英飞

Ganlecture9 squence generation

2022/07/11 GAN 共 523 字，约 2 分钟

GAN Lecture 9 (2018): Sequence Generation

问题

image-20231004195730630

回顾

image-20231004200206494

Improving Supervised Seq-to-seq Model

方法1 RL

image-20231004200309937

image-20231004200424315

image-20231004200612231

回顾 policy gradient

image-20231004201245600

image-20231004201537146

直觉解释

image-20231004201707978

实际应用

image-20231004202141891

对比

image-20231004202529959

Alpha Go style training

image-20231004202901848

方法2 GAN

image-20231004203006715

Algorithm

image-20231004203657514

sequence model

image-20231004203912250

文字生成 是一串 token ,没法微分，

怎么求解这个问题

image-20231004204114224

方法1 Gumbel-softmax

一个trick 使得不能微分能够微分

方法2

image-20231004204223924

存在问题

image-20231004204422717

方法3 RL

image-20231004204813453

image-20231004204949382

存在问题

image-20231004205328130

采样不够多

Solution

image-20231004205510409

方法3

image-20231004205610331

image-20231004205647257

More Application

image-20231004205901818

Unsupervised Conditional Sequence Generation

image-20231004210021003

1 Text Style Transfer

image-20231004210050622

做法1 Direct Transformation

image-20231004210246075

把word换成对应的embeding

做法2 Projection to Common Space

image-20231004210506829

2 Unsupervised Abstractive Summarization

image-20231004210924291hh

回顾

把文章和摘要当作不同的domain

image-20231004211030101

做法—类似cycle GAN

image-20231004211435150

另一个角度理解

image-20231004211602883

image-20231004211709011

Unsupervised Machine Translation

image-20231004212143943

Unsupervised Speech Recognition

image-20231004212343190

文档信息

Search

    Table of Contents