Typing of core classes

[simonottenhauskenbun]

This PR should make working with Annotation easier for those using type checkers.

In my own project I neede to wrap Annotation with another class and add ugly type-checked code like this:

class Diarization:
... 
    @classmethod
    def labeled_tracks(cls, diarization: Annotation) -> List[Tuple[Segment, TrackName, Label]]:
       return list(diarization.itertracks(yield_label=True))  # type: ignore

Now typed versions of yield_label=True and yield_label=False exist. Type checkers do not allow to "set the type" by parameter value.

Details

• fix type erros in pyannote.core.feature.py
• fix type erros in pyannote.core.utils.types.py
• Add named tuples SegmentTrack and SegmentTrackLabel as return types for itertracks in annotation.py

Notes on NamedTuples

SegmentTrack(NamedTuple) and SegmentTrackLabel(NamedTuple) are fully compatible with Tuple[Segment, TrackName]andTuple[Segment, TrackName, Label], i.e. they can be deconstructed via a,b,c = ...and accessed withtup[0], tup[1]In addition the fields can be accessed via dot-notation. This also works type-checked for the union type ofSegmentTrackandSegmentTrackLabel` for the common fields.

class SegmentTrack(NamedTuple):
    segment: Segment
    track: TrackName

class SegmentTrackLabel(NamedTuple):
    segment: Segment
    track: TrackName
    label: Label

s1 = SegmentTrack(...)
s2 = SegmentTrackLabel(...)

segment, track = s1 # OK!
segment, track, label = s2 # OK!

some_s: Union[SegmentTrack, SegmentTrackLabel] = s1 # or s2
print(some_s.segment) # OK!
print(some_s.track) # OK!

[evfinkn]

Okay this is kind of crazy because I was about to make a pull request for this same reason 🤯 Thanks for doing this!

[evfinkn]

I think this is what you meant:

Suggested change

	return self.itertracks_with_labels() # type: ignore
	return self.itertracks(yield_label=True) # type: ignore

However—while I'd be fine with having itertracks_with_labels and itertracks_without_labels—a better approach to getting typed itertracks might be using overload like I did here:

    @overload
    def itertracks(self, yield_label: Literal[False] = ...) -> Iterator[Track]: ...
    @overload
    def itertracks(self, yield_label: Literal[True]) -> Iterator[LabeledTrack]: ...
    @overload
    def itertracks(self, yield_label: bool) -> TrackIterator: ...

(Though this would obviously need to be slightly altered to use your NamedTuples). This has the added benefit of working with existing code. The same thing can probably be done with Timeline.crop.

[simonottenhauskenbun]

Thanks for the comments @evfinkn
I have tried overloaded typing with bool constants in the past, and then stumbled upon the note that overloading is only allowed for types, not values. However, I did not consider your approach of using Literal[False]. I will have to see if this works with pylance. if it does the overload approach is far more elegant 😀

edit: tested it and it works! Changed in 230cf3b

[hbredin]

A few months later, here I am looking at the PR.
That's great thanks.

Any chance you could the same for Timeline.crop_iter with and without returned mapping?

[hbredin]

FYI -- a bit busy right now but I did notice this PR.

I think pyannote.core API needs some love anyway and should probably be simplified...