EMUsort

Processing time

EMUsort, HD-EMG, 200 s, wgpu, warm (default nt = 61):

Stage	Time (s)
Fitting preprocessing	1.84
Finding clips	2.09
Learning templates	0.96
Detecting spikes	2.82
Clustering spikes	3.67
Matching templates	3.45
Clustering matched spikes	4.13
Merging units	0.61
Whole program	19.6: 10.2× real time

Peaks at 0.55 GB of GPU memory and 0.65 GB of host memory. Two runs of the same input give identical spikes and units.

Slower than October, 15.3 s then and 19.0 s now on the stages October timed. As for Kilosort4, the cause hasn’t been investigated.

Checks without ground truth

There is no reference sorting for the HD-EMG recording, so the checks are those a reference does not need (first 200 s, default nt = 61):


Spikes / units	126 153 / 69
Units with > 1% / > 5% of inter-spike intervals < 2 ms	33 / 9
Presence over the segment (median)	1.00

October’s run also compared nt = 121 (103 362 spikes, 51 units, 0.8% of template energy in the outer 10% of the window against 4.1% at nt = 61). That comparison hasn’t been repeated.

Two fixes came out of these checks: duplicate detections at bit-identical neighbouring centres (before: 94 000 spikes, 53% of a median unit’s intervals under 2 ms) and the template length (see EMUsort settings).