MRG, BUG: Make sure resample return at least one sample

[mscheltienne]

Issue spotted in xdawn decoding example, when replacing tmp_info['sfreq'] = 1. with epochs.copy().resample(1).

https://github.com/mne-tools/mne-python/blob/main/examples/decoding/decoding_xdawn_eeg.py#L105-L106

Original post:

• MRG: Add a locking/unlocking context manager to Info for error checking. #9867 (comment)
• MRG: Add a locking/unlocking context manager to Info for error checking. #9867 (comment)
• MRG: Add a locking/unlocking context manager to Info for error checking. #9867 (comment)

[agramfort]

not sure to get it. You expose a bug by adding a new test but the CIs pass. What do I miss?

[mscheltienne]

@agramfort I skipped the CI for now, wanted to push a fix first, probably this afternoon.
Maybe not the smartest thing to do..

[mscheltienne]

I gave it some thought, and IMO, it just doesn't make sense to allow resampling to yield 0 data points and to return an empty instance. I chose to raise an error when it happens. If you disagree and prefer to have resample always return at least one datapoint, then I can change it to return e.g. the first datapoint when new_freq < current_freq / current_n_samples.

I don't see a use case for resampling to a frequency yielding only 1 point, and 0 makes even less sense to me. Please let me know if you can think of a use case!

Previously, resample would return an instance with 0 datapoints, which would then raise errors when e.g. representing.

IN: raw_x_sample.times
array([], dtype=float64)

IN: repr(raw_x_sample)
Traceback (most recent call last):

  File "/var/folders/b1/kx2x00x95ls80zzkct4647sh0000gn/T/ipykernel_11717/1501997967.py", line 1, in <module>
    repr(raw_x_sample)

  File "/Users/scheltie/Documents/git/mne-python/mne/io/base.py", line 1751, in __repr__
    % (name, len(self.ch_names), self.n_times, self.times[-1],

IndexError: index -1 is out of bounds for axis 0 with size 0

[mscheltienne]

I'm guessing <compat / old / py3.7> runs with CUDA, and that in this case, numpy had this issue covered with a ValueError raised. I matched the error type/msg with numpy.

[larsoner]

@mscheltienne I pushed a commit to add one more max where it was needed. Feel free to add a line to latest.inc about this bugfix !

[mscheltienne]

Added changelog entry under your name for this one, as I didn't do much except create the PR 😄
CIs failure looked unrelated. Let's see if it comes back green this time.

[larsoner]

as I didn't do much except create the PR

And find the problem, and think about a solution, and work through it, and add a test... I think your name should be there, too :)

[larsoner]

Thanks @mscheltienne !