7月28日深度学习笔记——GAN

---

前言

本文为7月28日深度学习笔记，分为两个章节：

• Feature Extraction：InfoGAN、VAE-GAN、BiGAN；
• Intelligent Photo Editing：GAN + Autoencoder.

---

一、Feature Extraction

The colors represents the characteristics.

1、InfoGAN

2、VAE-GAN

(1)、Algorithm

• Intialize Encoder, Decoder, and Discriminator;
• In each iteration:
  • Sample M images { x 1 , x 2 , … , x M } \{x^1, x^2, …, x^M\} {x1,x2,…,xM} from database;
  • Generate M codes { z ~ 1 , z ~ 2 , … , z ~ M } \{\tilde{z}^1, \tilde{z}^2, …, \tilde{z}^M \} {z~1,z~2,…,z~M} from encoder: ${\tilde{z}}^i=En(x^i)$;
  • Generate M images { x ^ 1 , x ^ 2 , … , x ^ M } \{\hat{x}^1, \hat{x}^2, …, \hat{x}^M \} {x^1,x^2,…,x^M} from decoder: ${\hat{x}}^i=De({\tilde{z}}^i)$;
  • Sample M codes { z 1 , z 2 , … , z M } \{z^1, z^2, …, z^M \} {z1,z2,…,zM} from prior $P(z)$;
  • Generate M images { x ^ 1 , x ^ 2 , … , x ^ M } \{\hat{x}^1, \hat{x}^2, …, \hat{x}^M \} {x^1,x^2,…,x^M} from decoder: ${\hat{x}}^i=De(z^i)$;
  • Update En to decrese $||{\tilde{x}}^i-x^i||$, decrease $KL(P({\hat{z}}^i|x^i))||P(z)$;
  • Update De to decrese $||{\tilde{x}}^i-x^i||$, increse $Dis({\tilde{x}}^i)$ and $Dis({\hat{x}}^i)$;
  • Update Dis to increase $Dis(x^i)$, decrese $Dis({\tilde{x}}^i)$ and $Dis({\hat{x}}^i)$.

3、BiGAN

(1)、Algorithm

• Intialize Encoder, Decoder, and Discriminator;
• In each iteration:
  • Sample M images { x 1 , x 2 , … , x M } \{x^1, x^2, …, x^M\} {x1,x2,…,xM} from database;
  • Generate M codes { z ~ 1 , z ~ 2 , … , z ~ M } \{\tilde{z}^1, \tilde{z}^2, …, \tilde{z}^M \} {z~1,z~2,…,z~M} from encoder: ${\tilde{z}}^i=En(x^i)$;
  • Generate M images { x ^ 1 , x ^ 2 , … , x ^ M } \{\hat{x}^1, \hat{x}^2, …, \hat{x}^M \} {x^1,x^2,…,x^M} from decoder: ${\hat{x}}^i=De({\tilde{z}}^i)$;
  • Sample M codes { z 1 , z 2 , … , z M } \{z^1, z^2, …, z^M \} {z1,z2,…,zM} from prior $P(z)$;
  • Generate M codes { x ~ 1 , x ~ 2 , … , x ~ M } \{\tilde{x}^1, \tilde{x}^2, …, \tilde{x}^M \} {x~1,x~2,…,x~M} from decoder: ${\tilde{x}}^i=De(z^i)$;
  • Update Dis to increase $Dis(x^i,{\tilde{z}}^i)$, decrease $Dis(x^i,z^i)$;
  • Update En and De to decrease $Dis(x^i,{\tilde{z}}^i)$, increse $Dis({\tilde{x}}^i,z^i)$.

---

二、Intelligent Photo Editing

1、GAN + Autoencoder

(1)、Attribute Representation

$$z_{long}\frac{1}{N_1}\sum _{x\in long}En(x)-\frac{1}{N_2}\sum _{x^{\prime }\notin long}En(x^{\prime })$$

(2)、Basic Idea

(3)、Back to z

• Methode 1:
  $$z^{\ast }=argminL(G(z),x^T)$$

• Methode 2:
  

• Methode 3:
  Using the results from method 2 as the initialization of method 1.

(4)、Editing Photos

$z_0$ is the code of the input image.
$$z^{\ast }=argminU(G(z))+{\lambda }_1||z-z_0|{|}^2-{\lambda }_{2}D(G(z))$$

---