[pymvpa] Parameter Search SVM

[Yaroslav Halchenko]

Hi Julian,

Unfortunately we do not have yet a generic (working for any classifier)
way to search for optimal hyperparameters (it is in our TODO).
ModelSelector was crafted by Emanuele for GPR.

BUT it is quite easy to implement basic hyperparameters selection
via a grid search through the nested cross-validation procedure.

E.g. in 
http://github.com/PyMVPA/PyMVPA/blob/master/doc/examples/nested_cv.py
we have an example which sweeps through all classifiers in the warehouse
to determine the 'best' for every cross-validation split.  In your
scenario it is virtually the same -- just that you are using the same
classifier, and vary the hyperparameters.

Hope this of some help

Yaroslav

On Wed, 06 Oct 2010, Karch, Julian wrote:

> Hello,

> i'm currently working on a project where i have to select hyperparameters for the libSVM implementation which generate a model which fits a dataset best (compared to other hyperparameters). Is there any implementation available in pymvpa accomplishes this job? I found Modelselector but as i unterstood the logic it needs the classifier/model to have a method called compute_log_marginal_likelihood which the SVM clfers seem to lag.



> Julian Karch
> Research Assistant

-- 
                                  .-.
=------------------------------   /v\  ----------------------------=
Keep in touch                    // \\     (yoh@|www.)onerussian.com
Yaroslav Halchenko              /(   )\               ICQ#: 60653192
                   Linux User    ^^-^^    [175555]