custom_diffusion

---

Prepare

1. 서버 환경

   ssh -i [키 경로] [계정id]@[IP] -p [포트번호] 

• 필요한 데이터 및 weights SSD에 저장

  sftp -P [포트번호] [계정id]@[IP]
  put -r [/로컬/디렉토리/경로] [/원격/디렉토리/경로]

• git clone
• wandb install and login

2. 로컬 환경

• git clone
• wandb install and login

---

Test Inference

• Diffusion inference test
• text to image, inpainting, controlnet, lora 추론 진행 및 wandb 업로드

  python3 test/test_inference.py --root_path [pth 모델 경로] --project_name [wandb 프로젝트 이름]

---

Testtrain

• Diffusion train test
• text to image, text inversion, lora 학습 진행 및 wandb 업로드

  accelerate launch test/test_train.py --batch_size 2 --diffusion_model_path [pth 모델 경로] --train_data_path [test_data_dir 경로]

---

Train Controlnet

• Diffusion Controlnet train
• args
  • diffusion_model_path: diffusion.pth 모델 경로
  • controlnet_model_path: controlnet.pth 모델 경로
  • train_data_path: train data dir 경로
  • validation_prompts: validation prompt string(nargs)
  • validation_images: validation image 경로(nargs)
  • epochs: epochs
  • batch_size: batch_size
  • save_ckpt_step: 체크포인트 저장 step
  • validation_step: validation 진행 step
  • precision: mixed precision dtype string (choices=["no", "fp16", "bf16"])
  • report_to: wandb
  • device: cuda string (choices=["cpu", "cuda"])
  • lr: learning rate
  • seed: seed 값

  accelerate launch train/train_controlnet.py

---

Train Text Inversion

• Diffusion Text Inversion train
• args
  • diffusion_model_path: diffusion.pth 모델 경로
  • lora: store true Lora도 같이 훈련할 지 여부
  • train_data_path: train data dir 경로
  • validation_prompts: validation prompt string(nargs)
  • epochs: epochs
  • batch_size: batch_size
  • save_ckpt_step: 체크포인트 저장 step
  • validation_step: validation 진행 step
  • precision: mixed precision dtype string (choices=["no", "fp16", "bf16"])
  • report_to: wandb
  • device: cuda string (choices=["cpu", "cuda"])
  • lr: learning rate
  • seed: seed 값

  accelerate launch train/train_text_inversion.py

---

Train Lora

• Diffusion Text Inversion train
• args
  • diffusion_model_path: diffusion.pth 모델 경로
  • clip: store true clip(text inversion embedding) 같이 훈련할 지 여부
  • train_data_path: train data dir 경로
  • validation_prompts: validation prompt string(nargs)
  • epochs: epochs
  • batch_size: batch_size
  • save_ckpt_step: 체크포인트 저장 step
  • validation_step: validation 진행 step
  • precision: mixed precision dtype string (choices=["no", "fp16", "bf16"])
  • report_to: wandb
  • device: cuda string (choices=["cpu", "cuda"])
  • lr: learning rate
  • seed: seed 값

  accelerate launch train/train_lora.py

---