Dimension mismatch warning for loss calculation. about kernelgan HOT 4 CLOSED

Hi,
I don't recall these warnings, before I go about checking them, please verify packages version (specifically Pytorch) and that your image size is divisible by 2 (better by 4). Please let me know if this was the issue.
Regarding ZSSR - similarly to KernelGAN it is an image specific NN that trains on the input image only. Here KernelGAN uses ZSSR to train on the specific image and KERNEL! => No pre-training for ZSSR

from kernelgan.

Hi,
The size of the image I used was divisible by 4 (2178 x 1454). I am using torch==1.2.0, torchvision==0.4.0.
Thanks for the info on how ZSSR works.

Also, I used this for blind SR of a real image. There were no crazy artifacts in the output image but at the same time it was also very close to a naive upsampling from gimp. Any tips on getting better results on real images would be helpful?

from kernelgan.

I used torch 1.0.0, try switching to that and see if you still get the warnings. Any way, if you didn't change configurations then I don't recall this warning.

Regarding the SR quality - some images are much more 'difficult' to do SR. It could be for several reasons e.g. type of camera, motion blur, the single kernel assumption doesn't hold etc. Our method works well in many cases but not all.

from kernelgan.

Thanks for all the information.

from kernelgan.