questions about model input and output about meta-sr-pytorch HOT 7 CLOSED

@yfzcopy0702 I have same question. And do you know why there is a mask? When I first read the paper, it doesn't appear in paper.

from meta-sr-pytorch.

Good question!!
If we provide mask.mat for the forward function, maybe it can avoid this problems. I will spend some time to test the time study.
And thank your for my advice.

from meta-sr-pytorch.

@yfzcopy0702 my 1st idea is do like what you said, but it will slower than XuecaiHu's version. because mask will do in inchannels(maybe 64) but not 3 channels

from meta-sr-pytorch.

@HymEric Actually, the model super-resolves the LR inputs with integer upscaling factors， due to the mask, the model is capable of solving super-resolution of arbitrary scale factor (including noninteger scale factors) with a single model.

from meta-sr-pytorch.

@ZhangDY827 guy, your statement is false. We use the mask to implement the meta-upscale module with Matrix operations instead of using "for" loop in python.

from meta-sr-pytorch.

@XuecaiHu Hello, I have reproduced your paper recently and thanks for your great work firstly! In your metardn.py file, you use "scale_int = math.ceil(self.scale)" to transfer the non-integer upscaling factor to integer factor. So, in my opinion, the size of output HR image is the integral. The mask matrix contains 0 and 1 to sample the useful pixel through "torch.masked_select" function. Is my understanding correct?

from meta-sr-pytorch.

@ZhangDY827 Maybe you should try to implement the meta-upscale module yourself. And you will understand why i use the mask. We add the unmeaning pixels which make the upscaled feature maps suitable for matrix operations. And use the mask to remove the unmeaning pixels to generate the needed SR image. While, if you use two "for" loops to implement our meta-upscale layer as the Pseudo code in our paper, you donot need the mask but it is very time-consuming. It is the implementation details and it is unneccesary to introduce in our paper.

from meta-sr-pytorch.