<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->
wav2vec2-large-xlsr-finetune-es-col
This model is a fine-tuned version of facebook/wav2vec2-large-xlsr-53 on the None dataset. It achieves the following results on the evaluation set:
- Loss: 2.6514
- Wer: 0.9874
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: 0.0003
- train_batch_size: 16
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 2
- total_train_batch_size: 32
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
- num_epochs: 30
- mixed_precision_training: Native AMP
Training results
Training Loss | Epoch | Step | Validation Loss | Wer |
---|---|---|---|---|
4.9709 | 3.25 | 400 | 2.9673 | 1.0 |
2.9488 | 6.5 | 800 | 2.9075 | 0.9973 |
2.907 | 9.76 | 1200 | 2.8772 | 0.9688 |
2.886 | 13.01 | 1600 | 2.8245 | 0.9484 |
2.8043 | 16.26 | 2000 | 2.7134 | 0.9874 |
2.7288 | 19.51 | 2400 | 2.6750 | 0.9874 |
2.7072 | 22.76 | 2800 | 2.6651 | 0.9874 |
2.6892 | 26.02 | 3200 | 2.6573 | 0.9874 |
2.683 | 29.27 | 3600 | 2.6514 | 0.9874 |
Framework versions
- Transformers 4.11.3
- Pytorch 1.10.1+cu102
- Datasets 1.13.3
- Tokenizers 0.10.3