GAN section note two

conditional GAN:

text  ---> image

such as supervised approach ,lots of samples can not do average operation.

the works D does are evaluate the c and x

x is realistic or not , c and x are matched or not

matched are high scores

not matched low   images that is not real are low 

two case gave the low scores

in each iteration :

sample m positive example from database(word and image pair) high

sample m noise sample vecter from D obtain a condition

Discriminator 

object--->NN  condition--->NN then into NN --->output a scores

object x--->NN----->score

NN + condition ---> NN ---->match scores

one is to see object is real or not

one is to see both are matched or not

two task are divided into seperation 

stack GAN(thought that is divided into more)

divide into two

text---->G--->image(low Dimention)--->D----->match scores

|

that image--->G----->image(high Dimention)---->G---->reality score

image ---->image

#supervised approach  vagurous(average)

GAN: noise + image --> G(image)---->pair---->D

two goals: reality and match

patch GAN:

adjust patch size

speech enhancement

typical deep learning approach 

use CNN 

evaluation: real and match

vedio generator

D 's input is G's input and output match or not

thoughts:

conditional GAN

引入了条件控制的观念  给与人类能理解的输入的到人们想得到的输出。这是非常奇妙的。

引入条件匹配的概念

学习级联的思想 产出大规模的输出 那么先从小规模输出开始 级联分配任务 前级查匹配 后级查清晰度

级联评估的思想：评估清晰度与条件匹配时，分开两个D，一个D一个分数，用两个分数进行调整G，调整更加有针对性，调整第一项或者第二项就更有目标。

可以利用GAN 技术自动生成后续电影。

生成歌曲。

作诗。

生成音频动画。

GAN生成电影还是蛮有挑战性的。