|
export CUDA_VISIBLE_DEVICES="0,1" |
|
|
|
python -m torch.distributed.launch --nproc_per_node=2 run_train.py \ |
|
--dataset_name="imvladikon/hebrew_speech_???" \ |
|
--use_auth_token="???" \ |
|
--audio_column_name="audio" \ |
|
--text_column_name="sentence" \ |
|
--model_name_or_path="imvladikon/wav2vec2-xls-r-300m-hebrew" \ |
|
--tokenizer_name_or_path="./wav2vec2-xls-r-300m-hebrew" \ |
|
--output_dir="./wav2vec2-xls-r-300m-hebrew" \ |
|
--overwrite_output_dir \ |
|
--evaluation_strategy="steps" \ |
|
--length_column_name="input_length" \ |
|
--gradient_checkpointing \ |
|
--fp16 \ |
|
--group_by_length \ |
|
--num_train_epochs="100" \ |
|
--per_device_train_batch_size="8" \ |
|
--per_device_eval_batch_size="8" \ |
|
--gradient_accumulation_steps="4" \ |
|
--learning_rate="3e-4" \ |
|
--warmup_steps="1000" \ |
|
--save_steps="1000" \ |
|
--eval_steps="1000" \ |
|
--preprocessing_num_workers="$(nproc)" \ |
|
--logging_steps="2000" \ |
|
--layerdrop="0.0" \ |
|
--activation_dropout="0.1" \ |
|
--save_total_limit="3" \ |
|
--freeze_feature_encoder \ |
|
--feat_proj_dropout="0.0" \ |
|
--mask_time_prob="0.75" \ |
|
--mask_time_length="10" \ |
|
--mask_feature_prob="0.25" \ |
|
--mask_feature_length="64" \ |
|
--do_train --do_eval \ |
|
--print_samples \ |
|
--use_augmentations \ |
|
--push_to_hub |