Inquiry Regarding "CUDA Out of Memory" Error with Spitfire Filter

[rohud91]

Hello,
I am encountering a "CUDA out of memory" error when applying the Spitfire filter to large images. I was wondering if there is a way to process these images iteratively or by splitting them into smaller parts while maintaining their original size and resolution.
Specifically, is it possible to send smaller batches to CUDA for processing without downscaling the image? I have already tried clearing the CUDA cache and setting max_split_size_mb: 21, but this did not resolve the issue.
I would appreciate any insights you can provide. Thank you for your help and your excellent work in image analysis!
Best regards,
Roman Hudeček

[sylvainprigent]

Hello,

Yes, the most common approach is to split the image into smaller images by partitioning it using a grid of overlapping patches. The patch overlap must be at least half the size of the PSF in each dimension to avoid artifacts.
There is no routine included into the Spitfire code, but it exists libraries to do it. For example, if you are using the python version, you can use the PyTorch function torch.nn.Unfold. I think the dask library can also do something similar to apply a function with a map-reduce pattern.

Best regards,
Sylvain