Xarray: Interpolation using non-dimension coordinates

[equaeghe]

MCVE Code Sample

>>> da = xr.DataArray([1., 2.], coords=[('dim', [.5, 1.])])
>>> da.coords['nondim'] = ('dim', [0., 1.])
>>> da
<xarray.DataArray (dim: 2)>
array([1., 2.])
Coordinates:
  * dim      (dim) float64 0.5 1.0
    nondim   (dim) float64 0.0 1.0
>>> da.interp(dim=.75)
<xarray.DataArray ()>
array(1.5)
Coordinates:
    nondim   float64 0.5
    dim      float64 0.75
>>> da.interp(nondim=.5)
Traceback (most recent call last):

  File "<ipython-input-192-e3df34cff90f>", line 1, in <module>
    da.interp(nondim=.5)

  File "/usr/lib64/python3.6/site-packages/xarray/core/dataarray.py", line 951, in interp
    **coords_kwargs)

  File "/usr/lib64/python3.6/site-packages/xarray/core/dataset.py", line 1860, in interp
    indexers = OrderedDict(self._validate_indexers(coords))

  File "/usr/lib64/python3.6/site-packages/xarray/core/dataset.py", line 1316, in _validate_indexers
    raise ValueError("dimensions %r do not exist" % invalid)

ValueError: dimensions ['nondim'] do not exist

Expected Output

The same interpolation as for 'dim'.

Problem Description

Apparently, xarray currently cannot interpolate using non-dimension coordinates.

Output of xr.show_versions()

commit: None

python: 3.6.5.final.0
python-bits: 64
OS: Linux
OS-release: 4.19.72-gentoo
machine: x86_64
processor: Intel(R) Core(TM) i7-2620M CPU @ 2.70GHz
byteorder: little
LC_ALL: None
LANG: nl_BE.UTF-8
LOCALE: nl_BE.UTF-8

xarray: 0.10.8
pandas: 0.24.2
numpy: 1.14.5
scipy: 1.1.0
netCDF4: 1.5.1.2
h5netcdf: None
h5py: 2.9.0
Nio: None
zarr: None
bottleneck: 1.2.1
cyordereddict: None
dask: 1.2.0
distributed: None
matplotlib: 2.2.2
cartopy: None
seaborn: None
setuptools: 40.6.3
pip: 19.1
conda: None
pytest: 3.10.1
IPython: 5.4.1
sphinx: 1.7.5

[keewis]

you might be able to achieve what you want by using swap_dims:

da.swap_dims({"dim": "nondim"}).interp(nondim=.5)

[equaeghe]

you might be able to achieve what you want by using swap_dims

Yes, that is a nice workaround (I used another, less convenient one). But I would say that the current (lack of) functionality is a bug: the documentation talks about coordinates, not about dimensions. Moreover, it seems like something one would want to support. I do not know how complicated this would be to implement, but if reasonably feasible, I would request to keep this issue open.

[TomNicholas]

This is an example of a general UI problem which I was actually about to raise an issue about.

There are loads of places in xarray's API which require a dimension, but the operation would make just as much sense if a 1D coordinate were supplied instead (as that uniquely specifies a dimension).

I encountered the same kind of problem in #3334 with da.transpose().

I wonder if it would be possible to have a little helper function process the dim input to these functions and if it's a 1D coord, just demote it to the corresponding dim.

[dcherian]

More generally, this is about applying operations with non dimensional coordinates which need not be 1D. Thinking about API, how about adding a coord kwarg?

da.differentiate(dim="nx", coord="x")

This says use np.diff on da along dimension nx and divide by spacing calculated by np.diff on da.x along dimension nx

If x is 1D, then da.differentiate(coord="x") is equivalent to da.differentiate(dim="nx", coord="x") which is your proposal. i think this is cleaner and more explicit than automatically mapping 1D nondimensional coords to dimensions.