Ganlecture7 feature extraction

2022/05/11 GAN 共 851 字,约 3 分钟

GAN Lecture 7 (2018): Feature Extraction

https://colab.research.google.com/github/zuti666/generative-models/blob/master/Feature_Extraction.ipynb

Info GAN 要处理的问题

input 和 output 不同维度上没有什么关系,不是简单的确定关系

image-20231004181426978

Info GAN 是什么

image-20231004181819014

c 的维度代表某些特征

[1606.03657] InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets (arxiv.org)

image-20231006202100728

image-20231006205416967

image-20231006205720267

image-20231006200407995

VAE -GAN

image-20231004182527526

image-20231006210220678

image-20231006210233830

Algorithm

image-20231004182949821

image-20231006210246984

公式

image-20231006210354131

image-20231006210511349

image-20231006210532347

Autoencoding beyond pixels using a learned similarity metric (mlr.press)

BiGAN /ALi

image-20231004183116691

Algorithm

image-20231004183351746

原理

image-20231004183523389

只learn d/e 与学习 autoencodeer 有什么区别

image-20231004183752725

[BiGAN] Adversarial Feature Learning (arxiv.org)

encoder 和 decoder 联手骗过 Discrimniator

image-20231004184624047

image-20231007114039761

image-20231004184644231

Triple GAN

image-20231004190037023

Triple Generative Adversarial Nets (neurips.cc)

少量label data ,大量 unlabel data , semi-supervised learning

目标是学习一个好的 classifier

image-20231004190207829

公式

image-20231007120048135

image-20231007120108874

image-20231007120536075

Domain-adversarial training

抽取特征

image-20231004185043528

image-20231004185241798

Domain-adversarial

Feature Disentangle

Encoder 和 Decodeer 中间抽取的特征并不一定是自己想要的信息,并且不清楚特征对应什么

我们想要对应不同东西对应的特征

image-20231004185636668

怎么做—1控制不同特征

image-20231004185724281

怎么做 ,2训练一个分类器 GAN的思想

image-20231004185801904

Speaker-Invariant Training Via Adversarial Learning

image-20231004192009727

文档信息

Search

    Table of Contents