Changing parameter 'nt0' results in a crash

[DradeAW]

Hi,

I tried running Kilosort with a different value for nt0 (from 61 to 75), and I obtained this crash:

Unable to perform assignment because the size of the left side is 61-by-1-by-3 and the size of the right side is 75-by-3.

Error in template_learning (line 110)
    rez.W(:,t,:) = gather(wsign * w(:,1:3));

Error in main_kilosort3 (line 42)
rez                = template_learning(rez, tF, st3);

From what I saw, this is because in clustering/template_learning.m, line 103, rez.W is defined as:

rez.W = zeros(61,0, 3, 'single');

Shouldn't this 61 be replaced by ops.nt0?

[marius10p]

Sure, yes, but I think the CUDA code will crash with nt0 >61. Unless your data is recorded at >=50Khz, there shouldn't be a reason to change nt0.

[DradeAW]

I am recording in the cerebellar cortex, and some spikes are called "complex spikes" which impacts the extracellular voltage quite significantly for ~4/5 ms (see the image below, mean of about 5,000 complex spikes)

I was worried that with a short nt0, there would be a residual voltage which might impact Kilosort (as I understand, Kilosort tries to explain the extracellular voltage using the spikes).
Btw, my high-pass filter is set at 150 Hz (I find that 300 Hz makes these complex spikes harder to detect).

[marius10p]

Could you please show that neuron in Phy with multiple snippets? I am curious if multiple events have the same stereotypical pattern after the initial depolarization.

[DradeAW]

The waveforms aren't perfectly aligned (this dataset comes from MountainSort, but I can find one that's from Kilosort where waveforms are more aligned if you want).

The positive phase is usually pretty stereotypical for the same neuron.

[marius10p]

This example is very noisy overall. If the positive phase is stereotypical, but the rest of the spike is variable, then Kilosort won't be able to deal with that correctly. In that case, you are better off setting a small nt0, so the clusters don't get split by the variability in the complex part of the spike. It's possible the "tail" of the complex spike will be assigned to a noise unit, which is fine.

[DradeAW]

Alright, thank you!