run_grid_search.sh

#!/usr/bin/env bash

declare -a learning_rates=("3e-5" "1e-4" "3e-4")
declare -a batch_sizes=("8" "16")
declare -a gradient_accumulation_step_sizes=("4")

for learning_rate in "${learning_rates[@]}"; do
    for batch_size in "${batch_sizes[@]}"; do
        for gradient_accumulation_steps in "${gradient_accumulation_step_sizes[@]}"; do
	    python create_model.py
            CUDA_VISIBLE_DEVICES=1 python run_speech_recognition_seq2seq.py \
                --dataset_name="librispeech_asr" \
                --model_name_or_path="./" \
                --tokenizer_name="./" \
                --dataset_config_name="clean" \
                --train_split_name="train.100" \
                --eval_split_name="validation" \
                --output_dir="./" \
                --preprocessing_num_workers="1" \
                --length_column_name="input_length" \
                --overwrite_output_dir \
                --num_train_epochs="5" \
                --per_device_train_batch_size=$batch_size \
                --per_device_eval_batch_size=$batch_size \
                --gradient_accumulation_steps=$gradient_accumulation_steps \
                --generation_max_length="40" \
                --generation_num_beams="1" \
                --learning_rate=$learning_rate \
                --warmup_steps="1000" \
                --evaluation_strategy="steps" \
                --text_column_name="text" \
                --save_steps="1500" \
                --eval_steps="1500" \
                --logging_steps="1" \
                --save_total_limit="1" \
                --freeze_feature_encoder \
                --gradient_checkpointing \
                --fp16 \
                --group_by_length \
                --predict_with_generate \
                --do_lower_case \
                --do_train \
                --do_eval \
                --push_to_hub \
                --use_auth_token
        done
    done
done