Remove Compositor extra dims using composite_dims

[jmilloy]

Adds composite_dims trait and removes extra dimensions from the evaluation coordinates.

• autodetect, ignore, or fail if the composite_dims are not provided?
• tests

Autodetection could look like this at its simplest:

@tl.default("composite_dims")
    def _default_composite_dims(self):
        return self.sources[0].coordinates.udims

[jmilloy]

@mpu-creare Another option instead of autodetecting the composite dims would just be to only remove extra dimensions if the composite_dims was defined. Maybe this is what you implemented as well.

[coveralls]

Coverage increased (+0.004%) to 91.025% when pulling 91d8039 on feature/composite_dims into 2773e8d on develop.

[mpu-creare]

Yes, that was exactly my intent. YOu should be able to autodetect 90% of the time, but for cases like TerrainTiles you can just drop the extra dimenions... or perhaps TerrainTiles is a special case that should just be handled there?

[mlshapiro]

I don't think TerrainTiles is a necessarily a special case. It is an example of a dataset that does not have a time dimension, of which there may be many.

[mlshapiro]

But yes - agree that auto-detecting makes a lot of sense!

@jmilloy thanks for getting right on it - let me know if you want me to help test / writes tests if you don't have time to work on this

[mpu-creare]

I'm a little hesitant about the autodetection because of all the circular reference we run into whenever we have a default set from compositor.sources. Tagging dims as an attribute makes this fixable in any pipeline, so I'm happy with this. Works for my current case in any case.

Just needs tests then I'm happy to merge.

[jmilloy]

Note

The obvious way to remove extra dimensions would be something like this:

extra = [dim for dim in coordinates.udims if dim not in self.dims]
coordinates = coordinates.drop(extra)

But if the compositor dims are lat, lon and the requested dims are lat, lon, time_alt, you have to remove the time_alt dimension together. On the other hand, if the requested dims are lat_lon_alt, you don't need to remove alt at all. So the removal is implemented like this:

extra = [
    c.name
    for c in coordinates.values()
    if (isinstance(c, Coordinates1d) and c.name not in self.dims)
    or (isinstance(c, StackedCoordinates) and all(dim not in self.dims for dim in c.dims))
]
coordinates = coordinates.drop(extra)

It is also implemented this way in DataSource.eval. In both the Compositor and DataSource tests, the lat, lon, alt_time case is tested, but the lat_lon_time case is not tested (commented out) because the interpolator cannot handle this case yet.

[mpu-creare]

I have a few comments suggestions, but happy to merge.

[mlshapiro]

FYI I merged in my changes to Terrain Tiles and the TileCompositor here since it would conflict as a separate branch. See bc2bb16 for more information. Long story short, Terrain Tiles should now composite correctly. Whew. Thanks @jmilloy