The discriminators of the DiscoGAN would learn to distinguish the real images from the fake ones in a specific domain. We will have two discriminators: one for domain A, and one for domain B. The discriminators are also convolutional networks that can perform binary classification. Unlike the traditional classification-based convolutional networks, the discriminators don't have any fully connected layers. The input images are down-sampled using convolution with a stride of two until the final layer, where the output is 1 x 1. Again, we use leaky ReLU as the activation function and batch normalization for stable and fast convergence. The following code shows the discriminator build function implementation in TensorFlow:
def build_discriminator(self,image,reuse=False,name='discriminator'):
with tf.variable_scope(name):
...
