Skip to content

Latest commit

 

History

History
143 lines (65 loc) · 3.55 KB

wgan.md

File metadata and controls

143 lines (65 loc) · 3.55 KB
Raw

WGAN

⌚️: 2021年8月30日

📚参考

一、WGAN

image-20210830164832526

1. GAN的局限性

image-20210830164944257

image-20210830165154779

image-20210830165408725

image-20210830165516967

1.1 模式崩塌

image-20210830165741914

image-20210830165908651

image-20210830165945978

1.2 过生成

image-20210830170111772

image-20210830170235516

image-20210830170253444

1.3 梯度消失

image-20210830170607017

image-20210830170733507

image-20210830170835791

image-20210830170942748

分布一样为1，分布不一样为0，只能判断两个分布是否相似，但是不能判断程度。

1.4 Mode Dropping

image-20210901154736642

2. WGAN

image-20210830171113266

image-20210830171439547

image-20210830171511566

image-20210830171613802

image-20210830171711784

image-20210830171842077

image-20210901151332825

1-L限制条件的实现有很多种，下面的图展示了原来WGAN和WGAN-GP的实现原理

image-20210901151809724

image-20210830171950750

image-20210830172116317

神经网络函数有很多不满足1-L条件，所以梯度裁剪是解决方式之一。

3. WGAN-GP

clip范围需要自己设定，而且很难设置，所以有了WGAN-GP

image-20210830172506988

image-20210830172516909

image-20210830172652739

image-20210830172737627

image-20210830173040055

image-20210830173113242

image-20210830173200067

二、WGAN实现

image-20210830173557486

image-20210830173557486

image-20210830174042772

image-20210830174446621

image-20210830174506938

image-20210830174530389

image-20210830174638333

image-20210830174807793

image-20210830174905100

image-20210830175129708

image-20210830175233265

image-20210830175257955

image-20210830175523529