[SyneRBI/SIRF] list-mode obj_fun.get_subset_sensitivity(0) returns None (Issue #1251)
Kris Thielemans
It works fine for the "sinogram" objective function.
Found by @gschramm.
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.
Kris Thielemans
@evgueni-ovtchinnikov wrote in #1253 (comment)
which is due to the fact that
PoissonLogLikelihoodWithLinearModelForMeanAndListModeDataWithProjMatrixByBindoes not override emptyObjectiveFunction.get_subset_sensitivity().
Can you elaborate a bit? get_subset_sensitivity is implemented in PoissonLogLikelihoodWithLinearModelForMean. It also used in PoissonLogLikelihoodWithLinearModelForMeanAndListModeDataWithProjMatrixByBin::actual_compute_subset_gradient_without_penalty. It is of course entirely possible that there is a bug somewhere, but I'd like to understand what you saw.
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.
Evgueni Ovtchinnikov
In STIR.py we have
class PoissonLogLikelihoodWithLinearModelForMeanAndListModeDataWithProjMatrixByBin(ObjectiveFunction):
From what you write it looks like this class should have been derived from PoissonLogLikelihoodWithLinearModelForMean, should it not?
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.
Kris Thielemans
yes indeed. And it is.
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.
Evgueni Ovtchinnikov
yes we do. I have just pushed the corrected STIR.py - let George try it.
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.
Georg Schramm
@evgueni-ovtchinnikov @KrisThielemans was the update pushed to the master branch of SIRF?
Should I re-run the complete Super-Build with DEVEL_BUILD=ON?
I am have some time for tests this late afternoon / tomorrow morning.
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.
Kris Thielemans
Not yet merged, no. You'll need to set
cmake ... -DSIRF_TAG=origin/acc-hess
you can also try UCL/STIR#1418 by adding
-DSTIR_URL=https://github.com/KrisThielemans/STIR-1.git -DSTIR_TAG=origin/ListModeObjFuncExtras
(although TBH I have never tried switch the URL like this in an existing build, so git status and git log in sources/STIR will be essential).
I'm not 100% sure that SIRF will pick up the LM Hessian, but I think it will. Comments in the respective PRs please.
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.
Georg Schramm
list-mode obj_fun.get_subset_sensitivity(0) returns None is now fixed (now sure which PR fixed it).
Using the SIRF:acc-hess and STIR:master branch, the following test snippet runs without error using a
"realistic" Poisson logL objective function:
for i in range(num_subsets): assert np.all(np.isclose(lm_obj_fun.get_subset_sensitivity(i).as_array(), obj_fun.get_subset_sensitivity(i).as_array()))
If the listmode subsets are based on views (which I remember they should be), everything works as expected.
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.
Kris Thielemans
Indeed. Listmode subsets currently reproduce those from projection data.
Excellent news!
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.
Evgueni Ovtchinnikov
should be fixed now that PR #1253 is merged
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.
Evgueni Ovtchinnikov
Closed #1251 as completed.
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.