boost.m

一个关于adaboost和NN的matlab程序

function [betaT,net]=boost(in,B,validation,tB,outfile,bmax,S1)
% WBL 22 August 2002 Boost Matlab MLP Neural Network bmax times
%                          1 input neuron per attribute, size(in)
%                          1 hidden layer of S1 inputs
%                          2 output neurons 
versn = '$Revision: 1.10 $';  %$Date: 2002/10/31 11:14:02 $

%Implement Pseudo-Loss ADAboost.M2 Table 1 p1873
%"R" p1874, Boosting Neural Networks, 
%Schwenk and Bengio, Neural computation, 12 1869-1887(2000)
%but use MATLAB to decide number of training iterations
%(hence variable rather than fixed)
%Do NOT use fancy cost function, just use Matlab default (least squares)
%Resampled tests automatically given higher impact on cost function because they
%occur multiple times in test set.
%Only tests that appear in testset are updated (possibly multiple times)
%others retain their previous value.

%WBL 31 Oct 2002 based on P450boost2.m r1.9
%WBL 30 Sep 2002 Update in line with P450boost.m r1.7
%WBL 21 Sep 2002 Based on P450boost.m r1.6
%WBL 21 Sep 2002 Based on P450train.m r1.1
%WBL  9 Sep 2002 Based on read.m Try again but save just net not whole of Matlab
%WBL 27 Aug 2002 Try hidden layer form 1..15 nodes, instead of just 20

% expand B into targets. Each element becomes 2 element row,
% element corresponding to B(i) is one all others are zero.
% (already done for validation)

ntests = size(B,2);
S2 = 2; %number of output neurons
targets = zeros(S2,ntests);
for i=1:ntests
    targets(B(i)+1.0,i) = 1.0;
end

%weights used to decide which training data to use for boosted network
%(data split into use for training and use for deciding when to stop training)
weights = zeros(1,ntests);
for i=1:ntests
    weights(i) = 1/ntests;
end

%outfile = fopen('boost25-sep-2002.log','w');
fprintf(outfile,'#boost.m %s train %d, ver %d %s\n',versn,ntests,size(tB,2),datestr(now));

%bmax=10;
betaT=zeros(1,bmax);
t=1;
for b=1:bmax
    %Select training data for network using weights
    if(t>1)
        [index1,index2] = randselect(weights);
    else 
        %first time use all training data 
        index1=1:1:ceil(ntests/2); %test=in; y=B;
        index2=ceil(ntests/2)+1:1:ntests;
    end

    %gather statistics about data actually used for training
    [t1,t1p,t1n,u1,u1p,u1n] = numtests(targets,index1);
    n=hist(index1,[1:ntests]);
    fprintf(         '%d %s Range test(%d,%d,%d,%d,%d,%d) %d %d %d ',b,datestr(now),t1,t1p,t1n,u1,u1p,u1n,min(n),max(n),median(n));
    fprintf(outfile,'%4d %s Range test(%d,%d,%d,%d,%d,%d) %d %d %d ',b,datestr(now),t1,t1p,t1n,u1,u1p,u1n,min(n),max(n),median(n));
    [t2,t2p,t2n,u2,u2p,u2n] = numtests(targets,index2);
    n=hist(index2,[1:ntests]);
    fprintf(        'stop(%d,%d,%d,%d,%d,%d) %d %d %d ',t2,t2p,t2n,u2,u2p,u2n,min(n),max(n),median(n));
    fprintf(outfile,'stop(%d,%d,%d,%d,%d,%d) %d %d %d ',t2,t2p,t2n,u2,u2p,u2n,min(n),max(n),median(n));
    
    train     =      in(:,index1);
    trainans  = targets(:,index1);
    stopset.P =      in(:,index2);
    stopset.T = targets(:,index2);

    [net{t},epochs,rms]=trainnet(train,trainans,S1,stopset);
    net_tr = sim(net{t},in);
    %net_tr = simple(targets);
    [betaT(t),loss] = pseudoloss(weights,net_tr,targets);
    fprintf(        'epochs %3d RMS %f beta %f\n',epochs,rms,betaT(t));
    fprintf(outfile,'epochs %3d RMS %f beta %f ', epochs,rms,betaT(t));

    %NON STANDARD paper says stop if beta>1 but get such large stochastic effects on P450 data
    %would stop many time on first trail! Hence simply discard this trial - roll back and try again

    if(betaT(t)>=1) 
        fprintf(outfile,'Boosting trial %4d failed\n',b);
    else 
        [roc,net_tr] = ensemble(t,betaT,net,in,          B,  outfile);
        fprintf(outfile,'AUROC %f ',roc);
        [roc,net_vr] = ensemble(t,betaT,net,validation.P,tB',outfile);
        fprintf(outfile,'AUROC %f\n',roc);
        
        weights=update(weights,betaT(t),loss);
        t = t+1;
        fprintf(outfile,'#%4d %4d Range of weights min %f (%f) max %f (%f)\n',b,t,...
            min(weights), min(weights)/mean(weights),...
            max(weights), max(weights)/mean(weights));
    end
end
%save('boost','versn','betaT','net');
%fclose (outfile);