File size: 2,408 Bytes
a1be16b |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 |
#!/usr/bin/env bash
TCMALLOC_LARGE_ALLOC_REPORT_THRESHOLD=10000000000 python3 run_distillation.py \
--model_name_or_path "distil-whisper/large-32-2-15s-context" \
--teacher_model_name_or_path "openai/whisper-large-v2" \
--feature_extractor_name "openai/whisper-large-v2" \
--train_dataset_config_name "all+all+all+en+en+ihm+sdm+clean+release3+all+l+L" \
--train_dataset_samples "100+360+500+2300+450+90+90+12000+450+3600+2500+5000" \
--train_dataset_name "librispeech_asr+librispeech_asr+librispeech_asr+common_voice_13_0+voxpopuli+ami-ihm+ami-sdm+peoples_speech-clean+tedlium+switchboard-data+gigaspeech-l+spgispeech" \
--train_split_name "train.clean.100+train.clean.360+train.other.500+train+train+train+train+train+train+train+train+train" \
--eval_dataset_name "distil-whisper/gigaspeech-l+esb/diagnostic-dataset+esb/diagnostic-dataset+esb/diagnostic-dataset+esb/diagnostic-dataset+esb/diagnostic-dataset+esb/diagnostic-dataset+esb/diagnostic-dataset+esb/diagnostic-dataset+esb/diagnostic-dataset+esb/diagnostic-dataset+esb/diagnostic-dataset+esb/diagnostic-dataset" \
--eval_dataset_config_name "l+librispeech+librispeech+common_voice+common_voice+voxpopuli+voxpopuli+tedlium+tedlium+spgispeech+spgispeech+ami+ami" \
--eval_split_name "validation+clean+other+clean+other+clean+other+clean+other+clean+other+clean+other" \
--eval_text_column_name "text+ortho_transcript+ortho_transcript+ortho_transcript+ortho_transcript+ortho_transcript+ortho_transcript+ortho_transcript+ortho_transcript+ortho_transcript+ortho_transcript+ortho_transcript+ortho_transcript" \
--eval_steps 5000 \
--save_steps 5000 \
--warmup_steps 500 \
--learning_rate 0.0001 \
--lr_scheduler_type "linear" \
--logging_steps 25 \
--save_total_limit 1 \
--max_steps 80000 \
--wer_threshold 10 \
--per_device_train_batch_size 64 \
--per_device_eval_batch_size 64 \
--max_duration_in_seconds 15 \
--dataloader_num_workers 16 \
--cache_dir "/home/sanchitgandhi/.cache" \
--dataset_cache_dir "/home/sanchitgandhi/.cache" \
--dtype "bfloat16" \
--output_dir "./" \
--wandb_name "large-32-2-ts-28k-wer-10-context-15s" \
--wandb_dir "/home/sanchitgandhi/.cache" \
--wandb_project "distil-whisper" \
--do_train \
--do_eval \
--use_scan \
--gradient_checkpointing \
--overwrite_output_dir \
--predict_with_generate \
--streaming \
--use_auth_token \
--push_to_hub
|