jpqd-wav2vec2-base-ft-keyword-spotting
This model is a fine-tuned version of facebook/wav2vec2-base on the superb dataset, using superb/wav2vec2-base-superb-ks as a teacher model
It was compressed using NNCF with Optimum Intel following the JPQD image classification example.
It achieves the following results on the evaluation set:
- Loss: 0.5632
- Accuracy: 0.9756
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: 7e-05
- train_batch_size: 32
- eval_batch_size: 64
- seed: 42
- gradient_accumulation_steps: 4
- total_train_batch_size: 128
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_ratio: 0.5
- num_epochs: 12.0
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Accuracy |
|---|---|---|---|---|
| 2.2245 | 1.0 | 399 | 2.2351 | 0.6209 |
| 6.9856 | 2.0 | 798 | 7.0597 | 0.7354 |
| 10.013 | 3.0 | 1197 | 9.8779 | 0.8069 |
| 11.3484 | 4.0 | 1596 | 11.1949 | 0.8719 |
| 11.6849 | 5.0 | 1995 | 11.5479 | 0.9014 |
| 11.5921 | 6.0 | 2394 | 11.4193 | 0.9495 |
| 0.8911 | 7.0 | 2793 | 0.7334 | 0.9500 |
| 0.8965 | 8.0 | 3192 | 0.6553 | 0.9685 |
| 0.7198 | 9.0 | 3591 | 0.6213 | 0.9669 |
| 0.7372 | 10.0 | 3990 | 0.5929 | 0.9675 |
| 0.7004 | 11.0 | 4389 | 0.5720 | 0.9721 |
| 0.6195 | 12.0 | 4788 | 0.5632 | 0.9756 |
Framework versions
- Transformers 4.26.1
- Pytorch 1.13.1+cu117
- Datasets 2.8.0
- Tokenizers 0.13.2