|
--- |
|
license: apache-2.0 |
|
tags: |
|
- generated_from_trainer |
|
datasets: |
|
- common_voice |
|
model-index: |
|
- name: wav2vec2-large-xlsr-53-demo-colab |
|
results: [] |
|
--- |
|
|
|
<!-- 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-53-demo-colab |
|
|
|
This model is a fine-tuned version of [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) on the common_voice dataset. |
|
It achieves the following results on the evaluation set: |
|
- Loss: 0.6901 |
|
- Wer: 1.6299 |
|
|
|
## 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 | |
|
|:-------------:|:-----:|:----:|:---------------:|:------:| |
|
| 8.5034 | 3.42 | 400 | 3.5852 | 1.0 | |
|
| 1.7853 | 6.83 | 800 | 0.7430 | 1.6774 | |
|
| 0.5675 | 10.26 | 1200 | 0.6513 | 1.6330 | |
|
| 0.3761 | 13.67 | 1600 | 0.6208 | 1.6081 | |
|
| 0.2776 | 17.09 | 2000 | 0.6401 | 1.6081 | |
|
| 0.2266 | 20.51 | 2400 | 0.6410 | 1.6295 | |
|
| 0.1949 | 23.93 | 2800 | 0.6910 | 1.6287 | |
|
| 0.1672 | 27.35 | 3200 | 0.6901 | 1.6299 | |
|
|
|
|
|
### Framework versions |
|
|
|
- Transformers 4.11.3 |
|
- Pytorch 1.10.0+cu111 |
|
- Datasets 1.14.0 |
|
- Tokenizers 0.10.3 |
|
|
