Postprocessing module¶

After spike sorting, we can use the postprocessing module to further post-process the spike sorting output. Most of the post-processing functions require a SortingAnalyzer as input.

Extensions as AnalyzerExtensions¶

There are several postprocessing tools available, and all of them are implemented as a AnalyzerExtension. If the SortingAnalyzer is saved to disk, all computations on top of it will be saved alongside the SortingAnalyzer itself (sub folder, zarr path or sub dict). This workflow is convenient for retrieval of time-consuming computations (such as pca or spike amplitudes) when reloading a SortingAnalyzer.

AnalyzerExtension objects are tightly connected to the parent SortingAnalyzer object, so that operations done on the SortingAnalyzer, such as saving, loading, or selecting units, will be automatically applied to all extensions.

To check what extensions have already been calculated for a SortingAnalyzer named sorting_analyzer, you can use:

import spikeinterface as si

available_extension_names = sorting_analyzer.get_loaded_extension_names()
print(available_extension_names)

>>> ["principal_components", "spike_amplitudes"]

In this case, for example, principal components and spike amplitudes have already been computed. To load the extension object you can run:

ext = sorting_analyzer.get_extension("spike_amplitudes")
ext_data = ext.get_data()

Here ext is the extension object (in this case the SpikeAmplitudeCalculator), and ext_data will contain the actual amplitude data. Note that different extensions might have different ways to return the extension. You can use ext.get_data? for documentation.

To check what extensions spikeinterface can calculate, you can use the get_computable_extensions method.

all_computable_extensions = sorting_analyzer.get_computable_extensions()
print(all_computable_extensions)

>>> ['random_spikes', 'waveforms', 'templates', 'noise_levels', 'amplitude_scalings', 'correlograms', 'isi_histograms', 'principal_components', 'spike_amplitudes', 'spike_locations', 'template_metrics', 'template_similarity', 'unit_locations', 'quality_metrics']

There is detailed documentation about each extension below. Each extension comes from a different module. To use the postprocessing extensions, you’ll need to have the postprocessing module loaded.

Some extensions depend on another extension. For instance, you can only calculate principal_components if you’ve already calculated both random_spikes and waveforms. We say that principal_components is a child of the other two or that is depends on the other two. Other extensions, like isi_histograms, don’t depend on anything. It has no children and no parents. The parent/child relationships of all the extensions currently defined in spikeinterface can be found in this diagram:

Parent child relationships for the extensions in spikeinterface

If you try to calculate a child before calculating a parent, an error will be thrown. Further, when a parent is recalculated we delete its children. Why? Consider calculating principal_components. This depends on random selection of spikes chosen during the computation of random_spikes. If you recalculate the random spikes, a different selection will be chosen and your principal_components will change (a little bit). Hence your principal components are inconsistent with the random spikes. To avoid this inconsistency, we delete the children.

We can also delete an extension ourselves:

sorting_analyzer.delete_extension("spike_amplitudes")

This does not delete the children of the extension, since there are some cases where you might want to delete e.g. the (large) waveforms but keep the (smaller) postprocessing outputs.

Computing extensions¶

To compute extensions we can use the compute method. There are several ways to pass parameters so we’ll go through them here, focusing on the principal_components extension. Here’s one way to compute the principal components of a SortingAnalyzer object called sorting_analyzer with default parameters:

sorting_analyzer.compute("principal_components")

In this simple case you can alternatively use compute_principal_components(sorting_analyzer), which matches legacy syntax. You can also compute several extensions at the same time by passing a list:

sorting_analyzer.compute(["principal_components", "templates"])

You might want to change the parameters. Two parameters of principal_components are n_components and mode. We can choose these as follows:

sorting_analyzer.compute("principal_components", n_components=3, mode="by_channel_local")

As your code gets more complicated it might be easier to store your calculation in a dictionary, especially if you’re calculating more than one thing:

compute_dict = {
    'principal_components': {'n_components': 3, 'mode': 'by_channel_local'},
    'templates': {'operators': ["average"]}
}
sorting_analyzer.compute(compute_dict)

There are also hybrid options, which can be helpful if you’re mostly using default parameters:

# here `templates` will be calculated using default parameters.
extension_params = {
    'principal_components': {'n_components': 3, 'mode': 'by_channel_local'},
}
sorting_analyzer.compute(
    ["principal_components", "templates"],
    extension_params=extension_params
)

Extensions are generally saved in two ways, suitable for two workflows:

When the sorting analyzer is stored in memory, the extensions are only saved when the .save_as method is called. This saves the sorting analyzer and all it’s extensions in their current state. This is useful when trying out different parameters and initially setting up your pipeline.
When the sorting analyzer is stored on disk the extensions are, by default, saved when they are calculated. You calculate extensions without saving them by specifying save=False as a compute argument. (e.g. sorting_analyzer.compute('waveforms', save=False)).

