| python -m torch.distributed.launch \ | |
| --nproc_per_node=2 \ | |
| run_xtreme_s.py \ | |
| --model_name_or_path="facebook/wav2vec2-xls-r-300m" \ | |
| --task="minds14" \ | |
| --language="all" \ | |
| --output_dir="xtreme_s_xlsr_300m_minds14" \ | |
| --overwrite_output_dir \ | |
| --num_train_epochs=50 \ | |
| --per_device_train_batch_size=32 \ | |
| --per_device_eval_batch_size=8 \ | |
| --gradient_accumulation_steps=1 \ | |
| --learning_rate="3e-4" \ | |
| --warmup_steps=1500 \ | |
| --evaluation_strategy="steps" \ | |
| --max_duration_in_seconds=30 \ | |
| --save_steps=200 \ | |
| --eval_steps=200 \ | |
| --logging_steps=1 \ | |
| --layerdrop=0.0 \ | |
| --mask_time_prob=0.3 \ | |
| --mask_time_length=10 \ | |
| --mask_feature_prob=0.1 \ | |
| --mask_feature_length=64 \ | |
| --freeze_feature_encoder \ | |
| --gradient_checkpointing \ | |
| --fp16 \ | |
| --group_by_length \ | |
| --do_train \ | |
| --do_eval \ | |
| --do_predict \ | |
| --metric_for_best_model="f1" \ | |
| --greater_is_better=True \ | |
| --load_best_model_at_end \ | |
| --push_to_hub |