mirror of
https://github.com/TMElyralab/MuseTalk.git
synced 2026-02-05 01:49:20 +08:00
28 lines
726 B
Bash
28 lines
726 B
Bash
export VAE_MODEL="./sd-vae-ft-mse/"
|
|
export DATASET="..."
|
|
export UNET_CONFIG="./musetalk.json"
|
|
|
|
accelerate launch --multi_gpu train.py \
|
|
--mixed_precision="fp16" \
|
|
--unet_config_file=$UNET_CONFIG \
|
|
--pretrained_model_name_or_path=$VAE_MODEL \
|
|
--data_root=$DATASET \
|
|
--train_batch_size=8 \
|
|
--gradient_accumulation_steps=4 \
|
|
--gradient_checkpointing \
|
|
--max_train_steps=200000 \
|
|
--learning_rate=5e-05 \
|
|
--max_grad_norm=1 \
|
|
--lr_scheduler="cosine" \
|
|
--lr_warmup_steps=0 \
|
|
--output_dir="..." \
|
|
--val_out_dir='...' \
|
|
--testing_speed \
|
|
--checkpointing_steps=1000 \
|
|
--validation_steps=1000 \
|
|
--reconstruction \
|
|
--resume_from_checkpoint="latest" \
|
|
--use_audio_length_left=2 \
|
|
--use_audio_length_right=2 \
|
|
--whisper_model_type="tiny" \
|