Crate dspk_sorter_kilosort 

The Kilosort family on the GPU: Kilosort4 (Pachitariu, Sridhar, Pennington & Stringer, Nature Methods 2024; feature ks4) and its fork EMUsort (O’Connell et al., eLife 2026; feature emusort, emusort). Written from the papers and the published defaults, their behaviour matched to upstream’s code by reading it; no code copied. GPL-3.0, as both upstreams.

Implemented stages: universal templates (wPCA / wTEMP) learned from the recording or loaded from the predefined wTEMP.npz (templates), universal-template spike detection with wPCA features on the device (detect), and the first graph-based clustering of those spikes into units (clustering), learned templates (learned), learned-template matching (matching), the clustering of the matched spikes with the refractory criterion, global merges ([merges]), duplicate-spike removal and good / mua labels, driven over a whole recording by runner (preprocessing fit: CAR, high-pass, local whitening), with Kilosort4’s drift correction (drift) and artifact blanking (artifacts). See the documentation.

Re-exports

pub use clustering::cluster_spikes;

pub use clustering::ClusteringOptions;

pub use clustering::SpikeClusters;

pub use matching::TemplateMatcher;

pub use matching::MAX_PEELS;

pub use clustering::merges::global_merges;

pub use clustering::merges::GlobalMergeOptions;

pub use clustering::merges::MergedClusters;

pub use learned::learned_templates;

pub use learned::LearnedTemplates;

pub use learned::TemplateMergeOptions;

pub use emusort::Emusort;

pub use emusort::EmusortConfig;

pub use ks4::kilosort4_provenance;

pub use ks4::Kilosort4;

pub use drift::estimate_drift;

pub use drift::DriftCorrector;

pub use drift::DriftModel;

pub use drift::DriftOptions;

pub use drift::DriftSpike;

pub use detect::detect_universal;

pub use detect::CentreOptions;

pub use detect::TemplateCentres;

pub use detect::UniversalDetector;

pub use detect::UniversalSpike;

pub use runner::fit_kilosort4_preprocessing;

pub use runner::fit_preprocessing;

pub use runner::run_plan;

pub use runner::ChannelDelays;

pub use runner::FitSettings;

pub use runner::FittedPreprocessing;

pub use runner::Kilosort4Result;

pub use runner::RunPlan;

pub use templates::extract_clips;

pub use templates::learn_universal_templates;

pub use templates::learn_universal_templates_with_progress;

pub use templates::ClipOptions;

pub use templates::ClipScaling;

pub use templates::LearnOptions;

pub use templates::TemplateClustering;

pub use templates::UniversalTemplates;

Modules

artifacts

Kilosort4’s artifact_threshold: a window holding a value at or above the threshold is taken to contain a recording artifact and zeroed (upstream zeroes the batch). Off by default, as upstream (∞).

clustering

Kilosort4’s first clustering of the universal-template spikes into units, written from the paper (Pachitariu et al. 2024, Methods: Template learning, Graph-based clustering, Hierarchical merging tree, Split/merge criteria, Bimodality of regression axis).

delays

Channel delays on the device (EMUsort’s remove_chan_delays, a stage of the shared runner): estimation (ChannelDelayEstimator) and removal (ChannelAligner). Batches never leave the device; the estimator downloads its [channels, channels, 2·max_lag + 1] cross-correlation once, at the end.

detect

Kilosort4 universal-template spike detection on the device.

drift

Kilosort4’s drift correction (paper, Methods, Drift correction; behaviour matched to upstream’s datashift.py and get_drift_matrix by reading them; no code copied).

emusort

EMUsort (O’Connell et al., eLife 2026): Kilosort4 adapted to motor-unit action potentials from high-density intramuscular arrays (Myomatrix), feature emusort. Written from the paper and the published defaults, its behaviour matched to upstream’s code (snel-repo/EMUsort a06bb60, ks4mods) by reading it; no code copied (GPL-3.0, as upstream).

ks4

Kilosort4 (feature ks4): the sorter entry point and its provenance. The stages and settings (crate::Kilosort4Config) are shared with EMUsort and live in the crate root.

learned

Learned templates for template matching, from the first clustering’s units (Kilosort4 paper, Methods, Template learning, written from its description):

matching

Learned-template matching with matching pursuit (Kilosort4 paper, Methods: Spike detection with learned templates and matching pursuit, Extracting PC features with background subtraction; written from its description). Per preprocessed window, on the device:

runner

Kilosort4-family runner (Kilosort4, and EMUsort through RunPlan::emusort): out-of-core halo-window streaming over a recording, every window processed on the device.

templates

Kilosort4 universal templates: the single-channel temporal basis wPCA ([n_pcs, nt]) and the universal templates wTEMP ([n_templates, nt]), learned from the recording (learn_universal_templates, Kilosort4’s default) or loaded from Kilosort4’s predefined wTEMP.npz (UniversalTemplates::from_npz).

Structs

Kilosort4Config

Kilosort4 settings this crate uses, with upstream defaults (kilosort/parameters.py).

Constants

KILOSORT4_WTEMP_ID

dspk-hub catalog id of Kilosort4’s predefined wTEMP.npz (dspk_hub::catalog::synapse).

LINE_NOISE_HZ

Line-noise frequency of the notch (Hz): 60 (50 outside the Americas).

NOTCH_Q

Default quality factor of the line-noise notch (−3 dB bandwidth f / 30: 2 Hz at 60 Hz).

TH_LEARNED

TH_UNIVERSAL

Detection thresholds (whitened σ) for the default 300–6000 Hz band. Upstream Kilosort4 uses 9 and 8 with a high-pass only; the band’s upper edge lowers the whitened noise, so the same thresholds detect ~36% more spikes. 10 and 9 give the same agreement with Kilosort4’s saved results as 9 and 8 with the high-pass (test recording, 2026-10-09; 11 and 10 miss spikes).

UPSTREAM_TH_LEARNED

UPSTREAM_TH_UNIVERSAL

Upstream Kilosort4’s thresholds (high-pass only, bandpass_high_hz = None).