@fighto Well, I have not run it so I can't say that for 100%. Also, it is good hearing from you, so I don't mind the message :-).

@fighto Looking at https://github.com/zhangyi-3/KBNet/blob/main/Denoising/test_real_denoising_sidd.py. Just remove all the .cuda() methods and it should work fine. I have never worked with the basicsr library they are using, but looks like a pytorch model. The .cuda() methods means move the weights from the CPU to GPU.

@fighto I haven’t had time to dig into it, but pytorch and tensorflow models shouldn’t mind where the processing is done. Weather on a GPU, CPU, or TPU. Most of the academic work is coded on their particular system so there is probably some code in their files that just specifies that the data should be on the GPU device. Normally like xxxx.to_device(‘gpu’) or similar. Updating that to move the data to the CPU should resolve the problem.

@fighto Yeah, those models look super small. Do you need a GPU to run?

@fighto I have a Nvidia on my laptop, but it is only 4GB, so it can run only small models. To run anything larger, I need to use Google Colab (V100 - 16GB or A100 - 40GB). For anything larger, I normally spin up a cloud VM with multiple GPUs.

@fighto I will normally run Colab, we spin up a GPU cluster to run training

@fighto it should be able to run Cuda on most Nvidia GPUs. AFAIK, Cuda is just the bridging technology fot interfacing vector data to GPUs

@ajkohn Google has been shown before to exaggerate the importance of their performance in the quantum space. Eg the test that was later broken by cpus. I do think it is fascinating thought, to watch as it develops.