search

Maclory / SPSR

Pytorch implementation of Structure-Preserving Super Resolution with Gradient Guidance (CVPR 2020 & TPAMI 2021)
448 stars 83 forks source link

Why are the input and output of the gradient branch three channels instead of single? #25

Closed Feihong-cc closed 3 years ago

Feihong-cc commented 3 years ago

Hi, Why are the input and output of the gradient branch three channels instead of single?

Maclory commented 3 years ago

It's just for simplicity. In my opinion, two implementations are similar intrinsically.

Feihong-cc commented 3 years ago

Can I communicate with you in Chinese?（哈哈，我能和你用中文交流吗？）我有些疑问，SPSR的梯度分支本质上应该是求图像的梯度，按照数字图像处理的方法，应该是将彩色图转为灰度图，再求其的梯度图才是正常的流程。你在SPSR代码里面，是将RGB三个通道分别拆开，分别求其梯度图，这样每个通道的梯度图不一定一样，最后再concat到一起，会不会导致图像的边缘信息（高频信息）更凌乱or模糊？请问你有测试过单通道和三通道分别求的梯度图，有什么差别吗？

Maclory commented 3 years ago

concat的话能保留更多的信息，一般不会导致信息变差。两种实现我们也都试过，差别很小。

Feihong-cc commented 3 years ago

(⊙o⊙)~太棒了，我又可以省去做验证的时间了！十分感谢你~

Maclory commented 3 years ago

哈哈哈，不客气

Feihong-cc commented 3 years ago

hi！你好鸭，我有个小问题想请教你，你的SR分支和梯度分支的loss是相加起来，loss=sr_loss+grad_loss，这样同时求反向传播和分别backward应该是存在差别的，你有试过分开backward吗？

Feihong-cc commented 3 years ago

如果我需要将两个loss分开backward的话，需要两个optimizer吗？

Maclory commented 3 years ago

我觉得怎么实现都可以，合理就行

Feihong-cc commented 3 years ago

好的，十分感谢你~