Feature request `numpy.unique`

Platform: Ubuntu 16.04 Python version: 3.6 Numba version: 0.37.0

I was trying to get a speedup in my code where I use numpy.unique() to find unique elements in an n-dimensional array. If I use this:

x = np.unique(arr)

this works fine and returns the expected results. But if I do this:

@jit(nogil=True, nopython=True)
def find_unique(arr):
    return np.unique(arr)

x = find_unique(arr)

This throws an error with the following message:

UntypedAttributeError: Failed at nopython (nopython frontend)
Unknown attribute 'unique' of type Module(<module 'numpy' from '/opt/conda/lib/python3.6/site-packages/numpy/__init__.py'>)
File "<ipython-input-15-bd84a508a624>", line 5
[1] During: typing of get attribute at <ipython-input-15-bd84a508a624> (5)

[stuartarchibald]

Thanks for the report. You are seeing that message because Numba does not yet support numpy.unique. Functions that are supported are listed here. I’m editing the title and making this a feature request. Thanks.

[Rik-de-Kort]

I’m not sure how to work with keyword arguments using the highlevel API (I’d have to spend some time on it), but since it’s not going to make it into numba until at least 0.39, here’s code for an overloaded implementation which returns the counts as well.

@overload(np.unique)
def np_unique(a):
        def np_unique_impl(a):
            b = np.sort(a.flatten())
            unique = list(b[:1])
            counts = [1 for _ in unique]
            for x in b[1:]:
                if x != unique[-1]:
                    unique.append(x)
                    counts.append(1)
                else:
                    counts[-1] += 1
            return unique, counts
    return np_unique_impl