This repository contains a PyTorch implementation of the Wasserstein GAN with gradient penalty.
WGAN works to minimize the Wasserstein-1 distance between the generated data distribution and the real data distribution. This technique offers more stability than the original GAN.
WGAN-GP improves upon WGAN by using a gradient penalty heuristic rather than weight clipping to encourage the discriminator to be locally 1-Lipschitz near the data manifold.
For more details, see the original paper https://arxiv.org/pdf/1704.00028.pdf.
Some direction was taken from these repositories: https://github.com/arturml/pytorch-wgan-gp, https://github.com/EmilienDupont/wgan-gp. In particular, the Generator and Discriminator architectures were taken from the first repository.
To run the code, adjust the hyperparameters in main.py and run
python3 main.py
This will train the GAN on the MNIST dataset. The necessary dependencies are contained in requirements.txt.
Here are some samples from the distribution of the generator:
These are the plots for the generator loss, discriminator loss, and gradient penalty. These agree with the plots given in this repository: https://github.com/arturml/pytorch-wgan-gp.
