run.sh

#!/usr/bin/env bash
## IMPORTANT: This script was stopped after 1.5 epochs (2400 steps)
## because the training loss was exploding => the best checkpoint (2000 steps)
## was then taken.
## MAKE SURE TO DO HYPER-PARAMETER TUNING TO GET BETTER RESULTS
python -m torch.distributed.launch \
	--nproc_per_node 2 run_speech_recognition_ctc.py \
	--dataset_name="edinburghcstr/ami" \
	--model_name_or_path="facebook/wav2vec2-large-lv60" \
	--dataset_config_name="ihm" \
	--train_split_name="train" \
	--eval_split_name="validation" \
	--output_dir="./" \
	--preprocessing_num_workers="16" \
	--overwrite_output_dir \
	--num_train_epochs="2" \
	--per_device_train_batch_size="32" \
	--per_device_eval_batch_size="32" \
	--gradient_accumulation_steps="1" \
	--learning_rate="3e-4" \
	--warmup_steps="500" \
	--evaluation_strategy="steps" \
	--text_column_name="text" \
	--min_duration_in_seconds="0.25" \
	--save_steps="400" \
	--eval_steps="1000" \
	--logging_steps="1" \
	--layerdrop="0.0" \
	--save_total_limit="3" \
	--freeze_feature_encoder \
	--gradient_checkpointing \
	--chars_to_ignore , ? . ! - \; \: \" “ % ‘ ” \
	--fp16 \
	--group_by_length \
	--push_to_hub \
	--do_eval \
	--do_train --do_eval