On Thu, May 2, 2013 at 11:21 PM, Andreas Mueller
<amuel...@ais.uni-bonn.de>wrote:
> That is pretty neat. What was your motivation for reimplementing Liblinear
> stuff?
>
Since a kernel matrix can be seen as a dataset (n_samples of n_samples
dimensions, if you use the entire training set), I added to lightning a
KernelDataset object, based on the dataset abstraction, which computes and
caches kernel values on the fly. This allowed me to, for example, apply
coordinate descent with a squared hinge loss and an l1 penalty to obtain a
sparse number of "support vectors" (basis vectors would be a more accurate
term): each non-zero weight of the linear classifier corresponds to one
support vector / basis vector. I removed KernelDataset from lightning
because the cache implementation was not so good and due to maintenance
cost.
> Did you implement multinomial logistic regression or the ovr version?
>
Both. But multiclass logistic regression only supports l1 / l2 (group
lasso) penalty. If you want to use l2 or l1 penalty, you need to use binary
logistic regression + OVR.
Mathieu
------------------------------------------------------------------------------
Get 100% visibility into Java/.NET code with AppDynamics Lite
It's a free troubleshooting tool designed for production
Get down to code-level detail for bottlenecks, with <2% overhead.
Download for free and get started troubleshooting in minutes.
http://p.sf.net/sfu/appdyn_d2d_ap2
_______________________________________________
Scikit-learn-general mailing list
Scikit-learn-general@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general