Mutliplication behavior

[e-dub]

I am trying to understand the behavior of xarray (in comparison with numpy arrays and sparse matrices). I have simplified my case below to a 2D problem to show the problem. I think that there must be a simple solution but am breaking my head around it. Therefore, any help would be most appreciated!

import numpy as np
import sparse as sp
import xarray as xr

ndof = 10
nx = 5

a = np.ones((ndof, ndof))
b = np.ones((ndof, nx))

aSp = sp.COO(a)
bSp = sp.COO(b)

aXr = xr.DataArray(a)
bXr = xr.DataArray(b)

c = a@b  # wroks
cSp = aSp@bSp  # works!
cXr = aXr@bXr  # does not work!!!

[keewis]

if you look at the DataArray objects, the dimension names (arr.dims) are ("dim0", "dim1") for both objects. Since aXr and bXr have different shapes, the matrix multiplication fails. To fix this, you can add dimension names:

In [5]: aXr = xr.DataArray(a, dims=("a", "b"))
   ...: bXr = xr.DataArray(b, dims=("b", "c"))
   ...: aXr @ bXr
Out[5]: 
<xarray.DataArray (a: 10, c: 5)>
array([[10., 10., 10., 10., 10.],
       [10., 10., 10., 10., 10.],
       [10., 10., 10., 10., 10.],
       [10., 10., 10., 10., 10.],
       [10., 10., 10., 10., 10.],
       [10., 10., 10., 10., 10.],
       [10., 10., 10., 10., 10.],
       [10., 10., 10., 10., 10.],
       [10., 10., 10., 10., 10.],
       [10., 10., 10., 10., 10.]])
Dimensions without coordinates: a, c

note that having multiple dimensions with the same name is not supported (i.e. xr.DataArray(..., dims=("a", "a")) breaks)

[e-dub]

That worked well! Thanks for the quick answer!

Unfortunately this solution causes problems further downwards in my code. Is there a way to do this without explicitly naming the dimensions?

[max-sixty]

Dimensions are always named in xarray.

Could I ask why use xarray if not for named / labeled dimensions?

[e-dub]

I am trying to speed up my code where I have large multiple dimensional arrays (i.e. 2D+). Currently comparing xarray and sparse for this. Or might you have another suggestion?

Thanks!

[max-sixty]

Xarray mostly wraps numpy or dask or sparse. I find named & labeled dimensions extremely helpful. But if you don't want them, that's totally fine — you might want to use those array libraries directly rather than xarray.

[e-dub]

Xarray is giving me an error downstream that the package sparse does not, namely number of dimensions and number of named dimensions. I am assuming that xarray automatically names the dimensions if not explicitly given and this mix of named and unnamed (in the rest of the code) is causing problems.

What I really like about xarray is the index assignment!
A[i,j,k] = whatever
This is a pain for sparse...

Thanks again for the comments and help here!