run.sh

#!/usr/bin/env bash

python run_parler_tts_training.py \
  --model_name_or_path parler-tts/parler_tts_mini_v0.1 \
  --feature_extractor_name parler-tts/dac_44khZ_8kbps \
  --description_tokenizer_name parler-tts/parler_tts_mini_v0.1 \
  --prompt_tokenizer_name parler-tts/parler_tts_mini_v0.1 \
  --report_to wandb \
  --overwrite_output_dir true \
  --train_dataset_name reach-vb/expresso-tagged-mistral-7b-instruct-v0.2 \
  --train_metadata_dataset_name reach-vb/expresso-tagged-mistral-7b-instruct-v0.2 \
  --train_dataset_config_name read \
  --train_split_name train \
  --eval_dataset_name reach-vb/expresso-tagged-mistral-7b-instruct-v0.2 \
  --eval_metadata_dataset_name reach-vb/expresso-tagged-mistral-7b-instruct-v0.2 \
  --eval_dataset_config_name read \
  --eval_split_name train \
  --max_eval_samples 8 \
  --per_device_eval_batch_size 16 \
  --target_audio_column_name audio \
  --description_column_name text_description \
  --prompt_column_name text \
  --max_duration_in_seconds 20 \
  --min_duration_in_seconds 2.0 \
  --max_text_length 400 \
  --preprocessing_num_workers 2 \
  --do_train true \
  --num_train_epochs 10 \
  --gradient_accumulation_steps 4 \
  --gradient_checkpointing true \
  --per_device_train_batch_size 32 \
  --learning_rate 3e-5 \
  --adam_beta1 0.9 \
  --adam_beta2 0.99 \
  --weight_decay 0.01 \
  --warmup_steps 100 \
  --logging_steps 2 \
  --freeze_text_encoder true \
  --audio_encoder_per_device_batch_size 4 \
  --dtype bfloat16 \
  --seed 456 \
  --output_dir ./ \
  --temporary_save_to_disk ../audio_code_tmp/ \
  --save_to_disk ../tmp_dataset_audio/ \
  --dataloader_num_workers 4 \
  --do_eval \
  --predict_with_generate \
  --include_inputs_for_metrics \
  --group_by_length true \
  --push_to_hub