function model=svmlight(data,options)
% SVMLIGHT Interface to SVM^{light} software.
%
% Synopsis:
% model = svmlight(data)
% model = svmlight(data,options)
%
% Description:
% This function serves as an interface between Matlab
% and SVM^{light} (Version: 5.00) optimizer which trains
% the Support Vector Machines classifier.
%
% The executable file 'svm_learn' must be in the path.
% The SVM^{light} software can be downloaded from:
% http://svmlight.joachims.org/
%
% This function creates temporary files 'tmp_alphaXX.txt',
% 'tmp_examplesXX.txt', 'tmp_modelXX.txt' and 'tmp_verbXX.txt' for
% comunication with the SVM^{light} software. The XX=datestr(now)
% is string consisting of current date and time.
%
% Input:
% data [struct] Labeled binary data:
% .X [dim x num_data] Training vectors.
% .y [1 x num_data] Labels of training data (1 or 2).
%
% options [struct] Control parameters:
% .ker [string] Kernel identifier:
% 'linear' (default),'rbf' and 'poly'.
% .arg [1x1] Kernel argument (default []).
% .C [1x1] SVM regularization constant (default C=inf).
% .mC [1x1] if mC is given then C is set to mC/length(data.y).
% .j [1x1] Cost-factor, by which training errors on
% positive examples outweight errors on negative examples (default 1).
% .eps [1x1] Tolerance of KKT-conditions (default eps=0.001).
% .b [1x1] if 1 (default) then finds w'*x +b else b = 0;
% .keep_files [1x1] If ==1 then keeps temporary files otherwise
% erase them.
% .svm_command [string] Path to SVM^{light} solver (default "svm_learn")
%
% Output:
% model [struct] Binary SVM classifier:
% .Alpha [nsv x 1] Weights of support vectors.
% .b [1x1] Bias of decision function.
% .sv.X [dim x nsv] Support vectors.
% .sv.inx [1 x nsv] Indices of SVs (model.sv.X = data.X(:,inx)).
% .nsv [int] Number of Support Vectors.
% .kercnt [int] Number of kernel evaluations used by the SVM^{light}.
% .trnerr [real] Classification error on training data.
% .margin [real] Margin of found classifier.
% .options [struct] Copy of used options.
% .cputime [real] Used CPU time in seconds.
%
% Example:
% data=load('riply_trn');
% model=svmlight(data,struct('ker','rbf','C',10,'arg',1))
% figure; ppatterns(data); psvm(model);
%
% See also
% SVMCLASS, XY2SVMLIGHT.
%
% About: Statistical Pattern Recognition Toolbox
% (C) 1999-2003, Written by Vojtech Franc and Vaclav Hlavac
% <a href="http://www.cvut.cz">Czech Technical University Prague</a>
% <a href="http://www.feld.cvut.cz">Faculty of Electrical Engineering</a>
% <a href="http://cmp.felk.cvut.cz">Center for Machine Perception</a>
% Modifications:
% 09-sep-2007, VF, -b option added
% 21-may-2007, VF, -q 42 (size of QP subproblem) added based on Soeren's suggestion
% 20-nov-2006, VF, added optional parameter mC
% 10-oct-2006, VF, "svm_command" option added
% 09-feb-2006, VF, added date_str(findstr(date_str,':')) = '.'; based on
% M.Urban comment.
