This is the package implementing the spike detection algorithm in Herding Spikes, adapted from the codebase at HS2. The algorithm is described in the following paper:
J.-O. Muthmann, H. Amin, E. Sernagor, A. Maccione, D. Panas, L. Berdondini, U. S. Bhalla, M. H. Hennig. "Spike Detection for Large Neural Populations Using High Density Multielectrode Arrays". In: Frontiers in Neuroinformatics 9 (2015). doi: 10.3389/fninf.2015.00028.
This package is to be used as a backend for Spikeinterface's sortingcomponents. The implementation is further optimized based on HS2 to handle large-scale recordings and can achieve real-time detection and localization on 6144-channel, 32 kHz data on a Haswell server CPU (single core). Parallelization on multiple cores can deliver an additional 2x speedup.
The software was first developed as part of the MSc project "Re-implementation and Optimization of a Scalable Spike Detection Algorithm for Large-Scale Extracellular Recordings" by Rickey K. Liang (supervisor Matthias H. Hennig), and the code is released under the GPL-3.0 license, inheriting HS2.
This package is to be invoked from SpikeInterface; please refer to the installation guide there to set up the environment. Python 3.9+ is suggested for development and testing, but Python 3.8 is also supported for users invoking from SpikeInterface.
The parallelization is controlled by env var OMP_NUM_THREADS. Set to 1 for single core running. 2x speedup achieved by 4~6 (depending on platform).
Please note that a C++ compiler (requires C++17 compatibility) should be properly configured to build the C++ extension for Python.
pip install hs-detectiongit clone https://github.com/lkct/hs-detection.git
cd hs-detection
# pip install . # for install
pip install -e . --config-settings editable_mode=compat # for develop, config needed for static analyzers
cd ..The first version officially released to the community--the version for the MSc Dissertation with updated meta-info and readme.
The following are performance issues not addressed in the current implementation, but could lead to significant differences in corresponding cases.
- The common median reference (CMR) runs too slow compared to common average reference (CAR) which can reach real-time.
- Shortcuts are not set for code paths other than scaling&CAR (ref). A +20% speed in C++ code was gained for the scaling&CAR shortcut.
int16data without the need to rescale could be expected as input from SI, but extra type-casts to and fromfloat32are now enforced.
The code was developed when the author was an MSc student at the School of Informatics, University of Edinburgh. Please reach out to Dr MH Hennig who was the project supervisor.