Skip to content

Latest commit

 

History

History
119 lines (101 loc) · 7.59 KB

README.md

File metadata and controls

119 lines (101 loc) · 7.59 KB

Observation-Guided Diffusion Probabilistic Models

This is the codebase for Observation-Guided DIffusion Probabilistic Models. This repository is based on openai/guided-diffusion. The repository for EDM baseline can be found at Junoh-Kang/OGDM_edm

Dependencies

We share the environment of the code by docker.

docker pull snucvlab/ogdm:adm

To install all pacakges in this codebase along with their dependencies, run

pip install -e .

Model training

python -m torch.distributed.launch --nproc_per_node=[N] scripts/image_train.py --config [config.yaml] --batch_size [B] --schedule_sampler pair_T,[k] --lossG_weight [gamma]

# example
python -m torch.distributed.launch --nproc_per_node=4 scripts/image_train.py --config configs/celeba_64/01_scratch.yaml --batch_size 32 --schedule_sampler pair_T,0.10 --lossG_weight 0.01

Model resume training

python -m torch.distributed.launch --nproc_per_node=[N] scripts/image_train.py --batch_size [B] --resume_checkpoint [project_dir],[resume_step]

# example
python -m torch.distributed.launch --nproc_per_node=4 scripts/image_train.py --batch_size 32 --resume_checkpoint logs/celeba_64/ours_scratch_0.10,0.01,300000 --lr_anneal_steps 500000

Model finetuning

python -m torch.distributed.launch --nproc_per_node=[N] scripts/image_train.py --config [config_file] --batch_size [B] --finetune [model_dir] --schedule_sampler pair_T,[k] --lossG_weight [gamma]

# example
python -m torch.distributed.launch --nproc_per_node=4 scripts/image_train.py --config configs/celeba_64/02_finetune.yaml --batch_size 32 --finetune logs/celeba_64/baseline/model/ema_0.9999_300000.pt --schedule_sampler pair_T,0.20 --lossG_weight 0.025

Model Sampling

python -m torch.distributed.launch --nproc_per_node=[N] scripts/image_sample.py --project_dir [project_dir] --pt_name [pt_name] --num_samples [num_to_sample] --batch_size [B] --sampler [samplers_seperated_by_commas] --eta [eta]

# example
python -m torch.distributed.launch --nproc_per_node=4 scripts/image_sample.py --project_dir "logs/celeba_64/ours_scratch_0.10,0.01" --pt_name "ema_0.9999_300000.pt" --num_samples 50000 --batch_size 512 --sampler "ddim50,ddim20,ddim10,S-PNDM49,S-PNDM19,F-PNDM41,F-PNDM11"

Evaluation

We share the environment of the evaluation code by docker.

docker pull snucvlab/ogdm:fid

Then, run

# When activations of reference is not ready, evaluate and save activations at the same time by
python scripts/fid_prdc.py [reference] [sample] -save_act_path [ref.npz]

# When activations of reference is ready, evaluate 
python scripts/fid_prdc.py [reference] [sample]

to evaluate.

Download trained models

We provide trained models for CIFAR-10, and CelebA. The downloaded files should look like this:

└── project
    ├── config.yaml  
    └── models
        └── ema_xxxxxx.pt

Here are links to download and wget commands.

