MNIST Matlab 程序 识别率可到0.983

% A BP network for MNIST data % 隐含层ReLU激活函数+均方误差代价函数+L2规则化 % Author: Guo chengfeng % Version 1.1 % date: 2018-3-19 %神经网络的层次 ntrain = length(y_train); arch = [784,100,100,10]; nlayer = length(arch); mini_batch_size = 100; max_iteration = 50000; eta = 1; lambda = 5; % mini = 20, eta = 6 eps= 5000. 可以达到0.946 %神经网络参数初始化 weight = cell(1,nlayer); bias = cell(1,nlayer); nabla_weight = cell(1,nlayer); nabla_bias = cell(1,nlayer); a = cell(1,nlayer); z = cell(1,nlayer); rstep = 100; for in = 2:nlayer weight{in} = rand(arch(in),arch(in-1))-0.5; bias{in} = rand(arch(in),1)-0.5; nabla_weight{in} = rand(arch(in),arch(in-1)); nabla_bias{in} = rand(arch(in),1); end for in = 1:nlayer a{in} = zeros(arch(in),mini_batch_size); z{in} = zeros(arch(in),mini_batch_size); end accuracy = zeros(1,ceil(max_epochs/rstep)); iaa = 0; for ip = 1:max_iteration pos = randi(ntrain-mini_batch_size); x = x_train(:,pos+1:pos+mini_batch_size); y = y_train(:,pos+1:pos+mini_batch_size); %正向计算 a{1} = x; [a,z]=feedforward(@acti_relu,@acti_sigmoid,weight,bias,nlayer,mini_batch_size,a,z); [weight,bias] = SGD(@acti_relu_prime,@acti_sigmoid_prime,weight,bias,... nabla_weight,nabla_bias,nlayer,mini_batch_size,eta,a,z,y,lambda,ntrain); if mod(ip,rstep) == 0 iaa = iaa+1; accuracy(iaa) = evaluatemnist(@acti_relu,@acti_sigmoid,x_valid,y_valid,weight,bias,nlayer); plot(accuracy); title(['Accuracy:',num2str(accuracy(iaa))]); getframe; end end figure plot(accuracy,'r','linewidth',2); ylim([0.95,0.99]) legend(num2str(arch)) title(['ReLU激活函数+均方误差代价函数+L2规则化,最高识别率：',num2str(max(accuracy)*100),'%']) grid on