#!/usr/bin/env bash
deepspeed run_speech_recognition_seq2seq_streaming.py \
	--deepspeed="ds_config.json" \
	--model_name_or_path="openai/whisper-large-v2" \
	--dataset_name="librispeech_asr" \
	--dataset_config_name="all" \
	--train_split_name="train.clean.100+train.clean.360+train.other.500" \
	--eval_split_name="validation.clean" \
	--output_dir="./" \
	--max_steps="25000" \
	--per_device_train_batch_size="32" \
	--per_device_eval_batch_size="8" \
	--logging_steps="25" \
	--learning_rate="3e-6" \
	--warmup_steps="50" \
	--evaluation_strategy="steps" \
	--eval_steps="5000" \
	--save_strategy="steps" \
	--save_steps="5000" \
	--generation_max_length="225" \
	--length_column_name="input_length" \
	--max_duration_in_seconds="30" \
	--freeze_feature_encoder="False" \
	--report_to="tensorboard" \
	--language="en" \
	--task="transcribe" \
	--do_lower_case \
	--gradient_checkpointing \
	--fp16 \
	--overwrite_output_dir \
	--do_train \
	--do_eval \
	--predict_with_generate \
	--use_auth_token \
	--push_to_hub