Mastering Computer Vision with TensorFlow 2.x

Generator

The architecture of the generator 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 up-sampling in a generator. Note that max pooling is replaced by strided convolution.
• The first layer, which takes in the probability P(z) from the discriminator, is connected to the next convolutional layer through matrix multiplication. This means that no formal fully connected layer is used. However, the network serves its purpose.
• We apply batch normalization to all the layers to rescale ...