ProbGAN: Towards Probabilistic GAN with Theoretical Guarantees

This repository contains the PyTorch implementation of the ProbGAN. This paper appears at ICLR 2019. If you find this repo useful for your research, please consider citing our [paper].

@article{he2018probgan,
  title={ProbGAN: Towards Probabilistic GAN with Theoretical Guarantees},
  author={He, Hao and Wang, Hao and Lee, Guang-He and Tian, Yonglong},
  year={2018}
}

Results

Install

This codebase is tested with Ubuntu 16.04 LTS, Python 3.6.8, PyTorch 1.0.0, and CUDA 9.0.

Usage

To train ProbGAN on different dataset with different GAN objectives.

GAN objectives
NS: original GAN (Non-saturating version)
MM: original GAN (Min-max version)
W: Wasserstein GAN
LS: Least-Square GAN

python train.py --dataset [cifar10 | stl10] --gan_obj [NS | MM | W | LS]

Acknowledgement

We inspired by the code of Bayesian GAN to train probabilistic GAN with SGHMC.