Array types incompatible error in convolve.

I am trying to use the convolve function from the dask_image library. No matter what combination of inputs I try, I seem to keep running into the same error.

My inputs are dcube:

<xarray.DataArray 'A' (time: 2, y: 823, x: 1296)>
array([[[330., 226., 709., ..., 135.,  82.,  76.],
        [125., 615., 293., ..., 156., 207., 127.],
        [126., 506., 218., ..., 228., 240., 179.],
        ...,
        [ 73., 137., 145., ..., 375., 453., 542.],
        [ 86.,  78., 148., ..., 321., 227., 619.],
        [ 88., 155.,  42., ..., 262., 279., 518.]],

       [[330., 226., 709., ..., 135.,  82.,  76.],
        [125., 615., 293., ..., 156., 207., 127.],
        [126., 506., 218., ..., 228., 240., 179.],
        ...,
        [ 73., 137., 145., ..., 375., 453., 542.],
        [ 86.,  78., 148., ..., 321., 227., 619.],
        [ 88., 155.,  42., ..., 262., 279., 518.]]])
Coordinates:
    pol      <U2 'HH'
  * y        (y) float64 3.653e+06 3.653e+06 3.653e+06 ... 3.652e+06 3.652e+06
  * x        (x) float64 1.047e+06 1.047e+06 1.047e+06 ... 1.049e+06 1.049e+06
  * time     (time) datetime64[ns] 2020-11-24T22:24:50 2020-11-18T01:15:05
Attributes:
    orbit:       A
    transform:   (1.1987408006010196, 0.0, 1047472.9120438152, 0.0, -1.198740...
    crs:         {'proj': 'utm', 'zone': 10, 'datum': 'WGS84', 'units': 'm', ...

and a patch:

filter_wt =  (1 / (N * N)) * np.ones((N, N),dtype='float64')
>> array([[0.11111111, 0.11111111, 0.11111111],
       [0.11111111, 0.11111111, 0.11111111],
       [0.11111111, 0.11111111, 0.11111111]])

When I try to use the convolve function as follows:

from dask_image.ndfilters import convolve
convolve(dcube[0], xr.DataArray(filter_wt), mode='constant', cval=0.0).compute()

I end up with this error: *** ValueError: Array types must be compatible.

Even though, DaskArray with a numpy backend and a numpy array should be compatible. To get around this, I tried initializing both as dask arrays:

(Pdb) convolve(da.from_array(dcube[i]), da.from_array(filter_wt), mode='constant', cval=0.0).compute()
*** ValueError: Array types must be compatible.

(Pdb) type(da.from_array(dcube[0]))
<class 'dask.array.core.Array'>
(Pdb) type(da.from_array(filter_wt))
<class 'dask.array.core.Array'>

Any instructions on why casting both to ask arrays didn’t work would be very helpful!

[GenevieveBuckley]

I’m going to close this issue now since it seems like the problem has been resolved.

For anyone reading, we’d be very happy to accept a pull request aimed at making the documentation clearer & more helpful for this particular problem. I think that’s a very good idea.

[GenevieveBuckley]

Thanks for responding to this. I tried setting them both to xarrays, but ran into an error that xarray doesn't have map_blocks. That’s why I assumed that the two must be dask arrays? One suggestion I have is to provide an override flag for this check.

The first argument must be a dask array, yes. Your other array inputs (in this case patch) must match the type of array stored in the dask array chunks of your first argument (dcube).

Examples of compatible arrays:

• Two (or more) dask arrays with numpy chunks
• A dask array with numpy chunks, and a numpy array

Examples of incompatible arrays:

• A numpy array and a cupy array
• A numpy array and an xarray.DataArray

The reason this is failing is probably that one is a ndarray and the other is a xarray.DataArray. As you mention that the check is discovering that one of the chunks is ndarray and the other is xarray.

Yes, this is a good summary.