Implement select_sorting_periods in metrics

[alejoe91]

Depends on #4316

Slices a sorting object based on an array ov valid periods. Periods are defined as a structured dtype as:

base_period_dtype = [
    ("start_sample_index", "int64"),
    ("end_sample_index", "int64"),
    ("segment_index", "int64"),
    ("unit_index", "int64"),
]

EDIT:

Refactored computation of spike train metrics, to make sure that periods are consistently taken into account. Added 2 utils functions to compute durations per unit and bin edges per unit, that optionally use the provided periods

[samuelgarcia]

This is OK for me.
Make a clear documentation somwhere woule help ?

[alejoe91]

@chrishalcrow I refactored a few metrics to make sure durations, spike counts, and bins are properly accounted for when slicing with periods. Happy to discuss about this!

[yger]

This looks good to me, I think this is a great extension to be able to compute metrics on sub_periods only. A couple of minor comments, and I think everything is not too slow, but good work!

[yger]

do we assume not too many periods, just to be sure that all these masks won't take too long?

[yger]

assuming not too many periods, these masks would be fine. Otherwise, we'll need to optimize

[yger]

We don't want to extend these quality metrics to the extension() ? What if someone wants to only get the ISI, CCG or anything else only on the periods? Would it be easy to slice the sorting, and then compute only on the sub sorting? Are the extension robust w.r.t. periods?

[alejoe91]

We don't want to extend these quality metrics to the extension() ? What if someone wants to only get the ISI, CCG or anything else only on the periods? Would it be easy to slice the sorting, and then compute only on the sub sorting? Are the extension robust w.r.t. periods?

I think that extension by extension we could use valid_unit_periods when computed :)

[alejoe91]

@samuelgarcia @chrishalcrow tests added! all good now :)

[chrishalcrow]

When there are no good periods, then periods will be an empty array? Should we have a general strategy for when this is true - empty units often cause pain/nans...

[chrishalcrow]

When I choose periods so that only one unit is good, and get the num spikes, get

import spikeinterface.full as si
import numpy as np
from spikeinterface.core.base import unit_period_dtype

rec, sort = si.generate_ground_truth_recording(seed=1205)
sa = si.create_sorting_analyzer(sort, rec)
periods = np.array((0,1000,4000,0), dtype=unit_period_dtype)
quality_metrics = si.compute_quality_metrics(sa, periods=periods, metric_names = ['num_spikes'])

print(quality_metrics['num_spikes'].values)

>>> [162 173 135 154 158 144 151 141 153 162]

But this sorting only has spikes from unit index 0, right? Is count_num_spikes_per_unit below using an old cache or something?