Coupled Generative Adversarial Networks
M.Liu and O.Tuzel, NIPS 2016.
code
COGAN learns the joint distribution of multi-domain images without any corresponding images.This is done by sharing a subset of parameters during training which helps GANs to synthesize pairs of corresponding images without supervision. This paper deals with the case of two image domains. Image x1 and x2 were drawn from the marginal distribution of the 1st and 2nd domain respectively. Two GANs one for each domain maps a random vector input z to images.
Generative models decode information from more abstract concept to material details where the first layer decode high level semantics and last layer decodes low-level details. The information flow is opposite incase of discriminative networks. The first layers of GAN have identical structure and share the weights and the last layers of discriminators are same with sharing of weights. COGANs is trained by back propagation with alternating gradient updating.
Coupled GANs are used in applications like : unsupervised domain adaptation (UDA) and cross-domain image transformation.
COGAN learns the joint distribution of multi-domain images without any corresponding images.This is done by sharing a subset of parameters during training which helps GANs to synthesize pairs of corresponding images without supervision. This paper deals with the case of two image domains. Image x1 and x2 were drawn from the marginal distribution of the 1st and 2nd domain respectively. Two GANs one for each domain maps a random vector input z to images.
Generative models decode information from more abstract concept to material details where the first layer decode high level semantics and last layer decodes low-level details. The information flow is opposite incase of discriminative networks. The first layers of GAN have identical structure and share the weights and the last layers of discriminators are same with sharing of weights. COGANs is trained by back propagation with alternating gradient updating.
Coupled GANs are used in applications like : unsupervised domain adaptation (UDA) and cross-domain image transformation.