Hands-On Convolutional Neural Networks with TensorFlow

Techniques to improve GANs' trainability

Here, we will describe some techniques to make life easier when training GANs:

• Normalize inputs between -1/1
• Use BatchNorm
• Use Leaky Relu (discriminator)
• Use Relu (Generator), tanh on generator output
• For downsampling use average pooling or strided convolution
• Use Adam optimizer
• If Discriminator loss went down fast, something is wrong
• Use dropout on generator (both train and test phase)