[Numpy-discussion] Test error with ATLAS, Windows 64 bit

[Matthew Brett]

Hi,

On Mon, Apr 14, 2014 at 12:12 PM, Warren Weckesser
<warren.weckesser at gmail.com> wrote:
>
> On Mon, Apr 14, 2014 at 2:59 PM, Matthew Brett <matthew.brett at gmail.com>
> wrote:
>>
>> Hi,
>>
>> With Carl Kleffner, I am trying to build a numpy 1.8.1 wheel for
>> Windows 64-bit, and latest stable ATLAS.
>>
>> It works fine, apart from the following test failure:
>>
>> ======================================================================
>> FAIL: test_special (test_umath.TestExpm1)
>> ----------------------------------------------------------------------
>> Traceback (most recent call last):
>>   File "C:\Python27\lib\site-packages\numpy\core\tests\test_umath.py",
>> line 329, in test_special
>>     assert_equal(ncu.expm1(-0.), -0.)
>>   File "C:\Python27\lib\site-packages\numpy\testing\utils.py", line
>> 311, in assert_equal
>>     raise AssertionError(msg)
>> AssertionError:
>> Items are not equal:
>>  ACTUAL: 0.0
>>  DESIRED: -0.0
>>
>> Has anyone seen this?  Is it in fact necessary that expm1(-0.) return
>> -0 instead of 0?
>>
>
>
> What a cowinky dink.  This moring I ran into this issue in a scipy pull
> request (https://github.com/scipy/scipy/pull/3547), and I asked about this
> comparison failing on the mailing list a few hours ago.  In the pull
> request, the modified function returns -0.0 where it used to return 0.0, and
> the test for the value 0.0 failed.  My work-around was to use
> `assert_array_equal` instead of `assert_equal`.  The array comparison
> functions treat the values -0.0 and 0.0 as equal.  `assert_equal` has code
> that checks for signed zeros, and fails if they are not the same sign.

Yes, sorry, I should have mentioned that I saw your post too.  I'd
live to use that workaround, but it looks like the teste against -0 is
intentional, and I was hoping for help understanding.

The relevant two lines of the test are:

        assert_equal(ncu.expm1(0.), 0.)
        assert_equal(ncu.expm1(-0.), -0.)

Julian - I think this one is for you - as the author of these lines:

f53ab41a numpy/core/tests/test_umath.py     (Julian Taylor
2014-03-02 02:55:30 +0100  329)         assert_equal(ncu.expm1(-0.),
-0.)

Can you say why you wanted -0 specifically?  Any clue as to why ATLAS
64 bit may give 0 instead?

Cheers,

Matthew