Features:

• Generative Adversarial Network
• Weight regularization losses and dropout
  • I'm not sure yet whether they help or hurt...
  • at which places should i put dropout
• Conditioning on data attributes (labels etc.)
  • usually the conditioning vector is just the one-hot label
  • can also be dense vector calculated from several additional data
  • not sure where to best include that? each layer or just append once
• One-sided label smoothing
• Feature Matching
  • I hope my code here is correct.
• Default architecture as in DCGAN
  • ReLU in Generator (Tanh for final)
  • Leaky ReLU in Discriminator
  • No pooling layers
  • Batch Normalization
  • No fully connected layers

Not Yet Implemented

• Minbatch discrimination
• Historical Averaging
• Virtual Batch Normalization

Papers

• Generative Adversarial Networks
• Conditional Generative Adversarial Nets
• Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks
• Improved Techniques for Training GANs