BackGround

1. GAN的训练对设置（结构、参数）过于敏感

 

Motivation

1. 通过结合众多稳定GAN训练的方法来实现一个大型的稳定GAN

 

Model

Baseline：SAGAN[1]

 

Technique：

1. Weight decay of G - 防止过拟合
2. Orthogonal Initialization
3. Truncation trick - resample z
4. Ways to sample z - Bernoulli{0, 1} max(N(0, i), 0)

 

Reference

[1] Zhang, Han, Ian Goodfellow, Dimitris Metaxas, and Augustus Odena. "Self-Attention Generative Adversarial Networks." ArXiv:1805.08318 [Cs, Stat], June 14, 2019. http://arxiv.org/abs/1805.08318.