len() function for COO objects

[nils-werner]

Quite trivial really, implements the len() operation following NumPy's convention of

len(array) == array.shape[0]

[hameerabbasi]

Seems fine, however it would be really nice to have docstrings for these, and if possible, tests.

[nils-werner]

Added them

[hameerabbasi]

Well, this looks good. Let me know if you're finished working on it, and I'll review and merge (don't want to be too fast this time).

[nils-werner]

This is ready to be merged.

[mrocklin]

Agreed.

[mrocklin]

Thanks @nils-werner !

[nils-werner]

One unexpected sideeffect of this:

numpy.array(sparse_arr)

now magically works!

x = sparse.random((20, 30))
np.allclose(x.todense(), np.array(x))    # True

[hameerabbasi]

I'm going to assume it's really slow. Do we have any benchmarks for this?

[nils-werner]

Yes, it's very very slow.

If I understand it correctly, numpy.array(x) looks for buffer and __array_interface__ when creating an array from an object. If it doesn't find one it simply iterates over it.

What if we implement __array_interface__ that calls self.todense()?

@property
def __array_interface__(self):
    return {
        'shape': self.shape,
        'data': self.todense(),
        'typestr': self.dtype.str,
    }

Before

x = sparse.random((20, 30, 20))

%timeit numpy.array(x)
# 1.48 s ± 25.5 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
%timeit x.todense()
# 11.6 µs ± 40.3 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)
numpy.allclose(x, x.todense)
# True

After

%timeit numpy.array(x)
# 23 µs ± 392 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
numpy.allclose(x, x.todense)
# True

I can't explain where the 2x slowdown (11 vs 23 us) comes from though...

[hameerabbasi]

A pull request would be welcome. Also, your data is too small. Maybe it's just the overhead.