[SciPy-Dev] [SciPy-User] ANN: SciPy 0.11.0 release candidate 2

Matthew Brett matthew.brett@gmail....
Sat Sep 8 12:33:22 CDT 2012


On Sat, Sep 8, 2012 at 5:31 PM, Ralf Gommers <ralf.gommers@gmail.com> wrote:
> On Tue, Aug 14, 2012 at 3:35 PM, John Hassler <hasslerjc@comcast.net> wrote:
>> On 8/14/2012 7:21 AM, Pauli Virtanen wrote:
>> > Ralf Gommers <ralf.gommers <at> gmail.com> writes:
>> > [clip]
>> >> Does anyone have an idea about that test_singular failure?
>> > That's very likely some problem with the underlying LAPACK library.
>> > I think the problem solved is close to a numerical instability.
>> >
>> > The failing comparison compares eigenvalues computed by
>> >
>> >      eig(A, B)
>> >      eig(A, B, left=False, right=False)
>> >
>> > which differ solely in passing 'N' vs. 'V' to DGGEV. The eigenvalue
>> > property of the former is also checked and seems to pass. Interestingly,
>> > the result obtained from the two seems to differ (therefore, the latter
>> > is probably wrong), which appears to point to a LAPACK issue.
>> >
>> > Here, it would be interesting to know if the problem occurs with
>> > the official Scipy binaries, or something else.
>> >
>> I installed rc2 on Python 2.7.3.  Same problem.  I get the test_singular
>> error on some, but not all, of the runs.  Both are win32-superpack from
>> http://sourceforge.net/projects/scipy/files/scipy/0.11.0rc2/.
>> The error occurs on less than half but more than 1/3 (based on a very
>> small sample) of the runs on both 2.7 and 3.2.
>> I've been working on computers for more than 50 years.  Somehow, I had
>> developed the delusion that they were deterministic .....
>> john
> What are we going to do about this one? I'm tempted to open a ticket for it
> and mark it as knownfail on Windows for now, since it's a corner case.

I have noticed that windows SVD appears to give different answers from
repeated runs on the same matrix, differing in terms of sign flips,
but valid SVDs.  I've no idea why, but I had to adjust the tests in
our code to allow for this.

I guess we should make sure the returned results are correct, and fail
otherwise.  But maybe we do not require two runs to give the same
answer.  Could that explain the problem?



More information about the SciPy-Dev mailing list