[pymvpa] Seg on OSX when using libsvm

[Fabian Pedregosa]

Yaroslav Halchenko wrote:
> 
> On Mon, 08 Mar 2010, Fabian Pedregosa wrote:
>> test_basic_clf_train_predict
>> (mvpa.tests.test_svmkernels.SVMKernelTests) ... Segmentation fault
> 
>> Can any mac user reproduce this ?
> 
> tried to reproduce on one of the systems we have access to (see details
> below, not running Snow Leopard! still known  I believe to have problems with
> Python):
> 
> But what I know is that our libsvm bindings or libsvm itself does have some
> issues which can lead to segfaults (and/or memory leaks) -- I hadn't luck yet
> to localize it.  You can valgrind your script (if it is available on your box)
> to see if there are related warnings.


Thanks for looking at this. I solved the problem by uninstalling shogun. 
It seems the problem was due to name clashing between libsvm & shogun. 
The stack trace is interesting:


0   libshogun.4.dylib             	0x03f87656 SVC_Q::get_Q(int, int) 
const + 54
1   libsvm.dylib                  	0x0425ecd9 
Solver::select_working_set(int&, int&) + 777
2   libsvm.dylib                  	0x0426109a Solver::Solve(int, QMatrix 
const&, double const*, signed char const*, double*, double, double, 
double, Solver::SolutionInfo*, int) + 778


which means it 'jumps' from libsvm to shogun (I have no clue why this 
happens)


Thanks,

> 
> P.S. I had to add 'from mvpa.datasets.miscfx import *' to your script
> so we get .summmary available for the dataset.
> 
> In [1]: run t.py        
> Dataset: 2x2 at int32, <sa: targets>
> stats: mean=1.5 std=1.11803 var=1.25 min=0 max=3
> No per targets/chunks due to KeyError('chunks',)Sequence statistics for 2 entries from set [0, 1]
> Counter-balance table for orders up to 2:
> Targets/Order O1    |  O2    |
>       0:       0 1  |   0 0  |
>       1:       0 0  |   0 0  |
> Correlations: min=-1 max=-1 mean=-1 sum(abs)=1
> [1.0]
> 
> In [2]: import mvpa                                                                                                                                 
> 
> In [3]: print mvpa.wtf()                                                                                                                            
> Current date:   2010-03-08 13:37
> PyMVPA:
>  Version:       0.5.0.dev
>  Path:          /Users/yoh/proj/pymvpa/mvpa/__init__.pyc
>  Version control (GIT):
>   Status:
>    # On branch master
>    nothing to commit (working directory clean)
>   Reference:
>    d32bba93da5f02d0ef986e4cb40c4cdbed723455 HEAD
>    d32bba93da5f02d0ef986e4cb40c4cdbed723455 refs/remotes/origin/HEAD
>   Difference from last release 0.5.0.dev:
>     332 files changed, 46592 insertions(+), 10344 deletions(-)
> SYSTEM:
>  OS:            posix Darwin 9.8.0 Darwin Kernel Version 9.8.0: Wed Jul 15 16:55:01 PDT 2009; root:xnu-1228.15.4~1/RELEASE_I386
>  Distribution:   10.5.8/i386
> EXTERNALS:
>  Present:       cPickle, ctypes, good scipy.stats.rv_discrete.ppf, griddata, gzip, hcluster, libsvm, libsvm verbosity control, nose, numpy, reportlab, rpy2, running ipython env, scipy, weave
>  Absent:        atlas_fsl, atlas_pymvpa, elasticnet, glmnet, good scipy.stats.rdist, h5py, lars, lxml, matplotlib, mdp, mdp ge 2.4, nibabel, nifti, nifti ge 0.20090205.1, nipy, openopt, pprocess, pylab, pylab plottable, pywt, pywt wp reconstruct, pywt wp reconstruct fixed, rpy, sg ge 0.6.4, sg ge 0.6.5, sg_fixedcachesize, shogun, shogun.krr, shogun.lightsvm, shogun.mpd, shogun.svrlight
>  Versions of critical externals:
>   ipython     : 0.8.1
>   reportlab   : 2.1
>   scipy       : 0.7.0
>   numpy       : 1.3.0
>   rpy2        : 2.0.3
>   ctypes      : 1.0.3
>   numpy       : 1.3.0
> RUNTIME:
>  PyMVPA Environment Variables:
>   PYTHONPATH          : "/Users/yoh/proj/pymvpa"
> ....
> 
> 
> 
> 
> 
> ------------------------------------------------------------------------
> 
> _______________________________________________
> Pkg-ExpPsy-PyMVPA mailing list
> Pkg-ExpPsy-PyMVPA at lists.alioth.debian.org
> http://lists.alioth.debian.org/mailman/listinfo/pkg-exppsy-pymvpa