pocket_vccore.m

最新的模式识别分类工具箱,希望对朋友们有用!

% Learns classifier and classifies test set
%  using the perceptron learning algorithm
%  Works with 2 class labels, any number of features
%  when the class labels are 0 and 1.
%  Invoke using Perceptron_VC
%  
% Inputs:
% Usage
%      [trainError, testError, estTrainLabels, estTestLabels] = ...
%           Perceptron_VCcore(trainFeatures, trainLabels,algParam ,testFeatures, testLabels)
% where
%
% Inputs:
% 	trainFeatures  - the training set vectors, one vector per column
%	trainLabels    - the labels of the above
%       algParam       - algorithm parameters
%	                  Number of iterations, or weights vector, or [weights, number of iterations]
%       testFeatures   - test set, one column per vector
%       testLabels     - labels  for test set%
% Outputs
%	trainError     - the error rate on the training set (one entry per
%	                   class + total error)
%	testError      - the error rate on the test set (one entry per class
%	                  + total error)
%       estTrainLabels - the labels produced by the algorithm for the
%                          training samples
%       estTestLabels - the labels produced by the algorithm for the
%                          test samples

function   [trainError, testError, estTrainLabels, estTestLabels] = ...
    Pocket_VCcore(trainFeatures, trainLabels,algParam,testFeatures, testLabels)


[Nclasses, classes] = find_classes([trainLabels(:);testLabels(:)]);

if ( max(classes) > 1 | min(classes < 0)),
  fprintf('When invoking this classifier, use only labels 0 and 1\n');
  estTrainLabels = zeros(size(trainLabels));
  estTestLabels  = zeros(size(testLabels));
  trainError = computeError(classes, trainLabels, estTrainLabels);
  testError  = computeError(classes, testLabels , estTestLabels);
  return;
end

[Dim, r] = size(trainFeatures);

%Weighted Pocket or not?
switch length(algParam),
case r + 1,
    %Ada boost form
    p           = algParam(1:end-1);
    max_iter    = algParam(end);
case {r, 0},
    %No parameter given
    p           = ones(1,r);
    max_iter    = 500;
otherwise
    %Number of iterations given
    max_iter    = algParam;
    p           = ones(1,r);
end

trainFeatures = [trainFeatures ; ones(1,r)];
train_one      = find(trainLabels == 1);
train_zero     = find(trainLabels == 0);

%Preprocessing
processedFeatures = trainFeatures;
processedFeatures(:,train_zero) = -processedFeatures(:,train_zero);

%Initial weights
w_percept  = rand(Dim+1,1);
w_pocket   = rand(Dim+1,1);

correct_classified = 0;
n	 	   = length(trainLabels);
iter		   = 0;

while ((longest_run(w_percept, processedFeatures) < 0.9*n) ...
      & (iter < max_iter))

   iter = iter + 1;
   %Every 10 points, do the pocket switchover
   for i = 1:10,
      indice = 1 + floor(rand(1)*n);
      if (w_percept' * processedFeatures(:,indice) <= 0)
         w_percept = w_percept + p(indice) * processedFeatures(:,indice);
      end
   end
   % Find if it is neccessary to change weights:
   if (longest_run(w_percept, processedFeatures) > longest_run(w_pocket, processedFeatures)),
      w_pocket = w_percept;
   end
end

if (iter == max_iter)&(length(algParam)~= r + 1),
   disp(['Maximum iteration (' num2str(max_iter) ') reached']);
end

% Now computes
estTrainLabels = w_percept' * trainFeatures;
estTrainLabels = estTrainLabels > 0;

[Dim, Ntest]  = size(testFeatures);
testFeatures  = [testFeatures ; ones(1,Ntest)];

estTestLabels = w_percept' * testFeatures;
estTestLabels = (estTestLabels > 0);

trainError = computeError(classes, trainLabels, estTrainLabels);
testError  = computeError(classes, testLabels , estTestLabels);