rnnlm_libri.yaml

rnn_lm:
  optimizer: 
    type: 'Adam'                              # Optimizer used for training (adam)
    learning_rate: 0.0001                     # Learning rate for opt
  model_para:
    emb_dim: 650
    h_dim: 650
    layers: 2
    rnn: 'LSTM'
    dropout_rate: 0.5

solver:
  # Data options
  dataset: 'librispeech'                      # 
  data_path: 'data/libri_fbank80_subword5000' # Source data path
  n_jobs: 4                                   # Subprocess used for torch Dataloader
  max_label_len: 400                          # Max length for output sequence (0 for no restriction)
  # Training options
  train_set: ['train-clean-100','train-clean-360','train-other-500']              #
  batch_size: 32                              # training batch size
  apex: True                                  # Use APEX (see https://github.com/NVIDIA/apex for more details)
  total_steps: 3000000                         # total steps for training                         
  # Validation options
  dev_set: ['dev-clean']                      
  dev_batch_size: 32                             
  dev_step: 1000
  dev_metric: 'dev_loss'
  # Misc
  max_timestep: 0
  test_set: []