Code for the experiments in "Convergence Properties of Generative Adversarial Networks".
We implemented the small-scale experiments in PyTorch (v1.8.1) to explore different techniques fostering GAN convergence. The code base covers different GAN objectives including non-saturating GAN, WGAN, WGAN-GP, WGAN-LP and WGAN with a simple gradient regularization. The learning scenarios cover the DiracGAN, learning intervalls in 1D, different manifolds in 2D and mixture of Gaussians in 2D with respective visualizations.
Specify the required parameters for each experiment (Dirac: DiracGAN.py, 1D: GAN_1D.py, 2D: GAN_2D.py) and select whether you want singleruns, multiruns or experiments (multiruns on all combinations of two parameter lists).
Visualization of training a WGAN + simple Gradient Penalty in the DiracGAN setting.
Upper left: Wasserstein Distance (blue) between target distribution (Dirac at ) and generator distribution (Dirac at
)
Lower left: Trajectories of Generator (
) and Discriminator (
)
Right: Depiction of current learning progress
Learning of a uniform distribution with a shallow neural net as discriminator. Successful learning (simple gradient penalty) and failed learning (WGAN-GP).
With GAN_2D.py learning of different manifolds or a mixture of Gaussians
This repository uses the implementation of the Sinkhorn Loss from https://github.com/gpeyre/SinkhornAutoDiff