[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Integrating Machine Learning Software and Datasets withing debian

On Thu, 2010-06-17 at 16:46 +0200, Christian Kastner wrote:
> Hello,

Dear Christian,

> I found some time today to finish the packages.
> On 06/14/2010 05:36 PM, Christian Kastner wrote:
> >> On Sun, 2010-06-13 at 23:13 +0200, Christian Kastner wrote:
> > While I ponder a move to debian-science, you can get them here (you want
> > the "wip" branch):
> > 
> > git://scm.kvr.at/git/pkg-liblinear.git
> > git://scm.kvr.at/git/pkg-libocas.git
> I dropped the wip branch, there was too much noise to be useful. The
> proper history is in the master branch now.
> > Packages are here:
> > http://www.kvr.at/debian/pool/main/libl/liblinear/liblinear_1.51~dfsg-1.dsc
> > http://www.kvr.at/debian/pool/main/libo/libocas/libocas_0.93-1.dsc
> > Open issues:
> >   - Clean up the current Makefile and debian/rules (very crude ATM)
> Done.
> >   - Test the binaries
> All tests were successful.
> >   - Test the shared libraries against third-party code
> Haven't done this, but seeing as the above binaries compiled and worked
> fine using the libs, I see no reason why third party code shouldn't.
> >   - Test liblinear's BLAS dependency against the CUDA implementation
> Skipped for now.

I would claim that is useless anyways (except all data is in GPU mem).

> >   - Write man pages (currently just placeholders)
> Done.
> >   - Sanitize quilt patches (DEP3)
> There's only one patch per package, against the Makefile. I usually
> prefer to split my changes into smaller logical units, but this turned
> out to be too cumbersome, and there was no visible benefit.
> >   - Octave interface(?)
> Skipped for now. Help from an experienced octave packager would be
> appreciated.

I guess I should do that - but maybe for the initial version we don't
need all the interfaces.

> >   - Build static libraries(?)
> Upstream doesn't build any, but I added appropriate targets in the
> Makefiles.
> So, reviewers and sponsors welcome!

I've had a look at the packages. Good job. My only concern currently is
that the package description is too long and that I am not sure if it is
really worth to have extra tools etc packages (that might be esoteric

So I would really reduce the package description down to the point and
potentially add a README with more information (or put it in the


This library implements Optimized Cutting Plane Algorithm (OCAS) for
training linear Support Vector Machine (SVM) classifiers from
large-scale data. The computational effort of OCAS scales linearly with
the number of training examples. It is one of the fastest SVM solvers
around for solving linear and multiclass L2 regularized SVMs.


LIBLINEAR is a library for learning linear classifiers for large scale
applications. It supports Support Vector Machines (SVM) with L2 and L1
loss, logistic regression, multi class classification and also Linear
Programming Machines (L1-regularized SVMs). Its computational complexity
scales linearly with the number of training examples making it one of
the fastest SVM solvers around.


For the one fact about the future of which we can be certain is that it
will be utterly fantastic. -- Arthur C. Clarke, 1962

Attachment: signature.asc
Description: This is a digitally signed message part

Reply to: