training / flax /distillation_scripts /run_librispeech_streaming_dummy.sh
nullonesix's picture
Saving train state of step 1
a1be16b verified
raw
history blame
1.31 kB
#!/usr/bin/env bash
python run_distillation.py \
--model_name_or_path "distil-whisper/tiny-random-whisper-2-1" \
--teacher_model_name_or_path "distil-whisper/tiny-random-whisper" \
--train_dataset_name "distil-whisper/librispeech_asr+distil-whisper/librispeech_asr-timestamped" \
--train_dataset_config_name "all+all" \
--train_dataset_samples "100+360" \
--train_split_name "train.clean.100+train.clean.360" \
--eval_dataset_name "distil-whisper/gigaspeech-l+esb/diagnostic-dataset" \
--eval_dataset_config_name "l+librispeech" \
--eval_split_name "validation+clean" \
--eval_text_column_name "text+ortho_transcript" \
--max_train_samples 1024 \
--max_eval_samples 32 \
--cache_dir "/home/sanchitgandhi/.cache" \
--dataset_cache_dir "/home/sanchitgandhi/.cache" \
--wandb_dir "/home/sanchitgandhi/.cache" \
--output_dir "./" \
--do_train \
--do_eval \
--per_device_train_batch_size 2 \
--per_device_eval_batch_size 2 \
--max_steps 10 \
--eval_steps 5 \
--dataloader_num_workers 14 \
--save_steps 5 \
--wer_threshold 10 \
--wandb_project "distil-whisper-debug" \
--logging_steps 1 \
--use_scan \
--gradient_checkpointing \
--overwrite_output_dir \
--predict_with_generate \
--return_timestamps \
--timestamp_probability 1 \
--freeze_encoder