
On Fri, Jan 14, 2011 at 1:33 PM, Bruce Southey <bsouthey@gmail.com> wrote:
On 01/14/2011 12:06 PM, josef.pktd@gmail.com wrote:
On Fri, Jan 14, 2011 at 12:48 PM, Pauli Virtanen <pav@iki.fi> wrote:
Fri, 14 Jan 2011 12:09:54 -0500, josef.pktd wrote:
compiled against numpy 1.5.1
Josef
====================================================================== FAIL: Test singular pair ----------------------------------------------------------------------
[clip]
"...\scipy\linalg\tests\test_decomp.py", line 189, in _check_gen_eig err_msg=msg)
This is actually a bit strange. The test does this:
w, vr = eig(A,B) wt = eigvals(A,B) # ... some stuff that does not touch w and wt # The failing assertion: assert_array_almost_equal(sort(w[isfinite(w)]), sort(wt[isfinite(wt)]), err_msg=msg)
The code paths in `eig` and `eigvals` are identical as far as the computation of `w` and `wt` is concerned.
So perhaps this is another manifestation of the no-determinism of linear algebra which I recall occurred on your computer? It seems this particular is supposed to be ill-conditioned and susceptible to rounding error...
***
It would be nice to find out what exactly is causing the non-determinism, since it seems you are not the only one seeing it (other machines where this occurs were reported on the Numpy ML).
http://permalink.gmane.org/gmane.comp.python.numeric.general/42082
Are you able to reproduce it using only the `dot` function? What happens if you then use the non-BLAS `numpy.core.multiarray.dot` instead?
I can try a bit later, I'm trying to figure out a stats bug and don't want to mess with my scipy since it takes me more than an hour to build.
Looking at the test and try the non atlas dot should be fine.
Josef
-- Pauli Virtanen
_______________________________________________ SciPy-Dev mailing list
_______________________________________________ SciPy-Dev mailing list SciPy-Dev@scipy.org http://mail.scipy.org/mailman/listinfo/scipy-dev
_______________________________________________ SciPy-Dev mailing list SciPy-Dev@scipy.org http://mail.scipy.org/mailman/listinfo/scipy-dev
SciPy-Dev@scipy.org http://mail.scipy.org/mailman/listinfo/scipy-dev
_______________________________________________ SciPy-Dev mailing list SciPy-Dev@scipy.org http://mail.scipy.org/mailman/listinfo/scipy-dev
Presumably this previously passed under numpy 1.3 (I think) then the change in versions reflects the change in numpy/scipy release managers. So does this occur with Christoph's releases on your system? http://www.lfd.uci.edu/~gohlke/pythonlibs/
If so, then there is something missing in the binary generation.
In anycase, we also need more information on the OS (I do presume windows but not version etc), and how numpy and scipy were installed including specific version of the installers if used.
The failure didn't show up yesterday, when I tested, the official 0.9b1 installer. Either it's a recent change in trunk, or it is because my ATLAS, lapack, blas, is an older version. If I'm the only one seeing it, then I will look at my setup and the details later. Josef
Bruce
_______________________________________________ SciPy-Dev mailing list SciPy-Dev@scipy.org http://mail.scipy.org/mailman/listinfo/scipy-dev