GAN Generated Adversersial Network
Areas it can be used in:
1. image manipulation and generation:
example: Generative Visual Manipulation on the Natural Image Manifold
2. Facial image processing/manipulation:
3. Image coloring:
Deep Color: http://color.kvfrans.com/draw
4. Image-to-Image translation: Cycle GAN, real-time performance on videos.
affinelayer.com/pixsrv
5. image super-resolution restoration:
photo realistic single image super-resolution using a generative
context encoders: feawture learning by inpainting
From the aforementioned knowledge, what we can conclude from it?
Given observations we can conlude an educated estimation of its parameters, thus we can sample from this distribution and generate new samples.
To sum up , this problem can be boiled down to concluding a sample distribution from samples and generate more samples from this educated sample distribution.
这类问题可以用生成模型解决。
生成模型可以分为三类:
1.完全可见信念网络 Fully Visible Belief Networks
2. 变分自编码器 variational autoencoder
3. 生成式对抗网络 generative adversersial network
VA可以和GAN结合起来改善效果。
对抗是生成网络有判别器和生成器构成。
生成器用来生成一个可以以假乱真的sample,
判别器尝试区分真假样本。
这样一个网络的训练过程遵循min max L 原则,即先最大化判别器的loss,再极小化生成器的loss。
总而言之,先训练判别器,再训练生成器。
JS, KL and Wd 是几个很重要的概念,理清楚这几个概念对理解gan的各种改进很重要。
W-GAN: 传统GAN优化的是JS散度,W-GAN更改了衡量标准。
LS-GAN, EB-GAN等等