srsplit.m

一个非常经典的核统计学习工具箱。集成了kpca、kdr、ksri等。具有分类和回归双重功能。

function [RIndex EndIndex]=srsplit(mode, label, ratio, th)
%==========================================================================
% To extract reduced set for SVM or SVR, stratified over 
%   (1) classes (SVM), or slices of sorted responses (SVR); and
%   (2) cross-validation folds, if applicable.
%--------------------------------------------------------------------------
% Inputs:
% mode            [1 x 1] : learning algorithm. {'SVM', 'SVR'}
% label           [m x 1] : training data label or response
% ratio           [1 x 1] : ratio of reduced set size to the full data size 
% th              [1 x 1] : number of folds in cross-validation
%--------------------------------------------------------------------------
% Outputs:
% For SVM
% RIndex         [th, ?] : the ith row records extracted indices of
%                          reduced set for the ith cross-validation fold
%                          of all classes
%
% EndIndex       [? x z] : the end index of each class in a cross-validation fold
%
% For SVR
% RIndex         [th, ?] : the ith row records extracted indices of
%                           reduced set for the ith cross-validation fold
%==========================================================================

rand('state',sum(100*clock)); % set random number generator seed

if(~exist('th', 'var'))
    th =1;
end

Rows=length(label);
if(strcmpi(mode, 'svm'))
    % SVM
    label(find(label==-1)) = 2;
    Num_label = length(unique(label));
    EndIndex = zeros(Num_label+1,1);
    RIndex = [];
    for i = 1 : Num_label
        Temp_Index = find(label==i);
        Temp_Rows = length(Temp_Index);
        Temp_Size = fix(Temp_Rows*ratio);
        Temp_Index_R = Temp_Index(randperm(Temp_Rows));
        Temp=[];
        for j = 1 : th
            Temp(j,:) = Temp_Index_R((j-1)*Temp_Size+1 : j*Temp_Size);
        end
        RIndex = [RIndex Temp];
        EndIndex(i+1) = EndIndex(i)+Temp_Size;
    end
    EndIndex = EndIndex(2:end);
elseif(strcmpi(mode, 'svr'))
    % SVR
    [SLabel, SIndex]=sort(label);
    clear SLabel;
    
    size=fix(Rows*ratio);% set the number of slices
    num=fix(Rows/size);% size of each slice, 
                       % there might be some left-out elements for a partial slice
    boxes=zeros(size, num); 
    for i=1: size
        base=(i-1)*num;
        if(i<size)
            boxes(i, :)=base + randperm(num);%random permutation within each slice       
        else
            % insure that elements from the last partial slice get a chance
            % to enter the reduced set        
            tmp=base + randperm(Rows-(num*(size-1)));
            boxes(i, :)=tmp(1: num);            
        end
    end
    
    RIndex=zeros(num, size); 
    for i=1: num
        RIndex(i, :)=SIndex(boxes(:, i))'; % pick one index from each slice within each CV-fold
    end
    RIndex=RIndex(1:th, :);
else
    error([mode, ', unknown!!']);
end