[pymvpa] libsvm dense arrays

[Yaroslav Halchenko]

> This works quite well, everything is switched to Shogun and runs quite  
> nicely.
cool!

> But it doesn't seem to matter - at least with some preliminary tests  
> (looking at model selection, so the code is going back and forth between  
> python<->libsvm/shogun quite a bit) using Shogun nets exactly the same  
are you doing model selection 'manually' with some custom code? just
sweeping over parameters or involving some optimizer? or do you use our
preliminary ModelSelector (I think it is not capable of handling
anything besides GPR atm)

> accuracies using Shogun for gamma < 0.5 in an Rbf classifier, though the  
> results seem comparable otherwise.
that is interesting ;-) could you try to sacrifice some runtime by
decreasing tolerance on convergence (parameter epsilon)?
do you use libsvm backend in shogun or some other?
also, are you using C-SVM? (libsvm also has nu-SVM), what C values do
you use? ... may be there is some difference in how we treat gamma for
libsvm and shogun...

> If time performance is the same, does that indicate Shogun does *not*  
> use a dense data representation?
hard to say ;-)

> If model selection yields different optimal choices due to (sometimes)  
> different cross-validation rates, doesn't that suggest that libsvm and  
> shogun are using fairly different code bases?
well... many things could be different -- from the difference in kernel
compuation (treatment of parameters, kernel computation per se, etc) to
slight differences in optimization. Cross-validation estimates might be
quite sensitive to little variations in the results, especially if the
data is 'problematic', ie generalization performance is not far from 'by
chance'

> by no means have I tested this on a robust dataset, but this simply  
> doesn't feel right :)
if you could come up with some artificial data, or share the data you
have, and share the minimalistic analysis script, then we could dig into
the problem together. Otherwise it is hard to say exactly what goes
wrong (if anything)

-- 
Yaroslav Halchenko
Research Assistant, Psychology Department, Rutgers-Newark
Student  Ph.D. @ CS Dept. NJIT
Office: (973) 353-5440x263 | FWD: 82823 | Fax: (973) 353-1171
        101 Warren Str, Smith Hall, Rm 4-105, Newark NJ 07102
WWW:     http://www.linkedin.com/in/yarik