wav2vec2-conformer-rope-large-960h-ft-speech_commands
This model is a fine-tuned version of facebook/wav2vec2-conformer-rope-large-960h-ft on the None dataset. It achieves the following results on the evaluation set:
- Loss: 0.7777
- Accuracy: 0.9727
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: 3e-05
- train_batch_size: 128
- eval_batch_size: 128
- seed: 42
- gradient_accumulation_steps: 4
- total_train_batch_size: 512
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_ratio: 0.1
- num_epochs: 5
Training results
| Training Loss | Epoch | Step | Validation Loss | Accuracy |
|---|---|---|---|---|
| 2.6618 | 1.0 | 165 | 2.3731 | 0.8350 |
| 1.7932 | 2.0 | 330 | 1.4893 | 0.9278 |
| 1.3897 | 3.0 | 495 | 1.0531 | 0.9661 |
| 1.1663 | 4.0 | 660 | 0.8407 | 0.9714 |
| 1.1051 | 5.0 | 825 | 0.7777 | 0.9727 |
Framework versions
- Transformers 4.21.1
- Pytorch 1.12.1
- Datasets 2.4.0
- Tokenizers 0.12.1