|
|
|
|
|
accelerate launch run_distillation.py \ |
|
--model_name_or_path "./distil-large-v3-init" \ |
|
--teacher_model_name_or_path "openai/whisper-large-v3" \ |
|
--train_dataset_name "distil-whisper/ami-ihm+distil-whisper/ami-ihm" \ |
|
--train_dataset_config_name "ihm+ihm" \ |
|
--train_split_name "train+validation" \ |
|
--text_column_name "text+text" \ |
|
--train_dataset_samples "7+4" \ |
|
--eval_dataset_name "distil-whisper/ami-ihm" \ |
|
--eval_dataset_config_name "ihm" \ |
|
--eval_split_name "test" \ |
|
--eval_text_column_name "text" \ |
|
--eval_steps 1000 \ |
|
--save_steps 1000 \ |
|
--warmup_steps 50 \ |
|
--learning_rate 0.0001 \ |
|
--lr_scheduler_type "constant_with_warmup" \ |
|
--timestamp_probability 0.2 \ |
|
--condition_on_prev_probability 0.2 \ |
|
--language "hi" \ |
|
--task "transcribe" \ |
|
--logging_steps 25 \ |
|
--save_total_limit 1 \ |
|
--max_steps 5000 \ |
|
--wer_threshold 20 \ |
|
--use_pseudo_labels True \ |
|
--per_device_train_batch_size 16 \ |
|
--per_device_eval_batch_size 16 \ |
|
--dataloader_num_workers 8 \ |
|
--preprocessing_num_workers 8 \ |
|
--ddp_timeout 7200 \ |
|
--dtype "bfloat16" \ |
|
--attn_implementation "sdpa" \ |
|
--output_dir "./" \ |
|
--do_train \ |
|
--do_eval \ |
|
--gradient_checkpointing \ |
|
--overwrite_output_dir \ |
|
--predict_with_generate \ |
|
--freeze_encoder \ |
|
--freeze_embed_positions \ |
|
--streaming True \ |
|
--push_to_hub |
|
|
