command:
  - python3
  - ${program}
  - --use_auth_token
  - --do_eval
  - --group_by_length
  - --overwrite_output_dir
  - --fp16
  - --do_lower_case
  - --do_eval
  - --do_train
  - --fuse_loss_wer
  - ${args}
method: grid
metric:
  goal: minimize
  name: train/loss
parameters:
  config_path:
    value: conf/conformer_transducer_bpe_xlarge.yaml
  dataset_config_name:
    value: clean
  dataset_name:
    value: librispeech_asr
  max_steps:
    value: 50
  model_name_or_path:
    value: stt_en_conformer_transducer_xlarge
  output_dir:
    value: ./sweep_output_dir
  per_device_train_batch_size:
    value: 8
  fused_batch_size:
    value: 8
  per_device_eval_batch_size:
    value: 4
  gradient_accumulation_steps:
    value: 1  #TBD. + grad checkpointing?
  preprocessing_num_workers:
    value: 1
  train_split_name:
    value: train.100
  eval_split_name:
    value: validation
  tokenizer_path:
    value: tokenizer
  vocab_size:
    value: 1024
  learning_rate:
    values:
      - 3e-4
      - 1e-4
      - 3e-5
      - 1e-5
  num_train_epochs:
    value: 3
program: run_speech_recognition_rnnt.py
project: rnnt-debug