aailabkaist / dg Goto Github PK

Dongjun Kim *, Yeongmin Kim *, Se Jung Kwon, Wanmo Kang, and Il-Chul Moon
^{* Equal contribution}

| paper |

• Download edm-cifar10-32x32-uncond-vp.pkl at EDM for unconditional model.
• Download edm-cifar10-32x32-cond-vp.pkl at EDM for conditional model.
• Place EDM checkpoint at the directory specified below.

${project_page}/DG/
├── checkpoints
│   ├── pretrained_score/edm-cifar10-32x32-uncond-vp.pkl
│   ├── pretrained_score/edm-cifar10-32x32-cond-vp.pkl
├── ...

• To draw 50k unconditional samples, run:

python3 generate.py --network checkpoints/pretrained_score/edm-cifar10-32x32-uncond-vp.pkl --outdir=samples/cifar_uncond_vanilla --dg_weight_1st_order=0

• To draw 50k conditional samples, run:

python3 generate.py --network checkpoints/pretrained_score/edm-cifar10-32x32-cond-vp.pkl --outdir=samples/cifar_cond_vanilla --dg_weight_1st_order=0

• Download DG/data/true_data.npz for unconditional model.
• Download DG/data/true_data_label.npz for conditional model.
• Place real data at the directory specified below.

${project_page}/DG/
├── data
│   ├── true_data.npz
│   ├── true_data_label.npz
├── ...

• Download DG/checkpoints/ADM_classifier/32x32_classifier.pt
• We train 32 resolution classifier from ADM.
• Place 32x32_classifier.pt at the directory specified below.

${project_page}/DG/
├── checkpoints
│   ├── ADM_classifier/32x32_classifier.pt
├── ...

• Download pre-trained unconditional checkpoint DG/checkpoints/discriminator/cifar_uncond/discriminator_60.pt for the test.

• Download pre-trained conditional checkpoint DG/checkpoints/discriminator/cifar_cond/discriminator_250.pt for the test.

• Place pre-trained discriminator at the directory specified below.

${project_page}/DG/
├── checkpoints/discriminator
│   ├── cifar_uncond/discriminator_60.pt
│   ├── cifar_cond/discriminator_250.pt
├── ...

• To train the unconditional discriminator from scratch, run:

python3 train.py

• To train the conditional discriminator from scratch, run:

python3 train.py --savedir=/checkpoints/discriminator/cifar_cond --gendir=/samples/cifar_cond_vanilla --datadir=/data/true_data_label.npz --cond=1 

• To generate unconditional discriminator-guided 50k samples, run:

python3 generate.py --network checkpoints/pretrained_score/edm-cifar10-32x32-uncond-vp.pkl --outdir=samples/cifar_uncond

• To generate conditional discriminator-guided 50k samples, run:

python3 generate.py --network checkpoints/pretrained_score/edm-cifar10-32x32-cond-vp.pkl --outdir=samples/cifar_cond --dg_weight_1st_order=1 --cond=1 --discriminator_ckpt=/checkpoints/discriminator/cifar_cond/discriminator_250.pt --boosting=1

• Download stat files at DG/stats/cifar10-32x32.npz
• Place cifar10-32x32.npz at the directory specified below.

${project_page}/DG/
├── stats
│   ├── cifar10-32x32.npz
├── ...

• Run:

python3 fid_npzs.py --ref=/stats/cifar10-32x32.npz --num_samples=50000 --images=/samples/cifar_uncond/

python3 fid_npzs.py --ref=/stats/cifar10-32x32.npz --num_samples=50000 --images=/samples/cifar_cond/

Note that we use LSGM for Cifar-10 backbone, and Soft-Truncation for CelebA64 backbone.
See alsdudrla10/DG_imagenet for the results and released code on ImageNet256.

If you find the code useful for your research, please consider citing

@article{kim2022refining,
  title={Refining Generative Process with Discriminator Guidance in Score-based Diffusion Models},
  author={Kim, Dongjun and Kim, Yeongmin and Kang, Wanmo and Moon, Il-Chul},
  journal={arXiv preprint arXiv:2211.17091},
  year={2022}
}

This work is heavily built upon the code from

• Karras, T., Aittala, M., Aila, T., & Laine, S. (2022). Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364.
• Dhariwal, P., & Nichol, A. (2021). Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34, 8780-8794.
• Song, Y., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., & Poole, B. (2020). Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456.