NOTE: We recommend choosing a workflow and sticking with it. Either keep everything on disk or keep everything in memory until you’d like to save. A mixture can lead to unexpected behavior. For example, consider the following code

sorting_analyzer = create_sorting_analyzer(
    sorting=sorting,
    recording=recording,
    format="memory",
)

sorting_analyzer.save_as(folder="my_sorting_analyzer")
sorting_analyzer.compute("random_spikes", save=True)

Here the random_spikes extension is not saved. The sorting_analyzer is still saved in memory. The save_as method only made a snapshot of the sorting analyzer which is saved in a folder. Hence compute doesn’t know about the folder and doesn’t save anything. If we wanted to save the extension we should have started with a non-memory sorting analyzer:

sorting_analyzer = create_sorting_analyzer(
    sorting=sorting,
    recording=recording,
    format="binary_folder",
    folder="my_sorting_analyzer"
)

sorting_analyzer.compute("random_spikes", save=True)

Available postprocessing extensions¶

noise_levels¶

This extension computes the noise level of each channel using the median absolute deviation. As an extension, this expects the Recording as input and the computed values are persistent on disk.

noise = compute_noise_level(recording=recording)

principal_components¶

This extension computes the principal components of the waveforms. There are several modes available:

“by_channel_local” (default): fits one PCA model for each by_channel
“by_channel_global”: fits the same PCA model to all channels (also termed temporal PCA)
“concatenated”: concatenates all channels and fits a PCA model on the concatenated data

If the input SortingAnalyzer is sparse, the sparsity is used when computing the PCA. For dense waveforms, sparsity can also be passed as an argument.

pc = sorting_analyzer.compute(
    input="principal_components",
    n_components=3,
    mode="by_channel_local"
)

For more information, see compute_principal_components()

template_similarity¶

This extension computes the similarity of the templates to each other. This information could be used for automatic merging. Currently, the only available similarity method is the cosine similarity, which is the angle between the high-dimensional flattened template arrays. Note that cosine similarity does not take into account amplitude differences and is not well suited for high-density probes.

similarity = sorting_analyzer.compute(input="template_similarity", method='cosine_similarity')

For more information, see compute_template_similarity()

spike_amplitudes¶

This extension computes the amplitude of each spike as the value of the traces on the extremum channel at the times of each spike.

NOTE: computing spike amplitudes is highly recommended before calculating amplitude-based quality metrics, such as Amplitude cutoff (amplitude_cutoff) and Amplitude median (amplitude_median).

amplitudes = sorting_analyzer.compute(input="spike_amplitudes", peak_sign="neg")

For more information, see compute_spike_amplitudes()

spike_locations¶

This extension estimates the location of each spike in the sorting output. Spike location estimates can be done with center of mass (method="center_of_mass" - fast, but less accurate), a monopolar triangulation (method="monopolar_triangulation" - slow, but more accurate), or with the method of grid convolution (method="grid_convolution")

NOTE: computing spike locations is required to compute Drift metrics (drift_ptp, drift_std, drift_mad).

spike_locations = sorting_analyzer.compute(
    input="spike_locations",
    ms_before=0.5,
    ms_after=0.5,
    spike_retriever_kwargs=dict(
        channel_from_template=True,
        radius_um=50,
        peak_sign="neg"
    ),
    method="center_of_mass"
)

For more information, see compute_spike_locations()

unit_locations¶

This extension is similar to the spike_locations, but instead of estimating a location for each spike based on individual waveforms, it calculates at the unit level using templates. The same localization methods (method="center_of_mass" | "monopolar_triangulation" | "grid_convolution") are available.

unit_locations = sorting_analyzer.compute(input="unit_locations", method="monopolar_triangulation")

For more information, see compute_unit_locations()

template_metrics¶

This extension computes commonly used waveform/template metrics. By default, the following metrics are computed:

“peak_to_valley”: duration in \(s\) between negative and positive peaks
“halfwidth”: duration in \(s\) at 50% of the amplitude
“peak_to_trough_ratio”: ratio between negative and positive peaks
“recovery_slope”: speed to recover from the negative peak to 0
“repolarization_slope”: speed to repolarize from the positive peak to 0
“num_positive_peaks”: the number of positive peaks
“num_negative_peaks”: the number of negative peaks

The units of recovery_slope and repolarization_slope depend on the input. Voltages are based on the units of the template. By default this is \(\mu V\) but can be the raw output from the recording device (this depends on the return_scaled parameter, read more here: SortingAnalyzer). Distances are in \(\mu m\) and times are in seconds. So, for example, if the templates are in units of \(\mu V\) then: repolarization_slope is in \(mV / s\); peak_to_trough_ratio is in \(\mu m\) and the halfwidth is in \(s\).

Optionally, the following multi-channel metrics can be computed by setting: include_multi_channel_metrics=True

“velocity_above”: the velocity in \(\mu m/s\) above the max channel of the template
“velocity_below”: the velocity in \(\mu m/s\) below the max channel of the template
“exp_decay”: the exponential decay in \(\mu m\) of the template amplitude over distance
“spread”: the spread in \(\mu m\) of the template amplitude over distance

../_images/1d_waveform_features.png — Visualization of template metrics. Image from ecephys_spike_sorting from the Allen Institute.¶

tm = sorting_analyzer.compute(input="template_metrics", include_multi_channel_metrics=True)

For more information, see compute_template_metrics()

correlograms¶

This extension computes correlograms (both auto- and cross-) for spike trains. The computed output is a 3d array with shape (num_units, num_units, num_bins) with all correlograms for each pair of units (diagonals are auto-correlograms).

ccg = sorting_analyzer.compute(
    input="correlograms",
    window_ms=50.0,
    bin_ms=1.0,
    method="auto"
)

For more information, see compute_correlograms()

isi_histograms¶

This extension computes the histograms of inter-spike-intervals. The computed output is a 2d array with shape (num_units, num_bins), with the isi histogram of each unit.

isi =  sorting_analyer.compute(
    input="isi_histograms"
    window_ms=50.0,
    bin_ms=1.0,
    method="auto"
)

For more information, see compute_isi_histograms()