# See https://github.com/sanchit-gandhi/seq2seq-speech/issues/23#issuecomment-1122183173: do_lower_case should only be set to True for the tokenizer if the tokenizer has upper case letters in the vocab
# Let's also not add extra remove_punctuation
# And limit max duration to 25 seconds
WANDB_ENTITY=NbAiLab WANDB_PROJECT=wav2vec2 python run_flax_speech_recognition_ctc.py \
        --model_name_or_path="./" \
        --hub_model_id="NbAiLab/wav2vec2-1b-npsc-nst-tpu" \
        --tokenizer_name="./" \
        --output_dir="./" \
        --overwrite_output_dir \
        --num_train_epochs="40" \
        --skip_steps="740000" \
        --per_device_train_batch_size="2" \
        --per_device_eval_batch_size="2" \
        --gradient_accumulation_steps="1" \
        --precision="full_mixed" \
        --matmul_precision="bfloat16" \
        --multisteps \
        --learning_rate="4.906677872895671e-07" \
        --warmup_steps="2000" \
        --length_column_name="input_length" \
        --evaluation_strategy="steps" \
        --text_column_name="text" \
        --save_steps="5000" \
        --eval_steps="5000" \
        --logging_steps="100" \
        --layerdrop="0.041" \
        --attention_dropout="0.094" \
        --activation_dropout="0.055" \
        --hidden_dropout="0.047" \
        --save_total_limit="5" \
        --freeze_feature_encoder \
        --feat_proj_dropout="0.04" \
        --mask_time_prob="0.082" \
        --mask_time_length="10" \
        --mask_feature_prob="0.25" \
        --mask_feature_length="64" \
        --gradient_checkpointing \
        --min_duration_in_seconds="0.5" \
        --max_duration_in_seconds="25.0" \
        --use_auth_token \
        --seed="42" \
        --group_by_length \
        --do_train --do_eval \
        --push_to_hub \
        --preprocessing_num_workers="32" \
        --ctc_zero_infinity \
        --wandb_project="wav2vec2" \
        --wandb_name="wav2vec2-1b-npsc-nst-tpu (cont.)" \