CUDA out of memory about pd-denoising-pytorch HOT 5 CLOSED

Codes updated.

from pd-denoising-pytorch.

I am not sure but it should be fine with images of 512x512. The testing-on-full-image version is to crop larger images to smaller patches for denoising and then combine them back. It will solve the out-of-memory issue but increase the execution time. If needed, I will release it in a few days.

from pd-denoising-pytorch.

Wow thanks for taking the time out of your weekend to do that! I'll try to test it tonight or tomorrow evening.

from pd-denoising-pytorch.

Just tested with your larger example images. Seems to work! Thanks!

from pd-denoising-pytorch.

Thanks for the repo, works on first time. Initially I tried the full frame inference code, and that was burning over 8GB of memory to process a 672x344 frame!! (including 3gb of swap memory on my limited but capable embedded system)
But then I saw this issue thread and I continued to read until the end of the repo readme and tried the Demo_on_full_image.py instead of test.py : that worked brilliantly. Actually it ran much faster and is much more stable on my system given that it did not use any swap memory.

I think you should highlight the memory friendly option earlier in your readme, like a mention "For limited memory users, I recommend using the Demo_on_full_image.py, see down below."

I did not notice any clear artefact in the memory friendly inference, real nice job.
I attach an example of an original image, the full frame processed one, and the cropped processed one (wbin=256, did consume around 3GB of memory max) for anyone interested.

from pd-denoising-pytorch.