Hi I am using 8-bit greyscale as input, and note the usual /255 operations to put data in the range 0-1. However the operation

        if args.need_degradation:  # degradation process
            np.random.seed(seed=0)  # for reproducibility
            img_L += np.random.normal(0, args.noise_level_img/255., img_L.shape)

can result in values outside this range, as shown in the image below:

I note that the values are later clipped by util.single2uint so what is saved is not precisely what is processed. Furthermore I wonder the effect on inferencing - presumably values are also clipped to 0-1 range?

Very impressed with the results! However at times even the lowest noise reduction "15" model can remove too much detail. Will there be a "5" or "10" noise level model in the future?

Thank you for all the hard work!

Hello! Your demo on Replicate is extremely cool ✨ Would you be open to adding a demo on Hugging Face Spaces as well? We've got docs for doing that, but I'd be happy to help out with it!

If you'd like, you can also mirror your model weights over on the Hugging Face model hub as well, which I know our users would love 🤗 People have been building all kinds of cool things with the models on the Hub these days, like this demo that combines 3 different models 👀

scunet_color_real_gan could even be uploaded as a private or gated model, if that's something you're interested in.

cc: @osanseviero

Hey @cszn! 👋

This pull request makes it possible to run your model inside a Docker environment, which makes it easier for other people to run it. We're using an open source tool called Cog to make this process easier.

This also means we can make a web page where other people can try out your model! View it here: https://replicate.com/cszn/scunet. The docker file can be found under the tab ‘run model with docker’.

We have added some examples, but do claim the page so you can own the page, customise the Example gallery as you like, push any future update to the web demo, and we'll feature it on our website and tweet about it too. Moreover, when the page is claimed, it will be automatically linked to the arXiv website as well (under “Demos”)! You can find the 'Claim this model' button on the top of the page ~

In case you're wondering who I am, I'm from Replicate, where we're trying to make machine learning reproducible. We got frustrated that we couldn't run all the really interesting ML work being done. So, we're going round implementing models we like. 😊

  File "C:\Users\Miki\SCUNet\models\network_scunet.py", line 88, in forward
    if self.type!='W': output = torch.roll(output, shifts=(self.window_size//2, self.window_size//2), dims=(1,2))
RuntimeError: CUDA out of memory. Tried to allocate 486.00 MiB (GPU 0; 12.00 GiB total capacity; 10.76 GiB already allocated; 0 bytes free; 11.15 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation.  See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

@cszn Please, add minor changes to code, to support --tile.... in before opened issue, I asked for same question, followed your instructions, uncommented line # number 99th but same problem appears again.... No way to denoise photos larger than 3 Mpx...due cuda oom errors. It is pain doing in graphic editor manualy cropping & after denoise joinning big photo gallery I kindly ask for help... Cause this project has realy great potencial. Onothers like swinir, restormer etc... doesnt do joob like your, in terms on real denoising.... tnx very much in advance

Results look very good, better than the "scunet_color" at times. But it would be nice to control the noise reduction like the grayscale and color ones.

Thank you.

you have mention in paper that

It takes about three days to train a denoising model on four NVIDIA RTX 2080 Ti GPUs.

however, in my case, it takes about 5 days to train on 4 V100 GPUs for 1200000 iterations and pics are crop into 544*544. It is much longer than 3 days. How can i speed up? thanks.

https://github.com/NVIDIA-AI-IOT/torch2trt/issues/807

TensorRT not support move batch dim.

https://forums.developer.nvidia.com/t/tensorrt-does-not-support-permute-in-n-batch-dimension/69925