tbb-boost, by Benoit Favre <[email protected]> (c) 2010

This program implements the scored-text classifier from Boostexter. It supports
the lib_svm file format for examples (one example per line, the label followed
by couples of feature_id, value separated by a colon).

label1 id1:value1 id2:value2 id3:value3
label2 id4:value4 id5:value5
...

Building:

install the intel thread building blocks (tbb) and a recent version of gcc
(that supports c++0x) then run make.

Training:
./train iterations < examples > model
./train 1000 < 165/train > 165.model

Predict labels:
./predict model < examples > predictions
./predict 165.model < 165/test > 165.predictions

The file formats are compatible with http://mlcomp.org.

Parallel training:

On multicore machines, training can be parallelized by using tbb-train in place
of train.  At the moment, the evaluation of weak classifiers is parallelized at
the feature level, so don't expect large gain unless you have very long
iterations over a large enough set of features balanced between examples.

Parallel predictions:

tbb-predict is a naive parallel implementation: it loads all test examples in
memory and then processes them in parallel.

This program is distributed under the GPL v3+ license.