SpikeInterface
Avoid segmentation error during kilosort2 cluster splitting
Hi We have been experiencing segmentation errors during cluster splitting with kilosort2.5 We resolved this by installing a different version of the mkl library and, in our fork, setting the corresponding (system specific) paths as environment variables in kilosort_master.m as follows:
setenv('BLAS_VERSION', '/usr/lib/x86_64-linux-gnu/mkl/libblas.so');
setenv('LAPACK_VERSION', '/usr/lib/x86_64-linux-gnu/mkl/liblapack.so');
Allowing this in df spikeinterface is pretty much the last step we need before we can get rid of our fork Do you think that would be possible? Do you have a preference about how to do this? We could for instance add the paths to BLAS and LAPACK as kilosort parameters.
We'll be happy to submit a PR implementing whichever approach seems preferable. Best Tom
Hi @TomBugnon
Would setting them as env variables solve your issue, as suggested by @Dradeliomecus here?
If that is not propagated to MATLAB, then we could check in the KS wrappers if the env variable is set and add the setenv to the matlab script accordingly.
@alejoe91 : Maybe it would be more convinient to add some option in kilosort with theses env variable that could injected in the generated script as suggested. If we use the slurm laucher then the solution of @Dradeliomecus will not work but if it is inside the shell script then it should be OK.
Ok good for me!
@TomBugnon could you make a PR? Adding a blas_path and lapack_path parameters as parameters (with None as default) should work
I prefer also the approach of setting those as parameters (so that different users on the same system can expect the same results when using the same spikeinterface parameters). Will submit a PR asap.