Make passing a DataArray for the xarray.concat dim argument equivalent to passing a pandas Index

[ceridwen]

Extending from #839, if I'm concatenating some DataArrays using concat,

print(xarray.concat(data, xarray.DataArray(['foo1', 'foo2', 'foo3', 'foo4', 'foo5'], name='stat'))

I get an unnamed dimension without coordinates.

<xarray.DataArray (dim_0: 5, index: 2)>
array([[ 24.841064,   0.750451],
       [ 24.841064,   0.750451],
       [ 19.062874,   0.796722],
       [ 14.9631  ,   0.354273],
       [ 14.9631  ,   0.354273]])
Coordinates:
  * index    (index) object 'Intercept' 'Lvl'
         (dim_0) <U3 'foo1' foo2' 'foo3' 'foo4' 'foo5'
Dimensions without coordinates: dim_0

Using a pandas.Index,

print(xarray.concat(data, pandas.Index(['foo1', 'foo2', 'foo3', 'foo4', 'foo5'], name='stat'))

<xarray.DataArray (stat: 5, index: 2)>
array([[ 14.9631  ,   0.354273],
       [ 19.982272,   0.555708],
       [ 14.974026,   0.60658 ],
       [ 24.841064,   0.750451],
       [ 24.841064,   0.750451]])
Coordinates:
  * index    (index) object 'Intercept' 'Lvl'
  * stat     (stat) object 'foo1' 'foo2' 'foo3' 'foo4' 'foo5'

I want the latter, not the former, but I expected the latter when using a DataArray.

[shoyer]

Agreed, this should definitely work! (I think the fact that it doesn't is probably related to the relatively recent change that made coordinate labels along dimensions optional.)

[shoyer]

This is the location of the helper function that parsing the dim argument:

xarray/xarray/core/combine.py

Line 123 in 9763a66

def _calc_concat_dim_coord(dim):

[dcherian]

I have a simple fix in #1812 .

With that change, this works.

print(xarray.concat(data, xarray.DataArray(['foo1', 'foo2', 'foo3', 'foo4', 'foo5'], name='stat'))

But if you provided a dimension name like

print(xarray.concat(data, xarray.DataArray(['foo1', 'foo2', 'foo3', 'foo4', 'foo5'], dims=['new_dim'], name='stat'))

then new_dim is renamed to stat. What is the desired behaviour in this case viz. do we preserve dimension name new_dim or assign new name stat?

[shoyer]

What is the desired behaviour in this case viz. do we preserve dimension name new_dim or assign new name stat?

Oh, this is trickier than I thought!

The challenge is that once you make the DataArray, there is no good way to know if a default dimension name like 'dim_0' was intentional or not.

The way to handle this currently is to pass a 1-dimensional xarray.Variable object for the dim argument. These don't have separate names, so there's no ambiguity:

In [5]: xarray.concat([xarray.DataArray(1), xarray.DataArray(2)], dim=xarray.Variable('x', [3, 4]))
Out[5]:
<xarray.DataArray (x: 2)>
array([1, 2])
Coordinates:
  * x        (x) int64 3 4

But this is a little verbose. Potentially we could call xarray.as_variable() on tuple inputs (like in the Dataset constructor) so dim=('x', [3, 4]) works.

[stale]

In order to maintain a list of currently relevant issues, we mark issues as stale after a period of inactivity

If this issue remains relevant, please comment here or remove the stale label; otherwise it will be marked as closed automatically

[max-sixty]

One option here would be to take the name of the dataarray iff the dim name is dim_0. While I'm not a great fan of breaking that abstraction and hardcoding exceptions, it does solve the case quite well:

In [7]: xr.concat([da,da], xr.DataArray(['x', 'y'], name='stat'))
Out[7]:
<xarray.DataArray 'air' (dim_0: 2, time: 2920, lat: 25, lon: 53)>
array([[[[241.2    , 242.5    , 243.5    , ..., 232.79999, 235.5    ,
          238.59999],
         [243.79999, 244.5    , 244.7    , ..., 232.79999, 235.29999,
          295.19   ],
         [297.69   , 298.09   , 298.09   , ..., 296.49   , 296.19   ,
          295.69   ]]]], dtype=float32)
Coordinates:
  * lat      (lat) float32 75.0 72.5 70.0 67.5 65.0 ... 25.0 22.5 20.0 17.5 15.0
  * lon      (lon) float32 200.0 202.5 205.0 207.5 ... 322.5 325.0 327.5 330.0
  * time     (time) datetime64[ns] 2013-01-01 ... 2014-12-31T18:00:00
  * dim_0    (dim_0) <U1 'x' 'y'  # Very unlikely we even want `dim_0` here

It's very unlikely we even want dim_0 there rather than stat

[dcherian]

We could also make it only succeed if DataArray.name is None and raise an error otherwise.

Supporting the tuple form also seems like a good idea

[simonkeys]

Just pointing out that the documentation says "If dimension is provided as a DataArray or Index, its name is used as the dimension to concatenate along and the values are added as a coordinate." Which seems to currently be not true for a DataArray.