[神经网络]Matlab神经网络原理6.6.1节 - 使用BP网络通过身高体重分类性别(批量和工具箱)

main_script_feedforwardnet.m 工具箱代码

clear;close all;clc;
[data, label] = getdata('student.xls');
[traind,trainl,testd,testl] = divide(data,label);

%% 初始化网络
rng('default');
rng(0);

net = feedforwardnet(3); % 3个隐含层
net.trainFcn = 'trainbfg';

%% Train
net = train(net,traind',trainl);

%% Test
test_out = sim(net,testd');
test_out(test_out >= 0.5) = 1;
test_out(test_out < 0.5) = 0;
rate = sum(test_out == testl)/length(testl);
fprintf('正确率\n %f %%\n',rate*100);

main_script.m 手动训练代码

clear;close all;clc;
[data, label] = getdata('student.xls'); % 读取样本和训练输出
[traind,trainl,testd,testl] = divide(data,label);

%% 设置参数
rng('default');
rng(6);                 % 随机种子
nTrainNum   = 60;       % 训练的样本数
nSampDim    = 2;        % 样本数的向量维度

%% 构造网络
net.nIn     = 2;        % 输入层的向量数
net.nHidden = 3;        % 隐含层节点数
net.nOut    = 1;        % 输出节点数

% 初始化权值和偏置
w           = 2*(rand(net.nHidden, net.nIn)-0.5); 
b           = 2*(rand(net.nHidden, 1)-0.5);

W           = 2*(rand(net.nOut, net.nHidden)-0.5);
B           = 2*(rand(net.nOut, 1)-0.5);

net.w1 = [w,b];
net.w2 = [W,B];

%% 训练数据归一化
mm = mean(traind);

% 均值平移
for i = 1:2
    traind_s(:,i) = traind(:,i) - mm(i);
end

% 方差标注化
ml(1) = std(traind_s(:,1));
ml(2) = std(traind_s(:,1));
for i = 1:2
    traind_s(:,i) = traind_s(:,i) / ml(i);
end

%% 训练
SampInEx = [traind_s';ones(1,nTrainNum)]; %输入数据带偏置
expectedOut = trainl;

eb      = 0.01;     % 误差容限
eta     = 0.6;      % 学习率
mc      = 0.8;      % 动量因子
maxiter = 2000;     % 最大迭代次数
iteration = 0;      % 第一代

errRec = zeros(1,maxiter);
outRec = zeros(nTrainNum, maxiter);
NET= [];% 记录
dWEXOld = 0;
dwexOld = 0;

% 开始迭代
for i = 1:maxiter
    % 隐含层输出计算
    hid_input = net.w1 * SampInEx;
    hid_out = logsig(hid_input);
    
    % 输出层计算
    ou_input1   = [hid_out;ones(1,nTrainNum)];
    ou_input2   = net.w2 * ou_input1;
    out_out     = logsig(ou_input2);
    
    % 记录迭代过程中的输出值
    outRec(:,i) = out_out';
    
    % 计算误差
    err = expectedOut - out_out;
    sse = sumsqr(err);
    errRec(i) = sse;
    fprintf('第%d次迭代 误差：%f\n', i, sse);
    iteration = iteration + 1;
    
    % 判断是否收敛
    if sse <= eb % error bound
        break;
    end
    
    % 误差反向传播
    % 隐含层与输出层之间的局部梯度
    DELTA = err.*dlogsig(ou_input2, out_out);
    % 输入层与隐含层之间的局部梯度
    delta = net.w2(:,1:end-1)'*DELTA.*dlogsig(hid_input, hid_out);
    
    % 权值修改量
    dWEX = DELTA * ou_input1';
    dwex = delta * SampInEx';
    
    % 修改权值，如果不是第一次修改，则使用动量因子
    if i == 1
        net.w2 = net.w2 + eta * dWEX;
        net.w1 = net.w1 + eta * dwex;
    else
        net.w2 = net.w2 + (1 - mc) * eta * dWEX + mc * dWEXOld;
        net.w1 = net.w1 + (1 - mc) * eta * dwex + mc * dwexOld;
    end
    
    dWEXOld = dWEX;
    dwexOld = dwex;
    
    
end

%% 测试
% 测试数据归一化
for i = 1:2
    testd_s(:,i) = testd(:,i) - mm(i);
end

for i = 1:2
    testd_s(:,i) = testd_s(:,i)/ml(i);
end
    
% 计算测试输出
InEx = [testd_s';ones(1,260 - nTrainNum)];
hid_input = net.w1 * InEx;
hid_out = logsig(hid_input);
ou_input1 = [hid_out;ones(1,260 - nTrainNum)];
ou_input2 = net.w2 * ou_input1;
out_out = logsig(ou_input2);
out_out1 = out_out;

% 取整
out_out(out_out<0.5) = 0;
out_out(out_out>=0.5) = 1;
% 正确率
rate = sum(out_out == testl)/length(out_out);

%% 显示
% 显示训练样本
train_m = traind(trainl == 1,:);
train_m = train_m';
train_f = traind(trainl == 0,:);
train_f = train_f';

figure(1);
plot(train_m(1,:),train_m(2,:),'bo');
hold on;
plot(train_f(1,:),train_f(2,:),'r*');
xlabel('身高');
ylabel('体重');
title('训练样本分布');
legend('男生','女生');

figure(2);
axis on
hold on
grid 
[nRow, nCol] = size(errRec);
plot(1:nCol,errRec,'b-','LineWidth',1.5);
legend('误差平方和');
xlabel('迭代次数','FontName','Times','FontSize',10);
ylabel('误差');

fprintf('-----------------错误分类表---------------');
fprintf('   编号      标签      身高      体重\n')
ind = find(out_out ~= testl);

for i = 1:length(ind)
    fprintf('   %4d     %4d   %f    %f \n', ind(i), testl(ind(i)), testd(ind(i),1), testd(ind(i),2));
end

fprintf('最终迭代次数\n %d\n', iteration);
fprintf('正确率:\n %f%%n',rate*100);

数据集：