MSG-GAN: Multi-Scale Gradients for Generative Adversarial Networks
BackGround
- 基于流(flow-based)和自回归(autoregressive)的模型尚且未能与state-of-art GAN匹敌
- GAN虽然强大,但一直存在model collapse和training instability的问题
-
特别是在高分辨率(512/1024)图像生成任务上
Motivation
- progressive growing technique[1]虽然能够实现state-of-art的效果,但是该方法由于有众多的超参数需要调,因此十分难以训练
-
某些以往的方法(GMAN[2]、StackGAN[3, 4]、LapGAN[5]、Pix2PixHD[6])会在多个分辨率各布置一个Discriminator,各自有独立的下采样层,导致参数量剧增
Model
Technique
- 每个Discriminator的block的最前面使用MinibatchStdDev[1]
- 对每个随机向量进行hypersphere normalization[1],即使各元素平方和为向量维度。即ProGAN中的Pixel-wise Norm,但同时用于初始化z。
-
对模型的参数进行Kaiming/He Init
Conclusion
本文实现了一种简单、稳定、对Lr有鲁棒性的高分辨率图像生成方法,但显存消耗较大(训练256的Oxford Flowers用到了1xV100 32GB,1024的CelebA-HQ则用到了8xV100-16GB0)。Combine Function对实验结果会一定影响。
Reference
[1] Karras, Tero, Timo Aila, Samuli Laine, and Jaakko Lehtinen. "Progressive Growing of GANs for Improved Quality, Stability, and Variation." ArXiv:1710.10196 [Cs, Stat], February 26, 2018. http://arxiv.org/abs/1710.10196.
[2] Durugkar, Ishan, Ian Gemp, and Sridhar Mahadevan. "Generative Multi-Adversarial Networks." ArXiv:1611.01673 [Cs], March 2, 2017. http://arxiv.org/abs/1611.01673.
[3] Zhang, Han, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris Metaxas. "StackGAN: Text to Photo-Realistic Image Synthesis with Stacked Generative Adversarial Networks." ArXiv:1612.03242 [Cs, Stat], August 4, 2017. http://arxiv.org/abs/1612.03242.
[4] Zhang, Han, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris Metaxas. "StackGAN++: Realistic Image Synthesis with Stacked Generative Adversarial Networks." ArXiv:1710.10916 [Cs, Stat], June 27, 2018. http://arxiv.org/abs/1710.10916.
[5] Denton, Emily, Soumith Chintala, Arthur Szlam, and Rob Fergus. "Deep Generative Image Models Using a Laplacian Pyramid of Adversarial Networks." ArXiv:1506.05751 [Cs], June 18, 2015. http://arxiv.org/abs/1506.05751.
[6] Wang, Ting-Chun, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. "High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs." ArXiv:1711.11585 [Cs], August 20, 2018. http://arxiv.org/abs/1711.11585.