Longguang Wang

Hi @chenquan-hdu, thanks for your interest in our work. For pair comparison with StereoSR, we follow its evaluation protocol and crop 64 pixels before calculating PSNR scores.

Hi @Sun-Wenfeng. 可以在main.py的主函数外部（比如第8行）加上这句代码：os.environ['CUDA_VISIBLE_DEVICES'] = "4,5,6,7"，然后n_GPUs设置为4就可以了。

Hi @Gideon1998, `generate_indices()` aims to produce the coordinates of "important" pixels marked by the spatial mask. With these coordinates, we can perform sparse convolution to compute the output at only...

Hi sunyclj. 理论上超分都应该具有稀疏性的特点，因此个人觉得稀疏模块应该也可以用在GAN或者其他超分网络中。我们在附加材料里尝试过应用于SRResNet等网络，但我们没有尝试过在GAN网络中进行应用，欢迎尝试。

Hi @Jackyinuo, sorry for the late response. It seems the reason is that `len(self.images_hr)` equals 0. Please re-check the training images are correctly organized and confirm `dir_data` (the parent path...

@ymtupup 你好，感谢对我们工作的关注。 报错应该是LR图片命名方式的问题。可以用matlab重新生成一次LR图片，并在保存图片时去掉命名中最后的'x3'；或者根据你现在的命名方式，可以将[这一行代码](https://github.com/LongguangWang/SMSR/blob/056c034c4c98172d6b7b504b5b33a9efff07d3dc/data/multiscalesrdata.py#L97)改成`self.dir_lr, 'X{}/{}x3{}'.format(int(s), filename, self.ext[1])`应该也可以解决这一问题。

@ymtupup 同学你好，我目前也是个博士生。可以考虑从最近的一些综述文章入手，了解超分辨领域近期的发展情况。 [1] Deep Learning for Image Super-resolution: A Survey, TPAMI 2019. [2] Video Super Resolution Based on Deep Learning: A comprehensive survey, arXiv 2020

Hi @wusuhuang, it is recommanded to change the pytorch version to 1.1.0. Or you can try the solution [here](https://github.com/LongguangWang/DASR/issues/22), maybe it can also solve this issue.

Hi @haikunzhang95, Model 1可以简单理解为所有的spatial和channel mask都是1，变成全通掩膜，其余结构没有任何变化。具体实现的话可以将代码中涉及mask的部分都去掉即可。

Hi @cheun726, 直接在cmd命令行里执行这句代码即可。