epoch 29, batch 4850, loss[loss=0.1467, simple_loss=0.2313, pruned_loss=0.03109, over 4823.00 frames. ], tot_loss[loss=0.1643, simple_loss=0.2387, pruned_loss=0.045, over 952979.42 frames. ], batch size: 39, lr: 2.82e-03, grad_scale: 16.0 2023-04-28 05:06:05,345 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=165234.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 05:06:15,003 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7438, 1.3530, 1.3381, 1.5363, 1.9069, 1.5416, 1.3555, 1.3034], device='cuda:4'), covar=tensor([0.1559, 0.1456, 0.1818, 0.1332, 0.0712, 0.1439, 0.1661, 0.2183], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0307, 0.0348, 0.0286, 0.0324, 0.0303, 0.0299, 0.0375], device='cuda:4'), out_proj_covar=tensor([6.3530e-05, 6.2858e-05, 7.2719e-05, 5.7166e-05, 6.6027e-05, 6.2844e-05, 6.1752e-05, 7.9230e-05], device='cuda:4') 2023-04-28 05:06:29,992 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=165272.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:06:31,765 INFO [finetune.py:976] (4/7) Epoch 29, batch 4900, loss[loss=0.1708, simple_loss=0.251, pruned_loss=0.04528, over 4826.00 frames. ], tot_loss[loss=0.1642, simple_loss=0.239, pruned_loss=0.0447, over 952057.18 frames. ], batch size: 33, lr: 2.82e-03, grad_scale: 16.0 2023-04-28 05:06:35,317 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.134e+02 1.561e+02 1.845e+02 2.104e+02 4.657e+02, threshold=3.691e+02, percent-clipped=1.0 2023-04-28 05:06:37,062 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=165282.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:06:57,499 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=165314.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:07:04,963 INFO [finetune.py:976] (4/7) Epoch 29, batch 4950, loss[loss=0.1902, simple_loss=0.2596, pruned_loss=0.06041, over 4914.00 frames. ], tot_loss[loss=0.1651, simple_loss=0.2397, pruned_loss=0.04524, over 952225.79 frames. ], batch size: 42, lr: 2.82e-03, grad_scale: 16.0 2023-04-28 05:07:29,755 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=165362.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:07:38,074 INFO [finetune.py:976] (4/7) Epoch 29, batch 5000, loss[loss=0.1952, simple_loss=0.2665, pruned_loss=0.06197, over 4918.00 frames. ], tot_loss[loss=0.1643, simple_loss=0.2385, pruned_loss=0.04501, over 953338.54 frames. ], batch size: 33, lr: 2.82e-03, grad_scale: 16.0 2023-04-28 05:07:41,608 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.839e+01 1.556e+02 1.839e+02 2.298e+02 4.035e+02, threshold=3.677e+02, percent-clipped=3.0 2023-04-28 05:07:57,197 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=165402.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:08:08,233 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.64 vs. limit=2.0 2023-04-28 05:08:12,132 INFO [finetune.py:976] (4/7) Epoch 29, batch 5050, loss[loss=0.1728, simple_loss=0.2408, pruned_loss=0.0524, over 4786.00 frames. ], tot_loss[loss=0.1637, simple_loss=0.2372, pruned_loss=0.04513, over 954186.83 frames. ], batch size: 25, lr: 2.82e-03, grad_scale: 16.0 2023-04-28 05:08:29,765 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=165450.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:08:29,791 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2740, 1.2532, 3.8450, 3.6068, 3.4208, 3.6874, 3.6945, 3.4017], device='cuda:4'), covar=tensor([0.7362, 0.5653, 0.1224, 0.1953, 0.1146, 0.2016, 0.1619, 0.1596], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0312, 0.0410, 0.0414, 0.0354, 0.0417, 0.0320, 0.0366], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-28 05:08:40,115 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9824, 2.6185, 2.0089, 2.1998, 1.4472, 1.5023, 2.1453, 1.4083], device='cuda:4'), covar=tensor([0.1594, 0.1248, 0.1324, 0.1432, 0.2189, 0.1788, 0.0936, 0.2007], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0210, 0.0171, 0.0204, 0.0202, 0.0187, 0.0157, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 05:08:45,057 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1035, 2.5609, 2.1485, 2.5021, 1.8300, 2.2076, 2.1751, 1.6593], device='cuda:4'), covar=tensor([0.1680, 0.1039, 0.0791, 0.1103, 0.3197, 0.1049, 0.1647, 0.2453], device='cuda:4'), in_proj_covar=tensor([0.0280, 0.0299, 0.0215, 0.0274, 0.0311, 0.0252, 0.0245, 0.0261], device='cuda:4'), out_proj_covar=tensor([1.1156e-04, 1.1750e-04, 8.4254e-05, 1.0739e-04, 1.2507e-04, 9.8960e-05, 9.8693e-05, 1.0268e-04], device='cuda:4') 2023-04-28 05:08:45,561 INFO [finetune.py:976] (4/7) Epoch 29, batch 5100, loss[loss=0.1832, simple_loss=0.2429, pruned_loss=0.06172, over 4895.00 frames. ], tot_loss[loss=0.1624, simple_loss=0.2349, pruned_loss=0.0449, over 954435.34 frames. ], batch size: 35, lr: 2.82e-03, grad_scale: 16.0 2023-04-28 05:08:48,563 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=165479.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:08:49,074 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.177e+01 1.550e+02 1.878e+02 2.337e+02 3.681e+02, threshold=3.756e+02, percent-clipped=1.0 2023-04-28 05:09:36,575 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9371, 2.3629, 2.0381, 2.2945, 1.7098, 2.0760, 1.9750, 1.5755], device='cuda:4'), covar=tensor([0.1670, 0.1217, 0.0787, 0.1084, 0.3537, 0.1042, 0.1688, 0.2603], device='cuda:4'), in_proj_covar=tensor([0.0281, 0.0300, 0.0215, 0.0274, 0.0311, 0.0253, 0.0246, 0.0262], device='cuda:4'), out_proj_covar=tensor([1.1189e-04, 1.1777e-04, 8.4529e-05, 1.0777e-04, 1.2522e-04, 9.9178e-05, 9.9053e-05, 1.0299e-04], device='cuda:4') 2023-04-28 05:09:48,067 INFO [finetune.py:976] (4/7) Epoch 29, batch 5150, loss[loss=0.1586, simple_loss=0.229, pruned_loss=0.04407, over 4892.00 frames. ], tot_loss[loss=0.1629, simple_loss=0.2351, pruned_loss=0.04533, over 951935.15 frames. ], batch size: 32, lr: 2.82e-03, grad_scale: 16.0 2023-04-28 05:10:21,719 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=165551.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:10:52,972 INFO [finetune.py:976] (4/7) Epoch 29, batch 5200, loss[loss=0.1966, simple_loss=0.2673, pruned_loss=0.06293, over 4901.00 frames. ], tot_loss[loss=0.1663, simple_loss=0.2391, pruned_loss=0.04679, over 950746.06 frames. ], batch size: 43, lr: 2.82e-03, grad_scale: 16.0 2023-04-28 05:10:55,996 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.013e+02 1.620e+02 1.941e+02 2.269e+02 3.767e+02, threshold=3.883e+02, percent-clipped=2.0 2023-04-28 05:11:45,802 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=165612.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:11:59,705 INFO [finetune.py:976] (4/7) Epoch 29, batch 5250, loss[loss=0.1997, simple_loss=0.2851, pruned_loss=0.05718, over 4803.00 frames. ], tot_loss[loss=0.1667, simple_loss=0.2401, pruned_loss=0.04658, over 951025.42 frames. ], batch size: 45, lr: 2.82e-03, grad_scale: 16.0 2023-04-28 05:12:12,047 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2322, 1.3894, 3.7964, 3.5684, 3.3811, 3.6591, 3.6585, 3.4084], device='cuda:4'), covar=tensor([0.7570, 0.5614, 0.1265, 0.1978, 0.1277, 0.1884, 0.1906, 0.1496], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0309, 0.0408, 0.0412, 0.0351, 0.0414, 0.0318, 0.0363], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-28 05:13:07,360 INFO [finetune.py:976] (4/7) Epoch 29, batch 5300, loss[loss=0.2296, simple_loss=0.3, pruned_loss=0.07959, over 4839.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.2414, pruned_loss=0.04676, over 952146.35 frames. ], batch size: 47, lr: 2.82e-03, grad_scale: 16.0 2023-04-28 05:13:16,015 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.838e+01 1.557e+02 1.826e+02 2.262e+02 6.421e+02, threshold=3.651e+02, percent-clipped=2.0 2023-04-28 05:13:47,156 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5572, 0.6921, 1.4902, 1.9086, 1.6224, 1.4981, 1.5398, 1.5437], device='cuda:4'), covar=tensor([0.4565, 0.6848, 0.6031, 0.6196, 0.5567, 0.7699, 0.7524, 0.8747], device='cuda:4'), in_proj_covar=tensor([0.0447, 0.0425, 0.0520, 0.0509, 0.0474, 0.0514, 0.0515, 0.0530], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 05:14:12,920 INFO [finetune.py:976] (4/7) Epoch 29, batch 5350, loss[loss=0.2086, simple_loss=0.2771, pruned_loss=0.06998, over 4883.00 frames. ], tot_loss[loss=0.168, simple_loss=0.2422, pruned_loss=0.04686, over 952468.75 frames. ], batch size: 35, lr: 2.82e-03, grad_scale: 16.0 2023-04-28 05:14:43,452 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=165750.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:15:21,306 INFO [finetune.py:976] (4/7) Epoch 29, batch 5400, loss[loss=0.1706, simple_loss=0.2351, pruned_loss=0.05304, over 4815.00 frames. ], tot_loss[loss=0.167, simple_loss=0.2407, pruned_loss=0.04661, over 952004.07 frames. ], batch size: 30, lr: 2.82e-03, grad_scale: 16.0 2023-04-28 05:15:24,331 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=165779.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:15:24,845 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.026e+02 1.500e+02 1.780e+02 2.167e+02 4.402e+02, threshold=3.560e+02, percent-clipped=1.0 2023-04-28 05:15:45,779 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0026, 2.6092, 0.9927, 1.3888, 1.8751, 1.2512, 3.3786, 1.7655], device='cuda:4'), covar=tensor([0.0707, 0.0548, 0.0838, 0.1196, 0.0518, 0.1004, 0.0247, 0.0600], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0073, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-28 05:15:55,112 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=165803.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:16:07,218 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=165811.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:16:21,414 INFO [finetune.py:976] (4/7) Epoch 29, batch 5450, loss[loss=0.145, simple_loss=0.2167, pruned_loss=0.03664, over 4840.00 frames. ], tot_loss[loss=0.1644, simple_loss=0.2375, pruned_loss=0.04565, over 950489.89 frames. ], batch size: 47, lr: 2.82e-03, grad_scale: 16.0 2023-04-28 05:16:22,708 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=165827.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:16:35,147 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.64 vs. limit=5.0 2023-04-28 05:16:48,206 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=165864.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:16:55,273 INFO [finetune.py:976] (4/7) Epoch 29, batch 5500, loss[loss=0.1363, simple_loss=0.2106, pruned_loss=0.03094, over 4858.00 frames. ], tot_loss[loss=0.1613, simple_loss=0.2341, pruned_loss=0.04421, over 951085.39 frames. ], batch size: 31, lr: 2.82e-03, grad_scale: 16.0 2023-04-28 05:16:58,256 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.287e+01 1.513e+02 1.807e+02 2.218e+02 5.669e+02, threshold=3.614e+02, percent-clipped=2.0 2023-04-28 05:17:16,283 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=165907.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:17:28,021 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0140, 1.0227, 1.2263, 1.1451, 0.9951, 0.9362, 1.0074, 0.5465], device='cuda:4'), covar=tensor([0.0492, 0.0509, 0.0410, 0.0533, 0.0630, 0.1006, 0.0431, 0.0584], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0066, 0.0064, 0.0068, 0.0075, 0.0094, 0.0071, 0.0061], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-28 05:17:29,075 INFO [finetune.py:976] (4/7) Epoch 29, batch 5550, loss[loss=0.1915, simple_loss=0.2782, pruned_loss=0.0524, over 4833.00 frames. ], tot_loss[loss=0.1635, simple_loss=0.2368, pruned_loss=0.04509, over 951562.57 frames. ], batch size: 47, lr: 2.82e-03, grad_scale: 16.0 2023-04-28 05:18:17,442 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7507, 2.3921, 2.0040, 1.8561, 1.2907, 1.3926, 2.1728, 1.3183], device='cuda:4'), covar=tensor([0.1716, 0.1357, 0.1265, 0.1505, 0.2196, 0.1882, 0.0815, 0.1992], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0211, 0.0171, 0.0205, 0.0201, 0.0187, 0.0157, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 05:18:28,707 INFO [finetune.py:976] (4/7) Epoch 29, batch 5600, loss[loss=0.205, simple_loss=0.2712, pruned_loss=0.0694, over 4871.00 frames. ], tot_loss[loss=0.1658, simple_loss=0.2404, pruned_loss=0.04559, over 954446.82 frames. ], batch size: 34, lr: 2.82e-03, grad_scale: 16.0 2023-04-28 05:18:35,764 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.67 vs. limit=2.0 2023-04-28 05:18:37,366 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.927e+01 1.529e+02 1.829e+02 2.121e+02 4.998e+02, threshold=3.659e+02, percent-clipped=4.0 2023-04-28 05:18:57,609 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.84 vs. limit=2.0 2023-04-28 05:19:29,266 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4796, 3.4746, 1.0399, 1.9285, 1.8646, 2.4354, 1.9287, 1.0381], device='cuda:4'), covar=tensor([0.1526, 0.1140, 0.2054, 0.1245, 0.1141, 0.1091, 0.1565, 0.2052], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0237, 0.0136, 0.0120, 0.0131, 0.0152, 0.0117, 0.0118], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 05:19:32,709 INFO [finetune.py:976] (4/7) Epoch 29, batch 5650, loss[loss=0.1384, simple_loss=0.2053, pruned_loss=0.03579, over 4710.00 frames. ], tot_loss[loss=0.1662, simple_loss=0.2418, pruned_loss=0.04531, over 954084.34 frames. ], batch size: 23, lr: 2.82e-03, grad_scale: 16.0 2023-04-28 05:20:11,616 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=166054.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 05:20:13,377 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9928, 1.7826, 4.3888, 4.1577, 3.9418, 4.2639, 4.2337, 3.9418], device='cuda:4'), covar=tensor([0.5980, 0.5072, 0.1161, 0.1718, 0.1094, 0.1998, 0.1254, 0.1478], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0309, 0.0407, 0.0410, 0.0351, 0.0414, 0.0317, 0.0364], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-28 05:20:33,899 INFO [finetune.py:976] (4/7) Epoch 29, batch 5700, loss[loss=0.1305, simple_loss=0.2039, pruned_loss=0.02853, over 4296.00 frames. ], tot_loss[loss=0.1634, simple_loss=0.2381, pruned_loss=0.0444, over 936018.52 frames. ], batch size: 18, lr: 2.82e-03, grad_scale: 16.0 2023-04-28 05:20:42,336 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 6.981e+01 1.433e+02 1.713e+02 2.148e+02 4.581e+02, threshold=3.425e+02, percent-clipped=1.0 2023-04-28 05:20:53,302 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.7469, 2.2781, 2.2785, 2.3772, 2.3164, 2.3364, 2.3692, 2.3150], device='cuda:4'), covar=tensor([0.3501, 0.4462, 0.4251, 0.4514, 0.4849, 0.6154, 0.4779, 0.4638], device='cuda:4'), in_proj_covar=tensor([0.0347, 0.0377, 0.0333, 0.0345, 0.0353, 0.0396, 0.0365, 0.0336], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 05:21:19,169 INFO [finetune.py:976] (4/7) Epoch 30, batch 0, loss[loss=0.1424, simple_loss=0.2222, pruned_loss=0.03132, over 4782.00 frames. ], tot_loss[loss=0.1424, simple_loss=0.2222, pruned_loss=0.03132, over 4782.00 frames. ], batch size: 25, lr: 2.82e-03, grad_scale: 16.0 2023-04-28 05:21:19,169 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-28 05:21:20,974 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7995, 1.0661, 1.7623, 2.3056, 1.8833, 1.7105, 1.7220, 1.7017], device='cuda:4'), covar=tensor([0.4632, 0.7579, 0.6242, 0.5818, 0.6161, 0.8262, 0.7877, 1.0248], device='cuda:4'), in_proj_covar=tensor([0.0448, 0.0427, 0.0523, 0.0511, 0.0477, 0.0517, 0.0519, 0.0533], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 05:21:26,312 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2935, 1.5581, 1.8352, 1.9788, 1.9310, 1.9701, 1.8216, 1.8633], device='cuda:4'), covar=tensor([0.3721, 0.5414, 0.4487, 0.4351, 0.5518, 0.6681, 0.5306, 0.4866], device='cuda:4'), in_proj_covar=tensor([0.0347, 0.0377, 0.0333, 0.0346, 0.0354, 0.0396, 0.0365, 0.0336], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 05:21:34,681 INFO [finetune.py:1010] (4/7) Epoch 30, validation: loss=0.1551, simple_loss=0.2236, pruned_loss=0.04334, over 2265189.00 frames. 