wav2vec2-xls-r-300m_Mrbrown_finetune1
This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the uob_singlish dataset.
This time use self-made dataset(cut the audio of "https://www.youtube.com/watch?v=a2ZOTD3R7JI" into slices and write the corresponding transcript, totally 4 mins), don't know why the word-error-rate keep 1. But can know that much be the problem of dataset, because last time use the same pre-trained model and standard singlish corpus fine-tune get nice result. (can find it at:RuiqianLi/wav2vec2-large-xls-r-300m-singlish-colab)
It achieves the following results on the evaluation set:
- Loss: 3.0927
- Wer: 1.0
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: 0.01
- train_batch_size: 2
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 2
- total_train_batch_size: 4
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
- num_epochs: 100
- mixed_precision_training: Native AMP
Training results
Training Loss | Epoch | Step | Validation Loss | Wer |
---|---|---|---|---|
3.7943 | 20.0 | 200 | 3.0597 | 1.0 |
2.9902 | 40.0 | 400 | 3.1604 | 1.0 |
2.9696 | 60.0 | 600 | 3.1112 | 1.0 |
2.8885 | 80.0 | 800 | 3.0234 | 1.0 |
2.8154 | 100.0 | 1000 | 3.0927 | 1.0 |
Framework versions
- Transformers 4.11.3
- Pytorch 1.10.0+cu113
- Datasets 1.18.3
- Tokenizers 0.10.3
- Downloads last month
- 3
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social
visibility and check back later, or deploy to Inference Endpoints (dedicated)
instead.