config.yaml

audio:
  chunk_size: 485100
  dim_f: 1024
  dim_t: 1101
  hop_length: 882
  n_fft: 2048
  num_channels: 2
  sample_rate: 44100
  min_mean_abs: 0.000

model:
  dim: 384
  depth: 8
  stereo: true
  num_stems: 4
  linear_transformer_depth: 0
  time_transformer_depth: 1
  freq_transformer_depth: 1
  num_bands: 60
  dim_head: 64
  heads: 8
  attn_dropout: 0.0
  ff_dropout: 0.0
  flash_attn: true
  dim_freqs_in: 2049
  sample_rate: 44100  # needed for mel filter bank from librosa
  stft_n_fft: 4096
  stft_hop_length: 882
  stft_win_length: 4096
  stft_normalized: False
  mask_estimator_depth: 2
  multi_stft_resolution_loss_weight: 1.0
  multi_stft_resolutions_window_sizes: !!python/tuple
  - 4096
  - 2048
  - 1024
  - 512
  - 256
  multi_stft_hop_size: 147
  multi_stft_normalized: False
  mlp_expansion_factor: 4
  use_torch_checkpoint: False # it allows to greatly reduce GPU memory consumption during training (not fully tested)
  skip_connection: True # Enable skip connection between transformer blocks - can solve problem with gradients and probably faster training

training:
  batch_size: 1
  gradient_accumulation_steps: 1
  grad_clip: 0
  instruments: ['drums', 'bass', 'other', 'vocals']
  lr: 1.0
  patience: 3
  reduce_factor: 0.95
  target_instrument: null
  num_epochs: 1000
  num_steps: 1000
  q: 0.95
  coarse_loss_clip: false
  ema_momentum: 0.999
  optimizer: prodigy
  read_metadata_procs: 8 # Number of processes to use during metadata reading for dataset. Can speed up metadata generation
  normalize: false
  other_fix: false # it's needed for checking on multisong dataset if other is actually instrumental
  use_amp: true # enable or disable usage of mixed precision (float16) - usually it must be true

inference:
  batch_size: 4
  dim_t: 1101
  num_overlap: 4
  normalize: false