[SciPy-Dev] seeded randn gets different values on osx

Charles R Harris charlesr.harris@gmail....
Sun Jun 13 09:57:52 CDT 2010


On Sat, Jun 12, 2010 at 10:06 PM, <josef.pktd@gmail.com> wrote:

> On Sat, Jun 12, 2010 at 10:09 PM, Charles R Harris
> <charlesr.harris@gmail.com> wrote:
> >
> >
> > On Sat, Jun 12, 2010 at 7:44 PM, Vincent Davis <vincent@vincentdavis.net
> >
> > wrote:
> >>
> >> There was another tread (scipy 8.0b release) describing a problem with
> >> FAIL: test_stats.test_kstest see full test fail below.
> >>
> >> After a little discusion with Josef here are some results. It appears
> >> to me that osx gets different values from a seeded randn
> >>
> >> >From Josef, I assume windows
> >> >>>>> np.random.seed(0)
> >> >>>>> np.random.randn(3)
> >> >> array([ 1.76405235,  0.40015721,  0.97873798])
> >> >
> >> On my machine OSX py 2.6.5, numpy 1.4.0 scipy 8.0b
> >> >>>> np.random.seed(0)
> >> >>>> np.random.randn(3)
> >> > array([ 0.06897149,  1.32078057,  1.5997924 ])
> >>
> >> ubuntu 10.0.4  py 2.5.6 current scipy and numpy just built from the git
> >> source.
> >> >>> np.random.seed(0)
> >> >>> np.random.randn(3)
> >> array([ 1.76405235,  0.40015721,  0.97873798])
> >>
> >> FAIL: test_stats.test_kstest
> >> ----------------------------------------------------------------------
> >> Traceback (most recent call last):
> >>  File
> >>
> "/Library/Frameworks/EPD64.framework/Versions/6.2/lib/python2.6/site-packages/nose/case.py",
> >> line 186, in runTest
> >>   self.test(*self.arg)
> >>  File
> >>
> "/Library/Frameworks/EPD64.framework/Versions/6.2/lib/python2.6/site-packages/scipy/stats/tests/test_stats.py",
> >> line 1228, in test_kstest
> >>   assert_almost_equal( D, 0.12464329735846891, 15)
> >>  File
> >>
> "/Library/Frameworks/EPD64.framework/Versions/6.2/lib/python2.6/site-packages/numpy/testing/utils.py",
> >> line 459, in assert_almost_equal
> >>   raise AssertionError(msg)
> >> AssertionError:
> >> Arrays are not almost equal
> >>  ACTUAL: 0.093893737596468518
> >>  DESIRED: 0.12464329735846891
> >> ______
> >
> > There were some changes to the random normal generator in trunk, but I
> don't
> > think they should affect this. What versions of numpy is everyone using?
>
> I'm also using numpy 1.4.0, but the results for the random numbers in
> the tests have been hard coded for 19 months ( numpy 1.3.0 ?)
>
> OS X 10.5 ppc:darwin-ppc in 0.7.2 and 0.8.0b1 numpy 1.4.1(?) has a
> small divergence 10e-13 (Derek Homeier)
>
> OSX py 2.6.5, numpy 1.4.0 scipy 8.0b is pretty far off  (Vincent)
>
> This is a numpy issue that shows up in stats because of the tests.
>
>
I don't see anything obvious in the revision history. Does the uniform
distribution also show changes? It is also possible that compiler
optimizations are messing with the code, unlikely, but it would be nice to
know the compiler versions used and/or 64 bits vs 32 bits. There is at least
one numpy test that depends on random numbers from a given seed being
reproducible and it isn't showing any problems. Sounds like we need to add a
test specifically to test reproducibility.

Chuck
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.scipy.org/pipermail/scipy-dev/attachments/20100613/a2f2602e/attachment.html 


More information about the SciPy-Dev mailing list