| 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 |