2023-04-28 05:21:34,682 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6529MB 2023-04-28 05:21:37,613 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=166106.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:21:48,580 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=166115.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 05:22:23,783 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.75 vs. limit=2.0 2023-04-28 05:22:31,248 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=166147.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:22:35,228 INFO [finetune.py:976] (4/7) Epoch 30, batch 50, loss[loss=0.1686, simple_loss=0.2478, pruned_loss=0.04469, over 4781.00 frames. ], tot_loss[loss=0.1624, simple_loss=0.2378, pruned_loss=0.04348, over 214384.16 frames. ], batch size: 29, lr: 2.81e-03, grad_scale: 16.0 2023-04-28 05:22:42,114 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.86 vs. limit=2.0 2023-04-28 05:22:42,594 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.63 vs. limit=2.0 2023-04-28 05:22:51,880 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=166159.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:23:14,383 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 7.993e+01 1.354e+02 1.707e+02 2.159e+02 3.038e+02, threshold=3.414e+02, percent-clipped=0.0 2023-04-28 05:23:44,643 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0554, 1.7374, 1.9564, 2.2742, 2.3189, 1.9708, 1.4921, 2.0334], device='cuda:4'), covar=tensor([0.0733, 0.1154, 0.0662, 0.0551, 0.0569, 0.0865, 0.0749, 0.0568], device='cuda:4'), in_proj_covar=tensor([0.0183, 0.0202, 0.0183, 0.0170, 0.0177, 0.0177, 0.0149, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 05:23:47,367 INFO [finetune.py:976] (4/7) Epoch 30, batch 100, loss[loss=0.154, simple_loss=0.2157, pruned_loss=0.0461, over 4221.00 frames. ], tot_loss[loss=0.1588, simple_loss=0.2318, pruned_loss=0.04292, over 377699.18 frames. ], batch size: 65, lr: 2.81e-03, grad_scale: 16.0 2023-04-28 05:23:50,920 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=166207.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:23:56,825 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=166208.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:24:41,850 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.11 vs. limit=2.0 2023-04-28 05:24:53,824 INFO [finetune.py:976] (4/7) Epoch 30, batch 150, loss[loss=0.145, simple_loss=0.2234, pruned_loss=0.03328, over 4816.00 frames. ], tot_loss[loss=0.1563, simple_loss=0.2287, pruned_loss=0.04199, over 507968.04 frames. ], batch size: 33, lr: 2.81e-03, grad_scale: 16.0 2023-04-28 05:24:55,629 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=166255.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:25:28,830 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.799e+01 1.515e+02 1.784e+02 2.130e+02 3.221e+02, threshold=3.568e+02, percent-clipped=0.0 2023-04-28 05:25:59,861 INFO [finetune.py:976] (4/7) Epoch 30, batch 200, loss[loss=0.1608, simple_loss=0.2372, pruned_loss=0.04224, over 4767.00 frames. ], tot_loss[loss=0.1575, simple_loss=0.2293, pruned_loss=0.04284, over 609712.81 frames. ], batch size: 26, lr: 2.81e-03, grad_scale: 16.0 2023-04-28 05:26:02,452 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=166307.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:26:40,557 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.49 vs. limit=2.0 2023-04-28 05:27:02,969 INFO [finetune.py:976] (4/7) Epoch 30, batch 250, loss[loss=0.1844, simple_loss=0.2587, pruned_loss=0.05507, over 4825.00 frames. ], tot_loss[loss=0.1613, simple_loss=0.2342, pruned_loss=0.04422, over 686191.54 frames. ], batch size: 40, lr: 2.81e-03, grad_scale: 16.0 2023-04-28 05:27:19,836 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=166368.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:27:38,575 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.540e+01 1.580e+02 1.903e+02 2.256e+02 3.495e+02, threshold=3.807e+02, percent-clipped=0.0 2023-04-28 05:28:02,277 INFO [finetune.py:976] (4/7) Epoch 30, batch 300, loss[loss=0.1675, simple_loss=0.2457, pruned_loss=0.04466, over 4899.00 frames. ], tot_loss[loss=0.1624, simple_loss=0.236, pruned_loss=0.04435, over 745242.77 frames. ], batch size: 32, lr: 2.81e-03, grad_scale: 16.0 2023-04-28 05:28:07,517 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=166406.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:28:09,961 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=166410.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 05:29:01,972 INFO [finetune.py:976] (4/7) Epoch 30, batch 350, loss[loss=0.186, simple_loss=0.2627, pruned_loss=0.05463, over 4907.00 frames. ], tot_loss[loss=0.1648, simple_loss=0.2389, pruned_loss=0.04532, over 793628.15 frames. ], batch size: 37, lr: 2.81e-03, grad_scale: 16.0 2023-04-28 05:29:08,236 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=166454.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:29:11,871 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=166459.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:29:31,636 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.632e+01 1.563e+02 1.883e+02 2.303e+02 4.800e+02, threshold=3.766e+02, percent-clipped=2.0 2023-04-28 05:29:32,397 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=166481.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:29:46,147 INFO [finetune.py:976] (4/7) Epoch 30, batch 400, loss[loss=0.1551, simple_loss=0.2288, pruned_loss=0.04064, over 4783.00 frames. ], tot_loss[loss=0.1649, simple_loss=0.2397, pruned_loss=0.04508, over 830423.04 frames. ], batch size: 51, lr: 2.81e-03, grad_scale: 16.0 2023-04-28 05:29:46,222 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=166503.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:29:49,559 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=166507.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:29:50,825 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=166509.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:29:50,903 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.15 vs. limit=2.0 2023-04-28 05:30:06,155 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=166531.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:30:13,246 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=166542.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:30:19,819 INFO [finetune.py:976] (4/7) Epoch 30, batch 450, loss[loss=0.1706, simple_loss=0.2411, pruned_loss=0.05007, over 4927.00 frames. ], tot_loss[loss=0.164, simple_loss=0.2382, pruned_loss=0.04496, over 856521.01 frames. ], batch size: 33, lr: 2.81e-03, grad_scale: 16.0 2023-04-28 05:30:31,312 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=166570.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:30:37,711 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2233, 1.7095, 2.1046, 2.5942, 2.0000, 1.6566, 1.3177, 1.8815], device='cuda:4'), covar=tensor([0.3191, 0.3264, 0.1837, 0.1938, 0.2628, 0.2627, 0.4148, 0.1970], device='cuda:4'), in_proj_covar=tensor([0.0293, 0.0248, 0.0229, 0.0316, 0.0223, 0.0238, 0.0230, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-28 05:30:39,260 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.011e+02 1.427e+02 1.810e+02 2.217e+02 8.020e+02, threshold=3.620e+02, percent-clipped=2.0 2023-04-28 05:30:47,090 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=166592.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:30:53,783 INFO [finetune.py:976] (4/7) Epoch 30, batch 500, loss[loss=0.1946, simple_loss=0.2616, pruned_loss=0.0638, over 4888.00 frames. ], tot_loss[loss=0.1633, simple_loss=0.2372, pruned_loss=0.0447, over 878536.46 frames. ], batch size: 32, lr: 2.81e-03, grad_scale: 16.0 2023-04-28 05:31:07,090 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([0.9616, 1.1334, 3.2892, 3.0656, 2.9064, 3.2110, 3.2172, 2.9067], device='cuda:4'), covar=tensor([0.7643, 0.5406, 0.1621, 0.2334, 0.1487, 0.1669, 0.1574, 0.1919], device='cuda:4'), in_proj_covar=tensor([0.0312, 0.0311, 0.0409, 0.0412, 0.0353, 0.0416, 0.0319, 0.0365], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-28 05:31:27,594 INFO [finetune.py:976] (4/7) Epoch 30, batch 550, loss[loss=0.2125, simple_loss=0.2623, pruned_loss=0.08133, over 4911.00 frames. ], tot_loss[loss=0.1624, simple_loss=0.2353, pruned_loss=0.04476, over 894675.20 frames. ], batch size: 32, lr: 2.81e-03, grad_scale: 16.0 2023-04-28 05:31:34,203 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=166663.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:31:45,972 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.022e+02 1.489e+02 1.745e+02 2.160e+02 4.947e+02, threshold=3.489e+02, percent-clipped=2.0 2023-04-28 05:32:01,359 INFO [finetune.py:976] (4/7) Epoch 30, batch 600, loss[loss=0.1886, simple_loss=0.2455, pruned_loss=0.06584, over 4862.00 frames. ], tot_loss[loss=0.1621, simple_loss=0.2351, pruned_loss=0.04458, over 907678.16 frames. ], batch size: 31, lr: 2.81e-03, grad_scale: 16.0 2023-04-28 05:32:05,746 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=166710.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 05:32:07,163 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.44 vs. limit=2.0 2023-04-28 05:32:23,054 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=166735.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:32:34,882 INFO [finetune.py:976] (4/7) Epoch 30, batch 650, loss[loss=0.1654, simple_loss=0.246, pruned_loss=0.04235, over 4828.00 frames. ], tot_loss[loss=0.1642, simple_loss=0.238, pruned_loss=0.04523, over 918309.45 frames. ], batch size: 39, lr: 2.81e-03, grad_scale: 16.0 2023-04-28 05:32:37,944 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=166758.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 05:32:47,103 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=166764.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:32:55,018 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9312, 1.4411, 2.0042, 2.3880, 2.0338, 1.8743, 1.9181, 1.9058], device='cuda:4'), covar=tensor([0.4402, 0.6712, 0.6237, 0.5406, 0.5770, 0.8082, 0.8113, 0.8137], device='cuda:4'), in_proj_covar=tensor([0.0449, 0.0426, 0.0523, 0.0510, 0.0476, 0.0516, 0.0519, 0.0532], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 05:33:07,156 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.129e+02 1.523e+02 1.858e+02 2.285e+02 4.759e+02, threshold=3.715e+02, percent-clipped=1.0 2023-04-28 05:33:29,723 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=166796.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:33:39,478 INFO [finetune.py:976] (4/7) Epoch 30, batch 700, loss[loss=0.2245, simple_loss=0.2833, pruned_loss=0.08283, over 4816.00 frames. ], tot_loss[loss=0.1652, simple_loss=0.2395, pruned_loss=0.04549, over 926774.55 frames. ], batch size: 33, lr: 2.81e-03, grad_scale: 16.0 2023-04-28 05:33:39,585 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=166803.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:34:03,937 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=166825.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:34:22,683 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=166837.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:34:37,355 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=166851.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:34:43,122 INFO [finetune.py:976] (4/7) Epoch 30, batch 750, loss[loss=0.1944, simple_loss=0.2608, pruned_loss=0.06404, over 4868.00 frames. ], tot_loss[loss=0.1656, simple_loss=0.2404, pruned_loss=0.04544, over 933651.44 frames. ], batch size: 34, lr: 2.81e-03, grad_scale: 16.0 2023-04-28 05:34:51,920 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=166865.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:35:11,850 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.075e+02 1.441e+02 1.685e+02 2.007e+02 5.927e+02, threshold=3.370e+02, percent-clipped=1.0 2023-04-28 05:35:22,366 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=166887.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:35:44,637 INFO [finetune.py:976] (4/7) Epoch 30, batch 800, loss[loss=0.174, simple_loss=0.2568, pruned_loss=0.04566, over 4923.00 frames. ], tot_loss[loss=0.166, simple_loss=0.2409, pruned_loss=0.04552, over 938799.93 frames. ], batch size: 42, lr: 2.81e-03, grad_scale: 16.0 2023-04-28 05:36:17,519 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=166929.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:36:38,368 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0513, 1.9485, 1.8004, 1.7442, 2.1587, 1.6944, 2.7301, 1.6864], device='cuda:4'), covar=tensor([0.3763, 0.2175, 0.4896, 0.3026, 0.1791, 0.2643, 0.1138, 0.4402], device='cuda:4'), in_proj_covar=tensor([0.0342, 0.0356, 0.0427, 0.0353, 0.0388, 0.0377, 0.0373, 0.0427], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 05:36:49,753 INFO [finetune.py:976] (4/7) Epoch 30, batch 850, loss[loss=0.1841, simple_loss=0.2464, pruned_loss=0.06088, over 4827.00 frames. ], tot_loss[loss=0.1629, simple_loss=0.2376, pruned_loss=0.04413, over 943660.06 frames. ], batch size: 38, lr: 2.81e-03, grad_scale: 16.0 2023-04-28 05:36:56,498 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.89 vs. limit=5.0 2023-04-28 05:36:59,948 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=166963.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:37:11,780 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1386, 2.5936, 1.1433, 1.2817, 2.1616, 1.2115, 3.6904, 1.6476], device='cuda:4'), covar=tensor([0.0741, 0.0629, 0.0823, 0.1495, 0.0516, 0.1160, 0.0282, 0.0733], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0073, 0.0051], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-28 05:37:20,968 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.032e+02 1.548e+02 1.839e+02 2.172e+02 6.436e+02, threshold=3.678e+02, percent-clipped=1.0 2023-04-28 05:37:37,955 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=166990.