Reduction function returning unique value

[zmoon]

Along a dimension or in a groupby group, I want to perform a reduction that checks that there is only one unique value in the group and returns it. I couldn't figure out how to apply the below with xr.apply_ufunc() (nor with groupby).

import numpy as np
import xarray as xr

def the_unique(x):
    """Return _the_ non-NaN unique value or raise ValueError."""
    uns = np.unique(x)
    uns = uns[~np.isnan(uns)]
    if uns.size == 1:
        return uns[0]
    elif uns.size == 0:
        raise ValueError("no non-NaN values")
    else:
        raise ValueError("more than one unique value")

assert the_unique([1, 1, 1]) == 1

x = np.tile(np.arange(3)[:,np.newaxis], 3)  # each row has one unique value
np.testing.assert_equal(
    np.apply_along_axis(the_unique, 1, x),
    [0, 1, 2]
)

da = xr.DataArray(x, dims=["y", "x"])

# xr.apply_ufunc(...) ?

Example situation:

• going pandas -> xarray with table of time series at different points, in order to reduce duplication for variables that (should) only vary in space not time

[dcherian]

This should work with reduce? Also pd.unique is usually faster than np.unique and won't return multiple NaNs

For apply_ufunc:

1. input_core_dims should be the dimension you're reducing
2. I think you have no output_core_dims
3. To loop over an axis like apply_along_axis use vectorize=True.
   You might even try

def the_unique(x):
	x = x[~np.isnan(x)]
	uniq = x[0]
	assert (x == uniq).all()
	return x

[zmoon]

Thanks @dcherian, I was able to get it to work with apply_ufunc by adding vectorize=True (I had input_code_dims as the dimension to reduce already). And I was able to use with groupby reduce by just adding axis=None to the signature (+ using the pd functions).

And thanks for the pd.unique tip, using that and pd.isnull allows this to work with non-numeric.

I will compare perf a bit, the code you posted looks faster, and maybe easier to make into a vectorized version that would use axis.