search

google-research / maxim

[CVPR 2022 Oral] Official repository for "MAXIM: Multi-Axis MLP for Image Processing". SOTA for denoising, deblurring, deraining, dehazing, and enhancement.
https://arxiv.org/abs/2201.02973
Apache License 2.0
1.02k stars 110 forks source link

problem about maxim #4

Closed woaicv closed 2 years ago

woaicv commented 2 years ago

Hi, thank you very much for your work! I currently want to learn your approach, but because my code level is not very good, I need to run through it a little bit. I can follow your run_eval, but when I add some output in MAXIM, there is no corresponding output after execution. What I would like to ask you is how to perform forward propagation and print out some desired results during the forward propagation process, which is conveni

WeChat92cdd69c7905f13f59663c879c680e62

ent for learning. Hope to get your answer, thank you very much🙏

vztu commented 2 years ago

Hi, you should run setup again if any modifications have been made.

Run pip install . again. Then you will see the logs.

woaicv commented 2 years ago

Ok, thanks a lot for your reply. The frameworks I have seen before are written in the forward format, so looking at your code will cause some problems. The multiple Unsamples in the figure below, Con1x1, Con3x3, do not share weights, right? Looking forward to your repl 221649288149_ pic 231649288319_ pic y

vztu commented 2 years ago

Yes, they're not sharing weights. BTW, we've set self.use_cross_gating = True by default. So the else: part is not reached.

woaicv commented 2 years ago

Ok, thank you very much. Does pixelshuffle and pixelunshuffle for unsample downsample have little effect on performance?

vztu commented 2 years ago

I haven't tried those alternatives. But I feel that it doesn't matter that much whichever downsample/upsample methods you use, as it is a UNet architecture and has long skip connects. You may have to try and see how performance varies.