AdaBoost Demo

clear;clc offset = 0.1; pdata = [rand(300,2) + offset + 1; ... rand(300,2) + offset + repmat([1, 0], 300, 1);... rand(300,2) + offset + repmat([0, 1], 300, 1)]; ndata = rand(900,2) - offset;%-repmat([.5, .5], 60, 1) .* (rand(60,2) - 0.2); ptestdata = [rand(20,2) + offset + 1; ... rand(20,2) + offset + repmat([1, 0], 20, 1);... rand(20,2) + offset + repmat([0, 1], 20, 1)];%repmat([0.5, .5], 10, 1) .* (rand(10,2) - 0.2); ntestdata = rand(50,2) - offset;%-repmat([.5, .5], 10, 1) .* (rand(10,2) - 0.2); testdata = [ptestdata; ntestdata]; data = [pdata; ndata]; sampNum = size(data,1); pSampNum = size(pdata,1); nSampNum = size(ndata,1); hold on for i = 1 : pSampNum plot(pdata(i,1), pdata(i,2), '.') end for i = 1 : nSampNum plot(ndata(i,1), ndata(i,2), 'r.') end classifierNum = 4; % c1 x, c2 y, c3 x, c4 y xset = sort(data(:,1)); yset = sort(data(:,2)); weight = 1 / sampNum * ones(sampNum, 1); A = []; % 训练弱分类器3 for kk = 1 errrate = []; for i = 1 : length(xset) - 1 txsplit = (xset(i) + xset(i + 1)) / 2; xdata = data(:,1); xdata(1:pSampNum) = xdata(1:pSampNum) < txsplit; xdata(pSampNum+1:end) = xdata(pSampNum+1:end) > txsplit; errrate = [errrate; xdata' * weight, txsplit]; end [e, minpos] = min(errrate(:,1)); xsplit2 = errrate(minpos, 2); a = 0.5 * log(1 / e - 1); y = data(:,1); y = y > xsplit2; y = 2 * (y - 0.5); t = [ones(pSampNum,1); -1 * ones(nSampNum,1)]; tmp = weight .* exp(-a .* t .* y); weight = tmp / sum(tmp); A = [A, a]; e end % 训练弱分类器1 for kk = 1 errrate = []; for i = 1 : length(xset) - 1 txsplit = (xset(i) + xset(i + 1)) / 2; xdata = data(:,1); xdata(1:pSampNum) = xdata(1:pSampNum) < txsplit; xdata(pSampNum+1:end) = xdata(pSampNum+1:end) > txsplit; errrate = [errrate; xdata' * weight, txsplit]; end [e, minpos] = min(errrate(:,1)); xsplit = errrate(minpos, 2); a = 0.5 * log(1 / e - 1); y = data(:,1); y = y > xsplit; y = 2 * (y - 0.5); t = [ones(pSampNum,1); -1 * ones(nSampNum,1)]; tmp = weight .* exp(-a .* t .* y); weight = tmp / sum(tmp); A = [A, a]; e end % 训练弱分类器2 for kk = 1 errrate = []; for i = 1 : length(yset) - 1 tysplit = (yset(i) + yset(i + 1)) / 2; ydata = data(:,1); ydata(1:pSampNum) = ydata(1:pSampNum) < tysplit; ydata(pSampNum+1:end) = ydata(pSampNum+1:end) > tysplit; errrate = [errrate; ydata' * weight, tysplit]; end [e, minpos] = min(errrate(:,1)); ysplit = errrate(minpos, 2); a = 0.5 * log(1 / e - 1); y = data(:,1); y = y > ysplit; y = 2 * (y - 0.5); t = [ones(pSampNum,1); -1 * ones(nSampNum,1)]; tmp = weight .* exp(-a .* t .* y); weight = tmp / sum(tmp); A = [A, a]; e end % 训练弱分类器4 for kk = 1 errrate = []; for i = 1 : length(yset) - 1 tysplit = (yset(i) + yset(i + 1)) / 2; ydata = data(:,1); ydata(1:pSampNum) = ydata(1:pSampNum) < tysplit; ydata(pSampNum+1:end) = ydata(pSampNum+1:end) > tysplit; errrate = [errrate; ydata' * weight, tysplit]; end [e, minpos] = min(errrate(:,1)); ysplit2 = errrate(minpos, 2); a = 0.5 * log(1 / e - 1); y = data(:,1); y = y > ysplit2; y = 2 * (y - 0.5); t = [ones(pSampNum,1); -1 * ones(nSampNum,1)]; tmp = weight .* exp(-a .* t .* y); weight = tmp / sum(tmp); A = [A, a]; e end for kk = 1 % draw classifiers plot(repmat(xsplit, 1, 2), [min(yset), max(yset)]) plot([min(xset), max(xset)], repmat(ysplit, 1, 2), 'r') plot(repmat(xsplit2, 1, 2), [min(yset), max(yset)], 'linewidth', 2) plot([min(xset), max(xset)], repmat(ysplit2, 1, 2), 'r', 'linewidth', 2) % for i = 1 : sampNum % if 2 * weight(i) > 0.5 * max(weight) + min(weight) || true % text(data(i,1), data(i,2), [' (', num2str(ceil(-log(weight(i)))), ')']) % end % end % test y1 = testdata(:,1); y1 = y1 > xsplit; y1 = 2 * (y1 - 0.5); y2 = testdata(:,2); y2 = y2 > ysplit; y2 = 2 * (y2 - 0.5); y3 = testdata(:,1); y3 = y3 > xsplit2; y3 = 2 * (y3 - 0.5); y4 = testdata(:,2); y4 = y4 > ysplit2; y4 = 2 * (y4 - 0.5); Y = [y1, y2, y3, y4]; pre = Y * A' > 0; for i = 1 : size(testdata, 1) if i <= size(ptestdata, 1) plot(testdata(i,1), testdata(i,2), '.') else plot(testdata(i,1), testdata(i,2), 'r.') end if pre(i) > 0 plot(testdata(i,1), testdata(i,2), 'o') else plot(testdata(i,1), testdata(i,2), 'ro') end end end