GAN Lecture 3 (2018) Unsupervised Conditional Generation

2022/01/12 GAN 共 1912 字,约 6 分钟

GAN Lecture 3 (2018): Unsupervised Conditional Generation

问题

风格转换任务,没有label ,只有两堆data,machine学习其中的转换

image-20231003223132108

两类做法

image-20231003223507212

1 直接做

差距小,例如颜色,纹理

2 先encoder抽特征,然后decoder根据特征生成

差距大

做法一

image-20231003223559429

问题,G生成梵高风格画作,可以骗过D ,但是所生成的结果与原来的输入X无关

D 判别是否符合梵高风格 ,G 任务不仅仅是生成风格类似的结果,还要保持输入输出一致

做法1 无视这个问题,直接做

有可能work的原因,G的输入和输出差不多

[1709.00074] The Role of Minimal Complexity Functions in Unsupervised Learning of Semantic Mappings (arxiv.org)

论文解释了,直接学习是有可能的,当输入和输出的domain比较接近的时候,一个浅的网络就能学到,但是当domain差距过大就难以实现了

当G很sallow,work

做法2 找一个pre-train的network

image-20231003224248786

[Yaniv Taigman] Unsupervised Cross-Domain Image Generation (arxiv.org)

baseline of DTN :

两个风险:

风险1 :

image-20231004103336202

风险2: image-20231004103350485

用一个$f$函数对输入图像$x$和$G$生成的图像$G(x)$进行判别,要它们属于一个domain,两者的距离越小越好

DTN后续改进

image-20231004104151727

改进1

第一个改进是G的网络架构,在f的基础上添加一些层g

image-20231004104409486

改进2

第二个是,使用f来衡量x和生成结果g(f(x))的相似度,同时还要保证参考的图像不变

image-20231004104254753

image-20231004104358175

做法3 CYCLE GAN

image-20231003224920135

[Jun-Yan-Zhu] Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks (arxiv.org)

image-20231004131010937

image-20231004125423159

image-20231004125448576

image-20231004125514665

image-20231004125540206

image-20231005145821855

cycleGAN存在的问题

image-20231003225231202

[Casey Chu ] CycleGAN, a Master of Steganography (arxiv.org)

论文提到 cycle Gan 学习过程会学到藏东西

image-20231003225643414

DIsco GAN

Learning to Discover Cross-Domain Relations with Generative Adversarial Networks (mlr.press)

image-20231004130108040

image-20231004130921000

image-20231004130932727

image-20231004130942223

Dual GAN

[1704.02510v4] DualGAN: Unsupervised Dual Learning for Image-to-Image Translation (arxiv.org)

主要思想

image-20231004130748082

公式

image-20231005134141037

image-20231005134214469

image-20231005134229001

做法4 Star GAN

image-20231005195020408

image-20231005195044154

image-20231003230029942

image-20231003230159087

论文

CVPR 2018 Open Access Repository (thecvf.com)

公式

adversarial Loss

image-20231005195324992

Domain Classification Loss

image-20231005195436358

image-20231005195546687

Reconstruction Loss

image-20231005200447658

Full Object

image-20231005200522677

做法二

image-20231003230507079

问题:两者没有联系

image-20231003230630972

方法1 共享参数

image-20231003231048031

Couple GAN

Coupled Generative Adversarial Networks (neurips.cc)

image-20231005202121161

UNIT

Unsupervised Image-to-Image Translation Networks (neurips.cc)

方法2 加一个domain的discriminator

image-20231003231449189

[1711.00043] Unsupervised Machine Translation Using Monolingual Corpora Only (arxiv.org)

image-20231005205028172

image-20231005205212942

image-20231005205225996

公式

DENOISING AUTO-ENCODING

image-20231005205413081

CROSS DOMAIN TRAINING

image-20231005205435987

ADVERSARIAL TRAINING

image-20231005205536835

Final Objective function

image-20231005205555067

方法3 Cycle Consistency

image-20231003231742559

CVPR 2018 Open Access Repository (thecvf.com)

image-20231005212316433

image-20231005212153110

方法4 semantic consistency

image-20231003231832215

[1611.02200] Unsupervised Cross-Domain Image Generation (arxiv.org)

[XGAN: Unsupervised Image-to-Image Translation for Many-to-Many MappingsSpringerLink](https://link.springer.com/chapter/10.1007/978-3-030-30671-7_3)

image-20231005215131488

公式

Total Loss

image-20231005215405074

Reconstruction loss

image-20231005215438578

Domain-adversarial loss

image-20231005215502045

Semantic consistency loss

image-20231005215524332

generative adversarial loss

image-20231005215600410

Teacher loss

image-20231005215620444

文档信息

Search

    Table of Contents