11-785 HW4P2: Attention-based Speech Recognition
Please modify configs/path/server.yaml
for your own path
You can see Hyperparameters values in configs/
.
Or You can see the details and experimental results in WanDB.
- best model: cnn_clipping-07:16:09:19
- Model Architecture
- You can check in
src/models/seq2seq.py
- Based on LAS architecture
- Added cnn embedding in encoder
- You can check in
- Optimizer
- Adam
- scheduler: CosineAnnealing
- Regularize
- Locked Dropout for LSTM and pbLSTM layers
- Dropout for embedding cnn
- gradient clipping
- teacher forcing and scheduling along with the learning rate
- Data & Augmentation
- Cepstral Mean Normalize
$ python run.py save_name={name_for_submission&weight_file}