imvladikon's picture
init
f6079f4
raw
history blame
1.21 kB
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