[SyneRBI/PETRIC-MaGeZ] replace home-grown RDP (Issue #10)
Kris Thielemans
KrisThielemans created an issue (SyneRBI/PETRIC-MaGeZ#10)@paskino tells me that his GPU is running out of memory in computing the diagonal of the Hessian for some reason.
you could consider using
https://github.com/SyneRBI/PETRIC2/blob/97aa45ecb1f9dbccc7d146537f6f5d4bca818420/SIRF_data_preparation/run_LBFGSBPC.py#L68-L82
if STIR version is >= 6.3.0
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.
Georg Schramm
gschramm left a comment (SyneRBI/PETRIC-MaGeZ#10)Good point.
(1) once setup, how to I apply this the diagnonal of the hessian to an image?
(2) I guess this uses the GPU?
(3) Does @paskino have more diagnostic information?
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.
Georg Schramm
gschramm left a comment (SyneRBI/PETRIC-MaGeZ#10)I did a short comparison of my cupy reimplementation vs the STIR based RDP diag hessian:
https://github.com/gschramm/PETRIC2/blob/main/test_rdp_hess.py
- for the siemens data, my current implementation takes ca. 0.13s whereas Kris' stir implementaiton takes 0.65s.
- both give almost identical results, except for the 1st/last slice (where I am using heuristic padding which is prob. wrong)
- my implementation keeps all kapp_ij betwen neighboring voxels in memory and hence needs ca 7.8GB of GPU memory (depending on image size ...) @paskino is this causing your memory issues?
In princple these weights can be recomputed / stored on the host, but that will slow down things (but in the end probably not important since we evaluated the diag Hessian currently only 2-4 times.
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.