Mastering Computer Vision with TensorFlow 2.x

Discriminator

The architecture of the discriminator network of the DCGAN is shown in the following diagram:

From the preceding diagram, we can see the following:

• All convolutional networks with stride but without max pooling allow the network to learn its own downsampling in the discriminator.
• We eliminate the fully connected layer. The last convolutional layer is flattened and directly connected to a single sigmoid output.
• We apply batch normalization to all the layers except the discriminator input layer to improve stability in learning.