This blog post mainly introduces GAN（Generative Adversarial Network）.

Generation

Why does the output sometimes need to be a distribution
The same input has different outputs.

algorithm



train generator$\iff$ train discriminator

Theory behind GAN

Divergence How is it calculated

Tips for training

Generator and Discriminator needs to match each other.
take GAN To generate text is more troublesome ：

It can't be used here gradient descent Under the circumstances , You can try reinforcement learning（RL）

It didn't work in the past GAN To produce text , Now? OK Of

Evaluation of Generation

Generator The evaluation method of ：

1. Quality of Image
   
2. Diversity-Mode Collapse
   When it comes to the back ,Generator Produced image It's always close
   
3. Diversity-Mode Dropping
   The generated samples are only concentrated in a certain group .
   
   
   
   

Conditional Generation

Now? Generator Not only to generate realistic images, But also with the input text（condition） Match .

Learning from Unpaired Data

Other style conversion GAN
StarGan You can switch between multiple styles