Sampling rate and template alignment

[dimokaramanlis]

Hello,

thank you for your great tool.

We mainly use a sampling rate of 10k in our recordings. In my understanding, Kilosort was developed initially for 25k recordings. To accommodate for the different rate, we set:

• nt0 = round(61*10k/25k) in the configuration file, for setting the time window.
• dt =-round(21 * size(wPCA,1)/61)+ [1:size(wPCA,1)] in get_PCproj.m, to properly align our spikes for the initialization.

Kilosort seems to work fine, except for one thing. When loading Kilosort results in phy, the templates seem to be shifted relative to the spike waveforms by one time sample:

Why do you think we are getting this shift?

In my understanding, the parameter responsible for that is nt0min (defined in fitTemplates.m). While nt0min is corrected for the length of the temporal window (ceil(20 * nt0/61)), there is a difference of one time sample for 25k recordings (nt0 at 20 vs spike trough at 21). For 10k recordings, both numbers end up at the 8th time sample.

Best, Dimos

[marius10p]

Hi Dimos, looks like your nt0 ends up even (24). Can you please try 25 and let us know if you get the same problem? - Marius

[dimokaramanlis]

Hi Marius,

I tried running Kilosort with odd nt0=25, with nt0min=9. Now it appears that the templates are shifted two samples relative to the spikes:

Another thing I've noticed is that at the TraceView, one can see the actual waveform below the template (instead of the residual?), and both data and template are properly aligned:

[dimokaramanlis]

If I subtract 1 or 2 from the spike times (for nt0 = 24 or 25, respectively), before saving them in phy's format, the spike traces and the templates end up properly aligned in phy.

The batch buffer is another parameter that might be relevant and I have changed to match our10k recordings (ops.ntbuff=26).

[marius10p]

Hmm, maybe it is the batch buffer. That doesn't need to be changed, since it's just the overlap between consecutive batches that we need to resolve those boundary spikes. Can you please try without changing it? Did you also change ops.NT by any chance? That also shouldn't be necessary.

[dimokaramanlis]

Sure, I will try again with the original buffer.

For NT, I use ops.NT = 32*512+ops.ntbuff, since my GPU goes out of memory for larger batch sizes. GPU has only 2GB, recordings come from 252 electrodes, and I ask for 960 templates.

[marius10p]

Ok about ops.NT, that should be fine.

[dimokaramanlis]

Hi Marius,

returning ntbuff to 64 didn't affect the shifts. I still get shifts of 1 or 2 two time samples, depending on the value of nt0 (24 or 25).

My only worry is that there may be a misalignment during the matching pursuit pass, ultimately affecting spike detection.

[dimokaramanlis]

I managed to solve the problem, I think it was related to template initialization 'fromData'. Now the spike waveforms are aligned to the templates and the residual appears properly in TraceView:

Before, I was using the default principal components (PCs), calculated for spike waveforms of 61 samples, centered at sample 20. The algorithm tried to downsample the PCs to match my nt0 values, but this might have messed up with the actual trough alignment.

I set nt0=25, centered some of my spikes at 8 and calculated PCs from those waveforms. Then, I used the following values to run Kilosort:

• dt =-9+ [1:size(wPCA,1)] in get_PCproj.m.
• nt0min=8

I'm not entirely sure where to pinpoint the problem, but at least Kilosort works even better now. Proper alignment and PCs from my data actually led to better templates and higher spike amplitudes.

Thanks again for the help.

Best, Dimos

[brendonw1]

I'm another user on this forum and I generally like and use KiloSort. Thanks for bringing up this issue.

This sounds important. Is it something likely to be affecting all of us? Is there a way to instantiate a systemic change in Kilosort?

On Fri, Feb 1, 2019 at 5:38 AM Dimokratis Karamanlis < notifications@github.com> wrote:

I managed to solve the problem, I think it was related to template initialization 'fromData'. Now the spike waveforms are aligned to the templates and the residual appears properly in TraceView:

[image: solved_problem] https://user-images.githubusercontent.com/1114242/52116553-b5ecb800-2611-11e9-8410-7fc6f1c2f5a7.png

Before, I was using the default principal components (PCs), calculated for spike waveforms of 61 samples, centered at sample 20. The algorithm tried to downsample the PCs to match my nt0 values, but this might have messed up with the actual trough alignment.

I set nt0=25, centered some of my spikes at 8 and calculated PCs from those waveforms. Then, I used the following values to run Kilosort:

• dt =-9+ [1:size(wPCA,1)] in get_PCproj.m.
• nt0min=8

I'm not entirely sure where to pinpoint the problem, but at least Kilosort works even better now. Proper alignment and PCs from my data actually led to better templates and higher spike amplitudes.

Thanks again for the help.

Best, Dimos

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/cortex-lab/KiloSort/issues/177#issuecomment-459680308, or mute the thread https://github.com/notifications/unsubscribe-auth/ADXrTdkLWfQnnYU1vZvB-3JUnaezc0Rxks5vJBkogaJpZM4aV7uC .

[marius10p]

@dimos glad to hear it works now. The PCs are only used for initialization, and they are quickly forgotten during the template optimization. They are also used for visualization in Phy, but they play no role in the actual spike extraction and assignment.