config.yaml

default: &DEFAULT

  #data
  exp_dir: carnatic/ckpts
  metadata_labeled_path: labeled_small_wav_metadata.json
  metadata_unlabeled_path: unlabeled_data/unlabeled_mp3_metadata.json
  num_files_per_raga_path: metadata_small.json
  num_classes: 150
  use_frac: 1
  use_unlabeled_data: !!bool False
  labeled_data_dir: labeled_data_small
  clip_length: 30
  sample_rate: 8000
  normalize: !!bool True

  #training
  batch_size: 16
  num_data_workers: 16
  n_epochs: 100
  lr: 0.001
  class_imbalance_weights: !!bool False
  patience: 10
  train_frac: 0.8
  
  #model
  model: 'base'
  n_input: 2 #stereo
  stride: 16
  n_channel: 32
  max_pool_every: 1
  
  #logging
  save_checkpoint: !!bool False
  wandb_api_key: f7892f37dd96b5f1da5c85a410300bb661f3c4de
  log_to_wandb: !!bool False
  

default_0.7: &DEFAULT_0.7

  <<: *DEFAULT
  metadata_labeled_path: labeled_0.7_wav_metadata.json
  num_files_per_raga_path: metadata_0.7.json
  labeled_data_dir: labeled_data_0.7

default_0.9: &DEFAULT_0.9

  <<: *DEFAULT
  metadata_labeled_path: labeled_0.9_wav_metadata.json
  num_files_per_raga_path: metadata_0.9.json
  labeled_data_dir: labeled_data_0.9
  train_frac: 0.85
  num_classes: 200

  
resnet: &RESNET

  <<: *DEFAULT
  model: 'resnet'
  n_blocks: 5 #for resnet
  n_channel: 128

resnet_0.7: &RESNET_0.7

  <<: *DEFAULT_0.7
  model: 'resnet'
  n_blocks: 10 #for resnet
  n_channel: 300
  num_classes: 150

resnet_0.9: &RESNET_0.9

  <<: *DEFAULT_0.9
  model: 'resnet'
  n_blocks: 10 #for resnet
  n_channel: 350
  max_pool_every: 1 #downsample every other res block


wav2vec_0.7: &WAV2VEC_0.7
  <<: *DEFAULT_0.7

  model: 'wav2vec'
  n_input: 1 #mono

  #transformer parameters (this config leads to around 29M params)
  extractor_mode: "layer_norm"
  extractor_conv_layer_config: None #harcoded for now, fix this at some point
  extractor_conv_bias: !!bool True
  encoder_embed_dim: 512
  encoder_projection_dropout: 0
  encoder_pos_conv_kernel: 3
  encoder_pos_conv_groups: 32
  encoder_num_layers: 12
  encoder_num_heads: 16
  encoder_attention_dropout: 0
  encoder_ff_interm_features: 1024
  encoder_ff_interm_dropout: 0
  encoder_dropout: 0
  encoder_layer_norm_first: !!bool True
  encoder_layer_drop: 0


wav2vec_0.9: &WAV2VEC_0.9
  <<: *DEFAULT_0.9

  model: 'wav2vec'
  n_input: 1 #mono

  #transformer parameters (this config leads to around 29M params)
  extractor_mode: "layer_norm"
  extractor_conv_layer_config: None #harcoded for now, fix this at some point
  extractor_conv_bias: !!bool True
  encoder_embed_dim: 512
  encoder_projection_dropout: 0
  encoder_pos_conv_kernel: 3
  encoder_pos_conv_groups: 32
  encoder_num_layers: 12
  encoder_num_heads: 16
  encoder_attention_dropout: 0
  encoder_ff_interm_features: 1024
  encoder_ff_interm_dropout: 0
  encoder_dropout: 0
  encoder_layer_norm_first: !!bool True
  encoder_layer_drop: 0