The SpatialAverager.w property outputs an intuitive 3D weights array. However, the mechanism to guess output chunks (when None is passed, the default) is in the BaseRegridder and it assumed that w was 4D.
This PR changes how we guess those output chunks:
Default chunks at most as large as the out shape.
Don't use self.w internally. I take the stance that self.w is for the user, meaning we can make it as intuitive as possible. The regridder mechanism makes the non-intuiive choice of using an out shape of (1, N) for the SpatialAverager, so I use the self.weights array directly instead. Also, we didn't need the DataArray container, so it removed (minimal) overhead.
Accept len-1 output_chunks when it's a sequence_out, but add the dummy singleton dimension before passing to dask.
Add SpatialAverager tests with dask. We missed the error because we were lacking this test.
This PR also reactivates the large region tests, adding a segmentize call, so it passes. Not sure how useful that test is, but it also doesn't cost much to keep it.
The
SpatialAverager.wproperty outputs an intuitive 3D weights array. However, the mechanism to guess output chunks (when None is passed, the default) is in theBaseRegridderand it assumed thatwwas 4D.This PR changes how we guess those output chunks:
self.winternally. I take the stance thatself.wis for the user, meaning we can make it as intuitive as possible. The regridder mechanism makes the non-intuiive choice of using an out shape of (1, N) for the SpatialAverager, so I use theself.weightsarray directly instead. Also, we didn't need the DataArray container, so it removed (minimal) overhead.output_chunkswhen it's asequence_out, but add the dummy singleton dimension before passing to dask.SpatialAveragertests with dask. We missed the error because we were lacking this test.This PR also reactivates the large region tests, adding a
segmentizecall, so it passes. Not sure how useful that test is, but it also doesn't cost much to keep it.