Minimal Wasserstein GAN

This is a simple TensorFlow implementation of Wasserstein Generative Advesarial Networks applied to MNIST.

Some example generated digits:

How to run

Simply run the file wgan_mnist.py. Results will be displayed in real time, while full training takes about an hour using a GPU.

Implementation details

The implementation follows Improved Training of Wasserstein GANs, using the network from the accompanying code. In particular both the generator and discriminator uses 3 convolutional layers with 5x5 convolutions.