numpy.sign behaves oddly when dtype is numpy.object

[atomicpirate]

In [1]: import numpy

In [2]: foo = numpy.array([-.1, 0, .1])

In [3]: numpy.sign(foo)
Out[3]: array([-1., 0., 1.])

In [4]: numpy.sign(foo.astype(numpy.object))
Out[4]: array([1, -1, -1], dtype=object)

The underlying data is still just a float:
In [5]: type(foo.astype(numpy.object)[0])
Out[5]: float

This result when using an array of numpy.object does not seem desirable. Either it should operate correctly on the underlying numbers or if called with invalid data (strings, etc) it should throw an exception. This way I could populate an array with heterogeneous data, say a mix of integers and floats, and still get a valid result.

I tested this on:
Mac OS 10.6.8
Python 3.4.2
numpy 1.9.2

[ewmoore]

Looks like OBJECT_sign is missing the logic to return 0 for 0. Compare with BYTE_sign.

[atomicpirate]

Right--zero didn't get a zero result, negative results in +1, positive results in -1. The results for numpy.object seem incorrect on all counts.

[pv]

I can't reproduce your example. get:
.
In [7]: numpy.sign(foo.astype(numpy.object))
array([-1, 0, 1], dtype=object)
In [12]: numpy.version
'1.9.2'

[pv]

Ah ok, the incorrect results occur on Python 3 only.

[pv]

Hah, it's a stupid bug, the == 0 should be == 1 here https://github.com/numpy/numpy/blob/master/numpy/core/include/numpy/npy_3kcompat.h#L328

[jaimefrio]

There, and it the two other checks below it, right?

[pv]

Yes

[springcoil]

Is this still open - it looks like an easy fix could be done.

[charris]

Closed by #6269.