run.sh

python -m torch.distributed.launch \
    --nproc_per_node=8 \
    ./examples/research_projects/xtreme-s/run_xtreme_s.py \
    --model_name_or_path="facebook/wav2vec2-xls-r-300m" \
    --task="fleurs-asr" \
    --language="en_us" \
    --output_dir="xtreme_s_xlsr_300m_fleurs_asr_en_us_nomask" \
    --overwrite_output_dir \
    --num_train_epochs=30 \
    --per_device_train_batch_size=8 \
    --per_device_eval_batch_size=1 \
    --gradient_accumulation_steps=1 \
    --eval_accumulation_steps=10 \
    --learning_rate="3e-4" \
    --ctc_zero_infinity \
    --warmup_steps=200 \
    --evaluation_strategy="steps" \
    --max_duration_in_seconds=20 \
    --preprocessing_num_workers=16 \
    --logging_steps=1 \
    --eval_steps=200 \
    --layerdrop=0.0 \
    --mask_time_prob=0.05 \
    --mask_time_length=10 \
    --mask_feature_prob=0.05 \
    --mask_feature_length=64 \
    --freeze_feature_encoder \
    --gradient_checkpointing \
    --fp16 \
    --fp16_full_eval \
    --group_by_length \
    --do_train \
    --do_predict \
    --per_lang_metrics=False \
    --push_to_hub