Supervised learning neural networks belong discriminator
DCGAN
Cancel All pools G network layer, the transposed convolution (transposed convolutional layer) and the step size is equal to 2 to the sampling
D also with the addition of a convolutional network instead of pooling stride
are used in quantities of D and G network Net the normalization (batch normalization), while generally not used in the final batch normalizaiton layer, which is to ensure the correct model to learn the mean and variance of the data
removed FC, allowing network becomes full convolution network
G network ReLU used as activation function, the last layer used as activation function Tanh
used as the activation function D LeakyReLU network
DCGAN are replaced by two convolutional neural network ( CNN ) of G and D , DCGAN in training will be used Adam optimization algorithm
InfoGAN an information theory and GAN integration of the neural network, the network having an information enables interpretation function
InfoGAN : with the implicit information GAN