search

assafshocher / ZSSR

"Zero-Shot" Super-Resolution using Deep Internal Learning
Other
402 stars 102 forks source link

Can't reproduce the results in the paper #21

Open Joechann0831 opened 4 years ago

Joechann0831 commented 4 years ago

Hi! Thank you for your sharing the code first. Recently, I tried to run your code with python3.7, after some modification such as range and print functions, the code can run normally. But the results are weired, the PSNR values are far away from those results shown in the paper. I wonder that is there any tricks when I use python3.7 to run? Do I need to change the framework to python2?

BTW, I re-implemented ZSSR using Pytorch1.0, but there is about 0.4dB drop with Set14x2 from your paper and 0.2dB drop with BSD100x2 from your paper. Do you have any additional tricks to get such a high results? If so, could you please give me a hand for re-implementation?

Thank you very much!

assafshocher commented 4 years ago

Hi, The biggest danger when moving to py3 is the division sign "/". in some cases it should change to "//". If you are using numeric kernel then you should modify https://github.com/assafshocher/ZSSR/blob/d2948526fde799a0c111469b78cd9e1b8659cb0f/imresize.py#L186

You should check other divisions if they happen to be performed on integers. If they do then they probably need to b e "//". This is my best guess, I suggest that you try to run with py2 and compare. There are some forks for this repo who made the transition to py3.

zzd1992 commented 4 years ago

@assafshocher Can you provide the py3 transition for this project?