R1 regularization

09.18.2024

Gradient Penalty라고도 불리나 봄.

 

Applying a gradient penalty in GAN training (especially in WGAN-GP) helps stabilize the training process by enforcing the Lipschitz continuity condition. This regularizes the discriminator’s gradients, ensuring that they remain well-behaved, which in turn helps the generator receive more meaningful updates. It addresses critical issues like mode collapse, vanishing/exploding gradients, and the problems with weight clipping in GANs.

from AMP (2021 SIGGRAPH) paper

 

---

글보다 해석이 훨씬 직관적이네. 3D-aware GAN 읽다보니 자주 표현하는 텀이었음.

https://ai.stackexchange.com/questions/25458/can-someone-explain-r1-regularization-function-in-simple-terms

Can someone explain R1 regularization function in simple terms?