用MATLAB实现的LeNet-5网络，基于cifar-10数据库.zip

function CNN = TrainCNN(Train, Label, Test, Tag, alpha) %% 在给定的数据集上训练神经网络:对LeNet-5进行训练. % Train: 给定数据集，每一列代表一个input. % Label: 数据集归类标签，每一列代表一个output. % Test: 给定测试集，每一列代表一个input. % Tag: 测试集归类标签，每一列代表一个output. % alpha: 初始学习率. % CNN: cell数组，依次存放A1, A2, A3, ...和 Loss. % 袁沅祥，2019-7 %% 初始值 if nargin < 4 alpha = 1e-2; % 初始学习率 end iter = 1000; % 单次最大迭代次数 [CNN, state] = TrainRecovery();% 恢复训练 start = size(CNN{end}, 2); % 上一次迭代次数 if state fprintf('CNN:迭代[%g]次,精度%g.\n', start, state); return end fprintf('从第[%g]步开始迭代.\n', start); p = alpha * 0.99^start; lr = p * 0.99.^(0:iter); % 学习率随迭代次数衰减 %profile on; %profile clear; %% 开始迭代 sy = size(Label, 1); num = 200;%size(Train, 3); % 第一行存放误差，第二、三行存放准确率 errs = zeros(3, iter); count = 0; EarlyStopping = 10; %CNN早停条件 queue = cell(EarlyStopping+1, 1); %存放最近几次CNN网络 for i = 1:iter tic; alpha = lr(i); % 总误差 total = zeros(1, num); for k = 1 : num % 遍历元素 I = Train(:, :, k); % 卷积,池化 32 -> 28 -> 14 C1 = cell(1, 6); S2 = cell(1, 6); for n = 1:6 C1{n} = reLU(conv2(I, CNN{1}{n}, 'valid')); S2{n} = Sampling(C1{ n }); end % 卷积,池化 14 -> 10 -> 5 C3 = cell(1, 16); S4 = cell(1, 16); F0 = zeros(16 * 5*5, 1); for n = 1:16 Sum = zeros(10, 10); for j = 1:6 Sum = Sum+conv2(S2{j},CNN{2}{n,j},'valid'); end C3{n} = reLU(Sum); S4{n} = Sampling(C3{n}); s = 1 + (n-1) * 25; F0(s:s+24) = S4{n}(:); end % 全连接 400 -> 120 -> 84 -> 10 F1 = reLU(CNN{3} * [1; F0]); F2 = reLU(CNN{4} * [1; F1]); F3 = reLU(CNN{5} * [1; F2]); Out = softmax(F3); % 交叉熵 Loss = -sigma(target * ln(output)) total(:, k) = - dot(Label(:, k), log(Out)); % BP-误差反向传播 err3 = (Out-Label(:, k)) .* Grad(F3); B3 = CNN{5} - alpha * err3 * [1; F2]'; err2 = (CNN{5}(:, 2:end)' * err3) .* Grad(F2); B2 = CNN{4} - alpha * err2 * [1; F1]'; err1 = (CNN{4}(:, 2:end)' * err2) .* Grad(F1); B1 = CNN{3} - alpha * err1 * [1; F0]'; % 卷积、池化层 err0 = reshape(CNN{3}(:, 2:end)' * err1, 5, 5, 16); % 转为矩阵 A2 = CNN{2}; A1 = CNN{1}; bp = cell(1, 16); for n = 1:16 bp{n} = Upsampling(err0(:,:,n)) .* Grad(C3{n});% 上采样:S4->C3 for j = 1:6 grad = conv2(rot180(S2{j}), bp{n}, 'valid'); % C3->S2 A2{n,j} = A2{n,j} - alpha * grad; end end for n = 1:6 err = zeros(14, 14); for j = 1:16 err = err + conv2(bp{j}, rot180(A2{j,n}), 'full'); end err = Upsampling(err) .* Grad(C1{n}); % 上采样:S2->C1 grad = conv2(rot180(I), err, 'valid'); % C1->I A1{n} = A1{n} - alpha * grad; end CNN{3} = B1; CNN{4} = B2; CNN{5} = B3; CNN{2} = A2; CNN{1} = A1; end queue = circshift(queue, 1); queue{1} = CNN; e = mean(total); errs(1, i) = e; s = Accuracy(CNN, Train, Label, num); t = Accuracy(CNN, Test, Tag, num); best = max(errs(3, 1:i)); % 前i-1次最好的结果 errs(1, i) = e; errs(2, i) = s; errs(3, i) = t; if t <= best count = count + 1; if count == EarlyStopping CNN = queue{end}; Loss = SaveResult(CNN, CNN{end}, errs, i-EarlyStopping, 1); return end else count = 0; end % 保存权重 if t >= 0 Loss = SaveResult(CNN, CNN{end}, errs, i, 10); end fprintf('%g err=%g lr=%g acc=%g %g use %gs\n',i+start,e,alpha,s,t,toc); end %profile viewer; end