mehrhardt
commented
5 years ago I can confirm that my "issue" in CCPPETMR/SIRF#214 disappears if one chooses 16 subsets for 64 views (note: 64/4 = 16).
@mehrhardt, I'm guessing that your algorithms doesn't need balanced subsets...
Yes, these kind of algorithms can be run and will converge for any subset selection, however imbalanced.
This was originally https://github.com/CCPPETMR/SIRF/issues/214. See there for more info.
When you run
OSMAPOSL, STIR will check that subsets are balanced (in terms of #views only of course, not stricly in terms of equal subset-sensitivities). For other algorithms, this is not a requirement, so it won't do that check (I believe STIR writes a warning). @mehrhardt, I'm guessing that your algorithms doesn't need balanced subsets...The subset scheme in STIR was developed for when the number of subsets does divide the number of views. With default settings for the projector, it's actually a bit more complicated, as it puts views that are related by symmetry inside the same subset. This means that usually you have that
num_subsetsneeds to dividenum_views/4.I have never spend any time on checking what happens with the subsets for other cases. I know the code doesn't fall over, but that's about it! The actual code is here.
Clearly, what you see is not desirable, but this code is really is basic as it gets. I guess therefore that the effect that you see is due to surprises with the symmetries.
I'm not sure what we could do to resolve this.