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:37:42,690 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8854, 1.6172, 1.5124, 1.7668, 2.1779, 1.8048, 1.5956, 1.4335], device='cuda:4'), covar=tensor([0.1880, 0.1658, 0.1979, 0.1384, 0.0931, 0.1576, 0.1896, 0.2549], device='cuda:4'), in_proj_covar=tensor([0.0316, 0.0309, 0.0350, 0.0287, 0.0325, 0.0305, 0.0301, 0.0378], device='cuda:4'), out_proj_covar=tensor([6.4217e-05, 6.3167e-05, 7.3301e-05, 5.7265e-05, 6.6175e-05, 6.3244e-05, 6.1999e-05, 7.9834e-05], device='cuda:4') 2023-04-28 05:37:47,448 INFO [finetune.py:976] (4/7) Epoch 30, batch 900, loss[loss=0.159, simple_loss=0.2271, pruned_loss=0.04549, over 4813.00 frames. ], tot_loss[loss=0.161, simple_loss=0.2349, pruned_loss=0.0436, over 945616.66 frames. ], batch size: 41, lr: 2.81e-03, grad_scale: 16.0 2023-04-28 05:37:52,365 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=167011.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:38:21,333 INFO [finetune.py:976] (4/7) Epoch 30, batch 950, loss[loss=0.1738, simple_loss=0.2351, pruned_loss=0.05625, over 4901.00 frames. ], tot_loss[loss=0.1617, simple_loss=0.2345, pruned_loss=0.04443, over 948551.65 frames. ], batch size: 32, lr: 2.81e-03, grad_scale: 32.0 2023-04-28 05:38:31,136 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.14 vs. limit=5.0 2023-04-28 05:38:38,164 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.089e+02 1.627e+02 1.847e+02 2.138e+02 3.460e+02, threshold=3.694e+02, percent-clipped=0.0 2023-04-28 05:38:45,983 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=167091.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:38:52,932 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=167100.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:38:55,141 INFO [finetune.py:976] (4/7) Epoch 30, batch 1000, loss[loss=0.1768, simple_loss=0.2517, pruned_loss=0.05091, over 4812.00 frames. ], tot_loss[loss=0.1638, simple_loss=0.2368, pruned_loss=0.04543, over 948010.77 frames. ], batch size: 38, lr: 2.81e-03, grad_scale: 32.0 2023-04-28 05:38:55,896 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=167104.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:38:55,961 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.19 vs. limit=2.0 2023-04-28 05:39:06,123 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=167120.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:39:17,576 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=167137.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:39:28,662 INFO [finetune.py:976] (4/7) Epoch 30, batch 1050, loss[loss=0.1972, simple_loss=0.2673, pruned_loss=0.06355, over 4896.00 frames. ], tot_loss[loss=0.1664, simple_loss=0.2404, pruned_loss=0.0462, over 950299.50 frames. ], batch size: 43, lr: 2.81e-03, grad_scale: 32.0 2023-04-28 05:39:29,884 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7758, 2.8786, 2.3360, 2.5227, 2.9393, 2.6100, 3.8858, 2.3271], device='cuda:4'), covar=tensor([0.3427, 0.2276, 0.4180, 0.3368, 0.1722, 0.2415, 0.1173, 0.3740], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0355, 0.0425, 0.0353, 0.0386, 0.0375, 0.0371, 0.0425], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 05:39:34,164 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=167161.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:39:36,570 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=167165.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:39:36,615 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=167165.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:39:46,135 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.104e+02 1.567e+02 1.873e+02 2.278e+02 3.738e+02, threshold=3.746e+02, percent-clipped=2.0 2023-04-28 05:39:49,156 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=167185.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:39:50,863 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=167187.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:40:01,671 INFO [finetune.py:976] (4/7) Epoch 30, batch 1100, loss[loss=0.1767, simple_loss=0.2623, pruned_loss=0.04559, over 4867.00 frames. ], tot_loss[loss=0.1664, simple_loss=0.2407, pruned_loss=0.046, over 952543.01 frames. ], batch size: 34, lr: 2.81e-03, grad_scale: 32.0 2023-04-28 05:40:08,808 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=167213.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:40:08,876 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3989, 1.4681, 1.8214, 1.7756, 1.2664, 1.2540, 1.5108, 0.9347], device='cuda:4'), covar=tensor([0.0578, 0.0673, 0.0366, 0.0633, 0.0796, 0.1030, 0.0636, 0.0615], device='cuda:4'), in_proj_covar=tensor([0.0071, 0.0067, 0.0065, 0.0069, 0.0076, 0.0095, 0.0072, 0.0062], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-28 05:40:15,591 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9278, 1.4450, 1.7899, 1.7900, 1.7548, 1.4507, 0.9120, 1.4612], device='cuda:4'), covar=tensor([0.2917, 0.2902, 0.1472, 0.1823, 0.2340, 0.2367, 0.4059, 0.1787], device='cuda:4'), in_proj_covar=tensor([0.0291, 0.0246, 0.0228, 0.0315, 0.0222, 0.0236, 0.0229, 0.0185], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-28 05:40:20,281 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9450, 2.5584, 0.9641, 1.2137, 1.9665, 1.1323, 3.4119, 1.4955], device='cuda:4'), covar=tensor([0.0889, 0.0726, 0.0939, 0.1691, 0.0615, 0.1440, 0.0390, 0.0935], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0073, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-28 05:40:22,664 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=167235.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:40:34,490 INFO [finetune.py:976] (4/7) Epoch 30, batch 1150, loss[loss=0.1663, simple_loss=0.2486, pruned_loss=0.04201, over 4877.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.2415, pruned_loss=0.04598, over 953876.07 frames. ], batch size: 32, lr: 2.81e-03, grad_scale: 32.0 2023-04-28 05:41:03,290 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.007e+02 1.590e+02 1.832e+02 2.168e+02 3.581e+02, threshold=3.663e+02, percent-clipped=0.0 2023-04-28 05:41:11,482 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=167285.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:41:33,521 INFO [finetune.py:976] (4/7) Epoch 30, batch 1200, loss[loss=0.1736, simple_loss=0.2485, pruned_loss=0.04932, over 4841.00 frames. ], tot_loss[loss=0.1658, simple_loss=0.2401, pruned_loss=0.04572, over 951873.12 frames. ], batch size: 44, lr: 2.81e-03, grad_scale: 32.0 2023-04-28 05:41:43,434 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.70 vs. limit=2.0 2023-04-28 05:42:37,510 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.18 vs. limit=2.0 2023-04-28 05:42:37,942 INFO [finetune.py:976] (4/7) Epoch 30, batch 1250, loss[loss=0.169, simple_loss=0.2299, pruned_loss=0.05401, over 4902.00 frames. ], tot_loss[loss=0.1647, simple_loss=0.2384, pruned_loss=0.04553, over 952731.95 frames. ], batch size: 43, lr: 2.81e-03, grad_scale: 32.0 2023-04-28 05:42:38,255 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.41 vs. limit=2.0 2023-04-28 05:42:45,303 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0926, 1.3955, 4.9099, 4.6376, 4.2240, 4.6931, 4.4358, 4.3421], device='cuda:4'), covar=tensor([0.6845, 0.6075, 0.1104, 0.1917, 0.1219, 0.1413, 0.1539, 0.1717], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0310, 0.0407, 0.0409, 0.0352, 0.0414, 0.0316, 0.0364], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-28 05:43:19,228 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 7.674e+01 1.528e+02 1.825e+02 2.243e+02 4.898e+02, threshold=3.650e+02, percent-clipped=3.0 2023-04-28 05:43:32,153 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=167391.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:43:42,859 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1244, 2.5524, 1.1450, 1.4699, 2.0204, 1.2374, 3.3681, 1.7315], device='cuda:4'), covar=tensor([0.0699, 0.0638, 0.0727, 0.1193, 0.0480, 0.1037, 0.0316, 0.0619], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0064, 0.0047, 0.0046, 0.0049, 0.0051, 0.0073, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-28 05:43:45,348 INFO [finetune.py:976] (4/7) Epoch 30, batch 1300, loss[loss=0.1494, simple_loss=0.215, pruned_loss=0.04194, over 4816.00 frames. ], tot_loss[loss=0.1624, simple_loss=0.2356, pruned_loss=0.04459, over 954230.37 frames. ], batch size: 25, lr: 2.81e-03, grad_scale: 32.0 2023-04-28 05:44:12,872 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=167420.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:44:25,756 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=167432.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:44:35,811 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=167439.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:44:45,752 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4488, 1.4187, 1.8833, 1.8192, 1.3034, 1.2576, 1.5582, 1.0017], device='cuda:4'), covar=tensor([0.0564, 0.0636, 0.0341, 0.0536, 0.0718, 0.1002, 0.0501, 0.0608], device='cuda:4'), in_proj_covar=tensor([0.0071, 0.0067, 0.0065, 0.0068, 0.0075, 0.0094, 0.0072, 0.0062], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-28 05:44:56,435 INFO [finetune.py:976] (4/7) Epoch 30, batch 1350, loss[loss=0.169, simple_loss=0.2433, pruned_loss=0.04737, over 4927.00 frames. ], tot_loss[loss=0.1631, simple_loss=0.2364, pruned_loss=0.04492, over 953773.56 frames. ], batch size: 33, lr: 2.81e-03, grad_scale: 32.0 2023-04-28 05:44:58,348 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=167456.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:45:00,777 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=167460.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:45:17,748 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=167468.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:45:25,587 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.814e+01 1.503e+02 1.795e+02 2.104e+02 5.763e+02, threshold=3.590e+02, percent-clipped=1.0 2023-04-28 05:45:34,218 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=167493.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:45:41,215 INFO [finetune.py:976] (4/7) Epoch 30, batch 1400, loss[loss=0.1378, simple_loss=0.2183, pruned_loss=0.02868, over 4828.00 frames. ], tot_loss[loss=0.1647, simple_loss=0.2385, pruned_loss=0.04549, over 953080.81 frames. ], batch size: 33, lr: 2.81e-03, grad_scale: 32.0 2023-04-28 05:46:14,262 INFO [finetune.py:976] (4/7) Epoch 30, batch 1450, loss[loss=0.1676, simple_loss=0.2499, pruned_loss=0.04269, over 4750.00 frames. ], tot_loss[loss=0.1662, simple_loss=0.2407, pruned_loss=0.04588, over 954161.46 frames. ], batch size: 54, lr: 2.81e-03, grad_scale: 32.0 2023-04-28 05:46:46,039 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.74 vs. limit=5.0 2023-04-28 05:46:52,054 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.057e+02 1.458e+02 1.734e+02 2.083e+02 3.176e+02, threshold=3.469e+02, percent-clipped=0.0 2023-04-28 05:47:00,606 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=167585.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:47:21,142 INFO [finetune.py:976] (4/7) Epoch 30, batch 1500, loss[loss=0.155, simple_loss=0.242, pruned_loss=0.03395, over 4862.00 frames. ], tot_loss[loss=0.1664, simple_loss=0.2412, pruned_loss=0.0458, over 954987.99 frames. ], batch size: 34, lr: 2.81e-03, grad_scale: 32.0 2023-04-28 05:47:52,140 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.63 vs. limit=2.0 2023-04-28 05:47:56,734 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=167633.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:48:20,793 INFO [finetune.py:976] (4/7) Epoch 30, batch 1550, loss[loss=0.2172, simple_loss=0.2806, pruned_loss=0.07689, over 4705.00 frames. ], tot_loss[loss=0.1656, simple_loss=0.2405, pruned_loss=0.04533, over 954980.55 frames. ], batch size: 59, lr: 2.81e-03, grad_scale: 32.0 2023-04-28 05:48:40,251 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.357e+01 1.558e+02 1.802e+02 2.131e+02 3.986e+02, threshold=3.604e+02, percent-clipped=1.0 2023-04-28 05:48:54,272 INFO [finetune.py:976] (4/7) Epoch 30, batch 1600, loss[loss=0.1616, simple_loss=0.2412, pruned_loss=0.04103, over 4767.00 frames. ], tot_loss[loss=0.1655, simple_loss=0.2396, pruned_loss=0.04569, over 954850.39 frames. ], batch size: 27, lr: 2.81e-03, grad_scale: 32.0 2023-04-28 05:49:05,589 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=167719.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:49:28,177 INFO [finetune.py:976] (4/7) Epoch 30, batch 1650, loss[loss=0.1483, simple_loss=0.219, pruned_loss=0.03882, over 4769.00 frames. ], tot_loss[loss=0.1618, simple_loss=0.2355, pruned_loss=0.04405, over 955373.36 frames. ], batch size: 26, lr: 2.81e-03, grad_scale: 32.0 2023-04-28 05:49:30,087 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=167756.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:49:32,993 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=167760.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:49:38,373 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=167767.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:49:47,070 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.070e+02 1.494e+02 1.761e+02 2.054e+02 5.504e+02, threshold=3.522e+02, percent-clipped=6.0 2023-04-28 05:49:47,202 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=167780.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:49:51,118 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4289, 2.3072, 2.4097, 2.9203, 2.9197, 2.2728, 2.0984, 2.5453], device='cuda:4'), covar=tensor([0.0747, 0.0936, 0.0543, 0.0456, 0.0527, 0.0774, 0.0622, 0.0492], device='cuda:4'), in_proj_covar=tensor([0.0182, 0.0201, 0.0182, 0.0170, 0.0177, 0.0177, 0.0148, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 05:49:52,949 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=167788.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:50:02,177 INFO [finetune.py:976] (4/7) Epoch 30, batch 1700, loss[loss=0.1704, simple_loss=0.243, pruned_loss=0.0489, over 4898.00 frames. ], tot_loss[loss=0.1602, simple_loss=0.2336, pruned_loss=0.04342, over 956444.87 frames. ], batch size: 43, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 05:50:02,853 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=167804.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:50:05,284 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=167808.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:50:18,856 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=167828.