8000 test_pareto on 32-bit got even worse · Issue #18387 · numpy/numpy · GitHub
[go: up one dir, main page]
Skip to content

test_pareto on 32-bit got even worse #18387

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Open
bnavigator opened this issue Feb 10, 2021 · 6 comments
Open

test_pareto on 32-bit got even worse #18387

bnavigator opened this issue Feb 10, 2021 · 6 comments
Labels
component: numpy.random

Comments

@bnavigator
Copy link
Contributor

Hi,
NumPy 1.20.0 on openSUSE Tumbleweed 32-bit looks even worse than Ubuntu some years ago (#430):

[ 1044s] __________________________ TestRandomDist.test_pareto __________________________
[ 1044s] [gw4] linux -- Python 3.8.7 /usr/bin/python3.8
[ 1044s] 
[ 1044s] self = <numpy.random.tests.test_generator_mt19937.TestRandomDist object at 0xf06d9bc8>
[ 1044s] 
[ 1044s]     def test_pareto(self):
[ 1044s]         random = Generator(MT19937(self.seed))
[ 1044s]         actual = random.pareto(a=.123456789, size=(3, 2))
[ 1044s]         desired = np.array([[1.0394926776069018e+00, 7.7142534343505773e+04],
[ 1044s]                             [7.2640150889064703e-01, 3.4650454783825594e+05],
[ 1044s]                             [4.5852344481994740e+04, 6.5851383009539105e+07]])
[ 1044s]         # For some reason on 32-bit x86 Ubuntu 12.10 the [1, 0] entry in this
[ 1044s]         # matrix differs by 24 nulps. Discussion:
[ 1044s]         #   https://mail.python.org/pipermail/numpy-discussion/2012-September/063801.html
[ 1044s]         # Consensus is that this is probably some gcc quirk that affects
[ 1044s]         # rounding but not in any important way, so we just use a looser
[ 1044s]         # tolerance on this test:
[ 1044s] >       np.testing.assert_array_almost_equal_nulp(actual, desired, nulp=30)
[ 1044s] E       AssertionError: X and Y are not equal to 30 ULP (max is 31)
[ 1044s] 
[ 1044s] ../../../BUILDROOT/python-numpy-1.20.0-90.1.i386/usr/lib/python3.8/site-packages/numpy/random/tests/test_generator_mt19937.py:1520: AssertionError

(Pdb) (actual-desired)
array([[ 4.44089210e-16,  0.00000000e+00],
       [-2.22044605e-16,  0.00000000e+00],
       [ 0.00000000e+00, -2.30967999e-07]])
(Pdb) np.spacing(max(abs(actual).max(), abs(desired).max()))*30
2.2351741790771484e-07
@seberg seberg added this to the 1.20.2 release milestone Feb 10, 2021
@seberg
Copy link
Member
seberg commented Feb 10, 2021

I am surprised anything changed here, but marking for 1.20.2, just so that its not lost by accident. For what its worth, the mailing list link is slightly off for me, and this gives the right one:
https://mail.python.org/pipermail/numpy-discussion/2012-September/063764.html

Out of curiosity, do you know that this is a regression in that the 1.19 passed fine with the same compiler versions, etc?

@bnavigator
Copy link
Contributor Author

See #18388 for build logs. Here is a passing one for 1.19.5:
numpy-1.19-i586_log.txt

@bashtage
Copy link
Contributor

What is a little interesting is that it doesn't occur in the same test for RandomState. Even though the underlying code is the same, the seeding of the MT19937 generator in the failing file is different from that in the RandomState, and so the test values are different.

@charris charris removed this from the 1.20.2 release milestone Mar 27, 2021
@charris
Copy link
Member
charris commented Mar 27, 2021

I cleared the milestone here. @bashtage Seems this is a value issue/

@bashtage
Copy link
Contributor

I think it would be reasonable to relax the tolerance on 32-bit Linux as a pragmatic fix. It seems to be compiler-specific since the test doesn't fail on other 32-bit platforms including other 32-bit variants of Linux.

@WarrenWeckesser WarrenWeckesser mentioned this issue Dec 22, 2021
Closed
@Apteryks
Copy link

I also reproduce this on i686-linux (GNU Guix). The test_float_remainder_overflow test also failed there:

=================================== FAILURES ===================================
_________________ TestRemainder.test_float_remainder_overflow __________________
[gw15] linux -- Python 3.9.9 /gnu/store/iabliixhj5hirckni5wr3d75vjfcw74i-python-3.9.9/bin/python3

self = <numpy.core.tests.test_umath.TestRemainder object at 0xf1c7d958>

    def test_float_remainder_overflow(self):
        a = np.finfo(np.float64).tiny
        with np.errstate(over='ignore', invalid='ignore'):
            div, mod = np.divmod(4, a)
            np.isinf(div)
            assert_(mod == 0)
        with np.errstate(over='raise', invalid='ignore'):
            assert_raises(FloatingPointError, np.divmod, 4, a)
        with np.errstate(invalid='raise', over='ignore'):
>           assert_raises(FloatingPointError, np.divmod, 4, a)

a          = 2.2250738585072014e-308
div        = inf
mod        = 0.0
self       = <numpy.core.tests.test_umath.TestRemainder object at 0xf1c7d958>

