NIPS 2008: MLOSS workshop - 2008-12-12 17:55 PST - FracPete
I presented BMVW at the NIPS 2008 workshop on Machine Learning Open Source Software. At one point, the presentation will be available on

Version 0.0.4 released - 2008-12-04 11:24 NZ - FracPete
The releases come in quick succession now... Just released version 0.0.4. It contains a few bugfixes along some new functionality, e.g., re-running the last experiment setup that was used (handy, if one forgot to check "Submit"), selecting multiple datasets for displaying comments, data, etc.
Admins can now also delete datasets from the database (either from the command-line with the deleteDataset command or via the GUI in the datasets overview).
You can grab it from the downloads section as usual.

Version 0.0.3 released - 2008-12-02 16:10 NZ - FracPete
Version 0.0.3 is now officially released. It is mainly a bugfix release, with the most import fix: experiments are not longer submitted incorrectly on multi-core machines.
In addition to that, admins can now also delete experiments from the database (either from the command-line with the deleteExperiment command or via the GUI in the experiments overview).
You can grab it from the downloads section as usual.

Release 0.0.2 is out of the door! - 2008-10-01 11:22 NZ - FracPete
Just released version 0.0.2 of the framework!! You can grab it from the downloads section as usual. This release contains decent documentation for the user and the developer (how to implement new schemes/filter or clients in other languages).

I've also created an entry on Check it out! :-)

Basic support of MLPY classes - 2008-09-28 22:56 NZ - FracPete
Just added basic support for MLPY classes (Python-based machine learning classes) to the framework:


It's a bit of a hack, since the MLPY classes only take options via their constructor. This means, that one has to provide the constructor options via 'key=value' pairs as multiple options (= '-O key=value').

Next release is getting closer - 2008-09-26 13:33 NZ - FracPete
Man, processing XML in Python can be painfully slow. Using the 4Dom XML classes, it took 6min to load the anneal UCI dataset - compared to 3sec in Java. After switching to minidom instead and delaying the re-generation of the DOM document, I could cut it down to 10sec. Close enough, I'd say. ;-)

I've also added a few more classes:

  • FilteredScheme - a meta-scheme that filters the data before passing it on to the base scheme. A user-specified filter is used.
  • Support for all Weka filters (Java client only), through the "Weka" filter class.

Python client migrated - 2008-09-23 16:35 NZ - FracPete
Looks like I've migrated most of the Python code related to client-server communication. It now uses the same XML basis as the Java client. Now I can start on hunting down bugs... ;-)

Subversion repository with anonymous read access - 2008-09-22 09:50 NZ - FracPete
The subversion repository now has anonymous read access. You can access it via the following URLs:
  • trunk
  • tags
  • branches

Repository tagged - 2008-07-09 10:03 NZ - FracPete
Let's be honest: properties files as basis for the messages seemed to be an easy way of doing things. But it turned out to be rather infeasible with the ever increasing complexity and the accompanied nesting of props files. I've decided to switch to XML-based messages. Most major programming languages support XML and XML Schema anyway, so I don't see why it shouldn't work. I've just tagged the repository (20080709-before_switching_messages_to_XML) before I'm starting down this road...