config.yaml

tag: 'sound'
seed: 77
sample_len: 16384
data:
  shift: 4000
  dataset_train:
    file_path:  '/home/dev/other/denois/txt_datasets/train_big.txt'
    mode: 'train'

  dataset_val:
    file_path: '/home/dev/other/denois/txt_datasets/test_big.txt'    
    mode: 'validation'

  loader_train:
    shuffle: True
    num_workers: 3
    batch_size: 32

unet:
  n_layers: 12
  channels_interval:  24

demucs:
  chin: 1
  chout: 1
  hidden: 48
  max_hidden: 10000
  causal: true
  glu: true
  depth: 5
  kernel_size: 8
  stride: 4
  normalize: true
  resample: 4
  growth: 2
  rescale: 0.1

loss:
  stft_sc_factor: 0.5
  stft_mag_factor: 0.5

optim:
  lr: 0.00001

trainer:
  sample_rate: 16000  #16000
  resume:   # continue training resume_from
  epochs: 1200
  visualize_waveform_limit: 5
  visualize_audio_limit: 5
  visualize_spectrogram_limit: 5
  validation_interval: 5
  sample_length: 16384
  base_dir: '../results/'
  spectro_dir: 'speech'
  exp_name: 'demucs_w_cuts'
  cudnn_deterministic: True