QwenTest / pythonProject /diffusers-main /examples /research_projects /sana /train_sana_sprint_diffusers.sh
| your_local_path='output' | |
| hf download Efficient-Large-Model/SANA_Sprint_1.6B_1024px_teacher_diffusers --local-dir $your_local_path/SANA_Sprint_1.6B_1024px_teacher_diffusers | |
| # or Sana_Sprint_0.6B_1024px_teacher_diffusers | |
| python train_sana_sprint_diffusers.py \ | |
| --pretrained_model_name_or_path=$your_local_path/SANA_Sprint_1.6B_1024px_teacher_diffusers \ | |
| --output_dir=$your_local_path \ | |
| --mixed_precision=bf16 \ | |
| --resolution=1024 \ | |
| --learning_rate=1e-6 \ | |
| --max_train_steps=30000 \ | |
| --dataloader_num_workers=8 \ | |
| --dataset_name='brivangl/midjourney-v6-llava' \ | |
| --file_path data/train_000.parquet data/train_001.parquet data/train_002.parquet \ | |
| --checkpointing_steps=500 --checkpoints_total_limit=10 \ | |
| --train_batch_size=1 \ | |
| --gradient_accumulation_steps=1 \ | |
| --seed=453645634 \ | |
| --train_largest_timestep \ | |
| --misaligned_pairs_D \ | |
| --gradient_checkpointing \ | |
| --resume_from_checkpoint="latest" \ | |