Set num_histogram_bins in amplitude_cutoff (back) to 100

[ecobost]

#4353 changed the default num_histogram_bins in AmplitudeCutoff from 100 to 200 to get a more granular histogram/estimate.

spikeinterface/src/spikeinterface/metrics/quality/misc_metrics.py

Line 1034 in 797f0e8

"num_histogram_bins": 200,

But given the default amplitudes_bins_min_ratio=5, any cell with less than 1000 spikes will return amplitude_cutoff = NaN. That is (I think?) too big a hurdle (a unit with 0.1Hz firing rate will require 2.8 hours to spike 1000 times).

Assuming 500 spikes are enough to model the amplitude distribution, I change it back to 100 here. I also change the default value on the low-level amplitude_cutoff()to match this default.

EDIT: Even if we don't think 500 spikes are enough to model the distribution, the right parameter to change would be amplitudes_bins_min_ratio and not num_histogram_bins as we would rather have more samples per bin (lower variance)

[zm711]

If we go messing with AmplitudeCutoff and I don't get all NaN's I'll have no way to know my code is working :P

Isn't our default for random_spikes 500 see here. So if 500 is enough sample for a distribution for the core of the analyzer then I would agree with you that 500 should be enough here too!

[alejoe91]

If we go messing with AmplitudeCutoff and I don't get all NaN's I'll have no way to know my code is working :P

Isn't our default for random_spikes 500 see here. So if 500 is enough sample for a distribution for the core of the analyzer then I would agree with you that 500 should be enough here too!

Agreed! But on a related note we should maybe increase random spikes to get better estimates of templates #4389

[ecobost]

For what is worth, I use:

sa.create_sorting_analyzer(...,    num_spikes_for_sparsity=500) # default=100
analyzer.compute('random_spikes', max_spikes_per_unit=1000) # default 500
analyzer.compute('quality_metrics',  metric_params={
     'amplitude_cutoff': {'amplitudes_bins_min_ratio': 7.5}}) #i.e., 750 spikes, default is 5

Increasing the random_spikes to 1000 by default seems reasonable.