Longguang Wang

Hi @zhangjinyangnwpu, input LR images with ranges [0, 1] and [0, 255] are both okay for SR networks. In this work, we follow EDSR and RCAN to use images with...

Hi @sunyclj，看报错信息应该是cuda版本和显卡不匹配导致的。

Hi @zyl15936880370, 请参照readme文件按顺序操作即可。

Hi @dongling23, 我们是选定一张LR图片，然后在图片内随机扣取一些48*48图像块，送入encoder得到degradation representation，并将得到的representation和LR图片一起送到后面的超分辨网络中进行重建，最后统计图形块内的平均梯度与PSNR的关系。关于梯度我们就是简单计算了(dx^2+dy^2)^0.5。我们在多张图片上都进行了测试，整体上来看PSNR是跟梯度呈现正相关关系，我们在文中汇报了一个比较明显的图片场景，但是也有些图片上这一规律没有这么明显。

Hi @happy20200, thanks for your ineterest in our work. In our network, the learning of degradation representation is unsupervised while the learning of SR still requires the supervision from LR-HR...

Hi, please see `/experiment`.

Hi @1632325673，感谢对我们工作的关注。请在[这里](https://github.com/The-Learning-And-Vision-Atelier-LAVA/ArbSR/blob/master/utils/gen_test_data.m#L27)修改相应的放大倍率(比如你想要做x1.6放大，请把scale和scale2都设置为1.6)，然后就可以在对应路径下得到相应的LR图像了。

Hi @Macro03, thanks for your interest in our work. In our implementation, the LR images are generated online during training thus you can safely remove `self.dir_lr`.

Hi @Macro03, we also have this observation in our experiments on synthetic images. In my opinion, at early epochs, the [random samples](https://github.com/LongguangWang/DASR/blob/main/moco/builder.py#L34) in the queue is gradually replaced with the...

Hi @zyf1040895256, 对比实验里不同方法输入的LR图像都是用`trainer.py`里的[代码](https://github.com/LongguangWang/DASR/blob/main/trainer.py#L155)生成的。