|
|
|
|
|
|
|
|
|
python xla_spawn.py --num_cores=4 run_whisper_finetuning.py\ |
|
--model_name_or_path="openai/whisper-small" \ |
|
--output_dir="../whisper-NST-TPU-test4" \ |
|
--overwrite_output_dir=True \ |
|
--language="Norwegian" \ |
|
--task="transcribe" \ |
|
--dataset_name="NbAiLab/NST" \ |
|
--dataset_config="no-close" \ |
|
--do_train=True \ |
|
--do_eval=False \ |
|
--audio_column_name="audio" \ |
|
--text_column_name="text" \ |
|
--per_device_train_batch_size=16 \ |
|
--per_device_train_batch_size=16 \ |
|
--learning_rate=2e-5 \ |
|
--warmup_steps=0 \ |
|
--max_steps=20 \ |
|
--gradient_checkpointing=True \ |
|
--gradient_accumulation_steps=1 \ |
|
--group_by_length=False \ |
|
--evaluation_strategy="steps" \ |
|
--save_steps=20 \ |
|
--eval_steps=1 \ |
|
--max_eval_samples=2 \ |
|
--logging_steps=100 \ |
|
--load_best_model_at_end=False \ |
|
--metric_for_best_model="wer" \ |
|
--greater_is_better=False \ |
|
--report_to="tensorboard" \ |
|
--predict_with_generate=False \ |
|
--generation_max_length=225 \ |
|
--print_training_arguments=True \ |
|
--xla=True \ |
|
--push_to_hub=True |
|
|
|
|