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:50:35,264 INFO [finetune.py:976] (4/7) Epoch 30, batch 1750, loss[loss=0.1415, simple_loss=0.2287, pruned_loss=0.0272, over 4824.00 frames. ], tot_loss[loss=0.1641, simple_loss=0.2376, pruned_loss=0.04529, over 957388.75 frames. ], batch size: 33, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 05:50:40,301 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9982, 1.6536, 1.7370, 2.2025, 2.2226, 1.8417, 1.6289, 2.0258], device='cuda:4'), covar=tensor([0.0616, 0.1057, 0.0671, 0.0438, 0.0506, 0.0719, 0.0642, 0.0443], device='cuda:4'), in_proj_covar=tensor([0.0182, 0.0202, 0.0182, 0.0171, 0.0178, 0.0177, 0.0149, 0.0176], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 05:50:48,454 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1956, 2.4304, 0.8656, 1.4519, 1.5254, 1.8568, 1.6172, 0.9340], device='cuda:4'), covar=tensor([0.1272, 0.1084, 0.1603, 0.1127, 0.0976, 0.0866, 0.1382, 0.1343], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0241, 0.0138, 0.0122, 0.0134, 0.0155, 0.0120, 0.0119], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 05:50:53,216 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.596e+01 1.560e+02 1.890e+02 2.217e+02 7.846e+02, threshold=3.780e+02, percent-clipped=0.0 2023-04-28 05:51:05,311 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.01 vs. limit=5.0 2023-04-28 05:51:08,163 INFO [finetune.py:976] (4/7) Epoch 30, batch 1800, loss[loss=0.2045, simple_loss=0.279, pruned_loss=0.06501, over 4154.00 frames. ], tot_loss[loss=0.1666, simple_loss=0.2405, pruned_loss=0.04629, over 955143.78 frames. ], batch size: 65, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 05:51:32,742 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.97 vs. limit=2.0 2023-04-28 05:51:41,397 INFO [finetune.py:976] (4/7) Epoch 30, batch 1850, loss[loss=0.1554, simple_loss=0.2373, pruned_loss=0.03677, over 4812.00 frames. ], tot_loss[loss=0.1682, simple_loss=0.2424, pruned_loss=0.04698, over 954497.73 frames. ], batch size: 33, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 05:52:04,614 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.089e+02 1.506e+02 1.821e+02 2.216e+02 4.142e+02, threshold=3.642e+02, percent-clipped=2.0 2023-04-28 05:52:17,359 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=3.88 vs. limit=5.0 2023-04-28 05:52:25,060 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4268, 3.0872, 0.9350, 1.6344, 1.7760, 2.2842, 1.8370, 1.0246], device='cuda:4'), covar=tensor([0.1410, 0.0954, 0.1896, 0.1364, 0.1117, 0.1001, 0.1553, 0.1817], device='cuda:4'), in_proj_covar=tensor([0.0119, 0.0242, 0.0138, 0.0122, 0.0134, 0.0155, 0.0120, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 05:52:26,789 INFO [finetune.py:976] (4/7) Epoch 30, batch 1900, loss[loss=0.1604, simple_loss=0.233, pruned_loss=0.04391, over 4749.00 frames. ], tot_loss[loss=0.1668, simple_loss=0.2413, pruned_loss=0.0461, over 952795.85 frames. ], batch size: 54, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 05:53:09,285 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=168035.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:53:12,872 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3330, 1.9218, 2.1590, 2.5933, 2.6861, 2.1534, 1.9279, 2.3750], device='cuda:4'), covar=tensor([0.0724, 0.1180, 0.0717, 0.0537, 0.0550, 0.0841, 0.0718, 0.0501], device='cuda:4'), in_proj_covar=tensor([0.0181, 0.0201, 0.0182, 0.0170, 0.0177, 0.0176, 0.0148, 0.0175], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 05:53:14,060 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2509, 2.9204, 0.9231, 1.5677, 1.6040, 2.1927, 1.7097, 1.0282], device='cuda:4'), covar=tensor([0.1564, 0.1184, 0.2029, 0.1419, 0.1208, 0.1129, 0.1534, 0.2019], device='cuda:4'), in_proj_covar=tensor([0.0120, 0.0242, 0.0138, 0.0122, 0.0134, 0.0155, 0.0120, 0.0120], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 05:53:15,367 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.28 vs. limit=2.0 2023-04-28 05:53:31,845 INFO [finetune.py:976] (4/7) Epoch 30, batch 1950, loss[loss=0.1299, simple_loss=0.1981, pruned_loss=0.03091, over 4775.00 frames. ], tot_loss[loss=0.1641, simple_loss=0.2388, pruned_loss=0.04472, over 953315.17 frames. ], batch size: 28, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 05:53:56,262 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=168075.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:54:04,671 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.011e+02 1.564e+02 1.911e+02 2.576e+02 9.918e+02, threshold=3.821e+02, percent-clipped=1.0 2023-04-28 05:54:06,589 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3036, 2.0989, 2.4775, 2.7219, 2.3228, 2.2386, 2.3560, 2.2607], device='cuda:4'), covar=tensor([0.4425, 0.6968, 0.7351, 0.5323, 0.5972, 0.8478, 0.8639, 0.9374], device='cuda:4'), in_proj_covar=tensor([0.0450, 0.0428, 0.0523, 0.0510, 0.0477, 0.0517, 0.0518, 0.0533], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 05:54:12,979 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1011, 1.8596, 2.2452, 2.4323, 2.1122, 2.0081, 2.1681, 2.0615], device='cuda:4'), covar=tensor([0.4255, 0.6520, 0.6244, 0.5128, 0.5502, 0.7946, 0.7733, 1.0017], device='cuda:4'), in_proj_covar=tensor([0.0450, 0.0428, 0.0523, 0.0510, 0.0477, 0.0517, 0.0518, 0.0533], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 05:54:13,524 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=168088.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:54:24,498 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=168096.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 05:54:32,406 INFO [finetune.py:976] (4/7) Epoch 30, batch 2000, loss[loss=0.1471, simple_loss=0.2228, pruned_loss=0.03572, over 4894.00 frames. ], tot_loss[loss=0.1629, simple_loss=0.2366, pruned_loss=0.04455, over 953799.70 frames. ], batch size: 32, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 05:54:52,715 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7527, 1.9610, 1.8872, 1.5904, 1.2869, 1.3762, 1.9471, 1.2850], device='cuda:4'), covar=tensor([0.1641, 0.1417, 0.1274, 0.1557, 0.2264, 0.1891, 0.0840, 0.2074], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0210, 0.0170, 0.0205, 0.0201, 0.0188, 0.0157, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 05:54:55,110 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=168123.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:54:57,349 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.92 vs. limit=2.0 2023-04-28 05:55:14,210 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=168136.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:55:25,248 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1477, 1.9865, 2.5766, 2.7391, 1.9122, 1.7539, 2.0178, 1.1981], device='cuda:4'), covar=tensor([0.0568, 0.0635, 0.0300, 0.0577, 0.0734, 0.1072, 0.0633, 0.0637], device='cuda:4'), in_proj_covar=tensor([0.0071, 0.0067, 0.0065, 0.0069, 0.0075, 0.0095, 0.0072, 0.0062], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-28 05:55:30,828 INFO [finetune.py:976] (4/7) Epoch 30, batch 2050, loss[loss=0.159, simple_loss=0.2436, pruned_loss=0.03715, over 4939.00 frames. ], tot_loss[loss=0.1601, simple_loss=0.2334, pruned_loss=0.0434, over 955328.51 frames. ], batch size: 33, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 05:55:47,309 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.348e+01 1.497e+02 1.684e+02 2.075e+02 3.378e+02, threshold=3.368e+02, percent-clipped=0.0 2023-04-28 05:55:54,964 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.0706, 1.3753, 1.2639, 1.6597, 1.4849, 1.5218, 1.3179, 2.4703], device='cuda:4'), covar=tensor([0.0639, 0.0830, 0.0821, 0.1293, 0.0677, 0.0483, 0.0782, 0.0220], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0042, 0.0040, 0.0037, 0.0038, 0.0054], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 05:56:04,273 INFO [finetune.py:976] (4/7) Epoch 30, batch 2100, loss[loss=0.1605, simple_loss=0.2358, pruned_loss=0.04258, over 4814.00 frames. ], tot_loss[loss=0.1602, simple_loss=0.2332, pruned_loss=0.04359, over 955259.39 frames. ], batch size: 38, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 05:56:23,110 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=168233.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:56:37,574 INFO [finetune.py:976] (4/7) Epoch 30, batch 2150, loss[loss=0.1694, simple_loss=0.2426, pruned_loss=0.04811, over 4788.00 frames. ], tot_loss[loss=0.163, simple_loss=0.2365, pruned_loss=0.04478, over 952833.69 frames. ], batch size: 29, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 05:56:45,956 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.32 vs. limit=2.0 2023-04-28 05:56:54,474 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.048e+02 1.582e+02 1.868e+02 2.335e+02 3.831e+02, threshold=3.737e+02, percent-clipped=1.0 2023-04-28 05:57:04,562 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=168294.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:57:09,876 INFO [finetune.py:976] (4/7) Epoch 30, batch 2200, loss[loss=0.1673, simple_loss=0.2469, pruned_loss=0.04381, over 4785.00 frames. ], tot_loss[loss=0.1636, simple_loss=0.2377, pruned_loss=0.04478, over 952040.06 frames. ], batch size: 29, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 05:57:49,303 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0229, 2.3105, 0.8978, 1.3155, 1.5218, 1.3835, 2.5303, 1.4459], device='cuda:4'), covar=tensor([0.0630, 0.0532, 0.0614, 0.1189, 0.0456, 0.0871, 0.0390, 0.0651], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0064, 0.0046, 0.0046, 0.0049, 0.0051, 0.0072, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-28 05:58:09,637 INFO [finetune.py:976] (4/7) Epoch 30, batch 2250, loss[loss=0.1248, simple_loss=0.2116, pruned_loss=0.01894, over 4865.00 frames. ], tot_loss[loss=0.1667, simple_loss=0.2411, pruned_loss=0.04619, over 953012.53 frames. ], batch size: 31, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 05:58:40,768 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=168375.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:58:43,710 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.119e+02 1.481e+02 1.883e+02 2.203e+02 3.675e+02, threshold=3.765e+02, percent-clipped=0.0 2023-04-28 05:58:55,049 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=168391.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 05:59:13,988 INFO [finetune.py:976] (4/7) Epoch 30, batch 2300, loss[loss=0.1461, simple_loss=0.2165, pruned_loss=0.0379, over 4712.00 frames. ], tot_loss[loss=0.1671, simple_loss=0.2418, pruned_loss=0.04624, over 952534.32 frames. ], batch size: 23, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 05:59:38,058 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=168423.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 05:59:38,086 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=168423.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:00:16,725 INFO [finetune.py:976] (4/7) Epoch 30, batch 2350, loss[loss=0.1551, simple_loss=0.2218, pruned_loss=0.04418, over 4736.00 frames. ], tot_loss[loss=0.1648, simple_loss=0.2392, pruned_loss=0.04521, over 952827.35 frames. ], batch size: 59, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 06:00:30,226 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9671, 2.4754, 2.0498, 2.4140, 1.6773, 2.1923, 2.0298, 1.5652], device='cuda:4'), covar=tensor([0.1670, 0.0823, 0.0705, 0.0838, 0.3078, 0.0876, 0.1547, 0.2033], device='cuda:4'), in_proj_covar=tensor([0.0280, 0.0299, 0.0216, 0.0271, 0.0306, 0.0252, 0.0246, 0.0258], device='cuda:4'), out_proj_covar=tensor([1.1142e-04, 1.1728e-04, 8.4906e-05, 1.0651e-04, 1.2315e-04, 9.8853e-05, 9.8676e-05, 1.0149e-04], device='cuda:4') 2023-04-28 06:00:37,860 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=168471.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:00:48,858 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.790e+01 1.485e+02 1.765e+02 2.221e+02 4.572e+02, threshold=3.529e+02, percent-clipped=1.0 2023-04-28 06:00:59,970 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2342, 1.6186, 1.4974, 1.8016, 1.7341, 1.9194, 1.4785, 3.3458], device='cuda:4'), covar=tensor([0.0626, 0.0766, 0.0741, 0.1137, 0.0588, 0.0424, 0.0680, 0.0162], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0040, 0.0037, 0.0038, 0.0054], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0013, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 06:01:21,061 INFO [finetune.py:976] (4/7) Epoch 30, batch 2400, loss[loss=0.1482, simple_loss=0.2076, pruned_loss=0.04442, over 4839.00 frames. ], tot_loss[loss=0.1618, simple_loss=0.2356, pruned_loss=0.04403, over 954300.59 frames. ], batch size: 40, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 06:02:28,239 INFO [finetune.py:976] (4/7) Epoch 30, batch 2450, loss[loss=0.1604, simple_loss=0.226, pruned_loss=0.0474, over 4894.00 frames. ], tot_loss[loss=0.1607, simple_loss=0.2336, pruned_loss=0.04389, over 954096.42 frames. ], batch size: 35, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 06:02:59,361 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3894, 1.2926, 1.5883, 1.5962, 1.2957, 1.1526, 1.2215, 0.7928], device='cuda:4'), covar=tensor([0.0462, 0.0545, 0.0311, 0.0506, 0.0609, 0.1046, 0.0500, 0.0504], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0067, 0.0065, 0.0068, 0.0075, 0.0094, 0.0072, 0.0062], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-28 06:03:10,378 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.038e+02 1.524e+02 1.793e+02 2.156e+02 5.076e+02, threshold=3.587e+02, percent-clipped=1.0 2023-04-28 06:03:21,582 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=168589.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:03:36,318 INFO [finetune.py:976] (4/7) Epoch 30, batch 2500, loss[loss=0.1289, simple_loss=0.2093, pruned_loss=0.02428, over 4758.00 frames. ], tot_loss[loss=0.1641, simple_loss=0.2369, pruned_loss=0.04566, over 954362.91 frames. ], batch size: 28, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 06:04:19,677 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6983, 1.5596, 0.7183, 1.3742, 1.4927, 1.5570, 1.4512, 1.4822], device='cuda:4'), covar=tensor([0.0488, 0.0387, 0.0383, 0.0546, 0.0286, 0.0514, 0.0489, 0.0572], device='cuda:4'), in_proj_covar=tensor([0.0027, 0.0023, 0.0022, 0.0028, 0.0019, 0.0027, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0052, 0.0046, 0.0039, 0.0053, 0.0039, 0.0051, 0.0051, 0.0053], device='cuda:4') 2023-04-28 06:04:29,745 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.65 vs. limit=2.0 2023-04-28 06:04:49,842 INFO [finetune.py:976] (4/7) Epoch 30, batch 2550, loss[loss=0.1847, simple_loss=0.2652, pruned_loss=0.05213, over 4903.00 frames. ], tot_loss[loss=0.1666, simple_loss=0.2405, pruned_loss=0.04635, over 955295.81 frames. ], batch size: 37, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 06:05:23,464 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.119e+02 1.608e+02 1.886e+02 2.283e+02 4.830e+02, threshold=3.772e+02, percent-clipped=5.0 2023-04-28 06:05:27,128 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6131, 1.4317, 1.8403, 1.8555, 1.4494, 1.2967, 1.4710, 0.8943], device='cuda:4'), covar=tensor([0.0473, 0.0573, 0.0322, 0.0488, 0.0667, 0.1051, 0.0495, 0.0559], device='cuda:4'), in_proj_covar=tensor([0.0070, 0.0067, 0.0065, 0.0068, 0.0075, 0.0094, 0.0072, 0.0062], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-28 06:05:31,379 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=168691.