% 16-may-2004, VF
% 15-jan-2004, VF, handling argument of poly kernel repared
% 10-oct-2003, VF, computation of lin model added
% 29-aug-2003, VF, seconds are added to the name of temporary files
% 12-may-2003, VF, 1st 3 lines of verb_file are skiped
% 31-jan-2003, VF, added option 'j'
% 28-Jan-2003, VF
% 20-jan-2003, VF, temporary files are unique and are deleted at the end
% 14-Jan-2003, VF
% 26-sep-2002, VF
% 3-Jun-2002, V.Franc
tic;
data=c2s(data);
date_str=datestr(now);
date_str(findstr(date_str,' ')) = '-';
date_str(findstr(date_str,':')) = '.';
sec=clock;
date_str = [date_str '-' num2str(sec(end))];
examples_file = ['tmp_examples' date_str '.txt'];
model_file = ['tmp_model' date_str '.txt'];
verb_file = ['tmp_verb' date_str '.txt'];
alpha_file = ['tmp_alpha' date_str '.txt'];
model.name = 'SVM classifier';
if nargin < 2, options = []; else options=c2s(options); end
if isfield(options,'mC'), options.C = options.mC/length(data.y); end
if ~isfield(options,'ker'), options.ker = 'linear'; end
if ~isfield(options,'arg'), options.arg = '[]'; end
if ~isfield(options,'C'), options.C = inf; end
if ~isfield(options,'eps'), options.eps = 0.001; end
if ~isfield(options,'keep_files'), options.keep_files = 0; end
if ~isfield(options,'j'), options.j = 1; end
if ~isfield(options,'b'), options.b = 1; end
if ~isfield(options,'svm_command'), options.svm_command = 'svm_learn'; end
[dim,num_data ] = size(data);
switch options.ker
case 'linear'
ker='-t 0';
case 'rbf'
ker=['-t 2 -g ' num2str(1/(2*options.arg^2))];
case 'poly'
if length(options.arg) == 1,
ker=['-t 1 -r 1 -s 1 -d ' num2str(options.arg)];
else
ker=['-t 1 -s 1 -r ' num2str(options.arg(2)) ' -d ' ...
num2str(options.arg(1))];
end
end
command=[options.svm_command ' ' ...
'-c ' num2str(options.C) ' '...
ker ' ' ...
'-v 1 ' ...
'-m 40 ' ...
'-q 42 ' ...
'-j ' num2str(options.j) ' '...
'-e ' num2str(options.eps) ' '...
'-b ' num2str(options.b) ' ' ...
'-a ' alpha_file ' ' examples_file ' ' model_file ' > ' verb_file];
xy2svmlight(data,examples_file);
[a,b]=system(command);
checkfile(model_file); [lines]=textread(model_file,'%s');
for i=1:size(lines,1)
if strcmpi( lines(i), 'threshold' )==1,
model.b=-str2num( lines{i-2});
break;
end
end
checkfile(alpha_file); model.Alpha=textread(alpha_file,'%f');
model.Alpha=model.Alpha(:);
model.Alpha(find(data.y==2)) = -model.Alpha(find(data.y==2));
checkfile(verb_file);
[lines]=textread(verb_file,'%s',-1,'bufsize',5000000,'headerlines',3);
for i=1:size(lines,1)
if strcmpi( lines{i}, 'misclassified,' ),
model.trnerr=str2num( lines{i-1}(2:end));
model.trnerr=model.trnerr/length(model.Alpha);
end
if strcmpi( lines(i), 'vector:' ) & strcmpi( lines(i-1), 'weight' )==1,
tmp=str2num( lines{i+1}(5:end));
if tmp~=0, model.margin=1/tmp; else model.margin=[]; end
end
if strcmpi( lines(i), 'SV:' )==1,
model.nsv=str2num( lines{i+1});
end
if strcmpi( lines(i), 'evaluations:' )==1,
model.kercnt=str2num( lines{i+1});
end
end
model.nsv = length(find(model.Alpha~=0));
inx=find(model.Alpha);
model.sv.X = data.X(:,inx);
model.sv.y = data.y(inx);
model.sv.inx = inx;
model.Alpha = model.Alpha(inx);
model.Alpha(find(model.sv.y==2)) = -model.Alpha(find(model.sv.y==2));
if strcmp( options.ker, 'linear'),
model.W = model.sv.X * model.Alpha;
end
model.options = options;
model.fun = 'svmclass';
if options.keep_files == 0,
delete(examples_file);
delete(model_file);
delete(verb_file);
delete(alpha_file);
end
model.cputime=toc;
return;
function checkfile(fname)
attempts=0;
found = exist(fname);
while attempts < 5 && ~found
found = exist(fname);
attempts = attempts+1;
end
if found == 0,
error('File %s not found.\n', fname);
end
return;