Code for the paper Banach Wasserstein GAN.

Traditional WGAN uses an approximation of the Wasserstein metric to opimize the generator. This Wasserstein metric in turn depends upon an underlying metric on images which is taken to be the norm

The article extends the theory of WGAN-GP to any Banach space, while this code can be used to train WGAN over any Sobolev space with norm

The parameters p can be used to control the focus on outliers, with high p indicating a strong focus on the worst offenders. s can be used to control focus on small/large scale behaviour, where negative s indicates focus on large scales, while positive s indicates focus on small scales (e.g. edges).

Inception scores for the spaces and :

The code has some dependencies that can be easily installed

$ pip install https://github.com/adler-j/tensordata/archive/master.zip
$ pip install https://github.com/adler-j/adler/archive/master.zip

You also need a recent version of tensorflow in order to use the tf.contrib.gan functionality.