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:05:38,671 INFO [finetune.py:976] (4/7) Epoch 30, batch 2600, loss[loss=0.1791, simple_loss=0.2528, pruned_loss=0.05271, over 4863.00 frames. ], tot_loss[loss=0.1677, simple_loss=0.2419, pruned_loss=0.04673, over 955123.25 frames. ], batch size: 34, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 06:05:42,392 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7901, 2.3660, 1.9946, 1.9223, 1.2879, 1.3798, 2.0792, 1.2925], device='cuda:4'), covar=tensor([0.1732, 0.1380, 0.1442, 0.1607, 0.2326, 0.2032, 0.0923, 0.2145], device='cuda:4'), in_proj_covar=tensor([0.0200, 0.0210, 0.0170, 0.0205, 0.0201, 0.0188, 0.0158, 0.0188], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 06:05:55,263 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2685, 1.5095, 1.7693, 1.8753, 1.7891, 1.8273, 1.7693, 1.7984], device='cuda:4'), covar=tensor([0.3588, 0.4683, 0.3851, 0.4061, 0.5120, 0.6515, 0.4717, 0.4336], device='cuda:4'), in_proj_covar=tensor([0.0346, 0.0376, 0.0333, 0.0344, 0.0353, 0.0397, 0.0364, 0.0336], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 06:06:26,241 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=168739.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:06:29,400 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5392, 1.4780, 1.8242, 1.8557, 1.3574, 1.2474, 1.4944, 0.9400], device='cuda:4'), covar=tensor([0.0522, 0.0527, 0.0307, 0.0519, 0.0762, 0.1108, 0.0516, 0.0547], device='cuda:4'), in_proj_covar=tensor([0.0071, 0.0067, 0.0065, 0.0069, 0.0075, 0.0095, 0.0072, 0.0062], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-28 06:06:40,212 INFO [finetune.py:976] (4/7) Epoch 30, batch 2650, loss[loss=0.1723, simple_loss=0.2451, pruned_loss=0.04981, over 4774.00 frames. ], tot_loss[loss=0.1679, simple_loss=0.2424, pruned_loss=0.04673, over 955396.69 frames. ], batch size: 28, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 06:07:17,597 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.117e+02 1.537e+02 1.777e+02 2.075e+02 3.537e+02, threshold=3.553e+02, percent-clipped=0.0 2023-04-28 06:07:44,076 INFO [finetune.py:976] (4/7) Epoch 30, batch 2700, loss[loss=0.1533, simple_loss=0.2212, pruned_loss=0.04277, over 4810.00 frames. ], tot_loss[loss=0.1669, simple_loss=0.2413, pruned_loss=0.04626, over 954052.49 frames. ], batch size: 38, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 06:08:17,673 INFO [finetune.py:976] (4/7) Epoch 30, batch 2750, loss[loss=0.1709, simple_loss=0.2439, pruned_loss=0.04893, over 4759.00 frames. ], tot_loss[loss=0.1656, simple_loss=0.239, pruned_loss=0.04613, over 954214.31 frames. ], batch size: 27, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 06:08:25,363 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.5893, 4.3640, 3.0891, 5.3149, 4.5153, 4.5977, 1.8508, 4.6152], device='cuda:4'), covar=tensor([0.1636, 0.0980, 0.3692, 0.0915, 0.3212, 0.1835, 0.6145, 0.2084], device='cuda:4'), in_proj_covar=tensor([0.0248, 0.0217, 0.0253, 0.0303, 0.0299, 0.0250, 0.0275, 0.0274], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 06:08:26,944 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.73 vs. limit=2.0 2023-04-28 06:08:35,311 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.781e+01 1.528e+02 1.886e+02 2.380e+02 6.138e+02, threshold=3.773e+02, percent-clipped=2.0 2023-04-28 06:08:41,788 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=168889.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:08:48,449 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3516, 2.1277, 1.7970, 1.9178, 2.2852, 1.8618, 2.6176, 1.6303], device='cuda:4'), covar=tensor([0.3134, 0.1904, 0.4534, 0.2736, 0.1419, 0.2357, 0.1324, 0.4254], device='cuda:4'), in_proj_covar=tensor([0.0339, 0.0356, 0.0426, 0.0353, 0.0384, 0.0376, 0.0370, 0.0425], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 06:08:50,766 INFO [finetune.py:976] (4/7) Epoch 30, batch 2800, loss[loss=0.1687, simple_loss=0.2387, pruned_loss=0.04935, over 4894.00 frames. ], tot_loss[loss=0.163, simple_loss=0.2359, pruned_loss=0.04506, over 954611.46 frames. ], batch size: 35, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 06:09:12,959 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=168937.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:09:24,482 INFO [finetune.py:976] (4/7) Epoch 30, batch 2850, loss[loss=0.1478, simple_loss=0.2211, pruned_loss=0.03722, over 4144.00 frames. ], tot_loss[loss=0.1623, simple_loss=0.2348, pruned_loss=0.04492, over 954379.92 frames. ], batch size: 65, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 06:09:33,136 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=168967.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:09:34,354 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=168969.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 06:09:41,982 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.589e+01 1.497e+02 1.758e+02 2.128e+02 6.037e+02, threshold=3.517e+02, percent-clipped=2.0 2023-04-28 06:09:57,676 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.72 vs. limit=5.0 2023-04-28 06:09:57,757 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.40 vs. limit=2.0 2023-04-28 06:09:58,544 INFO [finetune.py:976] (4/7) Epoch 30, batch 2900, loss[loss=0.1813, simple_loss=0.2536, pruned_loss=0.05451, over 4840.00 frames. ], tot_loss[loss=0.1651, simple_loss=0.2382, pruned_loss=0.04606, over 955799.88 frames. ], batch size: 49, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 06:10:08,479 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2579, 1.8010, 2.1677, 2.2454, 2.1766, 1.7509, 1.1917, 1.8058], device='cuda:4'), covar=tensor([0.3238, 0.2865, 0.1634, 0.2310, 0.2317, 0.2666, 0.4104, 0.1864], device='cuda:4'), in_proj_covar=tensor([0.0292, 0.0245, 0.0228, 0.0314, 0.0223, 0.0235, 0.0229, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-28 06:10:14,823 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=169028.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:10:16,036 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=169030.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 06:10:31,851 INFO [finetune.py:976] (4/7) Epoch 30, batch 2950, loss[loss=0.1652, simple_loss=0.2483, pruned_loss=0.04107, over 4811.00 frames. ], tot_loss[loss=0.1666, simple_loss=0.2401, pruned_loss=0.04656, over 954478.11 frames. ], batch size: 39, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 06:10:49,917 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.313e+01 1.646e+02 1.860e+02 2.214e+02 4.123e+02, threshold=3.720e+02, percent-clipped=2.0 2023-04-28 06:10:50,649 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=169082.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:11:05,769 INFO [finetune.py:976] (4/7) Epoch 30, batch 3000, loss[loss=0.2334, simple_loss=0.2889, pruned_loss=0.08891, over 4123.00 frames. ], tot_loss[loss=0.1664, simple_loss=0.2406, pruned_loss=0.0461, over 955988.22 frames. ], batch size: 66, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 06:11:05,769 INFO [finetune.py:1001] (4/7) Computing validation loss 2023-04-28 06:11:13,897 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5055, 1.3519, 3.8333, 3.5824, 3.5026, 3.7050, 3.7638, 3.4472], device='cuda:4'), covar=tensor([0.6585, 0.4675, 0.1420, 0.2288, 0.1283, 0.1449, 0.0934, 0.1821], device='cuda:4'), in_proj_covar=tensor([0.0309, 0.0308, 0.0406, 0.0409, 0.0349, 0.0415, 0.0314, 0.0363], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-28 06:11:14,383 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7308, 1.6245, 1.8191, 2.1030, 2.1811, 1.6600, 1.4664, 1.9444], device='cuda:4'), covar=tensor([0.0772, 0.1216, 0.0782, 0.0532, 0.0585, 0.0848, 0.0700, 0.0566], device='cuda:4'), in_proj_covar=tensor([0.0187, 0.0207, 0.0187, 0.0174, 0.0181, 0.0182, 0.0152, 0.0180], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 06:11:15,204 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8199, 1.1104, 1.7830, 2.2791, 1.9286, 1.7929, 1.7635, 1.7708], device='cuda:4'), covar=tensor([0.5075, 0.6904, 0.6719, 0.5867, 0.6222, 0.7825, 0.8026, 0.9539], device='cuda:4'), in_proj_covar=tensor([0.0448, 0.0426, 0.0520, 0.0507, 0.0474, 0.0514, 0.0514, 0.0530], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 06:11:16,543 INFO [finetune.py:1010] (4/7) Epoch 30, validation: loss=0.1534, simple_loss=0.2215, pruned_loss=0.04259, over 2265189.00 frames. 2023-04-28 06:11:16,543 INFO [finetune.py:1011] (4/7) Maximum memory allocated so far is 6529MB 2023-04-28 06:11:24,725 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=169116.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:11:26,575 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-28 06:11:52,070 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=169143.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:12:04,396 INFO [finetune.py:976] (4/7) Epoch 30, batch 3050, loss[loss=0.1964, simple_loss=0.2606, pruned_loss=0.06606, over 4881.00 frames. ], tot_loss[loss=0.1663, simple_loss=0.2406, pruned_loss=0.04599, over 953414.13 frames. ], batch size: 43, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 06:12:13,913 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=169159.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:12:36,607 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=169177.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:12:42,081 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.026e+02 1.505e+02 1.782e+02 2.187e+02 4.062e+02, threshold=3.563e+02, percent-clipped=1.0 2023-04-28 06:12:54,191 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.40 vs. limit=5.0 2023-04-28 06:13:06,207 INFO [finetune.py:976] (4/7) Epoch 30, batch 3100, loss[loss=0.1404, simple_loss=0.2202, pruned_loss=0.03027, over 4796.00 frames. ], tot_loss[loss=0.1644, simple_loss=0.2384, pruned_loss=0.04521, over 953821.42 frames. ], batch size: 29, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 06:13:15,862 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.60 vs. limit=5.0 2023-04-28 06:13:29,659 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=169220.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:13:50,608 INFO [finetune.py:976] (4/7) Epoch 30, batch 3150, loss[loss=0.1613, simple_loss=0.2414, pruned_loss=0.04057, over 4796.00 frames. ], tot_loss[loss=0.1629, simple_loss=0.2362, pruned_loss=0.0448, over 954777.62 frames. ], batch size: 29, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 06:13:57,804 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1185, 2.4518, 1.0519, 1.3780, 1.8594, 1.3176, 3.3094, 1.7621], device='cuda:4'), covar=tensor([0.0665, 0.0677, 0.0773, 0.1280, 0.0498, 0.1001, 0.0359, 0.0611], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0046, 0.0046, 0.0049, 0.0051, 0.0072, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-28 06:14:10,086 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.051e+01 1.353e+02 1.739e+02 2.048e+02 6.582e+02, threshold=3.478e+02, percent-clipped=1.0 2023-04-28 06:14:23,996 INFO [finetune.py:976] (4/7) Epoch 30, batch 3200, loss[loss=0.1685, simple_loss=0.2404, pruned_loss=0.04828, over 4866.00 frames. ], tot_loss[loss=0.1607, simple_loss=0.2335, pruned_loss=0.04393, over 954123.88 frames. ], batch size: 44, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 06:14:31,367 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0489, 2.4685, 1.9625, 1.8510, 1.4527, 1.5429, 2.0307, 1.4180], device='cuda:4'), covar=tensor([0.1595, 0.1301, 0.1422, 0.1630, 0.2269, 0.1792, 0.0990, 0.1981], device='cuda:4'), in_proj_covar=tensor([0.0199, 0.0209, 0.0170, 0.0204, 0.0201, 0.0187, 0.0157, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 06:14:38,830 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=169323.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:14:40,577 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=169325.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 06:14:57,969 INFO [finetune.py:976] (4/7) Epoch 30, batch 3250, loss[loss=0.1663, simple_loss=0.2591, pruned_loss=0.03681, over 4873.00 frames. ], tot_loss[loss=0.1605, simple_loss=0.2333, pruned_loss=0.04382, over 953385.40 frames. ], batch size: 34, lr: 2.80e-03, grad_scale: 32.0 2023-04-28 06:14:59,377 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6361, 1.5697, 2.0053, 2.0246, 1.3977, 1.3576, 1.6009, 1.0294], device='cuda:4'), covar=tensor([0.0491, 0.0604, 0.0321, 0.0568, 0.0783, 0.1024, 0.0472, 0.0495], device='cuda:4'), in_proj_covar=tensor([0.0071, 0.0067, 0.0065, 0.0069, 0.0076, 0.0095, 0.0072, 0.0062], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-28 06:15:18,032 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.661e+01 1.547e+02 1.939e+02 2.313e+02 4.246e+02, threshold=3.878e+02, percent-clipped=2.0 2023-04-28 06:15:32,129 INFO [finetune.py:976] (4/7) Epoch 30, batch 3300, loss[loss=0.1816, simple_loss=0.2494, pruned_loss=0.05693, over 4876.00 frames. ], tot_loss[loss=0.1624, simple_loss=0.2363, pruned_loss=0.04423, over 953527.97 frames. ], batch size: 31, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:15:56,492 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=169438.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:16:01,486 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0530, 1.4766, 2.0113, 2.4698, 2.1554, 1.9945, 1.9977, 1.9241], device='cuda:4'), covar=tensor([0.4285, 0.6467, 0.6320, 0.5040, 0.5617, 0.7238, 0.7705, 0.7667], device='cuda:4'), in_proj_covar=tensor([0.0446, 0.0426, 0.0519, 0.0506, 0.0472, 0.0513, 0.0513, 0.0529], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 06:16:05,663 INFO [finetune.py:976] (4/7) Epoch 30, batch 3350, loss[loss=0.2149, simple_loss=0.2898, pruned_loss=0.06997, over 4912.00 frames. ], tot_loss[loss=0.1636, simple_loss=0.2383, pruned_loss=0.04442, over 954386.94 frames. ], batch size: 36, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:16:18,730 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=169472.