This is just an issue I don't know what to do about.
The current master after recent changes fails the long highdim test (below). for a 30d problem, the logz is outside 5sigma for the rwalk sampler.
It's borderline there (5.4 sigma). I know I could up the threshold to 6 sigma, but that'd defeat the point of the test.
I don't think there is any 'bug' per se that lead to the failure here, more of a random fluctuation. But it'd be good to address it in some way other than bump the error threshold.
I think if I make the bound single and increase the walks the problem does go away (i haven't investigated it further)
Ideas as are welcome. I currently do not release the next version before this is dealt with as I'd like all the tests to pass for every released version.
============================== slowest durations ===============================
9798.33s call tests/test_highdim.py::test_run[30-unif]
7041.08s call tests/test_highdim.py::test_run[30-rslice]
3521.69s call tests/test_highdim.py::test_run[30-rwalk]
1013.51s call tests/test_highdim.py::test_run[10-rslice]
789.92s call tests/test_highdim.py::test_run[10-rwalk]
349.81s call tests/test_highdim.py::test_run[10-unif]
(12 durations < 0.005s hidden. Use -vv to show these durations.)
=========================== short test summary info ============================
FAILED tests/test_highdim.py::test_run[30-rwalk] - AssertionError: assert 1.3369639712022945 < (5 * 0.2445641500808489)
I think if there are no ideas how to deal with this, I'll probably still bump the error threshold to 6sigma, but will keep this issue alive.
This is just an issue I don't know what to do about.
The current master after recent changes fails the long highdim test (below). for a 30d problem, the logz is outside 5sigma for the rwalk sampler. It's borderline there (5.4 sigma). I know I could up the threshold to 6 sigma, but that'd defeat the point of the test. I don't think there is any 'bug' per se that lead to the failure here, more of a random fluctuation. But it'd be good to address it in some way other than bump the error threshold. I think if I make the bound single and increase the walks the problem does go away (i haven't investigated it further)
Ideas as are welcome. I currently do not release the next version before this is dealt with as I'd like all the tests to pass for every released version.
I think if there are no ideas how to deal with this, I'll probably still bump the error threshold to 6sigma, but will keep this issue alive.