RSR

• Source code for "Run-Sort-ReRun: Escaping Batch Size Limitations in Sliced Wasserstein Generative Models, J. Lezama, W. Chen, Q. Qiu, ICML 2021"

• This repository contains PyTorch source code to run the toy example in Figure 1 of the paper and CIFAR-10 results in Tables 1 and 2 and Figure 3.

• Toy Example (Figure 1):

  • run rsr_toy.py
    • optionally, set large_batch_size to 4 to simulate Wasserstein distance on 4 samples (3rd subfigure)

• CIFAR10 Example:

  • Based on github repos: Yuantian013/E2GAN (included), mseitzer/pytorch-fid (included)
  • Download pretrained E2GAN baseline from this link
  • Execute run_training.sh or python rsr-features.py for training (inside cifar10)
  • E2GAN model fine-tuned with RunSort-ReRun available at this link

• StyleGAN2 Example:

  • Based on github repos: rosinality/stylegan2-pytorch (included)
  • Download converted pretrained StyleGAN2 baseline from this link. (It can also be obtained following the instruction of rosinality/stylegan2-pytorch)
  • Download LSUN church using this link, and place it at stylegan/.data/church/
  • Download the stats of LSUN church from this link
  • Execute run.sh or python rsr-features.py for training (inside stylegan)
  • StyleGAN2 model fine-tuned with RunSort-ReRun available at this link, place it at stylegan/weights/