linctrainwidro.m

一个关于数据聚类和模式识别的程序,在生物化学,化学中因该都可以用到.希望对大家有用,谢谢支持

function [recogRate, coef, allRecogRate, allCoef]=lincTrainWidro(DS, trainParam, plotOpt)
% lincTrainWidro: Linear classifier (Perceptron) training using (revised) Widro-Hoff method 
%	Usage: [recogRate, coef, allRecogRate, allCoef] = lincTrainWidro(DS, trainParam, plotOpt)
%
%	Type "lincTrainWidro" for a demo.

% Roger Jang, 20040910 

if nargin<1, selfdemo; return; end
[dim, dataNum]=size(DS.input);
if nargin<2 | isempty(trainParam),
	trainParam.eta=0.001;
	trainParam.maxIter=500;
	trainParam.useMre=1;
end
if nargin<3, plotOpt=0; end

if plotOpt==1 & dim==2,
	dcprDataPlot(DS);
	axis image
	limit=axis;
	lineH = line(limit(1:2), limit(3:4), 'linewidth', 2, 'erase', 'xor', 'color', 'k');
end

uniqueOutput=unique(DS.output);
if length(uniqueOutput)~=2, error('Must be 2-class problem!'); end
% 盢 DS.output эΘ -1 ┪ 1
index=find(DS.output==max(uniqueOutput));
DS.output=-ones(1, size(DS.output,2));
DS.output(index)=1;

allRecogRate=-ones(1, trainParam.maxIter);
allCoef=cell(1, trainParam.maxIter);

if trainParam.useMre==1
	[allRecogRate(1), coef]=lincTrainMre(DS);	% Initial parameters via mrecTrain
else
	coef=randn(dim+1,1);			% Initial parameters via random number
end

% The main loop
for i=1:trainParam.maxIter
	computed = lincEval(DS, coef);
	allCoef{i}=coef;
	allRecogRate(i)=sum(DS.output==computed)/dataNum;
	fprintf('%d/%d: %g%%\n', i, trainParam.maxIter, allRecogRate(i)*100);
	% Vectorized version, suitable for small data set
%	coef=coef+trainParam.eta*sum([DS.input; ones(1,dataNum)]*diag(DS.output-computed), 2);
	% For-loop version, suitable for big data set
	grad=0*coef;
	for j=1:dataNum
		grad=grad+(computed(j)-DS.output(j))*[DS.input(:,j); 1];
	end
	gradLength=norm(grad);
	if gradLength==0, break; end
	coef=coef-trainParam.eta*grad/norm(grad);		% Normalized version of steepest descent
	if plotOpt==1 & dim==2,
		set(lineH, 'ydata', (-coef(3)-coef(1)*limit(1:2))/coef(2));
		drawnow
	end
	% ====== Update step size
	if i>=5
		if all(diff(allRecogRate(i-4:i))>0)
			trainParam.eta = 1.1*trainParam.eta;
			fprintf('Increase trainParam.eta to %f\n', trainParam.eta);
		end
		if all(sign(diff(allRecogRate(i-4:i)))==[-1 1 -1 1]) | all(sign(diff(allRecogRate(i-4:i)))==[1 -1 1 -1])
			trainParam.eta = 0.9*trainParam.eta;
			fprintf('Decrease trainParam.eta to %f\n', trainParam.eta);
		end
	end
end

index=find(allRecogRate<0);
allRecogRate(index)=[];
allCoef(index)=[];

[recogRate, index]=max(allRecogRate);
coef=allCoef{index};

if plotOpt
	figure
	plot(1:length(allRecogRate), allRecogRate*100, 1:length(allRecogRate), allRecogRate*100, '.');
	xlabel('No. of iteration');
	ylabel('Recognition rate (%)');
	grid on
	line(index, allRecogRate(index)*100, 'marker', 'o', 'color', 'r');
	fprintf('Max. recog. rate = %.2f at %d iteration.\n', allRecogRate(index)*100, index);
end

% ====== Self demo
function coef=selfdemo
dataNum = 2000;
DS.input = rand(2, dataNum)*2-1;
DS.output = -ones(1, dataNum);
index = find(DS.input(1, :)+DS.input(2, :)>0);
DS.output(index)=1;
plotOpt=1;
coef=feval(mfilename, DS, [], plotOpt);