Repository navigation
__array_ufunc__ for xarray #1617
Description
Now that __array_ufunc__ has appeared in NumPy 1.13, we should be able to resolve a long-standing annoyance with xarray: the inability to use NumPy ufuncs like np.sin() directly on all xarray objects, such as xarray.Dataset.
I believe this could be straightforwardly accomplished by wrapping xarray.core.computation.apply_ufunc.
Some care should be taken to ensure that properly defer to other array types, as suggested in the NEP. This also could be a good time to resolve some of the xarray's broader consistency issues when interacting with and wrapping other array-like types.
Reacted by Maximilian Roos and Joe Hamman
Activity
shoyer commentedon Oct 29, 2017
A related task would be finishing up this PR:
ndarrayunless required #1118: Do not convert subclasses ofndarrayunless requiredUltimately, it would be nice to have some sort of registration system, especially for objects that aren't ndarray subclasses, that indicates that a class can safely be wrapped by xarray.
I'm thinking something like
xarray.register_data_type(ArrayWithUnits)which indicates that.datafor xarray objects can now be anArrayWithUnitsobject. This would also possibly require passing a reference to a set of function to use on data containing these objects (e.g., for stacking or concatenating).shoyer commentedon Mar 5, 2018
I'm starting to work on this.