run.sh

export WANDB_PROJECT="xls-r-1000-fr"
python run_speech_recognition_ctc.py \
    --activation_dropout="0.1" \
    --cache_dir="/workspace/container_0/.cache/huggingface/transformers" \
    --dataset_name="mozilla-foundation/common_voice_7_0" \
    --dataset_config_name="fr" \
    --eval_steps="1000" \
    --evaluation_strategy="steps" \
    --feat_proj_dropout="0.0" \
    --freeze_feature_encoder \
    --fp16 \
    --gradient_accumulation_steps="8" \
    --gradient_checkpointing \
    --group_by_length \
    --layerdrop="0.0" \
    --learning_rate="7.5e-5" \
    --length_column_name="input_length" \
    --load_best_model_at_end \
    --logging_steps="100" \
    --mask_feature_length="64" \
    --mask_feature_prob="0.25" \
    --mask_time_length="10" \
    --mask_time_prob="0.75" \
    --model_name_or_path="facebook/wav2vec2-xls-r-1b" \
    --num_train_epochs="5.0" \
    --output_dir="./" \
    --overwrite_output_dir \
    --per_device_train_batch_size="16" \
    --per_device_eval_batch_size="16" \
    --preprocessing_num_workers="4" \
    --push_to_hub \
    --report_to="wandb" \
    --save_steps="1000" \
    --save_total_limit="2" \
    --text_column_name="sentence" \
    --use_auth_token \
    --warmup_steps="2000" \
    --do_train --do_eval