dgx2_whisper_base_distil_att_loss_mozilla_epochs_40_batch_16_concat_dataset

This model is a fine-tuned version of rohitp1/dgx1_whisper_base_finetune_teacher_babble_noise_mozilla_100_epochs_batch_16 on the None dataset. It achieves the following results on the evaluation set:

• Loss: 2.1315
• Wer: 30.6873

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: 5e-05
• train_batch_size: 16
• eval_batch_size: 1
• seed: 42
• gradient_accumulation_steps: 128
• total_train_batch_size: 2048
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine_with_restarts
• lr_scheduler_warmup_ratio: 0.2
• num_epochs: 40
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
2.182	2.2	150	2.1102	31.0198
2.0124	4.41	300	2.0539	30.7134
1.9833	6.61	450	2.0528	30.7099
1.9666	8.82	600	2.0768	30.6490
1.9756	11.03	750	2.0959	30.6699
1.9659	13.23	900	2.1110	30.6612
1.9583	15.44	1050	2.1150	30.6403
1.9524	17.64	1200	2.1191	30.6507
1.9461	19.85	1350	2.1222	30.6716
1.9448	22.06	1500	2.1275	30.6890
1.9393	24.26	1650	2.1297	30.6646
1.935	26.47	1800	2.1315	30.6873

Framework versions

• Transformers 4.25.1
• Pytorch 1.12.1
• Datasets 2.8.0
• Tokenizers 0.13.2