Conditional Gan Summary

Worked on conditional GAN (cGAN) and image inpainting projects for a while. Here is a brief summary of the experience, thinking and lessons.

Inspired a lot by this blog!

Generative modeling is an unsupervised learning task in machine learning that involves automatically discovering and learning the patterns in input data in such a way that the model can be used to generate or output new examples that plausibly could have been drawn from the original dataset.

I also see the Discriminator as a special loss function learnt by neural network to measure the difference between two distributions.

---

DCGAN

DCGAN is the first GAN using CNN. Some good intro link and link

---

WGAN

WGAN is based on the fact that JS divergence is not a good measurement of two distribution difference, especially for high dimension data. However the tweak is very easy:

• In the D, remove the sigmoid from the last layer
• take the loss instead of log(loss) for both G and D
• for parameters in D, do weight clipping
• instead of momentum or Adam, RMSProp and SGD are recommended.

Good intro link, link

SNGAN

Weigth clipping is sometimes too brutal to satisfy the 1-Lipschitz restriction. Gradient clipping was an alternative way. Similar to WGAN, spectral norm is later introduced as a way to satisfy the 1-Lipschitz restriction.

Modulation GAN

This is related to AdaIn

Good intro link