Adjust eps for whitening in case of very small magnitude data

[magland]

Resolves #2064

I explain what is happening in the comments

I ran the following test:

import numpy as np
import spikeinterface.extractors as se
import spikeinterface.preprocessing as spre

# toy example
recording, sorting = se.toy_example(num_channels=4, duration=10, seed=0)
recording_whitened = spre.whiten(recording)
traces = recording_whitened.get_traces(segment_index=0)
print(np.var(traces, axis=0))

recording2 = spre.scale(recording, gain=1e-8)
recording2_whitened = spre.whiten(recording2)
traces2 = recording2_whitened.get_traces(segment_index=0)
print(np.var(traces2, axis=0))

And got the expected variances close to 1

[1.0054402  1.0075228  0.9955653  0.96718967]
[0.99692386 0.99832404 0.98711884 0.9610572 ]

And without the adjustment the second array is very wrong

[1.0054402  1.0075228  0.9955653  0.96718967]
[1.1976793e-08 1.1291137e-08 1.2642105e-08 1.6557108e-08]

[alejoe91]

@magland we discussed internally and we would like to expose epsilon and make a few changes. I cannot push to your branch. Can you allow me to do that?

Here's my working branch: https://github.com/alejoe91/spikeinterface/tree/adjust-eps-for-whitening

Needs another look

[magland]

@alejoe91 I have invited you to my fork of spikeinterface.

[magland]

@alejoe91 It seems that the implementation does not quite agree with the doc string. It says

If None, eps is estimated from the data.

But then the code says

if data.dtype.kind == "f" or eps is None:

Notice the "or"

IMO, the estimation of eps should only happen if eps is None.... AND ... the default should be None => effective default is 1e-8. That's somewhat different from what you have. If you do go the route of having default eps is None, which is what I am suggesting, then it needs to be updated throughout the codebase.

[alejoe91]

@magland done! Thanks for the feedback!

[magland]

@alejoe91 looks good to me!