Hands-On Convolutional Neural Networks with TensorFlow

Deep convolutional GAN

Deep Convolutional GAN (DCGAN) is an extension of the normal GAN that we saw previously. Rather than using fully connected layers, we work with convolution layers. The idea is that the use of convolution layers helps the generator form better images. Here's what an example of this kind of model might look like:

An example implementation of DCGAN would be the same as training a normal GAN, as earlier, but simply swapping out the discriminator and generator networks for some convolutional architectures, like in the following code. Note that the generator will make use of transposed convolutions to upsample:

def discriminator(x): ...