Model implementation

Our Discriminator and Generator implementations are inspired by the authors' torch implementation: https://github.com/reedscot/icml2016/blob/master/main_cls.lua.

Let's start by implementing the wrappers. We assume that the necessary imports are already in place.

Wrapper

We add a wrapper that combines a 2D Convolution with Batchnorm and an optional ReLU. This sequence of layers is very common in this model. In using this wrapper, the code becomes more compact and easier to read:

def ConvBatchnormRelu(x, n_filters, kernel_size, strides, padding, relu=True):
    x = Conv2D(n_filters, kernel_size=kernel_size, strides=strides,
               padding=padding, kernel_initializer=w_init)(x)
    x = BatchNormalization...