|
#!/usr/bin/env bash |
|
torchrun --nproc_per_node 3 examples/pytorch/speech-recognition/run_speech_recognition_ctc.py |
|
--dataset_name="librispeech_asr" \ |
|
--model_name_or_path="facebook/w2v-bert-2.0" \ |
|
--dataset_config_name="clean" \ |
|
--eval_split_name="test" \ |
|
--train_split_name="train.100" \ |
|
--output_dir="./wav2vec2-bert-CV16-en-libri" \ |
|
--num_train_epochs="7" \ |
|
--per_device_train_batch_size="12" \ |
|
--gradient_accumulation_steps="2" \ |
|
--per_device_eval_batch_size="12" \ |
|
--learning_rate="3e-5" \ |
|
--warmup_steps="10000" \ |
|
--evaluation_strategy="steps" \ |
|
--text_column_name="text" \ |
|
--save_steps="500" \ |
|
--eval_steps="250" \ |
|
--save_total_limit="3" \ |
|
--gradient_checkpointing \ |
|
--chars_to_ignore , ? . ! - \; \: \" “ % ‘ ” \ |
|
--fp16 --push_to_hub \ |
|
--do_train --do_eval \ |
|
--eval_metrics "wer" "cer" \ |
|
--freeze_feature_encoder false --logging_steps "5" \ |
|
--add_adapter true \ |
|
--preprocessing_num_workers "32" \ |
|
--mask_time_prob="0.0" --mask_feature_prob="0.0" \ |
|
--tokenizer_name_or_path "jonatasgrosman/wav2vec2-large-xlsr-53-english" \ |
|
--eval_accumulation_steps "2" --group_by_length --length_column_name="input_length" \ |
|
--layerdrop="0.0" \ |
|
--hidden_dropout="0.05" --activation_dropout="0.05" --feat_proj_dropout="0.05" |
|
|
|
|
|
|
