danieldeidda
commented
1 year ago The run time is mainly in the convolution between an image and the kernel (kernelise_image() function). each recon block in the notebook contains 3 reconstruction in sequence another thing could be just to do a very low and a very high parameter reconstruction and not do all the parameters as you say. Smaller data will definitely help as well
At present, the HKEM reconstructions take ~4 min each on the STFC cloud platform, even with the current
rebinstatement. It'd be nice to speed this up at some point (not now).Options:
Use smaller data. I guess theres no need to do this example of the fine-tuning with a full 127x344x344 image. I guess in-plane image cropping would already help a lot. Using a scanner with fewer rings could help as well as well.
@danieldeidda @ashgillman I guess most of the run-time sits in the kernel computations themselves? Do you know if these run-times are roughly expected? (we seem to have some slow-down with recent jupyter). I've had a quick look at stir::KOSMAPOSLReconstruction.cxx and see you've put openmp loops already. I even recall some profiling. Changing STIR would need a STIR issues of course.