error using gpuArray/subasgn: Subscript indices must either be real positive integers or logicals.

[jrnvndtslr]

Hi,

I've been running into this error, while trying to run kilosort2. It occurs in the isolated_peaks_new function, but I suspect the issue is caused much earlier in the code. From some earlier entries (45 & 142) I gathered that it may have something to do with the number of batches. However, setting ops.NT to a lower value does not resolve it. The file I'm trying to process has an ops.sampsToRead value of 68046020. I could try downsampling but I don't know if/when this will affect the quality of the outcome.

Any thoughts on how to mitigate this issue?

[marius10p]

It probably has to do with empty batches. If you really have batches with no spikes, then you should probably increase ops.NT substantially (by a factor of 8 or more). If you switch to kilosort3, then this shouldn't happen any more.

[jrnvndtslr]

It turned out to be (partially) due to the fact that the GPU on our cluster loads the wrong CUDA version by default. I can get my data through now but only with very lenient criteria, and the result is all noise. If I high-pass filter the data at 150Hz and inspect the trace I also can't identify any spikes. Given that one of the first steps in KS2 is a threshold-based detection of z-scored data, I'm afraid I have to conclude that my data quality is just too poor for spike detection.