src/run_persian.sh

#!/bin/bash

export LC_ALL=C.UTF-8
export LANG=C.UTF-8

export OUTPUT_DIR=/home/m3hrdadfi/code/wav2vec2-base-persian
export CACHE_DIR=/home/m3hrdadfi/data_cache/
export MODEL_NAME_OR_PATH=/home/m3hrdadfi/code/wav2vec2-base-persian


export TRAIN_FILE=/home/m3hrdadfi/data/fa/train_with_aug.csv
export VALIDATION_FILE=/home/m3hrdadfi/data/fa/test.csv
export SPEECH_FILE_COLUMN=path


#export MAX_EVAL_SAMPLES=5000
export PER_DEVICE_TRAIN_BATCH_SIZE=8
export PER_DEVICE_EVAL_BATCH_SIZE=8
#export GRADIENT_ACCUMULATION_STEPS=2
export NUM_TRAIN_EPOCHS=5.0
export LEARNING_RATE=5e-4
export WARMUP_STEPS=2000
export LOGGING_STEPS=500
#export EVAL_STEPS=2500
#export SAVE_STEPS=2500
export PREPROCESSING_NUM_WORKERS=4
export MAX_DURATION_IN_SECONDS=20.0
export ADAM_BETA_1=0.9
export ADAM_BETA_2=0.98
export WEIGHT_DECAY=0.01
export D_TYPE=bfloat16
export PAD_TO_MULTIPLE_OF=16384

python src/run_wav2vec2_pretrain_flax.py \
    --output_dir="$OUTPUT_DIR"  \
    --cache_dir="$CACHE_DIR" \
    --train_file="$TRAIN_FILE" \
    --validation_file="$VALIDATION_FILE" \
    --speech_file_column="$SPEECH_FILE_COLUMN" \
    --model_name_or_path="$MODEL_NAME_OR_PATH"  \
    --per_device_train_batch_size=$PER_DEVICE_TRAIN_BATCH_SIZE \
    --per_device_eval_batch_size=$PER_DEVICE_EVAL_BATCH_SIZE \
    --preprocessing_num_workers=$PREPROCESSING_NUM_WORKERS \
    --max_duration_in_seconds=$MAX_DURATION_IN_SECONDS \
    --num_train_epochs=$NUM_TRAIN_EPOCHS \
    --learning_rate=$LEARNING_RATE \
    --warmup_steps=$WARMUP_STEPS \
    --weight_decay=$WEIGHT_DECAY \
    --adam_beta1=$ADAM_BETA_1 \
    --adam_beta2=$ADAM_BETA_2 \
    --dtype="$D_TYPE" \
    --pad_to_multiple_of=$PAD_TO_MULTIPLE_OF \
    --logging_steps=$LOGGING_STEPS \
    --push_to_hub