Main API: dartsort() and the DARTsortUserConfig¶

Output spike trains.

Esimated motion

Set this to false if your data is super stable or already motion-corrected.

Is your data already preprocessed?

If so, the preprocessing flag below is ignored. By default, the recording is assumed to be preprocessed already (including standardization) if its data type is any floating point type.

dartsort will warn you if the data looks weird according to some simple checks (see check_recording in data_util and warn_about_preprocessing in preprocess_util).

If other than 'none', and following the already_preprocessed flag above, dartsort will apply some preprocessing to the recording. Be aware that dartsort expects its input to be standardized on each channel in addition to the usual highpass filtering, but that whitening is handled internally. See util/preprocess_util.py if you're curious about the details of the methods.

Options: 'ibllikecmr', 'ibllike', 'standardize', 'none'

If you have a lot of data and you're using a workflow where it is important to save a preprocessed copy of the recording, float16 is a good option. Only relevant if preprocessing != 'none'. If the recording isn't getting saved, stick to float32.

Detection steps before the final matching round will run until at least this many spikes are found or the whole recording is covered, to make sure that there is enough data for clustering. See also subsampling_presence. Set to None to disable subsampling.

Early detection steps which have already found subsampling_spikes spikes are only allowed to end early if they additionally cover this fraction of the recording, to make sure there's some coverage of conditions for template estimation.

By default, 1 template matching step is carried out using templates estimated from the initial detection round.

Whether to stop after initial localization and motion tracking.

Number of parallel workers to use when running on CPU. 0 means everything runs on the main thread; negative means #cpu - (val+1) so that -1 is all cores, -2 is all less 1, etc.

Number of parallel workers to use when running on GPU.

Like n_jobs_gpu, but for the motion tracking spike detection (cheaper jobs, so default to 2).

Max workers to use for small jobs.

Max workers to use for small jobs running on GPU.

The name of the PyTorch device to use. For example, 'cpu' or 'cuda' or 'cuda:1'. If unset, uses n_jobs_gpu of your CUDA GPUs if you have multiple, or else just the one, or your CPU.

Choose: 'threading_unless_multigpu', 'ThreadPoolExecutor', 'ProcessPoolExecutor', or some others.

Batch size for data processing.

If True, dartsort will store all temporary data in a scratch directory in tmpdir_parent or TMPDIR.

Save a copy of the preprocessed recording to a tmpdir?

'shutil' or 'rsync'

Control where tmpdirs are created.

Store all spike features from intermediate steps (for debugging)

Store the spike features from the final step (instead of just basic spike train outputs).

Length of time (ms) before trough (or peak) in waveform snippets. Default value corresponds to 42 samples at 30kHz.

Length of time (ms) after trough (or peak) in waveform snippets. Default value corresponds to 79 samples at 30kHz.

Largest time shift allowed when re-aligning events.

As a final postprocessing step, only the higher-scoring of any spikes within this time radius of each other are kept. If this is negative, it does nothing. If it's 0, exact duplicates are dropped.

Allow only troughs or events of both signs when detecting threshold crossings during initialization. Or positive only, if that's your thing.

Threshold in standardized (SNR) voltage units for initial detection; peaks or troughs larger than this value will be grabbed.

Template matching threshold. If subtracting a template leads to at least this great of a decrease in the norm of the residual, that match will be used. This is in the same units as the corresponding threshold in Kilosort and other sorters, and it represents reduction in Euclidean norm of standardized data due to matching a new event.

Initial detection's neural net matching threshold. Same as matching_threshold, except that a neural net is trying to guess the true waveforms here, rather than using cluster templates.

If subsampling, a quick thresholding will be run at this voltage threshold to grab spikes for motion estimation purposes.

Rank of temporal PCAs used in denoising and featurization.

As ms_before, but used only when computing PCA features in clustering.

As ms_after, but used only when computing PCA features in clustering.

Radius of neighborhoods around spike events extracted when denoising and subtracting NN-denoised events.

During initial detection, if two spike events occur at the same time within this radius, then the smaller of the two is ignored. But also all of the secondary channels of the big one, which is important.

Radius around detection channel or template peak channel used to extract spike features for clustering.

Extraction radius when fitting features like PCA; smaller than other radii to include less noise.

Radius around main channel used when localizing spikes.

Which neural net to use in initial detection? Set to Decollider (and set the pretrained path to None to train a brand-new unsupervised denoiser.

Path to a pytorch saved model (.pt file as dumped by torch.save()). If this is None, a new model will be fit.

Standard deviation of amplitude scaling regularization prior in template matching.

Boundaries on the amount of scaling allowed.

Upsampling of templates during matching to allow for temporal aliasing of waveforms.

Which steps run after the last matching iteration. Agglomeration merges oversplit units; cleaning deduplicates and orders units by depth.

Use rigid registration and ignore the window parameters.

Motion bins exceeding this speed will be replaced by interpolation.

Motion bins farther than this from the local median will be replaced by interpolation.

'full' or 'online'. Online helps with long recordings.

Chunk length in seconds for online motion tracking

Was do_motion_estimation set, or are we ignoring motion?

The recording's get_channel_locations() array.

The extended "registered geometry" array

Motion estimate from DREDge, if using.

Motion estimate from SpikeInterface, if using.

k-d tree for querying original geometry.

k-d tree for querying registered geometry.

k-d tree query bound.

Vertical probe geometry period.