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:16:26,057 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.059e+02 1.565e+02 1.911e+02 2.262e+02 9.005e+02, threshold=3.822e+02, percent-clipped=1.0 2023-04-28 06:16:39,414 INFO [finetune.py:976] (4/7) Epoch 30, batch 3400, loss[loss=0.189, simple_loss=0.2624, pruned_loss=0.05782, over 4866.00 frames. ], tot_loss[loss=0.165, simple_loss=0.2399, pruned_loss=0.04501, over 954505.44 frames. ], batch size: 34, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:16:47,268 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=169515.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:17:02,466 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5859, 1.7608, 1.5945, 2.0074, 1.9503, 2.0518, 1.6191, 4.4045], device='cuda:4'), covar=tensor([0.0531, 0.0779, 0.0795, 0.1222, 0.0603, 0.0570, 0.0742, 0.0107], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0039, 0.0043, 0.0040, 0.0037, 0.0038, 0.0055], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0014, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 06:17:09,191 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.1731, 2.6929, 3.0385, 3.6517, 3.0399, 2.5718, 2.6951, 3.0421], device='cuda:4'), covar=tensor([0.2603, 0.2552, 0.1343, 0.1886, 0.2119, 0.2299, 0.2945, 0.1426], device='cuda:4'), in_proj_covar=tensor([0.0294, 0.0248, 0.0230, 0.0317, 0.0225, 0.0238, 0.0231, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-28 06:17:12,722 INFO [finetune.py:976] (4/7) Epoch 30, batch 3450, loss[loss=0.1436, simple_loss=0.212, pruned_loss=0.03759, over 4802.00 frames. ], tot_loss[loss=0.1646, simple_loss=0.2395, pruned_loss=0.04487, over 953674.93 frames. ], batch size: 26, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:17:36,960 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.050e+02 1.504e+02 1.780e+02 2.155e+02 3.271e+02, threshold=3.560e+02, percent-clipped=0.0 2023-04-28 06:18:07,733 INFO [finetune.py:976] (4/7) Epoch 30, batch 3500, loss[loss=0.1356, simple_loss=0.2094, pruned_loss=0.03094, over 4767.00 frames. ], tot_loss[loss=0.1637, simple_loss=0.2379, pruned_loss=0.0448, over 954214.37 frames. ], batch size: 28, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:18:29,831 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=169621.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:18:31,005 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=169623.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:18:32,252 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=169625.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 06:18:52,350 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1362, 1.8877, 2.1692, 2.4851, 2.5126, 2.0586, 1.8496, 2.2496], device='cuda:4'), covar=tensor([0.0778, 0.1026, 0.0637, 0.0536, 0.0547, 0.0795, 0.0666, 0.0527], device='cuda:4'), in_proj_covar=tensor([0.0184, 0.0205, 0.0185, 0.0172, 0.0179, 0.0180, 0.0151, 0.0179], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 06:18:52,646 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.09 vs. limit=2.0 2023-04-28 06:19:13,502 INFO [finetune.py:976] (4/7) Epoch 30, batch 3550, loss[loss=0.1549, simple_loss=0.2385, pruned_loss=0.03568, over 4825.00 frames. ], tot_loss[loss=0.1626, simple_loss=0.2357, pruned_loss=0.04475, over 954757.32 frames. ], batch size: 30, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:19:22,159 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.1240, 3.1704, 2.4006, 3.5991, 3.2095, 3.2121, 1.3977, 3.0811], device='cuda:4'), covar=tensor([0.2133, 0.1327, 0.3209, 0.2417, 0.3049, 0.2228, 0.5363, 0.2517], device='cuda:4'), in_proj_covar=tensor([0.0251, 0.0219, 0.0255, 0.0306, 0.0303, 0.0252, 0.0277, 0.0277], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 06:19:30,802 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=169671.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:19:32,046 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=169673.0, num_to_drop=1, layers_to_drop={0} 2023-04-28 06:19:42,308 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.126e+01 1.430e+02 1.728e+02 2.038e+02 3.209e+02, threshold=3.455e+02, percent-clipped=0.0 2023-04-28 06:19:43,542 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=169682.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:20:12,671 INFO [finetune.py:976] (4/7) Epoch 30, batch 3600, loss[loss=0.2036, simple_loss=0.2751, pruned_loss=0.06609, over 4858.00 frames. ], tot_loss[loss=0.1615, simple_loss=0.2339, pruned_loss=0.04453, over 953922.89 frames. ], batch size: 44, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:20:45,859 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=169729.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:20:57,516 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=169738.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:21:10,902 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=169747.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:21:20,086 INFO [finetune.py:976] (4/7) Epoch 30, batch 3650, loss[loss=0.2182, simple_loss=0.2845, pruned_loss=0.07599, over 4192.00 frames. ], tot_loss[loss=0.1639, simple_loss=0.2368, pruned_loss=0.04552, over 955152.29 frames. ], batch size: 65, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:21:43,281 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=169772.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:21:53,879 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.801e+01 1.645e+02 1.930e+02 2.224e+02 4.571e+02, threshold=3.861e+02, percent-clipped=2.0 2023-04-28 06:22:03,120 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=169786.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:22:03,167 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=169786.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:22:11,322 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=169790.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:22:25,955 INFO [finetune.py:976] (4/7) Epoch 30, batch 3700, loss[loss=0.1914, simple_loss=0.2697, pruned_loss=0.05658, over 4901.00 frames. ], tot_loss[loss=0.1654, simple_loss=0.2392, pruned_loss=0.04578, over 955976.43 frames. ], batch size: 43, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:22:35,253 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=169808.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:22:44,957 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=169815.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:22:47,944 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=169820.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:23:00,451 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4061, 3.2291, 1.0373, 1.7486, 1.7397, 2.3187, 1.7714, 0.9783], device='cuda:4'), covar=tensor([0.1458, 0.1001, 0.1820, 0.1300, 0.1163, 0.1013, 0.1676, 0.1973], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0236, 0.0135, 0.0119, 0.0130, 0.0152, 0.0117, 0.0116], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 06:23:22,795 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=169847.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:23:29,211 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.1260, 0.7293, 0.9553, 0.8013, 1.2016, 0.9490, 0.8791, 0.9883], device='cuda:4'), covar=tensor([0.2137, 0.1693, 0.2254, 0.1770, 0.1276, 0.1766, 0.2059, 0.2634], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0307, 0.0350, 0.0286, 0.0324, 0.0305, 0.0300, 0.0375], device='cuda:4'), out_proj_covar=tensor([6.3566e-05, 6.2533e-05, 7.3114e-05, 5.7030e-05, 6.5881e-05, 6.3194e-05, 6.1828e-05, 7.9186e-05], device='cuda:4') 2023-04-28 06:23:31,472 INFO [finetune.py:976] (4/7) Epoch 30, batch 3750, loss[loss=0.1739, simple_loss=0.2529, pruned_loss=0.04749, over 4810.00 frames. ], tot_loss[loss=0.1658, simple_loss=0.2398, pruned_loss=0.04588, over 954381.05 frames. ], batch size: 33, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:23:42,232 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=169863.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:23:53,250 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7509, 0.7795, 1.6295, 2.0424, 1.7918, 1.6373, 1.6596, 1.6738], device='cuda:4'), covar=tensor([0.4219, 0.6585, 0.5653, 0.5563, 0.5714, 0.7218, 0.7512, 0.7493], device='cuda:4'), in_proj_covar=tensor([0.0447, 0.0428, 0.0523, 0.0509, 0.0476, 0.0516, 0.0516, 0.0532], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 06:24:04,807 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.130e+02 1.513e+02 1.804e+02 2.185e+02 4.010e+02, threshold=3.607e+02, percent-clipped=1.0 2023-04-28 06:24:15,654 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6536, 1.2614, 4.2445, 3.6668, 3.8133, 3.9889, 3.9224, 3.5270], device='cuda:4'), covar=tensor([0.8868, 0.8151, 0.1690, 0.3291, 0.2117, 0.3811, 0.2491, 0.3139], device='cuda:4'), in_proj_covar=tensor([0.0308, 0.0309, 0.0405, 0.0407, 0.0347, 0.0412, 0.0314, 0.0362], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-28 06:24:36,634 INFO [finetune.py:976] (4/7) Epoch 30, batch 3800, loss[loss=0.1478, simple_loss=0.2165, pruned_loss=0.03949, over 4743.00 frames. ], tot_loss[loss=0.1655, simple_loss=0.2396, pruned_loss=0.0457, over 953926.17 frames. ], batch size: 23, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:25:40,866 INFO [finetune.py:976] (4/7) Epoch 30, batch 3850, loss[loss=0.1589, simple_loss=0.2302, pruned_loss=0.04379, over 4819.00 frames. ], tot_loss[loss=0.1643, simple_loss=0.2383, pruned_loss=0.04518, over 951980.31 frames. ], batch size: 39, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:26:12,314 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=169977.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:26:14,680 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.720e+01 1.456e+02 1.655e+02 1.971e+02 4.065e+02, threshold=3.309e+02, percent-clipped=1.0 2023-04-28 06:26:45,749 INFO [finetune.py:976] (4/7) Epoch 30, batch 3900, loss[loss=0.1355, simple_loss=0.2013, pruned_loss=0.03484, over 4747.00 frames. ], tot_loss[loss=0.1624, simple_loss=0.236, pruned_loss=0.04446, over 952586.35 frames. ], batch size: 23, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:27:14,910 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7741, 1.3517, 1.8662, 2.1065, 1.7910, 1.6900, 1.7688, 1.8277], device='cuda:4'), covar=tensor([0.5522, 0.8312, 0.7183, 0.7959, 0.7147, 0.9898, 0.9359, 1.1535], device='cuda:4'), in_proj_covar=tensor([0.0448, 0.0428, 0.0522, 0.0509, 0.0476, 0.0516, 0.0516, 0.0532], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 06:27:43,950 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=170045.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:27:48,653 INFO [finetune.py:976] (4/7) Epoch 30, batch 3950, loss[loss=0.1519, simple_loss=0.2288, pruned_loss=0.0375, over 4846.00 frames. ], tot_loss[loss=0.1597, simple_loss=0.2328, pruned_loss=0.04333, over 951766.96 frames. ], batch size: 47, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:28:24,767 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.030e+02 1.459e+02 1.752e+02 2.144e+02 4.109e+02, threshold=3.505e+02, percent-clipped=2.0 2023-04-28 06:28:27,242 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=170085.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:28:46,916 INFO [finetune.py:976] (4/7) Epoch 30, batch 4000, loss[loss=0.1582, simple_loss=0.2237, pruned_loss=0.04633, over 4876.00 frames. ], tot_loss[loss=0.1589, simple_loss=0.2321, pruned_loss=0.04284, over 952592.58 frames. ], batch size: 31, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:28:46,999 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=170103.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:28:53,807 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=170106.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:29:04,565 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2740, 1.8823, 2.1853, 2.3401, 2.2029, 1.8871, 1.3734, 1.9511], device='cuda:4'), covar=tensor([0.3042, 0.2572, 0.1541, 0.1807, 0.2204, 0.2343, 0.3666, 0.1603], device='cuda:4'), in_proj_covar=tensor([0.0296, 0.0248, 0.0231, 0.0318, 0.0225, 0.0238, 0.0232, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-28 06:29:26,513 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.5680, 4.4745, 3.2517, 5.3016, 4.5881, 4.5365, 2.1023, 4.5648], device='cuda:4'), covar=tensor([0.1675, 0.1153, 0.3361, 0.0840, 0.3720, 0.1723, 0.5751, 0.2040], device='cuda:4'), in_proj_covar=tensor([0.0250, 0.0219, 0.0253, 0.0302, 0.0300, 0.0250, 0.0274, 0.0276], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 06:29:27,783 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7811, 2.2884, 1.7333, 1.6729, 1.2980, 1.2956, 1.7244, 1.2627], device='cuda:4'), covar=tensor([0.1587, 0.1251, 0.1351, 0.1592, 0.2216, 0.1830, 0.1020, 0.1982], device='cuda:4'), in_proj_covar=tensor([0.0197, 0.0207, 0.0168, 0.0202, 0.0199, 0.0185, 0.0155, 0.0186], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0003], device='cuda:4') 2023-04-28 06:29:39,710 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=170142.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:29:47,097 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2159, 1.7553, 1.4326, 2.1163, 2.2793, 1.8550, 1.8080, 1.5030], device='cuda:4'), covar=tensor([0.1855, 0.1594, 0.2084, 0.1434, 0.1146, 0.1793, 0.2116, 0.2451], device='cuda:4'), in_proj_covar=tensor([0.0314, 0.0307, 0.0350, 0.0287, 0.0324, 0.0306, 0.0300, 0.0376], device='cuda:4'), out_proj_covar=tensor([6.3686e-05, 6.2642e-05, 7.3265e-05, 5.7208e-05, 6.5860e-05, 6.3426e-05, 6.1795e-05, 7.9226e-05], device='cuda:4') 2023-04-28 06:29:47,766 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7512, 1.0339, 1.7296, 2.1921, 1.8229, 1.6762, 1.7399, 1.7187], device='cuda:4'), covar=tensor([0.4630, 0.7131, 0.6372, 0.5802, 0.5998, 0.8308, 0.7957, 0.8835], device='cuda:4'), in_proj_covar=tensor([0.0447, 0.0427, 0.0521, 0.0507, 0.0475, 0.0515, 0.0515, 0.0531], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 06:29:48,704 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.53 vs. limit=2.0 2023-04-28 06:29:50,076 INFO [finetune.py:976] (4/7) Epoch 30, batch 4050, loss[loss=0.1684, simple_loss=0.2566, pruned_loss=0.04008, over 4825.00 frames. ], tot_loss[loss=0.1608, simple_loss=0.2349, pruned_loss=0.04336, over 952782.91 frames. ], batch size: 33, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:30:09,290 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2024, 2.5707, 1.3701, 1.5898, 2.1336, 1.4380, 3.1210, 1.7714], device='cuda:4'), covar=tensor([0.0638, 0.0763, 0.0789, 0.1005, 0.0398, 0.0857, 0.0223, 0.0537], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0046, 0.0046, 0.0049, 0.0051, 0.0072, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-28 06:30:22,175 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.182e+02 1.558e+02 1.902e+02 2.221e+02 4.560e+02, threshold=3.805e+02, percent-clipped=3.0 2023-04-28 06:30:31,481 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.54 vs. limit=2.0 2023-04-28 06:30:36,789 INFO [finetune.py:976] (4/7) Epoch 30, batch 4100, loss[loss=0.1295, simple_loss=0.2075, pruned_loss=0.02572, over 4741.00 frames. ], tot_loss[loss=0.1623, simple_loss=0.2369, pruned_loss=0.04385, over 951076.62 frames. ], batch size: 27, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:30:42,958 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-28 06:30:45,623 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7271, 1.2591, 1.8112, 2.2211, 1.8165, 1.6962, 1.7814, 1.7310], device='cuda:4'), covar=tensor([0.4123, 0.6567, 0.5869, 0.5037, 0.5355, 0.7410, 0.7111, 0.8160], device='cuda:4'), in_proj_covar=tensor([0.0447, 0.0427, 0.0521, 0.0507, 0.0475, 0.0515, 0.0516, 0.0531], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 06:31:03,918 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=192, metric=1.76 vs. limit=2.0 2023-04-28 06:31:10,087 INFO [finetune.py:976] (4/7) Epoch 30, batch 4150, loss[loss=0.1764, simple_loss=0.2511, pruned_loss=0.0508, over 4891.00 frames. ], tot_loss[loss=0.1645, simple_loss=0.2392, pruned_loss=0.04489, over 951272.21 frames. ], batch size: 35, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:31:26,680 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=170277.