Fast annotation from mask

Describe the new feature or enhancement

In mne/preprocessing/artifact_detection.py there is a function called _annotations_from_mask. As its name suggests it creates an annotation from a numpy array mask. A for loop has been used in its implementation, which could be very time-consuming when the number of components is large.

Describe your proposed implementation

This could be implemented without any for loop. I have the following code, working in my tests:

def _annotations_from_mask(times, mask, annot_name):
    """Construct annotations from boolean mask of the data."""
    from scipy.ndimage.morphology import distance_transform_edt
    from scipy.signal import find_peaks
    mask_tf = distance_transform_edt(mask)
    midpoint_index = find_peaks(mask_tf)[0]
    onsets_index = midpoint_index - mask_tf[midpoint_index].astype(int) + 1
    ends_index = midpoint_index + mask_tf[midpoint_index].astype(int) - 1
    
    # Ensure onsets_index >= 0, otherwise the duration starts from the beginning
    onsets_index[onsets_index < 0] = 0
    # Ensure ends_index < len(times), otherwise the duration is to the end of times
    ends_index[ends_index >= len(times)] = len(times) - 1

    onsets = times[onsets_index]
    ends = times[ends_index]
    durations = ends - onsets
    desc = [annot_name]*len(durations)
    return Annotations(onsets, durations, desc)

This implementation doesn’t require any additional package. With a mask of num_comps equal to 24389, %timeit gives for the original implementation and the mine respectively:

2min 1s ± 640 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

590 ms ± 16.2 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

Describe possible alternatives

I do not have any other implementation in mind.

Please tell me how you think of this enhancement. If everything seems fine, I will be glad to create a PR.

[agramfort]

looks smart !

can you send us a PR?

thx a lot

[snwnde]

Can you send the pull request or shall we do it ?

@agramfort I created the PR #10089. Some checks fail though. It seems a precision problem.