[Feature]: waveform_mean and _sd should be ragged to enable writing sparse waveforms.

[alejoe91]

What would you like to see added to PyNWB?

We recently implemented in SpikeInterface the option to compute sparse waveforms, which is very convenient for high-density probes like Neuropixels.
The waveform_mean and waveform_sd columns of the units table are suitable to write templates computed on sparse channels. However, in order for this to happend, we should enable ragged arrays for them in the schema.

Currently, writing sparse waveforms to h5 or zarr result in an error for this reason.

Is your feature request related to a problem?

Yes, it's currently not possible to write waveform means and sds with uneven dimensions in the channel dimension.

What solution would you like?

To enable waveform_mean and waveform_sd to be ragged (index=True)

Do you have any interest in helping implement the feature?

Yes.

Code of Conduct

• I agree to follow this project's Code of Conduct
• Have you checked the Contributing document?
• Have you ensured this change was not already requested?

[alejoe91]

@rly @oruebel can we discuss about this?

[alejoe91]

@rly can I ping you on this? this is a feature that we would really like both for writing sparse templates and for writing waveforms from multiple probes for which different numbers of "bad" channels were removed during preprocessing.

[rly]

Thanks for the ping @alejoe91 and sorry for the delay in response. This is both an API and schema-level issue. Currently, the waveform_mean and waveform_sd columns are not ragged and have dimensions (num_units, num_samples) or (num_units, num_samples, num_electrodes). There must be at least one waveform provided for each unit, and the number of electrodes (channels) used for each unit must be the same. Despite that, the electrodes column is ragged, so one can specify a different number of electrodes per unit.

from pynwb.misc import Units
u = Units()
u.add_row(waveform_mean=[[1, 2, 3], [4, 5, 6]])
u.add_row(waveform_mean=[[7, 8, 9]])

@alejoe91 It sounds like you would like the ability to specify waveform_mean and waveform_sd with a different number of electrodes for each unit. Is that right?

To do that, we would add an "index" column to each of waveform_mean and waveform_sd in the schema, and then make those columns ragged in the API with index=True.

[alejoe91]

Thanks @rly

Yeah that's exactly what I'm suggesting :)

[rly]

This fix is pending a larger discussion on how to reconcile waveform_mean being a 2D array of samples x electrodes when there is a fixed number of electrodes per unit, and waveform_mean being a 2D array of samples x 1 with a VectorIndex when there is a variable number of electrodes per unit. Should the former set up be deprecated in favor of the more flexible latter option? Do we maintain both? That can be confusing for any downstream tool or person...