# Whisper Finetuning script for the very small Nynorsk Common Voice
# Test script. Currently this only runs if you uncomment the lines in the training code for casting this from 48K to 16K
# Currently for training on a 48GB 
# Reduce batch size and learning rate if training on smaller GPU

python run_whisper_finetuning.py \
	--model_name_or_path="openai/whisper-small" \
	--output_dir="../whisper-CC-nn" \
	--overwrite_output_dir=True \
	--language="Norwegian" \
	--task="transcribe" \
	--dataset_name="mozilla-foundation/common_voice_11_0" \
	--dataset_config="nn-NO" \
	--do_train=True \
	--do_eval=True \
	--audio_column_name="audio" \
	--text_column_name="sentence" \
	--per_device_train_batch_size=48 \
    --per_device_train_batch_size=48 \
	--learning_rate=4e-5 \
	--warmup_steps=500 \
	--max_steps=1000 \
	--gradient_checkpointing=True \
	--gradient_accumulation_steps=1 \
	--group_by_length=False \
	--evaluation_strategy="steps" \
	--save_steps=250 \
	--eval_steps=250 \
	--max_eval_samples=50 \
	--logging_steps=250 \
	--fp16=True \
	--load_best_model_at_end=True \
	--metric_for_best_model="wer" \
	--greater_is_better=False \
	--report_to="tensorboard" \
	--predict_with_generate=True \
	--generation_max_length=225 \
	--print_training_arguments=True \
	--push_to_hub=True