normalist-k/hw4p2

Instructions for HW4P2 (Youngin Kim, youngin2)

11-785 HW4P2: Attention-based Speech Recognition

0. Prerquisite

Please modify configs/path/server.yaml for your own path

1. Best Architecture

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
• 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

2. Run

$ python run.py save_name={name_for_submission&weight_file}