run_sweep.yaml

command:
  - python3
  - ${program}
  - --fp16
  - --fp16_full_eval
  - --do_train
  - --do_eval
  - --trust_remote_code
  - --overwrite_output_dir
  - ${args}
method: random
metric:
  goal: maximize
  name: eval/accuracy
parameters:
  model_name_or_path:
    value: facebook/mms-lid-4017
  train_dataset_name:
    value:  "stable-speech/concatenated-normalized-accent-dataset+stable-speech/concatenated-common-voice-15-accented"
  train_dataset_config_name:
    value: "default+default"
  train_split_name:
    value: "train+train"
  train_label_column_name:
    value: "labels+labels"
  eval_dataset_name:
    value: stable-speech/concatenated-normalized-accent-dataset
  eval_dataset_config_name:
    value: default
  eval_split_name:
    value: test
  eval_label_column_name:
    value: labels
  output_dir:
    value: "/raid/yoach/tmp/"
  remove_unused_columns:
    value: false
  learning_rate:
    distribution: log_uniform_values
    min: 3e-6
    max: 0.01
  lr_scheduler_type:
    value: constant
  max_length_seconds:
    value: 20  # give some data diversity for longer audio samples
  min_length_seconds:
    value: 5
  attention_mask:
    values:
    - true
  num_train_epochs:
    values:
      - 2
      - 5
      - 10
      - 20
      - 40
      - 60
  per_device_train_batch_size:
    value: 32
  per_device_eval_batch_size:
    value: 32
  preprocessing_num_workers:
    value: 8
  dataloader_num_workers:
    value: 8
  logging_strategy:
    value: steps
  logging_steps:
    value: 10
  evaluation_strategy:
    value: steps
  eval_steps:
    value: 2000
  save_strategy:
    value: "no"
  save_steps:
    value: 2000
  metric_for_best_model:
    value: accuracy
  push_to_hub:
    value: false
  use_weighted_layer_sum:
    value: false
  freeze_base_model:
    value: true
  max_samples_per_label:
    value: 10000
  save_to_disk:
    value: "/raid/yoach/tmp_dataset_accents/" 
  temporary_save_to_disk:
    value: "/raid/yoach/tmp_hidden_states/"
  use_last_embedding_layer:
    value: true
  filter_threshold:
    value: "0.001"
program: run_audio_classification.py
project: mms-lid-accent-classification-v2