# @package _group_

common:
  fp16: true
  fp16_no_flatten_grads: true
  log_format: json
  log_interval: 200
  user_dir: /private/home/abaevski/fairseq-py/examples/data2vec
#  tensorboard_logdir: tb

checkpoint:
  save_interval: 500
  save_interval_updates: 500
  keep_interval_updates: 1
  no_epoch_checkpoints: true
  best_checkpoint_metric: wer

task:
  _name: audio_finetuning
  data: /checkpoint/abaevski/data/speech/libri/10m/wav2vec/raw
  labels: ltr
  normalize: true

dataset:
  num_workers: 6
  max_tokens: 1000000
  skip_invalid_size_inputs_valid_test: true
  validate_after_updates: 100
  validate_interval: 500
  valid_subset: dev_other
  required_batch_size_multiple: 1

distributed_training:
  ddp_backend: legacy_ddp
  distributed_world_size: 4

criterion:
  _name: ctc
  zero_infinity: true
  post_process: letter
  wer_kenlm_model: /checkpoint/abaevski/data/speech/libri/4-gram.bin
  wer_lexicon: /checkpoint/abaevski/data/speech/libri/10h/wav2vec/raw/lexicon_ltr2.lst
  wer_lm_weight: 5
  wer_word_score: 2
  wer_sil_weight: -2

optimization:
  max_update: 10000
  lr: [2e-6]
#  lr: [1e-5]  # base 10h wer
  sentence_avg: true
  update_freq: [4]  # base 10h we -> 2/4

optimizer:
  _name: composite
  dynamic_groups: true
  groups:
    default:
      lr_float: 2e-6
      optimizer:
        _name: adam
        adam_betas: [0.9,0.95]
      lr_scheduler:
        _name: cosine
        warmup_updates: 1000

lr_scheduler: pass_through

model:
  _name: wav2vec_ctc
  w2v_path: ???
  apply_mask: true
  mask_prob: 0.4
  mask_length: 3
#  mask_prob: 0.65  # base 10h wer
  mask_channel_prob: 0.25
#  mask_channel_prob: 0.6  # base 10h wer
  mask_channel_length: 64
  layerdrop: 0.1
#  layerdrop: 0.05  # base 10h wer
  freeze_finetune_updates: 100

  zero_mask: true
  feature_grad_mult: 0.0
  activation_dropout: 0.1
  dropout: 0
  final_dropout: 0
  attention_dropout: 0
  update_alibi: false

#hydra:
#  job:
#    config:
#      override_dirname:
#        kv_sep: ':'
#        item_sep: '__'
#        exclude_keys:
#          - run_config
#          - distributed_training.distributed_port
#  sweep:
#    dir: /checkpoint/${env:USER}/${env:PREFIX}/${hydra.job.config_name}/${hydra.job.override_dirname}
#    subdir: ${hydra.job.num}
#  launcher:
#    submitit_folder: ${hydra.sweep.dir}
#    timeout_min: 3000
#    cpus_per_task: 10
#    gpus_per_node: 4
#    tasks_per_node: 4
#    mem_gb: 250
#    nodes: 1
#    name: ${env:PREFIX}_${hydra.job.config_name}
#    partition: devlab,learnlab,learnfair,scavenge
#    constraint: volta32gb
#    max_num_timeout: 30