test_affinities: RuntimeError: Factor is exactly singular on older scipy 64bit

[yarikoptic]

as reported on the list

======================================================================
ERROR: sklearn.cluster.tests.test_spectral.test_affinities
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/usr/lib/pymodules/python2.6/nose/case.py", line 183, in runTest
    self.test(*self.arg)
  File "/tmp/buildd/scikit-learn-0.12.0/debian/python-sklearn/usr/lib/python2.6/dist-packages/sklearn/cluster/tests/test_spectral.py", line 121, in test_affinities
    labels = sp.fit(X).labels_
  File "/tmp/buildd/scikit-learn-0.12.0/debian/python-sklearn/usr/lib/python2.6/dist-packages/sklearn/cluster/spectral.py", line 360, in fit
    random_state=self.random_state, n_init=self.n_init)
  File "/tmp/buildd/scikit-learn-0.12.0/debian/python-sklearn/usr/lib/python2.6/dist-packages/sklearn/cluster/spectral.py", line 218, in spectral_clustering
    mode=mode, random_state=random_state)
  File "/tmp/buildd/scikit-learn-0.12.0/debian/python-sklearn/usr/lib/python2.6/dist-packages/sklearn/cluster/spectral.py", line 122, in spectral_embedding
    sigma=1.0, which='LM')
  File "/tmp/buildd/scikit-learn-0.12.0/debian/python-sklearn/usr/lib/python2.6/dist-packages/sklearn/utils/arpack.py", line 1493, in _eigsh
    symmetric=True, tol=tol)
  File "/tmp/buildd/scikit-learn-0.12.0/debian/python-sklearn/usr/lib/python2.6/dist-packages/sklearn/utils/arpack.py", line 1031, in get_OPinv_matvec
    return SpLuInv(A.tocsc()).matvec
  File "/tmp/buildd/scikit-learn-0.12.0/debian/python-sklearn/usr/lib/python2.6/dist-packages/sklearn/utils/arpack.py", line 899, in __init__
    self.M_lu = splu(M)
  File "/usr/lib/python2.6/dist-packages/scipy/sparse/linalg/dsolve/linsolve.py", line 129, in splu
    diag_pivot_thresh, drop_tol, relax, panel_size)
RuntimeError: Factor is exactly singular

on amd64 (ok on 32bit) Debian squeeze, Ubuntu 10.04 and 10.10 where
scipy's are:

scikit-learn_0.12.0-1~nd10.04+1_amd64.build:Unpacking python-scipy (from .../python-scipy_0.7.0-2ubuntu0.1_amd64.deb) ...
scikit-learn_0.12.0-1~nd10.10+1_amd64.build:Unpacking python-scipy (from .../python-scipy_0.7.2-2ubuntu1_amd64.deb) ...
scikit-learn_0.12.0-1~nd60+1_amd64.build:Unpacking python-scipy (from .../python-scipy_0.7.2+dfsg1-1+squeeze1~nd60+1_amd64.deb) ...

[amueller]

Thanks for the report. Not sure what could cause that. It seems no-one else is experiencing this issue.
So you have this on two debian versions with three different scipy versions?

[amueller]

This is pretty much the setup that jenkins is running. I would have expected that to be 64bit.
The test should actually be pretty stable. Without having the platform I feel it is a bit hard to make it more robust.
Can anyone reproduce?
@yarikoptic Could you try using a simpler data set maybe?

[amueller]

This should be fixed by #1221 (which is not merged yet).

[amueller]

#1221 was merged now. @yarikoptic could you maybe check against current master? Thanks!

[GaelVaroquaux]

Hey @yarikoptic, I believe that we have really squished that one in master. I am closing it, but feel free to open if you can reproduce the problem with master.

We are releasing next week, so if you can reproduce, please do report the problem.

Cheers,