hubert-base-libri-demo-feature_extractor_not_frozen_v2

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: 0.1152
• Wer: 0.1118

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: 25
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
4.3761	1.12	500	3.4742	1.0
2.882	2.24	1000	3.6180	1.0
2.8637	3.36	1500	3.0941	1.0
1.2198	4.48	2000	0.3003	0.3313
0.3296	5.61	2500	0.1670	0.2085
0.2195	6.73	3000	0.1393	0.1615
0.1832	7.85	3500	0.1284	0.1445
0.166	8.97	4000	0.1227	0.1362
0.1491	10.09	4500	0.1201	0.1305
0.1157	11.21	5000	0.1141	0.1262
0.1175	12.33	5500	0.1311	0.1250
0.0939	13.45	6000	0.1227	0.1205
0.0919	14.57	6500	0.1234	0.1205
0.0871	15.7	7000	0.1141	0.1187
0.0792	16.82	7500	0.1154	0.1171
0.0746	17.94	8000	0.1118	0.1157
0.074	19.06	8500	0.1159	0.1147
0.077	20.18	9000	0.1172	0.1137
0.0662	21.3	9500	0.1158	0.1126
0.0652	22.42	10000	0.1146	0.1117
0.06	23.54	10500	0.1159	0.1117
0.0576	24.66	11000	0.1152	0.1118

Framework versions

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