%NULIBSVC Trainable classifier: LIBSVM, nu-algorithme
%
% [W,J,NU] = NULIBSVC(A,KERNEL,NU)
% [W,J,NU] = A*NULIBSVC([],KERNEL,NU)
% [W,J,NU] = A*NULIBSVC(KERNEL,NU)
%
% INPUT
% A Dataset
% KERNEL Mapping to compute kernel by A*MAP(A,KERNEL)
% or string to compute kernel by FEVAL(KERNEL,A,A)
% or cell array with strings and parameters to compute kernel by
% FEVAL(KERNEL{1},A,A,KERNEL{2:END})
% Default: linear kernel.
% NU nu value, upperbound error.
% Default NU is derived from 1-NN error.
%
% OUTPUT
% W Mapping: Support Vector Classifier
% J Object idences of support objects. Can be also obtained as W{4}
% NU Actual nu_value used
%
% DESCRIPTION
% Optimizes a support vector classifier for the dataset A by the libsvm
% package, see http://www.csie.ntu.edu.tw/~cjlin/libsvm/. LIBSVC calls the
% svmtrain routine of libsvm for training. Classifier execution for a
% test dataset B may be done by D = B*W; In D posterior probabilities are
% given as computed by svmpredict using the '-b 1' option.
%
% The kernel may be supplied in KERNEL by
% - an untrained mapping, e.g. a call to PROXM like W = LIBSVC(A,PROXM([],'R',1))
% - a string with the name of the routine to compute the kernel from A
% - a cell-array with this name and additional parameters.
% This will be used for the evaluation of a dataset B by B*W or PRMAP(B,W) as
% well.
%
% If KERNEL = 0 (or not given) it is assumed that A is already the
% kernelmatrix (square). In this also a kernel matrix should be supplied at
% evaluation by B*W or PRMAP(B,W). However, the kernel has to be computed with
% respect to support objects listed in J (the order of objects in J does matter).
%
% REFERENCES
% R.-E. Fan, P.-H. Chen, and C.-J. Lin. Working set selection using the second order
% information for training SVM. Journal of Machine Learning Research 6, 1889-1918, 2005
%
% SEE ALSO (<a href="http://37steps.com/prtools">PRTools Guide</a>)
% MAPPINGS, DATASETS, LIBSVC, SVC, PROXM
% Copyright: R.P.W. Duin, r.p.w.duin@37steps.com
% Faculty EWI, Delft University of Technology
% P.O. Box 5031, 2600 GA Delft, The Netherlands
function [W,J,NU] = nulibsvc(varargin)
checktoolbox('libsvm');
mapname = 'nuLIBSVM';
argin = shiftargin(varargin,{'prmapping','char','cell'});
argin = setdefaults(argin,[],proxm([],'p',1),[],[]);
if mapping_task(argin,'definition')
W = define_mapping(argin,'untrained',mapname);
elseif mapping_task(argin,'training') % Train a mapping.
[a,kernel,NU] = check_for_old_call(argin);
if isempty(NU), NU = 2*min(max(testk(a,1),0.01),(0.8*min(classsizes(a))/size(a,1))); end
%opt = ['-s 0 -t 4 -b 1 -c ',num2str(NU)];
opt = ['-s 1 -t 4 -b 1 -n ',num2str(NU), ' -q'];
islabtype(a,'crisp');
isvaldset(a,1,2); % at least 1 object per class, 2 classes
[m,k,c] = getsize(a);
nlab = getnlab(a);
K = compute_kernel(a,a,kernel);
K = min(K,K'); % make sure kernel is symmetric
K = [[1:m]' K]; % as libsvm wants it
% call libsvm
u = svmtrain(nlab,K,opt);
if isempty(u)
prwarning(1,'nulibsvc: no solution for SVM, pseudo-inverse will be used')
W = lkc(prdataset(K(:,2:end),getlabels(a)),0);
J = [1:m]';
return
end
% Store the results:
J = full(u.SVs);
if isequal(kernel,0)
s = [];
in_size = length(J);
% in_size = 0; % to allow old and new style calls
else
s = a(J,:);
in_size = k;
end
lablist = getlablist(a);
W = prmapping(mfilename,'trained',{u,s,kernel,J,opt},lablist(u.Label,:),in_size,c);
W = setname(W,mapname);
W = setcost(W,a);
else % Evaluation
[a,W] = deal(argin{1:2});
[u,s,kernel,J,opt] = getdata(W);
m = size(a,1);
K = compute_kernel(a,s,kernel);
k = size(K,2);
if k ~= length(J)
if isequal(kernel,0)
if (k > length(J)) & (k >= max(J))
% precomputed kernel; old style call
prwarning(1,'Old style execution call: The precomputed kernel was calculated on a test set and the whole training set!')
else
error('Inappropriate precomputed kernel!\nFor the execution the kernel matrix should be computed on a test set and the set of support objects');
end
else
error('Kernel matrix has the wrong number of columns');
end
else
% kernel was computed with respect to the support objects
% we make an approprite correction in the libsvm structure
u.SVs = sparse((1:length(J))');
end
K = [[1:m]' K]; % as libsvm wants it
%[lab,acc,d] = svmpredict(getnlab(a),K,u,'-b 1');
[lab,acc,d] = svmpredict(ones(m,1),K,u,'-b 1');
W = setdat(a,d,W);
end
return;
function K = compute_kernel(a,s,kernel)
% compute a kernel matrix for the objects a w.r.t. the support objects s
% given a kernel description
if isstr(kernel) % routine supplied to compute kernel
K = feval(kernel,a,s);
elseif iscell(kernel)
K = feval(kernel{1},a,s,kernel{2:end});
elseif ismapping(kernel)
K = a*prmap(s,kernel);
elseif kernel == 0 % we have already a kernel
K = a;
else
error('Do not know how to compute kernel matrix')
end
K = +K;
return
function [a,kernel,NU] = check_for_old_call(argin)
[a,kernel,NU,par] = deal(argin{:});
if ischar(kernel) && exist(kernel,'file') ~= 2
kernel = proxm(kernel,NU);
NU = par;
end