baseline.yaml

feature:
  data_root: midisynth_dataset-v1
  torchaudio_backend: sox_io  # use soundfile for windows
  feat_name: fbank
  feat_param:
    num_mel_bins: &n_feat 128

data:
  train:
    csv_path: midisynth_dataset-v1/leave-instrument-out/train.csv
    dataloader:
      batch_size: 64
      shuffle: True
      num_workers: 12
    
  val:
    csv_path: midisynth_dataset-v1/leave-instrument-out/test.csv
    dataloader:
      batch_size: 64
      num_workers: 12

  test:
    csv_path: midisynth_dataset-v1/leave-instrument-out/test.csv
    dataloader:
      batch_size: 64
      num_workers: 12

trainer:
  epochs: 100
  optimizer: adam
  learning_rate: 3e-4
  loss_fn: crossentropy
  checkpoint_dir: weights
  save_every: 5  # epoch

model:
  channel_first: False
  feat_dim: *n_feat
  n_class: &n_class 12

  cnn:
    - in_channel: *n_feat
      out_channel: 256
      kernel_size: 7
      stride: 1
      padding: "same"
      activation: "relu"

    - in_channel: 256
      out_channel: 512
      kernel_size: 7
      stride: 2
      padding: "valid"
      activation: "relu"

    - in_channel: 512
      out_channel: 512
      stride: 1
      kernel_size: 5
      padding: "same"
      activation: "relu"

    - in_channel: 512
      out_channel: 1024
      stride: 2
      kernel_size: 5
      padding: "valid"
      activation: "relu"

  pooling: 
    method: gap
    dropout: 0.2

  # rnn:
  #   name: lstm
  #   param:
  #     input_size: 512
  #     hidden_size: 512
  #     bidirectional: True

  linear:
    - feat_in: 1024
      feat_out: *n_class