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:31:29,028 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.851e+01 1.521e+02 1.781e+02 2.076e+02 3.722e+02, threshold=3.562e+02, percent-clipped=0.0 2023-04-28 06:31:30,429 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.7389, 1.3183, 1.7967, 2.2631, 1.8593, 1.6961, 1.7600, 1.7021], device='cuda:4'), covar=tensor([0.4083, 0.6337, 0.5824, 0.4669, 0.5354, 0.7346, 0.6940, 0.9081], device='cuda:4'), in_proj_covar=tensor([0.0447, 0.0426, 0.0520, 0.0507, 0.0474, 0.0515, 0.0515, 0.0531], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 06:31:35,286 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8435, 1.6887, 1.8752, 2.1555, 2.3082, 1.8900, 1.6240, 2.1069], device='cuda:4'), covar=tensor([0.0709, 0.1181, 0.0699, 0.0526, 0.0498, 0.0713, 0.0681, 0.0439], device='cuda:4'), in_proj_covar=tensor([0.0183, 0.0204, 0.0184, 0.0173, 0.0179, 0.0180, 0.0151, 0.0177], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 06:31:42,968 INFO [finetune.py:976] (4/7) Epoch 30, batch 4200, loss[loss=0.1756, simple_loss=0.2412, pruned_loss=0.05501, over 4817.00 frames. ], tot_loss[loss=0.1651, simple_loss=0.24, pruned_loss=0.04505, over 950126.55 frames. ], batch size: 30, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:31:54,044 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.44 vs. limit=2.0 2023-04-28 06:31:58,825 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=170325.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:32:16,768 INFO [finetune.py:976] (4/7) Epoch 30, batch 4250, loss[loss=0.1261, simple_loss=0.2009, pruned_loss=0.02561, over 4742.00 frames. ], tot_loss[loss=0.1636, simple_loss=0.2379, pruned_loss=0.04466, over 950758.64 frames. ], batch size: 27, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:32:21,662 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3489, 1.7112, 1.7800, 1.9102, 1.8068, 1.8754, 1.9078, 1.7715], device='cuda:4'), covar=tensor([0.3060, 0.4415, 0.4083, 0.3820, 0.4757, 0.6284, 0.4309, 0.4428], device='cuda:4'), in_proj_covar=tensor([0.0345, 0.0377, 0.0334, 0.0345, 0.0354, 0.0397, 0.0364, 0.0337], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 06:32:36,243 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.843e+01 1.455e+02 1.729e+02 2.061e+02 4.081e+02, threshold=3.459e+02, percent-clipped=1.0 2023-04-28 06:32:38,780 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=170385.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:32:48,572 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=170401.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:32:48,775 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.51 vs. limit=2.0 2023-04-28 06:32:50,168 INFO [finetune.py:976] (4/7) Epoch 30, batch 4300, loss[loss=0.1629, simple_loss=0.2314, pruned_loss=0.04724, over 4746.00 frames. ], tot_loss[loss=0.1621, simple_loss=0.2361, pruned_loss=0.04405, over 953071.81 frames. ], batch size: 54, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:32:50,284 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=170403.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:33:02,344 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5400, 3.0538, 1.1032, 1.9078, 1.9083, 2.1890, 1.9487, 1.1047], device='cuda:4'), covar=tensor([0.1135, 0.0961, 0.1711, 0.1095, 0.0904, 0.0959, 0.1378, 0.1570], device='cuda:4'), in_proj_covar=tensor([0.0117, 0.0237, 0.0136, 0.0120, 0.0131, 0.0153, 0.0117, 0.0117], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 06:33:10,858 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=170433.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:33:16,823 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=170442.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:33:22,276 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=170451.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:33:23,431 INFO [finetune.py:976] (4/7) Epoch 30, batch 4350, loss[loss=0.1678, simple_loss=0.2534, pruned_loss=0.04111, over 4749.00 frames. ], tot_loss[loss=0.1599, simple_loss=0.2334, pruned_loss=0.04318, over 952711.76 frames. ], batch size: 54, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:33:28,872 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=170461.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:33:44,608 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.0918, 1.5587, 1.9129, 2.1786, 1.9485, 1.5633, 1.1092, 1.7301], device='cuda:4'), covar=tensor([0.2627, 0.2811, 0.1507, 0.1844, 0.2225, 0.2355, 0.4223, 0.1711], device='cuda:4'), in_proj_covar=tensor([0.0295, 0.0248, 0.0231, 0.0317, 0.0226, 0.0237, 0.0231, 0.0187], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], device='cuda:4') 2023-04-28 06:33:52,312 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.003e+02 1.526e+02 1.784e+02 2.388e+02 5.022e+02, threshold=3.568e+02, percent-clipped=3.0 2023-04-28 06:34:04,608 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=170490.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:34:23,442 INFO [finetune.py:976] (4/7) Epoch 30, batch 4400, loss[loss=0.16, simple_loss=0.2289, pruned_loss=0.04555, over 4883.00 frames. ], tot_loss[loss=0.1637, simple_loss=0.2368, pruned_loss=0.04527, over 950432.33 frames. ], batch size: 32, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:34:47,091 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=170522.0, num_to_drop=1, layers_to_drop={1} 2023-04-28 06:35:01,112 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.7169, 2.0645, 2.0162, 2.1957, 1.9538, 1.9473, 2.0399, 1.9438], device='cuda:4'), covar=tensor([0.4310, 0.5822, 0.5001, 0.4463, 0.6055, 0.7101, 0.5923, 0.5915], device='cuda:4'), in_proj_covar=tensor([0.0342, 0.0375, 0.0331, 0.0343, 0.0351, 0.0394, 0.0362, 0.0335], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 06:35:08,312 INFO [finetune.py:976] (4/7) Epoch 30, batch 4450, loss[loss=0.1453, simple_loss=0.2274, pruned_loss=0.03162, over 4804.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.2409, pruned_loss=0.04699, over 951556.48 frames. ], batch size: 25, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:35:19,131 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=170569.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:35:26,203 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.085e+02 1.572e+02 1.800e+02 2.072e+02 3.635e+02, threshold=3.600e+02, percent-clipped=1.0 2023-04-28 06:35:41,960 INFO [finetune.py:976] (4/7) Epoch 30, batch 4500, loss[loss=0.2055, simple_loss=0.272, pruned_loss=0.06945, over 4805.00 frames. ], tot_loss[loss=0.1682, simple_loss=0.242, pruned_loss=0.04721, over 951124.71 frames. ], batch size: 45, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:35:59,514 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=170630.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:36:15,352 INFO [finetune.py:976] (4/7) Epoch 30, batch 4550, loss[loss=0.1375, simple_loss=0.2217, pruned_loss=0.02668, over 4798.00 frames. ], tot_loss[loss=0.1687, simple_loss=0.2432, pruned_loss=0.04716, over 952375.76 frames. ], batch size: 55, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:36:33,303 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.995e+01 1.537e+02 1.792e+02 2.273e+02 3.619e+02, threshold=3.584e+02, percent-clipped=1.0 2023-04-28 06:36:35,258 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3488, 1.6883, 1.4370, 1.8614, 1.6519, 2.0378, 1.4935, 3.5847], device='cuda:4'), covar=tensor([0.0574, 0.0777, 0.0782, 0.1142, 0.0637, 0.0504, 0.0744, 0.0118], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0043, 0.0040, 0.0037, 0.0038, 0.0054], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0014, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 06:36:47,827 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=170701.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:36:48,972 INFO [finetune.py:976] (4/7) Epoch 30, batch 4600, loss[loss=0.1563, simple_loss=0.2389, pruned_loss=0.03688, over 4761.00 frames. ], tot_loss[loss=0.1665, simple_loss=0.2414, pruned_loss=0.04585, over 953011.75 frames. ], batch size: 28, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:37:19,721 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=170749.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:37:22,626 INFO [finetune.py:976] (4/7) Epoch 30, batch 4650, loss[loss=0.1315, simple_loss=0.2047, pruned_loss=0.02915, over 4061.00 frames. ], tot_loss[loss=0.1646, simple_loss=0.2386, pruned_loss=0.04532, over 952504.01 frames. ], batch size: 17, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:37:39,521 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4382, 3.0250, 2.6015, 2.9046, 2.3071, 2.6503, 2.8018, 2.1511], device='cuda:4'), covar=tensor([0.2018, 0.1245, 0.0773, 0.1242, 0.2930, 0.1208, 0.1789, 0.2730], device='cuda:4'), in_proj_covar=tensor([0.0282, 0.0300, 0.0217, 0.0273, 0.0308, 0.0255, 0.0248, 0.0261], device='cuda:4'), out_proj_covar=tensor([1.1219e-04, 1.1770e-04, 8.4846e-05, 1.0687e-04, 1.2391e-04, 1.0031e-04, 9.9534e-05, 1.0252e-04], device='cuda:4') 2023-04-28 06:37:39,986 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.849e+01 1.520e+02 1.798e+02 2.217e+02 4.405e+02, threshold=3.597e+02, percent-clipped=2.0 2023-04-28 06:37:53,809 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.2178, 1.6731, 1.4798, 1.8734, 1.6796, 1.9782, 1.4780, 3.6005], device='cuda:4'), covar=tensor([0.0608, 0.0768, 0.0735, 0.1124, 0.0631, 0.0548, 0.0698, 0.0144], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0039, 0.0039, 0.0043, 0.0040, 0.0037, 0.0038, 0.0054], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0014, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 06:37:55,525 INFO [finetune.py:976] (4/7) Epoch 30, batch 4700, loss[loss=0.1738, simple_loss=0.2373, pruned_loss=0.05512, over 4859.00 frames. ], tot_loss[loss=0.1615, simple_loss=0.2347, pruned_loss=0.04415, over 953362.19 frames. ], batch size: 44, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:38:05,047 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=170817.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 06:38:29,293 INFO [finetune.py:976] (4/7) Epoch 30, batch 4750, loss[loss=0.1616, simple_loss=0.2353, pruned_loss=0.0439, over 4785.00 frames. ], tot_loss[loss=0.1609, simple_loss=0.2337, pruned_loss=0.04409, over 954589.88 frames. ], batch size: 29, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:38:47,640 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.061e+02 1.520e+02 1.829e+02 2.099e+02 3.984e+02, threshold=3.658e+02, percent-clipped=1.0 2023-04-28 06:38:59,653 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.3657, 2.8899, 1.1124, 1.6776, 2.3762, 1.4591, 4.0353, 2.0693], device='cuda:4'), covar=tensor([0.0675, 0.0883, 0.0848, 0.1225, 0.0480, 0.1006, 0.0230, 0.0570], device='cuda:4'), in_proj_covar=tensor([0.0051, 0.0064, 0.0046, 0.0046, 0.0049, 0.0051, 0.0072, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-28 06:39:16,798 INFO [finetune.py:976] (4/7) Epoch 30, batch 4800, loss[loss=0.1913, simple_loss=0.2666, pruned_loss=0.05799, over 4853.00 frames. ], tot_loss[loss=0.1637, simple_loss=0.2371, pruned_loss=0.04521, over 953415.76 frames. ], batch size: 49, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:39:45,900 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=170925.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:39:47,792 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=170928.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:39:56,363 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.4882, 1.9564, 1.7796, 2.2161, 1.9884, 2.1572, 1.7422, 4.6978], device='cuda:4'), covar=tensor([0.0538, 0.0746, 0.0769, 0.1160, 0.0632, 0.0494, 0.0724, 0.0103], device='cuda:4'), in_proj_covar=tensor([0.0037, 0.0038, 0.0039, 0.0042, 0.0040, 0.0037, 0.0038, 0.0054], device='cuda:4'), out_proj_covar=tensor([0.0012, 0.0013, 0.0014, 0.0015, 0.0014, 0.0013, 0.0013, 0.0016], device='cuda:4') 2023-04-28 06:40:18,964 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9751, 1.8237, 3.9613, 3.7750, 3.5497, 3.7308, 3.5564, 3.5380], device='cuda:4'), covar=tensor([0.6625, 0.4775, 0.1080, 0.1609, 0.1115, 0.1855, 0.4110, 0.1470], device='cuda:4'), in_proj_covar=tensor([0.0313, 0.0310, 0.0408, 0.0412, 0.0349, 0.0418, 0.0318, 0.0365], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001], device='cuda:4') 2023-04-28 06:40:21,166 INFO [finetune.py:976] (4/7) Epoch 30, batch 4850, loss[loss=0.205, simple_loss=0.2764, pruned_loss=0.06674, over 4832.00 frames. ], tot_loss[loss=0.1664, simple_loss=0.2407, pruned_loss=0.04607, over 953795.76 frames. ], batch size: 39, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:40:52,179 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 8.978e+01 1.534e+02 1.869e+02 2.233e+02 4.849e+02, threshold=3.739e+02, percent-clipped=1.0 2023-04-28 06:40:59,527 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=170989.