s3prl_s3prl_main/s3prl/downstream/asr/config.yaml

runner:
  total_steps: 200000
  gradient_clipping: 1
  gradient_accumulate_steps: 1
  log_step: 100
  eval_step: 2000
  save_step: 500
  max_keep: 1
  eval_dataloaders:
    - dev-clean

optimizer:
  name: TorchOptim
  torch_optim_name: Adam
  lr: 1.0e-4

# comment the whole scheduler config block
# to disable learning rate scheduling
# scheduler:
#   name: linear_schedule_with_warmup
#   num_warmup_steps: 1400
# comment the whole specaug config block
# to disable specaug on representation

specaug:
  adaptive: false
  adaptive_number_ratio: 0.04
  adaptive_size_ratio: 0.04
  max_n_time_masks: 20
  apply_time_warp: true
  apply_time_mask: true
  apply_freq_mask: true
  time_warp_window: 5
  time_mask_width_range: [0, 40]
  freq_mask_width_range: [0, 50]
  num_freq_mask: 4
  num_time_mask: 2

downstream_expert:
  datarc:
    train: ['train-clean-100']
    dev-clean: ['dev-clean']
    dev-other: ['dev-other']
    test-clean: ['test-clean']
    test-other: ['test-other']
    num_workers: 12
    train_batch_size: 32
    batch_size: 32
    eval_batch_size: 1
    libri_root: '/home/leo/d/datasets/LibriSpeech'
    bucket_file: './data/librispeech/len_for_bucket'
    dict_path: "./downstream/asr/char.dict"

    zero_infinity: True

    decoder_args:
      # See https://github.com/flashlight/text/blob/main/flashlight/lib/text/decoder/LexiconDecoder.h#L20-L30
      # for what the options mean. Python binding exposes the same options from C++.
      # KenLM is a fast LM query implementation, and it can be powered by:
      # 1. official LibriSpeech 4-gram LM: the 4-gram.arpa file on http://www.openslr.org/11
      # 2. fairseq style, letter-based lexicon: https://dl.fbaipublicfiles.com/fairseq/wav2vec/librispeech_lexicon.lst
      decoder_type: 'None'
      nbest: 1
      criterion: "ctc"
      beam: 5
      beam_threshold: 25
      kenlm_model: '/home/leo/d/datasets/4-gram.arpa'
      lexicon: '/home/leo/d/datasets/librispeech_lexicon.lst'
      lm_weight: 2
      word_score: -1
      unk_weight: -math.inf
      sil_weight: 0

  modelrc:
    project_dim: 1024
    select: RNNs
    Wav2Letter:
      total_rate: 320
    RNNs:
      total_rate: -1
      module: 'LSTM'                        # 'LSTM'/'GRU'
      bidirection: True
      dim: [1024, 1024]
      dropout: [0.2, 0.2]
      layer_norm: [False, False]
      proj: [False, False]              # Linear projection + Tanh after each rnn layer
      sample_rate: [1, 1]
      sample_style: 'concat'                  # 'drop'/'concat'