Tridesclous 2: pipeline

Read from the source (sorters/internal/tridesclous2.py, sortingcomponents/*).

1. Preprocessing. Bessel band-pass 150–6000 Hz (order 2, forward-backward), common median reference on ≥ 32 channels, local whitening (100 µm); noise levels (MAD on 20 chunks of 500 ms).
2. Detection. locally_exclusive (5 × noise, negative, 150 µm, ±1.5 ms) over the whole recording; max(5000 · channels, 20 000) peaks drawn uniformly.
3. Clustering (iterative-isosplit). Local SVD (5 components, 0.5 / 1.5 ms, 120 µm); starting from each peak’s channel, every cluster is split: its peaks’ features on the channels common to their 60 µm neighbourhoods, reduced to 6 dimensions, clustered by SpikeInterface’s isosplit (k-means start with 15 clusters, fewer for small sets; clusters under 10 points are noise), three levels deep. Templates: the mean of the SVD features mapped back; cleaning (channels with peak-to-peak / noise ≥ 1.5, SNR ≥ 3.5, trough within 0.2 ms); merging (l1 similarity > 0.8 over ±0.5 ms); clusters under 0.1 Hz dropped.
4. Templates for the peeler. Per unit, the channels within 100 µm of its peaks’ channel barycentre; the mean waveform (1.0 ms before, 2.5 ms after) of its clustered peaks; cleaning again.
5. Peeler. Per chunk, levels: peaks of the residual (two levels with the fast detector, 80 µm, ±0.8 ms; then one with a matched filter), largest first; each assigned to the closest short template (0.5 / 0.8 ms) among the units within 150 µm, shifted by up to ±2 samples, its amplitude fitted with its neighbours by least squares; amplitudes in [0.7, 1.4] are subtracted.
6. Final cleaning (auto_merge_units, template differences 0.05 … 0.35, lag 0.5 ms).

On the device

Preprocessing, detection candidates, SVD features and the mean templates on the device; the splits (isosplit), templates’ cleaning and merging, and the peeler on the host (the peeler subtracts spike after spike: sequential within a window; windows in parallel).

Choices of ours

Where	Upstream	Here	Why
Seed	none (seed=None: every run differs)	seed = 0	reproducible
k-means start of isosplit	SciPy kmeans2(minit="points"), NumPy’s stream	the same algorithm, our stream	same role
Whitening and noise chunks	random	evenly spaced	reproducible
Truncated SVDs	scikit-learn randomized	exact	equal up to the solver’s tolerance

Upstream quirks kept

• SpikeInterface’s isosplit adds each covariance diagonal term twice, labels the side below the cut point 2 (inverted), and skips a redistribution that would move a whole cluster’s worth of points: its clusters are rarely redistributed, only merged.
• The peeler’s matched-filter prototype is built in a loop that reads the last unit’s template only (a variable left from the previous loop).
• A spike whose fitted amplitude is above the upper limit keeps amplitude 1 and is not subtracted.
• The template similarity’s symmetric fill and lag sign (as in SpyKING CIRCUS 2).