Skip to content

HTTPS clone URL

Subversion checkout URL

You can clone with
or
.
Download ZIP
Branch: master

Fetching latest commit…

Cannot retrieve the latest commit at this time

..
Failed to load latest commit information.
.svn
data
html
AUTHORS
ChangeLog
Contents.m
INSTALL
LICENSE
Makefile
README
RELEASES
compute_errors_mex.c
lbpfilter_mex.c
lbppyr.m
lbppyr_features_mex.c
lbppyr_mex.c
lib_svmlight_format.c
lib_svmlight_format.h
liblbp.c
liblbp.h
libocas.c
libocas.h
libocas_test.m
libqp.h
libqp_splx.c
linclass.c
msvmocas.c
msvmocas.m
msvmocas_mex.c
msvmocas_test.m
ocas_helper.c
ocas_helper.h
ocas_lbp_helper.c
ocas_lbp_helper.h
sparse_mat.c
sparse_mat.h
svmlight_linclass.m
svmlight_linclass_mex.c
svmocas.c
svmocas.m
svmocas_lbp.m
svmocas_lbp_example.m
svmocas_lbp_mex.c
svmocas_mex.c
svmocas_parseout.m
version.h

README

LIBOCAS - Library implementing Optimized Cutting Plane Algorithm (OCAS) 
solver for training linear SVM classifiers

FEATURES
 - SVM solvers for training linear classifiers from large scale-data.
 - Binary (two-class) and genuine multi-class SVM formulations.
 - Optimized code written in C.
 - A stand alone application and MEX interface for Matlab.
 - Reads examples from SVM^light format.
 - Optimized for both sparse and dense features.
 - Parallelized version of the binary solver.
 - Allows using different C for each training example (Matlab's interace to binary solver).
 - Tools for classification.
 - Training translation invariant image classifiers from virtual examples.
 - Functions for computing image features based on Local Binary Patterns (LBP).


PROBLEM FORMULATION 

OCAS solver is currently implemented for training binary (two-class) and multi-class
SVM classifiers:

1. Binary case: OCAS solves the following unconstrained convex optimization task
    
   W^*,W0^* = argmin 0.5*(W'*W+W0^2) + C*sum max( 0, 1-y(i)*(W'*X(:,i)+W0*X0) )
                W,W0                  i=1:nData

where C is the regularization constant, X [nDim x nData] are training feature
vectors and y [nData x 1] are their binary labels (+1/-1). The result are 
parameters W^* [nDim x 1], W0^* [1 x 1] of the linear rule

      f(X) = sign( X'*W + W0 )

2. Multi-class case: OCAS solves the following unconstrained convex optimization task

   W^*  =  argmin  0.5*sum_y (W(:,y)'*W(:,y)) + C*  sum     max( (y~=y(i)) + (W(:,y) - W(:,y(i)))'*X(:,i)) 
             W                                    i=1:nData  y 

where C is the regularization constant, X [nDim x nData] are training feature
vectors and y [nData x 1] are their labels from 1 to Y. The result are 
parameters W^* [nDim x Y] of the linear rule

      f(X) = argmax X'*W(:,y)
               y

AVAILABILITY

LIBOCAS can be downloaded from 
    http://cmp.felk.cvut.cz/~xfrancv/ocas/html/index.html

PLATFORMS

GNU/Linux. 


LICENSE

LIBOCAS is licensed under the GPL version 3 (cf. LICENSE).


REFERENCES

V. Franc, S. Sonnenburg. Optimized Cutting Plane Algorithm for Large-Scale Risk Minimization. 
  The Journal of Machine Learning Research (JMLR), vol. 10, pp. 2157--2192. October 2009. 
  http://jmlr.csail.mit.edu/papers/volume10/franc09a/franc09a.pdf

V. Franc, S. Sonnenburg. OCAS optimized cutting plane algorithm for Support Vector 
  Machines. In Proceedings of ICML. Omnipress, 2008.
  http://cmp.felk.cvut.cz/~xfrancv/papers/Franc-OCAS-ICML08.pdf

S. Sonnenburg, V. Franc.  COFFIN: A Computational Framework for Linear SVMs.  
  In Proceedings of the 27nd International Machine Learning Conference (ICML'10). 
  Haifa 2010.
  http://cmp.felk.cvut.cz/~xfrancv/papers/Sonnenburg-COFFIN-ICML10.pdf
Something went wrong with that request. Please try again.