Gan network thought

GAN The network is composed of generators generator Discriminator discrimator form ：

• The generator inputs random noise of fixed length , Output fixed size image （ That is, fake pictures ）;
• Discriminator input image , The probability that the output image is a true image .

Connect the generator and the discriminator end to end to form a large network , The input is random noise , The probability that the output image is a true image （ This image can be generated by the generator , It can also be an image manually input to the discriminator ）.

The training process ：

• First, fix it generator , Input random noise , Input the false graph output by the generator and the same number of true graphs into the discriminator , The training tab corresponds to the input ： False picture -0, True picture -1; Training discriminator only , The result of training is that the discriminator can distinguish the true graph from the very poor false graph ;
• then , Hold on Judging device , Input noise to the generator , Send the generated false graph to the discriminator , The label of the discriminator is ： False picture -1, Training network （ Because the discriminator is fixed , Therefore, the training result is that the generator gradually improves its ability to fake , So that the discriminator that can only distinguish very poor false images cannot distinguish that the output image of the generator is a false image ）;
• Repeat the above two steps , Generators and discriminators iterate continuously to improve performance , Finally, the output image of the generator reaches the point that human eyes cannot distinguish .
  
  Image source ：https://www.leiphone.com/news/201706/ty7H504cn7l6EVLd.html