vincent-hoo / knowledge-distillation-for-super-resolution Goto Github PK

你好~FAKD这篇工作的论文可以分享出来吗?

I see there are several method called MSDataLoader and _MSDataLoaderIter you defined in "dataloader.py", which _MSDataLoaderIter is inherited from _DataLoaderIter in pytorch v1.1.0. However, as we know there is no "_DataLoaderIter " after pytorch v1.2.0, I tried to change it to derived from "_MultiProcessingDataLoaderIter" but failed.

I tried to install pytorch 1.1.0 but it's too old to be installed and not compatible with other libs.

Could you give me a clue how to change the code to adapt newer version of pytorch? Thanks

Hi.

How did you estimate Flops(G)??

Can you share any code or library name which you used??

Hi Vincent,

This repo seems super interesting! Would you mind sharing your paper? I would love to read it! My email is [email protected]. Thanks very much!

What`s resolution of input image for measuring FLOPs?

需要读取权限

同学你好，我是清华深研院MEAI2020级的同学，我对你的这份工作非常感兴趣，想follow你的工作。不知道能不能交个朋友，我们一起搞点新的东西。我之前有发过顶会的。

I am interesting wheter video super resolution like EDVR could be applied using your distillation work?

Thx,
Lei

I am unable to download the tar file for the testing data. If it is not possible to share, can you help me understand how I can prepare the test data on my own?

Thanks.

very thanks for your job! and I still have a little question, in fakd, even if the feat map size is [256x256], the size of affinity matrix is [1,1,65536,65536], and only batch size =1 (1 teacher feat and 1 student feat) will use almost 33 GiB memory! How to run the method correctly?

def forward(self, x):
feature_maps = []
x = self.sub_mean(x)
x = self.head(x)
feature_maps.append(x)

    res = self.body1(x)
    feature_maps.append(x)  <===  SHOULD BE feature_maps.append(res) 
    res = self.body2(res)
    feature_maps.append(x)  <===  SHOULD BE feature_maps.append(res)
    res = self.body3(res)
    feature_maps.append(res)
    
    res += x
    

    x = self.tail(res)
    x = self.add_mean(x)

    return feature_maps, x