search

jmiller656 / EDSR-Tensorflow

Tensorflow implementation of Enhanced Deep Residual Networks for Single Image Super-Resolution
MIT License
330 stars 107 forks source link

What are the changes compared to the original paper? #40

Closed jennyzu closed 5 years ago

jennyzu commented 5 years ago

Hi,

Thank you for the code. Can you please elaborate on what are your changes compared to the original article? I didn't understand the explanation in the README. I am trying to understand what to change in the code, if I want the same implementation as in the paper.

luxuriance19 commented 5 years ago

Hi,

Thank you for the code. Can you please elaborate on what are your changes compared to the original article? I didn't understand the explanation in the README. I am trying to understand what to change in the code, if I want the same implementation as in the paper.

I think the change of the implementation is the first layer, the paper is sub_mean of the div2k data set, and the upsampler layer is changed, in the paper, first pixshuffler then conv, here is first conv, then shuffler. but when I change the code into the orignal paper setting. The network can't converge. I think it's a little weird here. Do you have other implementations?

jennyzu commented 5 years ago

No, this is the only implementation I found. That's why it's important for me to understand the difference and understand why the difference exists.

jmiller656 commented 5 years ago

Hey, here's a link to an implementation by the original author: https://github.com/LimBee/NTIRE201