chainer-fast-neuralstyle icon indicating copy to clipboard operation
chainer-fast-neuralstyle copied to clipboard

Published 20 hours ago • verified publisher icon yusuketomoto

Why the output image substracts 120 before vgg forwarding, while the style image not?

Open crazylyf opened this issue 8 years ago • 3 comments

Hello, When I was diving into the code, I'm confusing why the output image of ITN network substract 120 at each pixel, while not for the style image. What is the reason behind this? Best,

crazylyf avatar Jul 27 '16 09:07 crazylyf

Style image also subtract 120 at each pixel. You can find: https://github.com/yusuketomoto/chainer-fast-neuralstyle/blob/master/train.py#L80 https://github.com/yusuketomoto/chainer-fast-neuralstyle/blob/master/net.py#L93

yusuketomoto avatar Jul 27 '16 12:07 yusuketomoto

Hi, to follow up on this, what exactly is 120? If its the mean pixel over all train images, shouldn't you be subtracting the mean image for preprocessing? @yusuketomoto @crazylyf

hashbangCoder avatar Aug 04 '16 16:08 hashbangCoder

Nevermind. I figured that its the mean pixel (as is needed) from the VGG documentation

hashbangCoder avatar Aug 09 '16 22:08 hashbangCoder