Deep Convolutional GAN

DCGAN was an early GAN architecture specialized for image generation.

• The input to the generator $g[z,\theta ]$ is a 100D latent variable $z$ sampled form a uniform distribution.
• This is then mapped to a $4\times 4$ spatial representation with 1024 channels using a linear transformation.
• Four convolutional layers follow, each of which uses a convolution that doubles the resolution (stride of 0.5).
• At the final layer, the $64\times 64\times 3$ signal is passed through a $\tanh$ function to generate an image $x^{\ast }$ in the range $[-1,1]$.
• The discriminator $f[\bullet ,\varphi ]$ is a standard convolutional network where the final convolutional layer reduces the size to $1\times 1$ with one channel. This single number is passed through a sigmoid function to create the output probability.
• After training, the discriminator is discarded. To create new samples, latent variables $z$ are drawn from the base distribution and passed through the generator.