NAME¶

msvmocas - train a multi-class linear SVM classifier

SYNOPSIS¶

msvmocas [options] example_file model_file

DESCRIPTION¶

msvmocas is a program that trains a multi-class linear SVM classifier using the Optimized Cutting Plane Algorithm for Support Vector Machines (OCAS) and produces a model file. example_file is a file with training examples in SVM^light format, and model_file is the file in which to store the learned linear rule f(x)=W'*x. model_file contains M columns and D lines, where M is the number of classes and D the number of dimensions, corresponding to the elements of the matrix W [D x M].

OPTIONS¶

A summary of options is included below. General options:

-h

Show summary of options.

-v (0|1)

Set the verbosity level (default: 1)

Learning options:

-c float

Regularization constant C. (default: 1)

-n integer

Use only the first integer examples for training. By default, integer equals the number of examples in example_file.

Optimization options:

-m (0|1)

Solver to be used:

 

-s integer

Cache size for cutting planes. (default: 2000)

Stopping conditions:

-a float

Absolute tolerance TolAbs: halt if QP-QD <= TolAbs. (default: 0)

-r float

Relative tolerance TolAbs: halt if QP-QD <= abs(QP)*TolRel. (default: 0.01)

-q float

Desired objective value QPValue: halt is QP <= QPValue. (default: 0)

-t float

Halts if the solver time (loading time is not counted) exceeds the time given in seconds. (default: infinity)

EXAMPLES¶

Train the multi-class SVM classifier from example file example4_train.light, with the regularization constant C=10, verbosity switched off, and save model to msvmocas.model: Compute the testing error of the classifier stored in msvmocas.model with linclass(1) using testing examples from example4_test.light and save the predicted labels to example4_test.pred:

SEE ALSO¶

svmocas(1), linclass(1).

AUTHORS¶

msvmocas was written by Vojtech Franc <xfrancv@cmp.felk.cvut.cz> and Soeren Sonnenburg <Soeren.Sonnenburg@tu-berlin.de>. This manual page was written by Christian Kastner <debian@kvr.at>, for the Debian project (and may be used by others).