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:41:19,848 INFO [finetune.py:976] (4/7) Epoch 30, batch 4900, loss[loss=0.1409, simple_loss=0.213, pruned_loss=0.03444, over 4735.00 frames. ], tot_loss[loss=0.1675, simple_loss=0.2417, pruned_loss=0.04669, over 952512.92 frames. ], batch size: 59, lr: 2.79e-03, grad_scale: 32.0 2023-04-28 06:41:42,005 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.36 vs. limit=2.0 2023-04-28 06:42:25,541 INFO [finetune.py:976] (4/7) Epoch 30, batch 4950, loss[loss=0.1766, simple_loss=0.2529, pruned_loss=0.05019, over 4895.00 frames. ], tot_loss[loss=0.1666, simple_loss=0.241, pruned_loss=0.04607, over 952362.05 frames. ], batch size: 43, lr: 2.78e-03, grad_scale: 64.0 2023-04-28 06:42:32,641 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.9750, 2.2933, 0.9692, 1.3733, 1.6628, 1.1487, 2.9928, 1.6920], device='cuda:4'), covar=tensor([0.0692, 0.0549, 0.0732, 0.1260, 0.0504, 0.1036, 0.0332, 0.0617], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0046, 0.0046, 0.0049, 0.0051, 0.0072, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-28 06:42:45,067 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.189e+02 1.630e+02 1.843e+02 2.166e+02 3.655e+02, threshold=3.685e+02, percent-clipped=0.0 2023-04-28 06:42:47,395 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.54 vs. limit=5.0 2023-04-28 06:42:52,360 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=171092.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:42:59,424 INFO [finetune.py:976] (4/7) Epoch 30, batch 5000, loss[loss=0.1778, simple_loss=0.2448, pruned_loss=0.0554, over 4731.00 frames. ], tot_loss[loss=0.1651, simple_loss=0.2391, pruned_loss=0.04555, over 952187.74 frames. ], batch size: 54, lr: 2.78e-03, grad_scale: 64.0 2023-04-28 06:43:00,148 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.5812, 1.6505, 0.7484, 1.3233, 1.6770, 1.4586, 1.3761, 1.4525], device='cuda:4'), covar=tensor([0.0514, 0.0366, 0.0340, 0.0583, 0.0278, 0.0517, 0.0476, 0.0566], device='cuda:4'), in_proj_covar=tensor([0.0028, 0.0023, 0.0021, 0.0028, 0.0019, 0.0027, 0.0027, 0.0028], device='cuda:4'), out_proj_covar=tensor([0.0052, 0.0047, 0.0039, 0.0054, 0.0039, 0.0051, 0.0051, 0.0053], device='cuda:4') 2023-04-28 06:43:09,446 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=171117.0, num_to_drop=1, layers_to_drop={2} 2023-04-28 06:43:13,833 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.29 vs. limit=2.0 2023-04-28 06:43:18,352 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8635, 1.5729, 1.9847, 2.0815, 1.6197, 1.3820, 1.6749, 1.0773], device='cuda:4'), covar=tensor([0.0400, 0.0715, 0.0345, 0.0586, 0.0625, 0.1082, 0.0562, 0.0536], device='cuda:4'), in_proj_covar=tensor([0.0071, 0.0067, 0.0065, 0.0069, 0.0076, 0.0095, 0.0072, 0.0062], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0003, 0.0003, 0.0003, 0.0003, 0.0004, 0.0003, 0.0003], device='cuda:4') 2023-04-28 06:43:33,016 INFO [finetune.py:976] (4/7) Epoch 30, batch 5050, loss[loss=0.1636, simple_loss=0.2314, pruned_loss=0.04795, over 4752.00 frames. ], tot_loss[loss=0.1632, simple_loss=0.2368, pruned_loss=0.04477, over 953076.97 frames. ], batch size: 27, lr: 2.78e-03, grad_scale: 64.0 2023-04-28 06:43:33,145 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=171153.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:43:34,295 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.2175, 2.6563, 1.1127, 1.5407, 1.9374, 1.2805, 3.5706, 2.0455], device='cuda:4'), covar=tensor([0.0647, 0.0697, 0.0794, 0.1183, 0.0506, 0.1025, 0.0285, 0.0550], device='cuda:4'), in_proj_covar=tensor([0.0050, 0.0064, 0.0046, 0.0046, 0.0049, 0.0051, 0.0071, 0.0050], device='cuda:4'), out_proj_covar=tensor([0.0008, 0.0010, 0.0007, 0.0008, 0.0008, 0.0008, 0.0010, 0.0008], device='cuda:4') 2023-04-28 06:43:41,327 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=171165.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:43:52,460 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.935e+01 1.498e+02 1.792e+02 2.110e+02 3.798e+02, threshold=3.585e+02, percent-clipped=1.0 2023-04-28 06:44:02,204 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=171196.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:44:06,504 INFO [finetune.py:976] (4/7) Epoch 30, batch 5100, loss[loss=0.1588, simple_loss=0.23, pruned_loss=0.04383, over 4873.00 frames. ], tot_loss[loss=0.1608, simple_loss=0.2337, pruned_loss=0.0439, over 952296.29 frames. ], batch size: 34, lr: 2.78e-03, grad_scale: 64.0 2023-04-28 06:44:21,437 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=171225.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:44:40,214 INFO [finetune.py:976] (4/7) Epoch 30, batch 5150, loss[loss=0.1425, simple_loss=0.2224, pruned_loss=0.03129, over 4763.00 frames. ], tot_loss[loss=0.1608, simple_loss=0.2338, pruned_loss=0.04386, over 950986.76 frames. ], batch size: 28, lr: 2.78e-03, grad_scale: 64.0 2023-04-28 06:44:43,277 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=171257.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:44:44,179 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.12 vs. limit=2.0 2023-04-28 06:44:53,818 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=171273.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:45:05,608 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.060e+02 1.555e+02 1.961e+02 2.367e+02 3.631e+02, threshold=3.922e+02, percent-clipped=1.0 2023-04-28 06:45:07,337 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=171284.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:45:35,546 INFO [finetune.py:976] (4/7) Epoch 30, batch 5200, loss[loss=0.2219, simple_loss=0.2894, pruned_loss=0.07719, over 4817.00 frames. ], tot_loss[loss=0.1637, simple_loss=0.2376, pruned_loss=0.04495, over 949197.56 frames. ], batch size: 38, lr: 2.78e-03, grad_scale: 32.0 2023-04-28 06:45:39,245 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.5007, 1.8183, 1.9452, 2.0815, 1.9489, 1.9615, 1.9929, 2.0006], device='cuda:4'), covar=tensor([0.3716, 0.5224, 0.4258, 0.4050, 0.5168, 0.7070, 0.4988, 0.4501], device='cuda:4'), in_proj_covar=tensor([0.0340, 0.0374, 0.0331, 0.0343, 0.0351, 0.0394, 0.0361, 0.0334], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 06:46:20,592 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.9668, 2.2533, 2.2256, 2.3735, 2.0842, 2.2422, 2.2508, 2.2363], device='cuda:4'), covar=tensor([0.3757, 0.5836, 0.4981, 0.4329, 0.5617, 0.6752, 0.5726, 0.5048], device='cuda:4'), in_proj_covar=tensor([0.0341, 0.0374, 0.0331, 0.0344, 0.0352, 0.0394, 0.0362, 0.0334], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 06:46:29,956 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([4.0552, 3.9956, 2.9568, 4.6598, 4.1468, 4.0138, 1.7798, 4.0549], device='cuda:4'), covar=tensor([0.1760, 0.1200, 0.3053, 0.1424, 0.3095, 0.2029, 0.5973, 0.2331], device='cuda:4'), in_proj_covar=tensor([0.0250, 0.0220, 0.0253, 0.0302, 0.0301, 0.0250, 0.0275, 0.0275], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 06:46:31,222 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=171345.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:46:36,593 INFO [finetune.py:976] (4/7) Epoch 30, batch 5250, loss[loss=0.1609, simple_loss=0.2341, pruned_loss=0.04381, over 4380.00 frames. ], tot_loss[loss=0.1638, simple_loss=0.2382, pruned_loss=0.04469, over 950212.59 frames. ], batch size: 19, lr: 2.78e-03, grad_scale: 32.0 2023-04-28 06:46:49,539 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.1043, 1.8594, 2.2850, 2.6422, 2.1317, 1.9898, 2.1190, 2.1378], device='cuda:4'), covar=tensor([0.5124, 0.7809, 0.7664, 0.5905, 0.6707, 0.9914, 0.9885, 0.9636], device='cuda:4'), in_proj_covar=tensor([0.0449, 0.0428, 0.0521, 0.0510, 0.0475, 0.0518, 0.0517, 0.0533], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 06:46:59,737 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([3.2991, 3.3264, 2.5667, 3.8494, 3.3286, 3.2800, 1.3894, 3.2670], device='cuda:4'), covar=tensor([0.2133, 0.1506, 0.3337, 0.2181, 0.3607, 0.2119, 0.6261, 0.2871], device='cuda:4'), in_proj_covar=tensor([0.0250, 0.0220, 0.0254, 0.0303, 0.0301, 0.0251, 0.0276, 0.0276], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 06:47:18,734 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.056e+02 1.554e+02 1.778e+02 2.227e+02 5.005e+02, threshold=3.556e+02, percent-clipped=1.0 2023-04-28 06:47:43,263 INFO [finetune.py:976] (4/7) Epoch 30, batch 5300, loss[loss=0.1671, simple_loss=0.2534, pruned_loss=0.0404, over 4907.00 frames. ], tot_loss[loss=0.165, simple_loss=0.2398, pruned_loss=0.04503, over 949542.65 frames. ], batch size: 36, lr: 2.78e-03, grad_scale: 32.0 2023-04-28 06:47:50,505 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=171406.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:48:41,620 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=171448.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:48:44,661 INFO [finetune.py:976] (4/7) Epoch 30, batch 5350, loss[loss=0.1605, simple_loss=0.233, pruned_loss=0.04403, over 4800.00 frames. ], tot_loss[loss=0.1643, simple_loss=0.2393, pruned_loss=0.04467, over 949327.97 frames. ], batch size: 25, lr: 2.78e-03, grad_scale: 32.0 2023-04-28 06:49:03,278 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.027e+02 1.512e+02 1.839e+02 2.201e+02 3.854e+02, threshold=3.679e+02, percent-clipped=1.0 2023-04-28 06:49:15,613 INFO [zipformer.py:1188] (4/7) warmup_begin=2666.7, warmup_end=3333.3, batch_count=171498.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:49:18,573 INFO [finetune.py:976] (4/7) Epoch 30, batch 5400, loss[loss=0.1382, simple_loss=0.2129, pruned_loss=0.03172, over 4810.00 frames. ], tot_loss[loss=0.1639, simple_loss=0.2383, pruned_loss=0.04476, over 948938.58 frames. ], batch size: 39, lr: 2.78e-03, grad_scale: 32.0 2023-04-28 06:49:31,068 INFO [scaling.py:679] (4/7) Whitening: num_groups=1, num_channels=384, metric=4.50 vs. limit=5.0 2023-04-28 06:49:49,501 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([2.4394, 1.9273, 1.8799, 2.0236, 1.9148, 1.9941, 1.9529, 1.9196], device='cuda:4'), covar=tensor([0.3874, 0.4385, 0.3914, 0.3442, 0.4983, 0.5982, 0.4166, 0.4451], device='cuda:4'), in_proj_covar=tensor([0.0343, 0.0375, 0.0331, 0.0345, 0.0352, 0.0394, 0.0362, 0.0335], device='cuda:4'), out_proj_covar=tensor([0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002], device='cuda:4') 2023-04-28 06:49:51,783 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=171552.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:49:52,349 INFO [finetune.py:976] (4/7) Epoch 30, batch 5450, loss[loss=0.1338, simple_loss=0.2204, pruned_loss=0.02359, over 4769.00 frames. ], tot_loss[loss=0.1609, simple_loss=0.2345, pruned_loss=0.04358, over 950319.73 frames. ], batch size: 27, lr: 2.78e-03, grad_scale: 32.0 2023-04-28 06:49:56,145 INFO [zipformer.py:1188] (4/7) warmup_begin=3333.3, warmup_end=4000.0, batch_count=171559.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:50:07,768 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.38 vs. limit=2.0 2023-04-28 06:50:10,994 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 9.901e+01 1.444e+02 1.753e+02 2.019e+02 4.846e+02, threshold=3.505e+02, percent-clipped=2.0 2023-04-28 06:50:12,288 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=171584.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:50:25,850 INFO [finetune.py:976] (4/7) Epoch 30, batch 5500, loss[loss=0.1649, simple_loss=0.23, pruned_loss=0.04989, over 4905.00 frames. ], tot_loss[loss=0.1595, simple_loss=0.2322, pruned_loss=0.04338, over 951483.05 frames. ], batch size: 32, lr: 2.78e-03, grad_scale: 32.0 2023-04-28 06:50:27,785 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.6338, 1.9863, 1.7084, 2.0326, 1.5843, 1.8410, 1.7460, 1.3173], device='cuda:4'), covar=tensor([0.1882, 0.1584, 0.0983, 0.1226, 0.3810, 0.1185, 0.2016, 0.2867], device='cuda:4'), in_proj_covar=tensor([0.0283, 0.0301, 0.0218, 0.0274, 0.0309, 0.0256, 0.0249, 0.0261], device='cuda:4'), out_proj_covar=tensor([1.1284e-04, 1.1813e-04, 8.5450e-05, 1.0724e-04, 1.2437e-04, 1.0052e-04, 9.9906e-05, 1.0255e-04], device='cuda:4') 2023-04-28 06:50:32,582 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.3083, 3.0611, 0.8880, 1.7175, 1.7243, 2.1210, 1.8267, 1.0510], device='cuda:4'), covar=tensor([0.1397, 0.0956, 0.1706, 0.1147, 0.1037, 0.0969, 0.1456, 0.1583], device='cuda:4'), in_proj_covar=tensor([0.0116, 0.0235, 0.0135, 0.0120, 0.0130, 0.0151, 0.0116, 0.0117], device='cuda:4'), out_proj_covar=tensor([0.0003, 0.0005, 0.0003, 0.0002, 0.0003, 0.0003, 0.0003, 0.0003], device='cuda:4') 2023-04-28 06:50:32,650 INFO [zipformer.py:2441] (4/7) attn_weights_entropy = tensor([1.8244, 1.1467, 1.7673, 2.2851, 1.8869, 1.7298, 1.7742, 1.7409], device='cuda:4'), covar=tensor([0.4314, 0.6476, 0.6083, 0.5321, 0.5408, 0.7444, 0.7006, 0.8144], device='cuda:4'), in_proj_covar=tensor([0.0448, 0.0426, 0.0519, 0.0508, 0.0474, 0.0515, 0.0515, 0.0531], device='cuda:4'), out_proj_covar=tensor([0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001], device='cuda:4') 2023-04-28 06:50:45,014 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=171632.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:50:59,646 INFO [finetune.py:976] (4/7) Epoch 30, batch 5550, loss[loss=0.187, simple_loss=0.2649, pruned_loss=0.05459, over 4742.00 frames. ], tot_loss[loss=0.1612, simple_loss=0.2342, pruned_loss=0.04412, over 951870.96 frames. ], batch size: 54, lr: 2.78e-03, grad_scale: 32.0 2023-04-28 06:51:38,775 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.059e+02 1.632e+02 1.883e+02 2.258e+02 4.653e+02, threshold=3.766e+02, percent-clipped=3.0 2023-04-28 06:52:00,706 INFO [zipformer.py:1188] (4/7) warmup_begin=1333.3, warmup_end=2000.0, batch_count=171701.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:52:01,823 INFO [finetune.py:976] (4/7) Epoch 30, batch 5600, loss[loss=0.1964, simple_loss=0.2839, pruned_loss=0.0545, over 4724.00 frames. ], tot_loss[loss=0.163, simple_loss=0.237, pruned_loss=0.04444, over 949384.51 frames. ], batch size: 59, lr: 2.78e-03, grad_scale: 32.0 2023-04-28 06:52:55,591 INFO [zipformer.py:1188] (4/7) warmup_begin=2000.0, warmup_end=2666.7, batch_count=171748.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:52:58,356 INFO [finetune.py:976] (4/7) Epoch 30, batch 5650, loss[loss=0.2196, simple_loss=0.2817, pruned_loss=0.07878, over 4899.00 frames. ], tot_loss[loss=0.1651, simple_loss=0.2399, pruned_loss=0.04521, over 949853.38 frames. ], batch size: 35, lr: 2.78e-03, grad_scale: 32.0 2023-04-28 06:53:37,490 INFO [optim.py:369] (4/7) Clipping_scale=2.0, grad-norm quartiles 1.071e+02 1.430e+02 1.735e+02 1.948e+02 3.086e+02, threshold=3.470e+02, percent-clipped=0.0 2023-04-28 06:53:47,105 INFO [scaling.py:679] (4/7) Whitening: num_groups=8, num_channels=96, metric=1.22 vs. limit=2.0 2023-04-28 06:53:50,970 INFO [zipformer.py:1188] (4/7) warmup_begin=666.7, warmup_end=1333.3, batch_count=171796.0, num_to_drop=0, layers_to_drop=set() 2023-04-28 06:54:00,718 INFO [finetune.py:976] (4/7) Epoch 30, batch 5700, loss[loss=0.1368, simple_loss=0.2085, pruned_loss=0.03258, over 3526.00 frames. ], tot_loss[loss=0.1624, simple_loss=0.2361, pruned_loss=0.0443, over 934006.97 frames. ], batch size: 15, lr: 2.78e-03, grad_scale: 32.0 2023-04-28 06:54:33,222 INFO [finetune.py:1241] (4/7) Done!