BIGGAN: LARGE SCALE GAN TRAINING FORHIGH FIDELITY NATURAL IMAGE SYNTHESIS

BackGround

1. GAN的训练对其各方面的设置（参数、模型）都很敏感

    

Motivation

1. 结合以外众多稳定GAN训练的方法来构建一个庞大的稳定的GAN

    

Model

Baseline：SAGAN[1]

Technique

1. Hinge loss
2. Class-conditional BatchNorm[1]，BN后的系数γ、β由condition通过全连接（每个BN共用FC层）后提供
3. Discriminator with projection[3]
4. Weights decay of Generator
5. Orthogonal Initialization
6. Truncation trick

conclusion

用大量的技术去稳定了一个拥有大量参数的cGAN的训练

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.

[2] Vries, Harm de, Florian Strub, Jérémie Mary, Hugo Larochelle, Olivier Pietquin, and Aaron Courville. "Modulating Early Visual Processing by Language." ArXiv:1707.00683 [Cs], December 18, 2017. http://arxiv.org/abs/1707.00683.

[3] Miyato, Takeru, and Masanori Koyama. "CGANs with Projection Discriminator." ArXiv:1802.05637 [Cs, Stat], August 14, 2018. http://arxiv.org/abs/1802.05637.

BIGGAN: LARGE SCALE GAN TRAINING FORHIGH FIDELITY NATURAL IMAGE SYNTHESIS

原文：https://www.cnblogs.com/JunzhaoLiang/p/13886458.html