  • CIFAR-10 baseline: cifar10_baseline.tar.gz 
    wget --load-cookies /tmp/cookies.txt "https://docs.google.com/uc?export=download&confirm=$(wget --quiet --save-cookies /tmp/cookies.txt --keep-session-cookies --no-check-certificate 'https://docs.google.com/uc?export=download&id=1F7deiE3_hAITp-G74B4s61PWyWnKcjT5' -O- | sed -rn 's/.*confirm=([0-9A-Za-z_]+).*/\1\n/p')&id=1F7deiE3_hAITp-G74B4s61PWyWnKcjT5" -O cifar10_baseline.tar.gz && rm -rf /tmp/cookies.txt
  • CIFAR-10 OGDM (scratch): cifar10_scratch.tar.gz 
    wget --load-cookies /tmp/cookies.txt "https://docs.google.com/uc?export=download&confirm=$(wget --quiet --save-cookies /tmp/cookies.txt --keep-session-cookies --no-check-certificate 'https://docs.google.com/uc?export=download&id=1JPYc1hPJRD6aksvkAKVzr5QBW8AF56nq' -O- | sed -rn 's/.*confirm=([0-9A-Za-z_]+).*/\1\n/p')&id=1JPYc1hPJRD6aksvkAKVzr5QBW8AF56nq" -O cifar10_scratch.tar.gz && rm -rf /tmp/cookies.txt
  • CIFAR-10 OGDM (ft,k=0.15): cifar10_ft_0.15.tar.gz 
    wget --load-cookies /tmp/cookies.txt "https://docs.google.com/uc?export=download&confirm=$(wget --quiet --save-cookies /tmp/cookies.txt --keep-session-cookies --no-check-certificate 'https://docs.google.com/uc?export=download&id=1wsMu1nZ9q4efzUWmmnThy0XcSuwt3Y81' -O- | sed -rn 's/.*confirm=([0-9A-Za-z_]+).*/\1\n/p')&id=1wsMu1nZ9q4efzUWmmnThy0XcSuwt3Y81" -O cifar10_ft_0.15.tar.gz && rm -rf /tmp/cookies.txt
  • CIFAR-10 OGDM (ft,k=0.20): cifar10_ft_0.20.tar.gz 
    wget --load-cookies /tmp/cookies.txt "https://docs.google.com/uc?export=download&confirm=$(wget --quiet --save-cookies /tmp/cookies.txt --keep-session-cookies --no-check-certificate 'https://docs.google.com/uc?export=download&id=1HzFaFIhtd10GyU_jXqZN5UQw84HC9FGZ' -O- | sed -rn 's/.*confirm=([0-9A-Za-z_]+).*/\1\n/p')&id=1HzFaFIhtd10GyU_jXqZN5UQw84HC9FGZ" -O cifar10_ft_0.20.tar.gz && rm -rf /tmp/cookies.txt
  • CelebA baseline: celeba_baseline.tar.gz 
    wget --load-cookies /tmp/cookies.txt "https://docs.google.com/uc?export=download&confirm=$(wget --quiet --save-cookies /tmp/cookies.txt --keep-session-cookies --no-check-certificate 'https://docs.google.com/uc?export=download&id=1rVF8gO7QPF5E_2fZBSywMLAKrtt3ddt3' -O- | sed -rn 's/.*confirm=([0-9A-Za-z_]+).*/\1\n/p')&id=1rVF8gO7QPF5E_2fZBSywMLAKrtt3ddt3" -O celeba_baseline.tar.gz && rm -rf /tmp/cookies.txt
  • CelebA OGDM (scratch): celeba_scratch.tar.gz 
    wget --load-cookies /tmp/cookies.txt "https://docs.google.com/uc?export=download&confirm=$(wget --quiet --save-cookies /tmp/cookies.txt --keep-session-cookies --no-check-certificate 'https://docs.google.com/uc?export=download&id=1giccdeqoAURBuXvg0XfXW25vj1UtaTte' -O- | sed -rn 's/.*confirm=([0-9A-Za-z_]+).*/\1\n/p')&id=1giccdeqoAURBuXvg0XfXW25vj1UtaTte" -O celeba_scratch.tar.gz && rm -rf /tmp/cookies.txt
  • CelebA OGDM (ft, k=0.15): celeba_ft_0.15.tar.gz 
    wget --load-cookies /tmp/cookies.txt "https://docs.google.com/uc?export=download&confirm=$(wget --quiet --save-cookies /tmp/cookies.txt --keep-session-cookies --no-check-certificate 'https://docs.google.com/uc?export=download&id=1jop4N4Y5lwz3k4mW7I3g6s218wi3hgr7' -O- | sed -rn 's/.*confirm=([0-9A-Za-z_]+).*/\1\n/p')&id=1jop4N4Y5lwz3k4mW7I3g6s218wi3hgr7" -O celeba_ft_0.15.tar.gz && rm -rf /tmp/cookies.txt
  • CelebA OGDM (ft, k=0.20): celeba_ft_0.20.tar.gz 
    wget --load-cookies /tmp/cookies.txt "https://docs.google.com/uc?export=download&confirm=$(wget --quiet --save-cookies /tmp/cookies.txt --keep-session-cookies --no-check-certificate 'https://docs.google.com/uc?export=download&id=11A1YubnEtNxXKPllCaBh3YmPLq6nz1gh' -O- | sed -rn 's/.*confirm=([0-9A-Za-z_]+).*/\1\n/p')&id=11A1YubnEtNxXKPllCaBh3YmPLq6nz1gh" -O celeba_ft_0.20.tar.gz && rm -rf /tmp/cookies.txt

Citation

@inproceedings{kang2023odgm,
  author    = {Junoh Kang and Jinyoung Choi and Sungik Choi and Bohyung Han},
  title     = {Observation-Guided Diffusion Probabilistic Models},
  booktitle = {},
  year      = {2023}
}

Acknowledgments