`0.9.1` tests fail with 4 Assertion Errors in `nterfaces/fmriprep/tests/test_load_confounds.py`

[TheChymera]

This is the full build and test log: https://ppb.chymera.eu/4a909d.log
The values seem to be off by quite a bit. Any idea what could be wrong?

[htwangtw]

It's likely due to how we simulated a short time series as test data.

nilearn/nilearn/interfaces/fmriprep/tests/test_load_confounds.py

Lines 46 to 55 in 932eee9

	# create random noise and a random mixture of confounds standardized
	# to zero mean and unit variance
	if sys.version_info < (3, 7): # fall back to random state for 3.6
	np.random.RandomState(42)
	beta = np.random.rand(nx * ny * nz, X.shape[1])
	tseries_rand = scale(np.random.rand(nx * ny * nz, nt), axis=1)
	else:
	randome_state = np.random.default_rng(0)
	beta = randome_state.random((nx * ny * nz, X.shape[1]))
	tseries_rand = scale(randome_state.random((nx * ny * nz, nt)), axis=1)

I will have a look soon.

[sanjayankur31]

We're also seeing these 4 fail on Fedora builds, but only for certain arches: s390x, ppc64le, aarch64. So I've skipped these tests on these arches for the time being. These tests pass on x86_64 and i686.

[htwangtw]

@sanjayankur31 what's the python version you are running the test on?

Note to self - need to understand why the tests only failed with options including standardize_signal=True, but fine with standardize_signal=False

[sanjayankur31]

This is on Python 3.10 currently.

Edit: I just realised I'd bungled this up---we're on 0.9.0 on Fedora, and these 4 tests fail there too. I'll go update to 0.9.1 now and see how that fares. Sorry about that!

[htwangtw]

These tests didn't change between the two versions so I wouldn't expect any changes. Still, thanks for letting me know!

[sanjayankur31]

Yeh, the same tests fail on the same arches on Fedora rawhide (37): aarch64, s390x, ppc64le.
Here's a test build with logs:

https://koji.fedoraproject.org/koji/taskinfo?taskID=86638657

And here's one on F36 with armv7hl also enabled:

https://koji.fedoraproject.org/koji/taskinfo?taskID=86643728

[htwangtw]

Currently the length of the simulated signal is 300 volumes.
I can try two things:

1. increase the length of the simulated signal
2. investigate the collinearity in the simulated signal mixture and start from there...

@sanjayankur31 I will let you know if I will need your help with testing!

[sanjayankur31]

Sure, sounds good 👍

[Remi-Gau]

This refers to a fairly old issue and we do not seen to still have the problem on more recent version so I will close.

reopen if you feel otherwise