GAN

• I want to try it once.

• https://www.imagazine.co.jp/gan%EF%BC%9A敵対的生成ネットワークとは何か%E3%80%80%EF%BD%9E%E3%80%8C教師/

  • 
  • z (noise) is like a seed value.
    • When you change the seed value, the image gradually changes.
    • It’s often seen in images generated by GAN.

Understanding the Basic Structure of GAN (Part 1) - Qiita

• It feels like the target readers are right in the middle.
• Discriminator
  • Well, this feels like a typical (?) classification problem with a Neural Network.
• Generator
  • It performs upsampling.
  • Instead of using reverse CNN, it simply reshapes the values.
    • Reverse CNN is used in something called DCGAN (blu3mo).
• Training
  • “Train only the Discriminator using real data.”
  • “Train the Generator using the Discriminator.”
  • They seem to alternate between these two steps.
• Loss Function
  • 
  • I don’t understand the meaning of $\mathbb{E}$.
  • In the training of the Discriminator, we want to maximize V(D, G).
    • We want to maximize $\log D(x)$ = make it output 1 when it’s real data (x).
    • We want to maximize $\log (1-D(G(z)))$ = make it output 0 when it’s fake data (G(z)).
  • In the training of the Generator, we want to minimize V(D, G).
    • We want to minimize $\log (1-D(G(z)))$ = make it output 1 when it’s fake data (G(z)).
    • In other words, we train in the opposite direction of the Discriminator.