issue_comments

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

Unfortunately neither of your suggestions work. With the second, I get the error:

• ValueError: parameter 'value': expected array with shape (10000, 100), got (10000, 245)

With the first:

• ValueError: operands could not be broadcast together with shapes (5000,100,245) (100,)

It's okay. I have something that works. And it's deterministic :D

I think this would also work (due to the intrinsic broadcasting behavior of numpy functions): python def get_chebyshev_distances_xarray_ufunc(array, dye_array): return abs(array - dye_array).max(axis=-2)

OK, I think you're doing this right. You want an output with shape ['all_sites', 'dye_sites'] right?

My suggestion would be to add an explicit call to np.broadcast_arrays() at the start of your applied function. This will make the dimensions a little easier to understand. ```python def get_chebyshev_distances_xarray_ufunc(array, dye_array): array, dye_array = np.broadcast_arrays(array, dye_array) # array is a 3D numpy array with logical dimensions ['all_sites', 'dye_sites', 'dim_1'] # dye_array is a 3D numpy array with logical dimensions ['all_sites', 'dye_sites', 'dim_1']

# compute the distance matrix
# return a numpy array with logical dimensions ['all_sites', 'dye_sites']

``` (edit note: fixed dimensions)

actually, scratch that

The behavior is definitely deterministic, if hard to understand!

phew!

Maybe it would help to describe what you were trying to do here.

Sure - thanks!

I have a dataset that's long, the sample code shown below is 200k rows, but the full dataset will be much larger. I'm interested in pairwise distances except not for all rows, just the distances for few thousand rows, wrt to the full 200k.

Here's how I hack this together:

My starting array

```python

df_array = xr.DataArray(df) df_array = df_array.rename({PIVOT: 'all_sites'}) df_array

<xarray.DataArray (all_sites: 185084, dim_1: 245)> array([[0., 0., 0., ..., 0., 0., 0.], [0., 0., 0., ..., 0., 0., 0.], [0., 0., 0., ..., 0., 0., 0.], ..., [0., 0., 0., ..., 0., 0., 0.], [0., 0., 0., ..., 0., 0., 0.], [0., 0., 0., ..., 0., 0., 0.]]) Coordinates: * all_sites (all_sites) object '0.gravatar.com||gprofiles.js||Gravatar.init' ... 'кÑ\x83Ñ\x80Ñ\x81Ñ\x8b.1Ñ\x81енÑ\x82Ñ\x8fбÑ\x80Ñ\x8f.Ñ\x80Ñ\x84||store.js||store.set' * dim_1 (dim_1) object 'AnalyserNode.connect' ... 'HTMLCanvasElement.previousSibling' ```

My slice of the array

python sites_of_interest = [sub list of all sites] df_dye_array = xr.DataArray(df.loc[sites_of_interest]) df_dye_array = df_dye_array.rename({PIVOT: 'dye_sites'})

Chunk

python df_array_c = df_array.chunk({'all_sites': 10_000}) df_dye_array_c = df_dye_array.chunk({'dye_sites': 100})

Get distances

```python def get_chebyshev_distances_xarray_ufunc(df_array, df_dye_array): chebyshev = lambda x: np.abs(df_array[:,0,:] - x).max(axis=1) result = np.apply_along_axis(chebyshev, 1, df_dye_array).T return result

distance_array = xr.apply_ufunc( get_chebyshev_distances_xarray_ufunc, df_array_c, df_dye_array_c, dask='parallelized', output_dtypes=[float], input_core_dims=[['dim_1'], ['dim_1']], ) ```

What I get out is an array with the length of my original array and the width of my sites of interest where each number is the chebyshev distance between their respective rows of the original dataset (which are 245 long).

The notion of "core dimensions" in apply_ufunc() is definitely quite tricky to understand.

That said, I think this is (mostly) doing the right thing: - Your inputs have dimensions ['row_a', 'dim_1'] and ['row_b', 'dim_1'] - Xarray broadcasts over dimensions that aren't included in "core dimensions" , so inputs are broadcast to have dimensions like ['row_a', 'row_b', 'dim_1'] and ['row_a', 'row_b', 'dim_1'].

This is probably especially confusing because the unlabeled versions of da and db are given "broadcastable" shapes (1000, 1, 100) and (1000, 100) rather than the fully "broadcast" shapes of (1000, 1000, 100) and (1000, 1000, 100), which would make it more obvious what is going on.

For your specific use case: maybe you meant to specify input_core_dims=[['row_a'], ['row_b']] instead? That version would give inputs with dimensions like ['dim_1', 'row_a'] and ['dim_1', 'row_b'].

More generally: I think we really need a version of apply() that doesn't do this confusing broadcasting and dimension reordering. See https://github.com/pydata/xarray/issues/1130 for discussion about that.

unlabeled versions of da and db are given "broadcastable" shapes (1, 1000, 100) and (1000, 100)

Is it (1000, 1, 100) as my code seems to return, or, as you said (1, 1000, 100)? Is it deterministic?

That was a typo (I'll fix it). It should be (1000, 1, 100).

The behavior is definitely deterministic, if hard to understand!

Can you clarify one thing in your note.

unlabeled versions of da and db are given "broadcastable" shapes (1, 1000, 100) and (1000, 100)

Is it (1000, 1, 100) as my code seems to return, or, as you said (1, 1000, 100)? Is it deterministic?

Maybe it would help to describe what you were trying to do here. What should "one unit" of your calculation look like?

Hi, I will have to think about your response a lot more to see if I can wrap my head around it.

In the meantime I'm not sure I have my input_core_dims correct, but that's the only configuration I could get to work.

I chunk along row_a, and row_b and I output a new array with the dims [row_a, row_b].

By trial and error, the above configuration is the only one I could find where I got out the dims I was expecting and didn't get an error.

I should add, if I pass in plain numpy arrays then I do not have this problem. But ultimately I want to pass in a chunked DataArray, as described here: http://xarray.pydata.org/en/stable/dask.html#automatic-parallelization (this is my whole reason for using xarray).

The work around is easy I just use da[:,0,:] but it's odd!