ENH: Support obects with __array__ method

[snowman2]

https://docs.scipy.org/doc/numpy/reference/arrays.interface.html

Had the thought, putting here to look into.

[mraspaud]

Yes, it would be great to have support for xarray.DataArray or dask.array.Array objects!

[snowman2]

I did some looking into xarray.DataArray and is seems it does not have __array_interface__. However, it does have __array__. @mraspaud did you have any thoughts on what behavior you would like to see there?

[mraspaud]

Oh, sorry, I hadn't checked that xarray had this interface. Anyway, our use case is that we have xarray.DataArray structures containing dask.array.Array objects. Passing any of these to a Proj object just fails at the moment:

 In [1]: import xarray as xr                                                     

In [2]: from pyproj import Proj                                                 

In [6]: prj(xr.DataArray([1, 2, 3]), xr.DataArray([1, 2, 3]))                   

TypeError: input must be an array, list, tuple or scalar

In [7]: import dask.array as da                                                 

In [8]: prj(da.from_array([1, 2, 3]), da.from_array([1, 2, 3]))                 

TypeError: input must be an array, list, tuple or scalar

It would be nice if pyproj wouldn't just choke on these. Just calling eg np.asarray should suffice. Now I realise pyproj uses the underlying proj library and thus converts array to C structures at one point, so it would loose the nice features of xarray or dask along the way, but it would work.

[snowman2]

Okay, so if I interpreted that correctly, you would be okay with getting a numpy array returned?

Next step is to check if dask has the same __array__ method.

[mraspaud]

I'd be ok with that. Maybe @djhoese has an opinion on this ?

[djhoese]

Part of me wishes pyproj could magically know how to handle dask arrays in a dask-friendly way (map_blocks wrapper), but I think converting to a numpy array and returning a numpy array matches xarray behavior. For example, when there are methods in xarray that dask doesn't support I believe xarray just converts to numpy and continues on processing.

[snowman2]

Looks like the dask array also has __array__, so I think the code could be updated to just call that function and it would work for both.

[snowman2]

Implemented in #625

[snowman2]

Looks like this will also work with pandas.Series.__array__