[Paper Review] 08.Improved Consistency Regularization for GANs
Contents
- Abstract
- Introduction
- ICR (Improved Consistency Regularization)
0. Abstract
increase GAN performance by..
\(\rightarrow\) forcing a consistency cost on the discriminator!
This paper …
- 1) shows that CR (Consistency Regularization) can introduce artifacts into GAN samples
- 2) propose several modifications to CR procedure
1. Introduction
CR-GAN
-
(1) real images & (2) corresponding augmented coutnerparts are fed into the DISCRIMINATOR
-
discriminator is encouraged to produce similar outpus for both!
( via auxiliary loss term )
Limitations of CR in CR-GAN
-
augmentations are only applied to real images, NOT to generated samples
( imbalanced )
-
regularize ONLY the discriminator
\(\rightarrow\) by constraining the mapping from the prior to the generated samples..
can achieve further performance gains!
ICR (Improved Consistency Regularization)
applies forms of consistency regularization to the ..
- 1) generated images ( bCR = balanced CR)
- 2) latent vector space ( zCR = latent CR )
- 3) generator
ICR = bCR + zCR
achieve SOTA ( best known FID scores on various GANs )
2. ICR (Improved Consistency Regularization)
Intuition of CR
-
encode some prior knowledge to model
( = model should produce consistence predictions )
Augmentations ( or transformations )
- ex) image flipping / rotating / sentence back-translating / adversarial attacks…
Penalizing inconsistencies via L2 loss, KL-div
