Generative Adversarial Networks

GANs (mainly in image synthesis)

Survey Papers / Repos

• Are GANs Created Equal? A Large-Scale Study [1711.10337]

• Which Training Methods for GANs do actually Converge? [1801.04406]

• A Large-Scale Study on Regularization and Normalization in GANs [1807.04720]

• hindupuravinash/the-gan-zoo

Resources

• google/compare_gan

• TF-GAN: TensorFlow-GAN

• torchgan

• PyTorch-GAN

• lzhbrian/metrics: IS, FID implementation in TF, PyTorch

Models

Loss functions

• Vanilla GAN [1406.2661]

• EBGAN [1609.03126]

• LSGAN [1611.04076]

• WGAN [1701.07875]

• BEGAN [1703.10717]

• Hinge Loss [1705.02894]

Regularization

• Gradient Penalty [1704.00028]

• DRAGAN [1705.07215]

• SNGAN [1802.05957]

• Consistency Regularization [1910.12027]

Architecture

• Deep Convolution GAN (DCGAN) [1511.06434]

• Progressive Growing of GANs (PGGAN) [1710.10196]

• Self Attention GAN (SAGAN) [1805.08318]

• BigGAN [1809.11096]

• Style based Generator (StyleGAN) [1812.04948]

• Mapping Network (StyleGAN) [1812.04948]

• LOGAN: Latent Optimisation for Generative Adversarial Networks [1912.00953]

Conditional GANs

• Vanilla Conditional GANs [1411.1784]

• Auxiliary Classifer GAN (ACGAN) [1610.09585]

Others

Tricks

• Two time-scale update rule (TTUR) [bioinf-jku/TTUR] [1706.08500]

• Self-Supervised GANs via Auxiliary Rotation Loss (SS-GAN) [1811.11212]

Metrics (my implementation: lzhbrian/metrics)

• Inception Score [1606.03498] [1801.01973]

  • Assumption

    • MEANINGFUL: The generated image should be clear, the output probability of a classifier network should be [0.9, 0.05, ...] (largely skewed to a class). $p(y|\mathbf{x})$ is of low entropy.

    • DIVERSITY: If we have 10 classes, the generated image should be averagely distributed. So that the marginal distribution $p(y) = \frac{1}{N} \sum_{i=1}^{N} p(y|\mathbf{x}^{(i)})$ __is of high entropy.

    • Better models: KL Divergence of $p(y|\mathbf{x})$ and $p(y)$ should be high.

  • Formulation

    • $\text{IS} = \exp (\mathbb{E}_{\mathbf{x} \sim p_g} D_{KL} [p(y|\mathbf{x}) || p(y)] )$

    • where

      • $\mathbf{x}$ is sampled from generated data

      • $p(y|\mathbf{x})​$ is the output probability of Inception v3 when input is $\mathbf{x}​$

      • $p(y) = \frac{1}{N} \sum_{i=1}^{N} p(y|\mathbf{x}^{(i)})$ is the average output probability of all generated data (from InceptionV3, 1000-dim vector)

      • $D_{KL} (\mathbf{p}||\mathbf{q}) = \sum_{j} p_{j} \log \frac{p_j}{q_j}$, where $j$ is the dimension of the output probability.

  • Reference

    • Official TF implementation is in openai/improved-gan

    • Pytorch Implementation: sbarratt/inception-score-pytorch

    • TF seemed to provide a good implementation

    • scipy.stats.entropy

    • zhihu: Inception Score 的原理和局限性

    • A Note on the Inception Score

• FID Score [1706.08500]

  • Formulation

    • $\text{FID} = ||\mu_r - \mu_g||^2 + Tr(\Sigma_{r} + \Sigma_{g} - 2(\Sigma_r \Sigma_g)^{1/2})​$

    • where

      • $Tr$ is trace of a matrix (wikipedia)

      • $X_r \sim \mathcal{N}(\mu_r, \Sigma_r)$ and $X_g \sim \mathcal{N}(\mu_g, \Sigma_g)$ are the 2048-dim activations the Inception v3 pool3 layer

      • $\mu_r$ is the mean of real photo's feature

      • $\mu_g$ is the mean of generated photo's feature

      • $\Sigma_r$ is the covariance matrix of real photo's feature

      • $\Sigma_g$ is the covariance matrix of generated photo's feature

  • Reference

    • Official TF implementation: bioinf-jku/TTUR

    • Pytorch Implementation: mseitzer/pytorch-fid

    • TF seemed to provide a good implementation

    • zhihu: Frechet Inception Score (FID)

    • Explanation from Neal Jean