fix PSD weights handling when bad annotations present

[drammock]

follow-up to #12536

• corrects bug where weights were discarded for spans shorter than n_per_seg (the wrapped function actually just shrinks n_per_seg in those cases)
• reduces weights for long spans to the number of used samples (where we expect trailing samples to be discarded)

[cbrnr]

I can confirm that this fixes the error from #11413. I do get a warning though:

UserWarning: nperseg = 2048 is greater than input length  = 1201, using nperseg = 1201

I understand that this is expected and technically correct, but from a user perspective we are just trying to compute a spectrum (using data with annotations), so I'm not sure if this warning causes more confusion than anything else. Would an info message not be sufficient here?

[drammock]

I'm not sure if this warning causes more confusion than anything else. Would an info message not be sufficient here?

message comes from scipy, not from MNE. We could catch and swallow the warning maybe? Or replace it with an info log message that is a bit more informative.

[larsoner]

We could catch and swallow the warning maybe? Or replace it with an info log message that is a bit more informative.

I'm fine with swallowing the warning or re-emitting. Instead of re-emitting it's probably even cleaner to just detect the too-short condition in code (since you have direct access to all the lengths) and emit a message that way rather than parse the obtained warnings or whatever.

[cbrnr]

Somehow I still see the UserWarning:

>>> raw.plot_psd()
NOTE: plot_psd() is a legacy function. New code should use .compute_psd().plot().
Setting 1201 of 6007 (19.99%) samples to NaN, retaining 4806 (80.01%) samples.
Effective window size : 3.410 (s)
At least one good data span is shorter than n_per_seg, and will be analyzed with a shorter window than the rest of the file.
/Users/clemens/Projects/mne-python/mne/time_frequency/psd.py:52: UserWarning: nperseg = 2048 is greater than input length  = 1201, using nperseg = 1201
  _, _, spect = func(epoch)
<stdin>:1: FutureWarning: The value of `amplitude='auto'` will be removed in MNE 1.8.0, and the new default will be `amplitude=False`.
Plotting power spectral density (dB=True).
<MNELineFigure size 2000x700 with 2 Axes>

For reference, here's the MWE again:

import mne
from mne.datasets import sample

fname = sample.data_path() / "MEG" / "sample" / "sample_audvis_raw.fif"
raw = mne.io.read_raw(fname).pick_types(eeg=True).crop(0, 10).load_data()

raw.plot_psd()  # works
raw.set_annotations(mne.Annotations(2, 2, "bad"))
raw.plot_psd()  # UserWarning

[drammock]

Argh, I didn't see that when I tried. I'll look again. The warnings catcher comes with a caveat that its behavior is undefined in multi threaded contexts, so I may need to handle the warning at a lower level (maybe wrap the scipy func directly)

[drammock]

OK, CIs are green and the warning-catcher now immediately wraps scipy.signal.spectrogram so it should be immune to multi-threading issues.

[larsoner]

Thanks @drammock !