Fix issue with partially failing unit-test test_sort_real_schur() #59
Conversation
|
Ok, so my assumption was indeed correct: the problem occured with changes from numpy 1.24.4 to 1.25.0. |
| @@ -1,3 +1,3 @@ | |||
| docrep>=0.3.1 | |||
| numpy>=1.19.0 | |||
| numpy>=1.19.0,<1.25.0 | |||
There was a problem hiding this comment.
I think this is too restrictive; based on your comment, this is hardware related (works on my M1), would rather skip the test for certain architectures and investigate further.
There was a problem hiding this comment.
It is OS related I would say (not hardware related): with MacOS the test never fails but with Linux...
Or did you try it with Linux?
There was a problem hiding this comment.
(M1 to M5 are the test matrices used... nothing to do with hardware...)
There was a problem hiding this comment.
Yes, sorry, just went too quickly through this PR. Seems really like a problem with the combination numpy>=1.25.0 and ubuntu, numpy==1.24.4 on ubuntu seems to work well (based on the 3.8 job).
Would consider skipping the combination using this combination.
There was a problem hiding this comment.
I looked at the 1.25.0 release notes, it wasn't immediately obvious which change would be affecting this.
There was a problem hiding this comment.
Actually, 1.25.0 was OK here: https://github.com/msmdev/pyGPCCA/actions/runs/5306789772/job/14363166673
There was a problem hiding this comment.
The only thing that looked suspicious to me when going through the numpy changelog was the change in np.r_[] and np.c_[]. We are using these in the troubled test.
There was a problem hiding this comment.
It is indeed strange that numpy 1.25.0 worked with python 3.9 on ubuntu one time but on the other hand pinning numpy below 1.25.0 solved the problem in all cases. So there should be a connection...
There was a problem hiding this comment.
Taking a closer look, I recognized that the case were 1.25.0 worked was the test using slepc, while 1.25.0 never succeeded without using slepc. But I don't see how this should affect the outcome of this test.
The unit-test test_sort_schur() partially fails (works for M1 but not for M2 to M5) for some combinations of OS, python version, and slepc/noslepc. Strangely, the pattern of this failures is not constant over CI #378 to CI #382. Thus, my first assumption is that the issue might be related to changes in numpy introduced with version 1.25.0.