Re: OCTAVE: LIBSVM and LIBSVM Tools [Win 8 64 bit Pro]

[Carnë Draug]

On 18 December 2013 07:22, Hans Lodder <address@hidden> wrote:
> Hi!
>
> Together with more than 200.000 other students I am following the Machine
> Learning course on Coursera from prof. Andrew Ng (4th edition). Ng
> encourages us to use OCTAVE. He advises to use the LIBSVM package and its
> tools for exploring the possibilities of Support Vector Machines, this being
> a very good and versatile implementtaion.
>
> More on LIBSVM:
>
>   http://www.csie.ntu.edu.tw/~cjlin/libsvm/#download
>
>
> In the view of Ng SVM's are the best ML bet, and provide the fastest way to
> an excellent ML result:
>
> - run an excellent SVM standard implementation like LIBSVM.
> - perform error analysis, and tune the SVM, or try a more appropriate
> algorithm, as indicated by the error analysis.
>
> It would therefore be great if an excellent SVM was available in OCTAVE out
> of the box.
>
> Now I understand that LIBSVM is part of the NaN package, which is one of the
> Octave-Forge packages. However, only 2 of the many functions are supported.
>
> Therefore, I searched on the Internet if anyone took the opportunity of
> installing LIBSVM and Tools directly in OCTAVE. They did, and it seems to
> work.
>
> I looked into the README.MinGW file, found that many links are dead, and
> also found your names. I also saw that there is complete MinGW installation
> in the OCTAVE directory tree.
>
> My questions:
>
> - What is the easiest way for a non techie to install the complete LIBSVM
> and tools on my OCTAVE C:\Octave\Octave3.6.4_gcc4.6.2 standard install?
> - Is there anything I can do to facilitate adding LIBSVM to the OCTAVE FORGE
> packages?
>
> If this works for me, I am going to place a message on the ML forum!
>
> I would appreciate hearing from you.
>
> Kind regards,
>
> Hans Lodder

Hi Hans

I'm CC'ing the Octave help mailing list so others can pitch in, as
well as the maintainers of libsvm.

The preferred way of adding stuff to Octave is through the
installation of Octave packages. I have recently finished a script
that prepares an Octave package from the matlab interface to libsvm,
and I am just waiting for a reply from the libsvm developers about the
pull request [1]. Until this happens, the best method is to download
my clone of libsvm [2] and run the "octave/release_octave_pkg.py"
script. The generated package can be installed from the Octave prompt
with "pkg install path_to_generated_tarball". It can then be loaded
and unloaded like any other package.

A few notes:

* the interface uses mex files which does not allow for help text. The
hack used by libsvm around this is to print the help text when called
incorrectly. Octave creates a cache of help text to speed up searching
it. Because there's no real help text for this functions (accessible
through the function help), installation will issue an warning about
unusable help text which can be safely ignored.

* the package assumes the user has libsvm installed somewhere on the
system which may not be true. Plus, the libsvm version available on
the system needs to be recent enough to install the package (the
version currently available in Debian, 3.12, is not recent enough by
the way). You can work around this by adding the svm.h and svm.cpp
files to the tarball, adjusting the Makefile accordingly, and removing
the configure file. But the right way to do it is to have the right
version of libsvm on your system.

Carnë

[1] https://github.com/cjlin1/libsvm/pull/5
[2] https://github.com/carandraug/libsvm