train.yaml

# config for training

trainer:
  checkpoint: "2spk_pit_a"
  optimizer: "adam"
  # 1e-3 => str
  # 1.0e-3 => float
  lr: 1.0e-3
  momentum: 0.9
  weight_decay: 0
  clip_norm: 200
  min_lr: 1.0e-10
  patience: 1
  factor: 0.7
  disturb_std: 0

model:
  num_spks: 2
  rnn: "blstm"
  num_layers: 3
  hidden_size: 896
  dropout: 0.5
  non_linear: "relu"
  bidirectional: true

spectrogram_reader:
  frame_shift: 128
  frame_length: 256
  window: "hann"
  transpose: true
  # NOTE: do not apply log here
  # false => produce psm
  # true  => produce irm
  apply_abs: false

train_scp_conf:
  mixture: "create_scp/tr_mix.scp"
  spk1: "create_scp/tr_s1.scp"
  spk2: "create_scp/tr_s2.scp"

valid_scp_conf:
  mixture: "create_scp/cv_mix.scp"
  spk1: "create_scp/cv_s1.scp"
  spk2: "create_scp/cv_s2.scp"

debug_scp_conf:
  mixture: "create_scp/tr_mix.scp"
  spk1: "create_scp/tr_s1.scp"
  spk2: "create_scp/tr_s2.scp"

dataloader:
  shuffle: true
  batch_size: 10
  drop_last: false
  apply_log: true
  mvn_dict: "cmvn.dict"
  num_workers: 1