39 lines
1.2 KiB
Bash
39 lines
1.2 KiB
Bash
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 |