default.yaml

common:
  fp16: true
  disable_tqdm: true # set true when using nohup
  seed: 0

path:
  checkpoint: "models/s5-hubert"
  quantizer1: "models/s5-hubert/quantizer1.joblib"
  quantizer2: "models/s5-hubert/quantizer2.npy"
  segment_dir: "segments/s5-hubert"
  result: "results/s5-hubert.json"

dataset:
  root: "data" # ${root}/LibriSpeech/train-clean-100, train-clean-360, ...
  download: false
  max_sample_size: 80160 # 5.01 [s]
  train_file: "src/sdhubert/files/librispeech_train_10Ksubset.txt"
  dev_file: "src/sdhubert/files/librispeech_val.txt"
  test_file: "src/sdhubert/files/librispeech_test.txt"
  dev_alignment: "src/sdhubert/files/librispeech_syllable_val.json"
  test_alignment: "src/sdhubert/files/librispeech_syllable_test.json"

dataloader:
  batch_size: 72 # work with single 24GB VRAM GPU
  num_workers: 30

model:
  model_type: "s5hubert"
  model_name_or_path: "facebook/hubert-base-ls960"
  init_last_layer: 3
  head_out_size: 256
  head_hidden_size: 2048
  ema_decay: 0.999
  segmentation_layer: 8

optim:
  epoch: 15
  lr: 0.0001
  lr_min: 0.00001
  stage_ratio: [0.03, 0.47, 0.50] # tri-stage lr schedule
  weight_decay: 0.01
  max_norm: null

n_clusters:
  step1: 16384
  step2: 4096