目录

• 概
• 主要内容

Zhang H., Zhang Z., Odena A. and Lee H. CONSISTENCY REGULARIZATION FOR GENERATIVE ADVERSARIAL NETWORKS. ICLR, 2020.

Zhao Z., Singh S., Lee H., Zhang Z., Odena A. and Zhang H. Improved Consistency Regularization for GANs. AAAI, 2020.

让GAN训练稳定的方法主要有normalization 和 regularization.

这两篇文章介绍了 consistency regularization.

如上图所示, \(T\)是augmentation,

CR-GAN的思路是, 希望\(D(T(x)), D(x)\)彼此接近,

bCR-GAN在此基础上, 还希望\(D(G(z)), D(T(G(z)))\)也彼此接近.

zCR-GAN则是将\(T\)直接作用在\(z\)上:

1. \(G(z), G(T(z))\)彼此远离, 即增加多样性;
2. \(D(G(z)), D(G(T(z)))\)彼此靠近, 即生成的图片应该有共同的主体特征.

至于ICR-GAN, 是bCR和zCR的结合.

注: 如果\(z\)是隐向量, \(T\)采取高斯噪声\(T(z) \sim \mathcal{N}(z, \sigma_{noise})\).

注: 远离和靠近的度量, 文中采用的是

\[\|\cdot \|^2.
\]