alan-xw / realsr Goto Github PK

Hello, what dataset is used for training？

Hi, @Alan-xw
I didn't use your repo to train realsr, I used original edsr-pytorch instead.
I first pre-upscale the input x so that the output size is original size * args.scale.
the training process is ok, but when test PSNR on Set5 each epoch, something wrong happened.

model/common.py", line 39, in _space_to_channel
    x = x.contiguous().view(b, C, hout, scale, wout, scale)
RuntimeError: shape '[1, 4, 65, 4, 67, 4]' is invalid for input of size 280800

Hi, thanks for your codes. However, in origional paper, the model (k = 5) parameters is 5.731M, the model parameters you implemented with pytorch contains only 1545765. I check the model carefully and can not find the reason. Do you notice that...

I have not found MeanBlur() in common.py

Hi, the author() uses the y channel to calculate psnr. Doesn't your calc_psnr function calculate the y channel?

Dear @Alan-xw , thank you for your code.
I want to train LP-KPN model for superresolution task in which the input lr images are 4 times smaller than GT. I wonder which part of the model I should change. Could you please give me any advice on this?

In the line 213~219 of "model/common.py":

        gaussian_down = self.Prdown(Gaussian_lists[-1])
        Gaussian_lists.append(gaussian_down)
        size_lists.append(gaussian_down.size()[2:])
        Gaussian_lists.append(gaussian_down)
        Lap = Gaussian_lists[-2]-self.PrUp(Gaussian_lists[-1],size_lists[-2])
        Lap = lap.mul(1./255.)
        Laplacian_lists.append(Lap)

I think the 2nd "Gaussian_lists.append(gaussian_down)" need to be commented, and replace "lap" with "Lap".

Hey,
I implement the SRFlow based on the paper with Pytorch. I saw your issue on the official repository.
I would be very happy if you could share your code with me because I have some problems with implementation.
Looking forward to your reply.!

        Gaussian_lists.append(gaussian_down)
        size_lists.append(gaussian_down.size()[2:])
        Gaussian_lists.append(gaussian_down)
        Lap = Gaussian_lists[-2]-self.PrUp(Gaussian_lists[-1],size_lists[-2])
        Lap = lap.mul(1./255.)

--->why append gaussian_down two times? and Lap = lap.mul the second lap should be Lap?

how about the performance of your implementation?
Can it achieve similar performance to the original paper?

self.sub_mean = common.MeanShift(rgb_range, vgg_mean, vgg_std)

I can't find Where the function "MeanShift" is.