[Numpy-discussion] [Announce] Numpy 1.3.0b1

Charles R Harris charlesr.harris@gmail....
Thu Mar 19 12:01:40 CDT 2009


On Thu, Mar 19, 2009 at 10:21 AM, Robert Pyle <rpyle@post.harvard.edu>wrote:

>
> On Mar 19, 2009, at 11:35 AM, Charles R Harris wrote:
>
> >
> >
> > On Thu, Mar 19, 2009 at 9:17 AM, Robert Pyle
> > <rpyle@post.harvard.edu> wrote:
> > I'm getting one test failure with 1.3.0b1 ---
> >
> > FAIL: test_umath.TestComplexFunctions.test_loss_of_precision(<type
> > 'numpy.complex256'>,)
> > ----------------------------------------------------------------------
> > Traceback (most recent call last):
> >   File "/Library/Frameworks/Python.framework/Versions/4.1.30101/lib/
> > python2.5/site-packages/nose-0.10.3.0001-py2.5.egg/nose/case.py", line
> > 182, in runTest
> >     self.test(*self.arg)
> >   File "/Library/Frameworks/Python.framework/Versions/4.1.30101/lib/
> > python2.5/site-packages/numpy/core/tests/test_umath.py", line 498, in
> > check_loss_of_precision
> >     check(x_series, 2*eps)
> >   File "/Library/Frameworks/Python.framework/Versions/4.1.30101/lib/
> > python2.5/site-packages/numpy/core/tests/test_umath.py", line 480, in
> > check
> >     assert np.all(d < rtol), (np.argmax(d), x[np.argmax(d)], d.max())
> > AssertionError: (0, nan, nan)
> >
> > Yes, that test fails on some architectures. What type of cpu do you
> > have? It would help if you could track down the cause of the nans,
> > see ticket #1038.
>
> CPU is PPC (G5).  I added a print statement in the test to pin things
> down a bit.  The failing test appears to be
>
>             d = np.absolute(np.arcsinh(x)/np.arcsinh(x+0j).real - 1)
>              assert np.all(d < rtol), (np.argmax(d), x[np.argmax(d)],
> d.max())
>
> with dtype =  <type 'numpy.complex256'>
>
> It passes with dtype =  <type 'numpy.complex64'> and dtype =  <type
> 'numpy.complex128'>
>
> Is that any help?
>

Not yet ;) I think there is a problem with the range of values in x that
might have their source in the finfo values. So it would help if you could
pin down just where x goes wrong by printing it out. That is what the short
script that a included in the ticket comments does. Mind, I think you will
need to do a bit of exploration. I don't think the failures are significant
in that it probably doesn't need to test the range of values that it does,
but it would be nice to understand precisely why it fails.

Chuck
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.scipy.org/pipermail/numpy-discussion/attachments/20090319/a324e32e/attachment.html 


More information about the Numpy-discussion mailing list