GANLib

Tensorflow based GANs library

Dependencies

Tensorflow 1.10.1
Numpy 1.14.2
Matplotlib 2.1.1

Examples

Simple GAN: simple_gan.py
Progressive Growing of GANs: pg_gan.py
And slightly faster version of it: pg_gan_fast.py

Available GANs

Vanilla GAN. https://arxiv.org/pdf/1406.2661.pdf
Conditional GAN. https://arxiv.org/pdf/1411.1784.pdf
Disco GAN. https://arxiv.org/pdf/1703.05192.pdf
Adversarial Autoencoder. https://arxiv.org/pdf/1511.05644.pdf

Optimization Distances

Minmax (Original GAN optimization distance). https://arxiv.org/pdf/1406.2661.pdf
Crossentropy (Same as original but written in terms of cross entropy cost function)
Wasserstein (Earth mover's distance applyed to GAN) https://arxiv.org/pdf/1701.07875.pdf
Wasserstein GP (Improved version of Wasserstein distance) https://arxiv.org/pdf/1704.00028.pdf
Cramer (Most advanced, but difficult to calculate distance) https://openreview.net/pdf?id=S1m6h21Cb !!! discriminator output layer should have more that one neuron, and bigger the number the better !!!

Metrics

Magic Distance. (Estimate probability that points from one set appear in the other. Totally ignorant to mode collapse problem.)
Iception score. (Working only with image batches containing 3 channels in last channel format)