A Review on Generative Adversarial Networks: Algorithms, Theory, and Applications

A Review on Generative Adversarial Networks: Algorithms, Theory, and Applications

 

1 Inttroduction

  GANs由两个模型组成:生成器和鉴别器。生成器试图捕获真实示例的分布,以便生成新的数据样本。鉴别器通常是一个二值分类器,尽可能准确地将生成样本与真实样本区分开来。GANs的优化问题是一个极大极小优化问题。优化终止于相对于生成器的最小值和相对于鉴别器的最大值的鞍点。

 

2.1 Generative algorithms

  生成算法可分为两类:显式密度模型和隐式密度模型。

 

2.1,1 Explicit density model

  显式密度模型假设分布,利用真实数据训练包含分布或拟合分布参数的模型。完成后,使用所学习的模型或分布生成新的示例。

 

2.1.2 Implicit density model

  隐式密度模型不能直接估计或拟合数据分布。它在没有明确假设[101]的情况下从分布中生成数据实例,并利用生成的实例修改模型。GANs属于有向隐式密度模型范畴。

 

3 Algorithm

3.1 Generative Adversarial Nets (GANs)

3.1.1.1 Original minimax game:

A Review on Generative Adversarial Networks: Algorithms, Theory, and Applications

 

 

3.2 GANs’ representative variants

3.2.1 InfoGAN

https://zhuanlan.zhihu.com/p/55945164

A Review on Generative Adversarial Networks: Algorithms, Theory, and Applications

 

 从损失函数的角度来看,infoGAN的损失函数变为:A Review on Generative Adversarial Networks: Algorithms, Theory, and Applications

 

 3.2.2 Conditional GANs (cGANs)

https://blog.csdn.net/taoyafan/article/details/81229466

Conditional GAN 结构图:

A Review on Generative Adversarial Networks: Algorithms, Theory, and Applications

 

 判别网络两种形式:A Review on Generative Adversarial Networks: Algorithms, Theory, and Applications

 

ACGAN (Auxiliary Classifier GANs):

https://zhuanlan.zhihu.com/p/91592775

A Review on Generative Adversarial Networks: Algorithms, Theory, and Applications

 

 3.2.3 CycleGAN

https://www.jianshu.com/p/5bf937a0d993

 

3.3.3.6 BigGANs and StyleGAN:

StyleGAN:

https://zhuanlan.zhihu.com/p/62119852

A Review on Generative Adversarial Networks: Algorithms, Theory, and Applications

 

A Review on Generative Adversarial Networks: Algorithms, Theory, and Applications

上一篇:Android config Android.mk parameter


下一篇:[Debug]IOS微信浏览器不支持form表单的target=_blank