LinAlgError: SVD did not converge when plotting data (raw.plot())

[FriederikeScholten]

I am currently going through the tutorials provided by MNE and when I try to use the built-in plot function, I often get an error that the SVD did not converge (see attached picture). Two other colleagues have the same problem (using Spyder & Visual Studio Code). One of them made the experience that it does work sometimes after re-running the cell multiple times. Any idea what the issue might be here?

[agramfort]

with what data is that?

[FriederikeScholten]

@agramfort , we used the 'sample' dataset that is used in the tutorial

[agramfort]

what does mne.sys_info() give you?

…

[larsoner]

If you're using NumPy with OpenBLAS it could be OpenMathLib/OpenBLAS#3044 which won't be fixed until an OpenBLAS release then NumPy release. Or if you're using NumPy with MKL you could be caught by https://community.intel.com/t5/Intel-Distribution-for-Python/Bug-in-GESDD-but-not-GESVD/m-p/1105241 which has been around for 5 years and stopped getting replies from Intel folks. We can work around either one probably with a from mne.fixes import _safe_svd and using that instead of scipy.linalg.svd, can you give it a try? If it works, it would be awesome if you could open a quick PR for this!

If it does not fix the problem, you could be hitting this issue because you have nan or inf in your projector, and we should add some check for the projector being invalid

[FriederikeScholten]

This is what the mne.sys_info() gives me:

[larsoner]

@FriederikeScholten can you double check that all(np.isfinite(proj['data']['data']).all() for proj in raw.info['projs']) evaluates to True? If it is then you're probably hitting the Intel bug. If you can check your CPU architecture, try out some of the cases I posted in that issue, and tell them you're hitting the issue, too? Maybe more pings/reports will get them to fix it.

And in the meantime, feel free to try using _safe_svd, it should work around the issue

[FriederikeScholten]

Thanks for the suggestions. I do not know how (to my knowledge I did not change anything), but the plot function works fine now, also for my colleague. Just to follow up, the code: all(np.isfinite(proj['data']['data']).all() for proj in raw.info['projs']) evaluates True.