Hands-On Generative Adversarial Networks with Keras

Discriminator

The Discriminator architecture used in this experiment is very similar to DCGAN. The most important aspects of this implementation are the dense layer on the text embedding and its concatenation with a projection of the image input that goes through multiple 2D Convolutions with Batchnorm and LeakyReLU. This concatenation will go through a couple more Convolutions and produce the final output of the model, which will have a Sigmoid non-linearity, indicating that we are working with the standard GAN framework. We define input for the images and text embeddings using the following code:

def build_discriminator(image_input_shape=(64, 64, 3), text_input_shape=(1024,),                        embedding_dim=128, ndf=64, activation='linear'): # define inputs ...