Interesting Paper
Reference:
github domain-transfer-network
DTN:Unsupervised Cross-Domain Image Generation

Reference:
overview of CycleGAN
Reference:
Tips for TF CycleGAN
Cycle GAN: no label no pairing, two generator, two Discriminator

what is GAN?
overview of GAN
generative network learns to map from a latent space
discriminative network discriminates between instances from the true data distribution and candidates produced by the generator.
so that means:
generator increase the error rate of the discriminative network

by the way:
variational autoencoders (VAE) is also a way to generate the images without supervision.

GAN VS VAE

because of the limited time, so I only study GAN hier:

Loss Function of Generator

the part of fake data from generator, which is judged as real by Discriminative network

Loss Function of Discriminative network

the part of fake data from generator, which is judged as fake by Discriminative network as well as the part of real data, which is judged as real by Discriminative network

note that: the loss function is in form of cross entropy

some new GAN:

LSGAN

LSGAN

WGAN

WGAN

ps: WGAN-GP and DRAGAN also use clip by value, but on gradient.

SGAN

SGAN

now go back to Cycle GAN

New GAN

the biggest difference between traditional GAN and new GAN is that, the generator do not generate fake image just from the noise signal, but from the feature of the original image.

we already know Conv layer can separate higher-order features from original image. And then Transpose Conv layer can in turn generate ein image based on the features.
So that we use a number of Conv- layers to build a encoded and transposed Conv layers to build a decoder.
ans we use a deep network to ** make changes ** to act as a feature converter!
and Cycle GAN use ResNet50 as the feature converter!

Details of Cycle GAN

Details of Cycle GAN

according to the graph, we know there are 2 generator. G Net A2B is used as generating Class B from A, B2A is in turn.
and 2 Discriminator in the same way.