# Generative or Contrastive
###### tags: `Reid paper survey`

# InfoMax-GAN: Improved Adversarial Image Generation via Information Maximization and Contrastive Learning
- https://openaccess.thecvf.com/content/WACV2021/papers/Lee_InfoMax-GAN_Improved_Adversarial_Image_Generation_via_Information_Maximization_and_Contrastive_WACV_2021_paper.pdf
* unknow
- Note that variational的思想就是針對某個分佈很難求解的時候，採用另外一個分佈來近似這個分佈的做法，並使用變分信息最大化(論文：The IM algorithm: A variational approach to information maximization)的方法求解變分下界 https://blog.csdn.net/winycg/article/details/105297089

* tip
- Contrastive learning
- 最大化編碼的局部和全局特徵之間的互信息
* Abstract
* nfoMax-GAN
- projection to high dim(Reproducing Kernel Hilbert Space
(RKHS) [2],)， which exploits the value of linear evaluation
in capturing similarities between the global and local
features
- 

# Self-supervised Learning: Generative or Contrastive
- https://zhizhou-yu.github.io/2020/06/26/Self-supervised-Learning-Generative-or-Contrastive.html

* 自監督表示學習利用輸入數據本身作為監督信號

* Self-Supervised vs un-Supervised
- 

# ContraGAN: Contrastive Learning for Conditional Image Generation
- https://arxiv.org/abs/2006.12681
* 一般GAN的z 試real image+noise?
* Tip
- 同一個batch下做contrastive
- An epoch is comprised of one or more batches
- 一般GAN (unsupervised)
- https://medium.com/%E9%9B%9E%E9%9B%9E%E8%88%87%E5%85%94%E5%85%94%E7%9A%84%E5%B7%A5%E7%A8%8B%E4%B8%96%E7%95%8C/%E6%A9%9F%E5%99%A8%E5%AD%B8%E7%BF%92-ml-note-generative-adversarial-network-gan-%E7%94%9F%E6%88%90%E5%B0%8D%E6%8A%97%E7%B6%B2%E8%B7%AF-c672125af9e6
- https://zhuanlan.zhihu.com/p/337723355
- Conditional GAN (supervised)
- G(pair train)
- Conditional Discriminator
- 除了要分辨真假，還要判斷c&x pair 有多合適
- 
- 對於生成器而言，也希望能夠分類正確，當時希望判別器不能正確分辨假數
據。

* Abstract
- ContraGAN的判別器判別給定樣本的真實性，並最大程度地減少對比目標，以學
習訓練圖像之間的關係。
- ContraGAN that considers data-to-class relations by using
a conditional contrastive loss

* Introduction
- motivated by ACGAN and ProjGAN，一般的GAN只考慮 data-to-class
- In this manner, the discriminator can capture not only
data-to-class but also data-to-data relations between
samples.
- 結合對比學習就是如何構造正負樣本的問題
* Background
- Generative Adversarial Networks
- a discriminator to synthesize realistic images
- whether the given images are synthesized or not,
- the generator (G) tries to fool the discriminator by
generating realistic images from noise vectors
- 
- D(x)表示D網路判斷真實圖片是否真實的概率，對於D來說，這個值越接近1
越好。
- D(G(z))是D網路判斷G生成的圖片的是否真實的概率，對於D來說，這個值
越接近0越好，對於G來說，這個值越接近1越好。
- GAN中有二個Neural Network需要去Train

- Conditional GANs
- synthesize realistic images is utilizing class label
information
* method
- preal(x) : ) is the real data distribution
- pz(z) is a predefined prior distribution, typically
multivariate Gaussian
- X : a randomly sampled minibatch of training images