search

tuvovan / MIRNet-Keras

Keras Implementation of MIRNet - SoTA in Image Denoising, Super Resolution and Image Enhancement - CVPR 2020
MIT License
24 stars 7 forks source link

SR won't work on real world images #5

Open ofirkris opened 3 years ago

ofirkris commented 3 years ago

Hi, thanks for sharing this implementation. I'm researching SR for real world cases - and get out of memory issues with an empty 16GB P100 GPU

2020-11-26 09:03:19.928246: W tensorflow/core/framework/cpu_allocator_impl.cc:81] Allocation of 2488320000 exceeds 10% of free system memory.

I've tested SR with 1200px and this crashes each time

tuvovan commented 3 years ago

hi, how much memory do you have in your GPU? The number of this network is quite huge (34M) so that kind of big image will cause the OOM like that. If you have any idea on reducing the number or params, please feel free to make a pull request!

ofirkris commented 3 years ago

I have 16GB of GPU (P100) The implementation here - https://github.com/swz30/MIRNet does work on such images - but the Super-Resolution output effect is not good compared to the input image.

Is Mirnet currently still the SoTa compared to "Deep Plug-and-Play Image Restoration" https://github.com/cszn/DPIR

P.S - the link to Google Drive - for weights of the lighting model isn't correct.

tuvovan commented 3 years ago

let share to me the link to the image that you want to test. I will try with my machine.

And thank you for your contributions! :D

tuvovan commented 3 years ago

And about SOTA or not: I would say that depends on the number provided by the original paper. I was not able to reach the number stated in the paper anyway.

ofirkris commented 3 years ago

Thanks! I tried with this image https://i.ytimg.com/vi/agLZhd98Sus/maxresdefault.jpg