cmd.sh

python run_speech_recognition_ctc.py \
	--model_name_or_path="facebook/w2v-bert-2.0" \
	--dataset_name="CLEAR-Global/luo_19h" \
	--train_split_name="train" \
	--eval_split_name="validation" \
	--output_dir="./w2v-bert-2.0-luo_cv_fleurs_19h" \
	--max_steps="100000" \
	--per_device_train_batch_size="32" \
	--per_device_eval_batch_size="32" \
    --gradient_accumulation_steps="2" \
	--freeze_feature_encoder=false \
	--add_adapter=true \
	--mask_time_prob="0.0" \
	--final_dropout="0.1" \
	--attention_dropout="0.05" \
	--feat_proj_dropout="0.05" \
	--hidden_dropout="0.05" \
	--ctc_zero_infinity=true \
	--learning_rate="3e-5" \
	--warmup_ratio="0.1" \
	--eval_strategy="steps" \
	--save_steps="1000" \
	--eval_steps="1000" \
	--logging_steps="1" \
	--eval_metrics wer cer \
	--save_total_limit="1" \
	--load_best_model_at_end \
    --max_duration_in_seconds="30" \
	--gradient_checkpointing \
	--fp16 \
	--group_by_length \
	--length_column_name "input_length" \
	--do_train --do_eval \
	--preprocessing_num_workers="22" \
	--dataloader_num_workers="22" \
	--push_to_hub \
	--push_to_hub_organization="CLEAR-Global" \
	--hub_strategy="checkpoint"