hubert-base-libri-demo-feature_extractor_not_frozen_v1_30epochs_weight_decay

This model is a fine-tuned version of facebook/hubert-base-ls960 on the None dataset. It achieves the following results on the evaluation set:

• Loss: 3.8346
• Wer: 1.0000

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: 0.00015
• train_batch_size: 64
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 3000
• num_epochs: 30
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
4.3231	1.12	500	3.4739	1.0000
2.879	2.24	1000	3.5811	1.0000
2.8709	3.36	1500	3.9899	1.0000
2.871	4.48	2000	4.0805	1.0000
2.8706	5.61	2500	3.8464	1.0000
2.8672	6.73	3000	3.7938	1.0000
2.8709	7.85	3500	3.8578	1.0000
2.8708	8.97	4000	3.7691	1.0000
2.8678	10.09	4500	3.8619	1.0000
2.8662	11.21	5000	3.8804	1.0000
2.8664	12.33	5500	3.8169	1.0000
2.8662	13.45	6000	3.6758	1.0000
2.8654	14.57	6500	3.7314	1.0000
2.8658	15.7	7000	3.8113	1.0000
2.8647	16.82	7500	3.8938	1.0000
2.8653	17.94	8000	3.9268	1.0000
2.8652	19.06	8500	3.9288	1.0000
2.8649	20.18	9000	3.9164	1.0000
2.8654	21.3	9500	3.8781	1.0000
2.8652	22.42	10000	3.8628	1.0000
2.8642	23.54	10500	3.8646	1.0000
2.8655	24.66	11000	3.8467	1.0000
2.8658	25.78	11500	3.8157	1.0000
2.8647	26.91	12000	3.8274	1.0000
2.8648	28.03	12500	3.8022	1.0000
2.8644	29.15	13000	3.8346	1.0000

Framework versions

• Transformers 4.30.0.dev0
• Pytorch 2.0.1
• Datasets 2.12.1.dev0
• Tokenizers 0.13.3