Little fix for 'eig did not converge' crash (Kilosort 2.0)

[DradeAW]

These changes do not change the default Kilosort 2.0 behaviour, but adds alternatives in case of a crash to try to make the run successful.

If these changes still provide a crash, in my experience increasing the 'NT' parameter helps :)

[abisi]

Hello, I've had this issue ("EIG did not converge at index ") just recently in runTemplates. Did you get why this occurs?
Also, from what value to what value did you increase the NT parameter?

Thanks!
Axel

[DradeAW]

Hi!

My understanding is that computing the eigenvectors and eigenvalues for such large arrays is complicated.
Sometimes the method employed by Kilosort fails (especially for long and heavy-channel recordings)

So trying an alternative way can be useful!

Here is the default value for NT:

Kilosort/configFiles/configFile384.m

Line 43 in 1a030bf

ops.NT = 64*1024+ ops.ntbuff; % must be multiple of 32 + ntbuff. This is the batch size (try decreasing if out of memory).

I changed it to 3x this value, and for my recording it worked. You can also try 10x the value, it shouldn't have too much of an effect on the sorting efficacy.

[abisi]

Hello,

Thank you for your answer! I'll try both your proposed changes and changing ops.NT to see what would fix this issue in my case (2h long recordings with all bank channels).

So far, going from Matlab 2023a back to Matlab 2021b - which I was using - fixed the issue.

Best,
Axel