run_all.sh

#!/usr/bin/env bash

accelerate launch run_pseudo_labelling.py \
  --model_name_or_path "openai/whisper-large-v3" \
  --dataset_name "mozilla-foundation/common_voice_16_1" \
  --dataset_config_name "de" \
  --dataset_split_name "train+validation+test" \
  --text_column_name "sentence" \
  --id_column_name "path" \
  --output_dir "../common_voice_16_1_de_pseudo_labelled" \
  --wandb_project "distil-whisper-labelling" \
  --per_device_eval_batch_size 64 \
  --dtype "bfloat16" \
  --attn_implementation "flash_attention_2" \
  --logging_steps 500 \
  --max_label_length 256 \
  --concatenate_audio \
  --preprocessing_batch_size 500 \
  --preprocessing_num_workers 8 \
  --dataloader_num_workers 8 \
  --report_to "wandb" \
  --language "de" \
  --task "transcribe" \
  --return_timestamps \
  --streaming False \
  --generation_num_beams 1 \
  --push_to_hub

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 "../common_voice_16_1_de_pseudo_labelled" \
  --train_split_name "train" \
  --text_column_name "sentence" \
  --eval_dataset_name "mozilla-foundation/common_voice_16_1" \
  --eval_split_name "validation" \
  --eval_text_column_name "sentence" \
  --eval_steps 5000 \
  --save_steps 5000 \
  --warmup_steps 500 \
  --learning_rate 0.0001 \
  --timestamp_probability 0.2 \
  --condition_on_prev_probability 0.2 \
  --language "de" \
  --task "transcribe" \
  --logging_steps 25 \
  --save_total_limit 1 \
  --max_steps 50000 \
  --wer_threshold 10 \
  --per_device_train_batch_size 32 \
  --per_device_eval_batch_size 32 \
  --dataloader_num_workers 8 \
  --preprocessing_num_workers 8 \
  --ddp_timeout 7200 \
  --dtype "bfloat16" \
  --attn_implementation "flash_attention_2" \
  --output_dir "./" \
  --do_train \
  --do_eval \
  --gradient_checkpointing \
  --overwrite_output_dir \
  --predict_with_generate \
  --freeze_encoder \
  --freeze_embed_positions \
  --streaming False \
  --push_to_hub