Generative adversarial network

From David's Wiki
\( \newcommand{\P}[]{\unicode{xB6}} \newcommand{\AA}[]{\unicode{x212B}} \newcommand{\empty}[]{\emptyset} \newcommand{\O}[]{\emptyset} \newcommand{\Alpha}[]{Α} \newcommand{\Beta}[]{Β} \newcommand{\Epsilon}[]{Ε} \newcommand{\Iota}[]{Ι} \newcommand{\Kappa}[]{Κ} \newcommand{\Rho}[]{Ρ} \newcommand{\Tau}[]{Τ} \newcommand{\Zeta}[]{Ζ} \newcommand{\Mu}[]{\unicode{x039C}} \newcommand{\Chi}[]{Χ} \newcommand{\Eta}[]{\unicode{x0397}} \newcommand{\Nu}[]{\unicode{x039D}} \newcommand{\Omicron}[]{\unicode{x039F}} \DeclareMathOperator{\sgn}{sgn} \def\oiint{\mathop{\vcenter{\mathchoice{\huge\unicode{x222F}\,}{\unicode{x222F}}{\unicode{x222F}}{\unicode{x222F}}}\,}\nolimits} \def\oiiint{\mathop{\vcenter{\mathchoice{\huge\unicode{x2230}\,}{\unicode{x2230}}{\unicode{x2230}}{\unicode{x2230}}}\,}\nolimits} \)

GANs are generative adversarial networks. They were developed by Ian Goodfellow.
Goal: Learn to generate examples from the same distribution as your training set.

Structure

GANs consist of a generator and a discriminator, both of which are usually CNNs.

For iteration i
  For iteration j
    Update Discriminator
  Update Generator

Generator

Two popular types of CNNs used in GANs are Resnets and UNets.
In both cases, we have convolutional blocks which consist of a conv2d layer, a batch norm, and an activation (typically Relu or leakyrelu).


Discriminator

A popular discriminator is the PatchGAN discriminator.
These are typically several convolutional blocks stacked together. Each convolutional layer in the conv block typically has a kernel size of (3x3) or (4x4) and a stride of 1-2.

Variations

Conditional GAN

Paper
Feed data y to both generator and discriminator

Wasserstein GAN

Paper
Medium post
This new WGAN-GP loss function improves the stability of training.
Normally, the discriminator is trained with a cross-entropy with sigmoid loss function.
The WGAN proposes using Wasserstein distance which is implemented by removing the cross-entropy+sigmoid and clipping (clamp) the weights on the discriminator to a range \(\displaystyle [-c, c]\).
However, weight clipping leads to other issues which limit the critic.
Instead of clipping, WGAN-GP proposes gradient penalty to enforce 1-Lipschitz .

Progressive Growing of GANs (ProGAN)

Paper
Progressively add layers to the generator and the discriminator of the GAN.
At the beginning, the generator makes a 4x4 image and the discriminator takes input the 4x4 image. Then, another layer is faded in the generator and the discriminator for and 8x8 image,...

Stacked Generative Adversarial Networks

Paper

Applications

CycleGan

InfoGAN

SinGAN

Paper
Website
Github Official PyTorch Implementation
SinGAN: Learning a Generative Model from a Single Natural Image

MoCoGAN

Paper
MoCoGAN: Decomposing Motion and Content for Video Generation

Video Prediction

  • Dual Motion GAN (Liang et al. 2017)
    • Have a frame generator and a motion generator
    • Combine the outputs of both generators using a fusing layer
    • Trained using a frame discriminator and a motion discriminator. (Each generator are trained with both discriminators)

Image and Video Compression

Object Segmentation

StyleGAN

Important Papers

Latent Space Exploration

Inversion

How to go from an image to a latent space vector

  • Image2StyleGAN
    • Mostly showing off applications using StyleGAN: morphing, style transfer, expression transfer
    • Invert StyleGAN to get style vectors \(\displaystyle w\) but with a different style vector per layer.
    • Able to get StyleGAN trained on faces to output cats, dogs, cars, ...
    • Followup Papers: Image2StyleGAN++ adds Activation Tensor Manipulation

Activation Tensor Manipulation

  • GAN Dissection: Visualizing and Understanding Generative Adversarial Networks
    • Authors: David Bau
    • Basically, individual "units" or channels of the intermediate representations correspond to some features like windows or trees in the output
    • Dissection: Identifying which units correspond to features can be done by visualizing each channel as a heatmap. Then threshold the heatmap so each value is binary 0/1. Calculate the IOU between the heatmap and the segmented feature in the generated picture.
    • Intervention: By zeroing out channels, you can remove windows or trees from the generated image. Alternatively you can add windows or trees at specific locations by activating the neurons at that location of the corresponding window/tree channel.
    • This is fairly specific to CNN architectures where there is a locality correspondence between the intermediate representations and the output image.
    • Followup Papers: Semantic photo manipulation

Resources