“`” 参考回答：
CycleGAN其实就是一个A→B单向GAN加上一个B→A单向GAN。两个GAN共享两个生成器，然后各自带一个判别器，所以加起来总共有两个判别器和两个生成器。一个单向GAN有两个loss，而CycleGAN加起来总共有四个loss。CycleGAN论文的原版原理图和公式如下，其实理解了单向GAN那么CycleGAN已经很好理解。
<img alt=""img"" referrerpolicy=""no-referrer"" src=""https://uploadfiles.nowcoder.com/images/20190318/311436_1552884579198_DEA8EA3F87FBBFD436227E74FB3FC8B0"">
<img alt=""img"" referrerpolicy=""no-referrer"" src=""https://uploadfiles.nowcoder.com/images/20190318/311436_1552884600879_3F1A546C067512AFCD7B936256784527"">
下面放一张网友们自制的CycleGAN示意图，比论文原版的更加直观，出处见水印。
<img alt=""img"" referrerpolicy=""no-referrer"" src=""https://uploadfiles.nowcoder.com/images/20190318/311436_1552884627496_B3727CA35DB639FED01C85BE930A4D24"">
<img alt=""img"" referrerpolicy=""no-referrer"" src=""https://uploadfiles.nowcoder.com/images/20190318/311436_1552884645667_BBBC197AAAF6B72372EC0FBE02BD26C9"">
<pre><code> "“`
Was this helpful?
0 /
0