/gnu/store/vbhm2gxkfhmdx085yy2kkrwc34y40778-python-numpy-1.21.6/lib/python3.9/site-packages/numpy/core/tests/test_umath.py:607: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
/gnu/store/iabliixhj5hirckni5wr3d75vjfcw74i-python-3.9.9/lib/python3.9/unittest/case.py:739: in assertRaises
    return context.handle('assertRaises', args, kwargs)
        args       = (<ufunc 'divmod'>, 4, 2.2250738585072014e-308)
        context    = None
        expected_exception = <class 'FloatingPointError'>
        kwargs     = {}
        self       = <numpy.testing._private.utils._Dummy testMethod=nop>
/gnu/store/iabliixhj5hirckni5wr3d75vjfcw74i-python-3.9.9/lib/python3.9/unittest/case.py:201: in handle
    callable_obj(*args, **kwargs)
        args       = [4, 2.2250738585072014e-308]
        callable_obj = <ufunc 'divmod'>
        kwargs     = {}
        name       = 'assertRaises'
        self       = None
/gnu/store/iabliixhj5hirckni5wr3d75vjfcw74i-python-3.9.9/lib/python3.9/unittest/case.py:223: in __exit__
    self._raiseFailure("{} not raised by {}".format(exc_name,
        exc_name   = 'FloatingPointError'
        exc_type   = None
        exc_value  = None
        self       = <unittest.case._AssertRaisesContext object at 0xf1c7dcd0>
        tb         = None
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

self = <unittest.case._AssertRaisesContext object at 0xf1c7dcd0>
standardMsg = 'FloatingPointError not raised by divmod'

    def _raiseFailure(self, standardMsg):
        msg = self.test_case._formatMessage(self.msg, standardMsg)
>       raise self.test_case.failureException(msg)
E       AssertionError: FloatingPointError not raised by divmod

msg        = 'FloatingPointError not raised by divmod'
self       = <unittest.case._AssertRaisesContext object at 0xf1c7dcd0>
standardMsg = 'FloatingPointError not raised by divmod'

/gnu/store/iabliixhj5hirckni5wr3d75vjfcw74i-python-3.9.9/lib/python3.9/unittest/case.py:163: AssertionError
__________________________ TestRandomDist.test_pareto __________________________
[gw2] linux -- Python 3.9.9 /gnu/store/iabliixhj5hirckni5wr3d75vjfcw74i-python-3.9.9/bin/python3

self = <numpy.random.tests.test_generator_mt19937.TestRandomDist object at 0x8fee2340>

    def test_pareto(self):
        random = Generator(MT19937(self.seed))
        actual = random.pareto(a=.123456789, size=(3, 2))
        desired = np.array([[1.0394926776069018e+00, 7.7142534343505773e+04],
                            [7.2640150889064703e-01, 3.4650454783825594e+05],
                            [4.5852344481994740e+04, 6.5851383009539105e+07]])
        # For some reason on 32-bit x86 Ubuntu 12.10 the [1, 0] entry in this
        # matrix differs by 24 nulps. Discussion:
        #   https://mail.python.org/pipermail/numpy-discussion/2012-September/063801.html
        # Consensus is that this is probably some gcc quirk that affects
        # rounding but not in any important way, so we just use a looser
        # tolerance on this test:
>       np.testing.assert_array_almost_equal_nulp(actual, desired, nulp=30)
E       AssertionError: X and Y are not equal to 30 ULP (max is 31)

actual     = array([[1.03949268e+00, 7.71425343e+04],
       [7.26401509e-01, 3.46504548e+05],
       [4.58523445e+04, 6.58513830e+07]])
desired    = array([[1.03949268e+00, 7.71425343e+04],
       [7.26401509e-01, 3.46504548e+05],
       [4.58523445e+04, 6.58513830e+07]])
random     = Generator(MT19937) at 0xEEDC8028
self       = <numpy.random.tests.test_generator_mt19937.TestRandomDist object at 0x8fee2340>

/gnu/store/vbhm2gxkfhmdx085yy2kkrwc34y40778-python-numpy-1.21.6/lib/python3.9/site-packages/numpy/random/tests/test_generator_mt19937.py:1528: AssertionError
=========================== short test summary info ============================
FAILED ../../core/tests/test_umath.py::TestRemainder::test_float_remainder_overflow
FAILED ../../random/tests/test_generator_mt19937.py::TestRandomDist::test_pareto
= 2 failed, 15026 passed, 623 skipped, 18 xfailed, 4 xpassed in 189.89s (0:03:09) =

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees
No one assigned
Labels
component: numpy.random
Projects
None yet
Milestone
No milestone
Development

No branches or pull requests
5 participants
@seberg @charris @bnavigator @bashtage @Apteryks
0