pytorch unrolled gan

This repository implements UNROLLED GENERATIVE ADVERSARIAL NETWORKS (https://arxiv.org/pdf/1611.02163.pdf).

Install

docker pull pytorch/pytorch:1.3-cuda10.1-cudnn7-devel
pip install -r requirements.txt
git clone git@github.com:facebookresearch/higher.git
cd higher
pip install .

Run

You can run unrolled gan with 10 steps using the following code.

python main.py --config yes_higher_unroll_10

• yes_higher_unroll_10 uses higher library to create gradient of gradient.
• no_higher_unroll_10 option is to show what happens if the gradient is only first order approximation.
• unroll_0 option does not unroll gradient at all (simple GAN).

All running script provided in run.sh

run.sh

Result

unroll_0___________	no_higher_unroll_10_	yes_higher_unroll_10

Caveats

This implementation is based on another implementation of unrolled gan https://github.com/andrewliao11/unrolled-gans. The reason for the reimplementation is that the previous implementation did not unroll Discriminator's gradient, but only used the simulated Discriminator as the new discriminator to update Generator.

In this implementation, I used facebook's higher order gradient library higher https://github.com/facebookresearch/higher to create unrolled gradients.