深度学习-【draft】生成对抗网络
1. 介绍GAN
- GAN的基本思想
- 为什么生成器不自己学？
- 为什么判别器不自己做?
- 具体算法
- 笔记： 李宏毅学习笔记30.GAN.01. 李宏毅GAN教程（1）
2. Gan的数学原理（GAN背后的理论）
3. Conditional GAN (条件GAN)
研讨厅思路：
背景意义
Generative Adversarial Network（GAN）被引次数：26083次
Ian J. Goodfellow:phD
https://github.com/hindupuravinash/the-gan-zoo
GAN的应用
GAN的原理
调节Generator和Discrimnator的训练次数比。一般来说，Discrimnator要训练的比Genenrator多。比如训练五次Discrimnator，再训练一次Genenrator(WGAN论文 是这么干的)。这一条不一定对！
GAN与VAE的比较
GAN的发展
http://nooverfit.com/wp/独家｜gan大盘点，聊聊这些年的生成对抗网络-lsgan-wgan-cgan-info/
10个必读的GAN
WGAN，DCGAN，CGAN,Improved Techniques for Training GANs
优化GAN的方法
（1) 结构上的改进CGAN
（2）除了结构上的改进还有，loss, 模型初始化和权重上的改进
（3）GAN的重要实现
目前各领域最先进的GAN
GAN研究方向汇总（附源码） - 清华阿罗的文章 - 知乎 https://zhuanlan.zhihu.com/p/69305310
Level 0: Definition of GANs
| Level | Title | authors | Publication | Links |
|---|---|---|---|---|
| Beginner | GAN : Generative Adversarial Nets | Goodfellow & et al. | NeurIPS (NIPS) 2014 | link |
| Beginner | GAN : Generative Adversarial Nets (Tutorial) | Goodfellow & et al. | NeurIPS (NIPS) 2016 Tutorial | link |
| Beginner | CGAN : Conditional Generative Adversarial Nets | Mirza & et al. | -- 2014 | link |
| Beginner | InfoGAN : Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets | Chen & et al. | NeuroIPS (NIPS) 2016 |
Level 1: Improvements of GANs training
然后看看 loss、参数、权重的改进：
| Level | Title | Co-authors | Publication | Links |
|---|---|---|---|---|
| Beginner | LSGAN : Least Squares Generative Adversarial Networks | Mao & et al. | ICCV 2017 | link |
| Advanced | Improved Techniques for Training GANs | Salimans & et al. | NeurIPS (NIPS) 2016 | link |
| Advanced | WGAN : Wasserstein GAN | Arjovsky & et al. | ICML 2017 | link |
| Advanced | WGAN-GP : improved Training of Wasserstein GANs | 2017 | link | |
| Advanced | Certifying Some Distributional Robustness with Principled Adversarial Training | Sinha & et al. | ICML 2018 | linkcode |
Level 2: Implementation skill
GAN的实现
| Title | Co-authors | Publication | Links | size | FID/IS |
|---|---|---|---|---|---|
| Keras Implementation of GANs | Linder-Norén | Github | link | ||
| GAN implementation hacks | Salimans paper & Chintala | World research | linkpaper | ||
| DCGAN : Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks | Radford & et al. | 2015.11-ICLR 2016 | linkpaper | 64x64 human | |
| ProGAN:Progressive Growing of GANs for Improved Quality, Stability, and Variation | Tero Karras | 2017.10 | paperlink | 1024x1024 human | 8.04 |
| SAGAN：Self-Attention Generative Adversarial Networks | Han Zhang & Ian Goodfellow | 2018.05 | paperlink | 128x128 obj | 18.65/52.52 |
| BigGAN:Large Scale GAN Training for High Fidelity Natural Image Synthesis | Brock et al. | 2018.09-ICLR 2019 | demopaperlink | 512x512 obj | 9.6/166.3 |
| StyleGAN:A Style-Based Generator Architecture for Generative Adversarial Networks | Tero Karras | 2018.12 | paperlink | 1024x1024 human | 4.04 |
GAN的应用 Level 3： GANs Applications
| 图像翻译 (Image Translation) | 超分辨率 (Super-Resolution) | 图像上色(Colourful Image Colorization) |
|---|---|---|
| 图像修复(Image Inpainting) | 图像去噪(Image denoising) | 交互式图像生成 |
JS散度在没有重合的时候，是常数log2，分辨不出来
LSGAN
WGAN
