regularisation

audio_classification_scripts/run_dropout_sweep.yaml

+command:
+  - python3
+  - ${program}
+  - --fp16
+  - --fp16_full_eval
+  - --do_train
+  - --do_eval
+  - --trust_remote_code
+  - --overwrite_output_dir
+  - --ignore_mismatched_sizes
+  - --gradient_checkpointing
+  - ${args}
+method: random
+metric:
+  goal: maximize
+  name: eval/accuracy
+parameters:
+  model_name_or_path:
+    value: facebook/mms-lid-126
+  train_dataset_name:
+    value: sanchit-gandhi/vctk+facebook/voxpopuli+sanchit-gandhi/edacc
+  train_dataset_config_name:
+    value: default+en_accented+default
+  train_split_name:
+    value: train+test+validation
+  train_label_column_name:
+    value: accent+accent+accent
+  eval_dataset_name:
+    value: sanchit-gandhi/edacc
+  eval_dataset_config_name:
+    value: default
+  eval_split_name:
+    value: test
+  eval_label_column_name:
+    value: accent
+  output_dir:
+    value: ./
+  remove_unused_columns:
+    value: false
+  learning_rate:
+    value: 1e-4
+  lr_scheduler_type:
+    value: constant_with_warmup
+  max_length_seconds:
+    value: 20
+  min_length_seconds:
+    value: 8
+  attention_mask:
+    value: true
+  warmup_steps:
+    value: 50
+  max_steps:
+    value: 1000
+  per_device_train_batch_size:
+    value: 32
+  per_device_eval_batch_size:
+    value: 32
+  preprocessing_num_workers:
+    value: 16
+  dataloader_num_workers:
+    value: 4
+  logging_strategy:
+    value: steps
+  logging_steps:
+    value: 10
+  evaluation_strategy:
+    value: steps
+  eval_steps:
+    value: 1000
+  save_strategy:
+    value: steps
+  save_steps:
+    value: 1000
+  freeze_base_model:
+    value: false
+  push_to_hub:
+    value: false
+  filter_threshold:
+    values:
+      - 1
+      - 2.5
+      - 5
+  feat_proj_dropout:
+    values:
+      - 0.0
+      - 0.1
+      - 0.2
+  attention_dropout:
+    values:
+      - 0.0
+      - 0.1
+      - 0.2
+  activation_dropout:
+    values:
+      - 0.0
+      - 0.1
+      - 0.2
+  hidden_dropout:
+    values:
+      - 0.0
+      - 0.1
+      - 0.2
+  final_dropout:
+    values:
+      - 0.0
+      - 0.1
+      - 0.2
+  mask_time_prob:
+    values:
+      - 0.0
+      - 0.1
+      - 0.2
+  mask_time_length:
+    values:
+      - 10
+      - 15
+      - 20
+  mask_feature_prob:
+    values:
+      - 0.0
+      - 0.1
+      - 0.2
+  mask_feature_length:
+    values:
+      - 10
+      - 15
+      - 20
+program: run_audio_classification.py
+project: mms-lid-accent-classification
\ No newline at end of file