Skip to main content

GAN

Neural Network which generates realistic data.

Motivation​

  • Stackgan generates images from text
  • Pix2Pix generates realistic images from sketches
  • Turn day scenes into night scenes
  • CycleGAN
  • Simulated eye images to Real images

How it works​

  • Consists of Generator and Discriminator
  • Generator takes in noise and generates image
  • Discriminator is a classifier which takes a real and a generated image and has to decide whether it's real or fake.

Tips and Tricks​

  • Use Leaky ReLu, important for gradient flow
  • At least one hidden layer in G and D
  • Generator output has tanh activation
  • Discriminator has sigmoid activation
  • Use Adam as optimizer
  • Use logits for cross entropy
  • Multiply labels times 0.9, helps discriminator to generalize better
  • Generator gets random noise z as input
  • Use Batchnorm everywhere except output layer of generator and input layer of descriminator
  • Paper: Improved Techniques for Training GANs