Building a Deep Convolutional GAN (DCGAN)
Although Vanilla GAN has proven itself as a generative model, it suffers from a few training problems. One of them is the difficulty in scaling networks to make them deeper in order to increase their capacities. The authors of DCGAN incorporated a few recent advancements in CNNs at that time to make networks deeper and stabilize the training. These include the removal of the maxpool layer, replacing it with strided convolutions for downsampling, and the removal of fully connected layers. This has since become the standard way of designing a new CNN.
Architecture guidelines
DCGAN is not strictly a fixed neural network that has layers pre-defined with a fixed set of parameters such as kernel size and the number of layers. Instead, it is more like architecture design guidelines. The use of batch normalization, activation, and upsampling in DCGAN has influenced the development of GANs. We will therefore look into them more, which should...
