whisper-dpv-finetuned-WITH-AUGMENTATION-LOWER-LR-WEIGHT-DECAY

This model is a fine-tuned version of openai/whisper-medium on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.8435
• Wer: 35.0215

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 4
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 100
• num_epochs: 4
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.5839	0.62	1000	0.5726	37.4633
0.2068	1.25	2000	0.5799	36.4911
0.1451	1.87	3000	0.6284	36.0389
0.0606	2.49	4000	0.7208	36.4006
0.0081	3.12	5000	0.8024	34.9537
0.0131	3.74	6000	0.8435	35.0215

Framework versions

• Transformers 4.26.0.dev0
• Pytorch 1.12.1
• Datasets 2.7.1
• Tokenizers 0.13.2