【测试环境】
matlab2023a
【源码文件截图】
【实现部分代码】
mlp_test.m
%% MLP 2-layer to test XOR
clear;
clc;Mode = 'MNIST'
%Mode = 'XOR'if (strcmp(Mode,'MNIST'))% Load the digits into workspace (MNIST Test, from% http://yann.lecun.com/exdb/mnist/)num_train = 1000;[train_IMG,train_labels,test_IMG,test_labels] = readMNIST(num_train);input =cell(num_train,1);output =cell(num_train,1);test_input=cell(length(test_IMG),1);test_output=cell(length(test_labels),1);for i=1:num_train%input_img = double(train_IMG{i});%Pre processing - prewittinput_img = edge(train_IMG{i},'prewitt');[width height] = size(input_img);img_vec = reshape(input_img,1,width*height);input{i}=double(img_vec);labels_arr = zeros(1,10);labels_arr(train_labels(i)+1)=1;output{i} = labels_arr;endfor i=1:length(test_input)%input_img = double(test_IMG{i});%Pre processing - prewittinput_img = edge(test_IMG{i},'prewitt');[width height] = size(input_img);img_vec = reshape(input_img,1,width*height);test_input{i} = double(img_vec);labels_arr = zeros(1,10);labels_arr(test_labels(i)+1)=1;test_output{i} = labels_arr;end
elseif (strcmp(Mode,'XOR'))num_train = 4;input =cell(num_train,1);output =cell(num_train,1);input{1} = [0 0];input{2} = [0 1];input{3} = [1 0];input{4} = [1 1];output{1} = [0];output{2} = [1];output{3} = [1];output{4} = [0];endif (length(input)~=length(output))error('len_input does not equal to len_output');
end%% Determine # of nodes in hidden layer & output layer
num_node_il = length(input{1});
%num_node_hl = [num_node_il*2];
num_node_hl = [num_node_il];
num_node_ol = length(output{1});set_node =[num_node_il num_node_hl num_node_ol];%% Init. template (random)
rand('state',sum(100*clock));num_layer = length(set_node);
W=cell(num_layer-1,1);
B=cell(num_layer-1);for i=1:num_layer-1%% [Xavier10] shows that the interval ~ from https://deeplearning.net/tutorial/mlp.htmlmin_W = -4*sqrt(6/(set_node(i)+set_node(i+1)));max_W = 4*sqrt(6/(set_node(i)+set_node(i+1)));W{i} = min_W+(2*max_W).*rand(set_node(i),set_node(i+1));B{i} = rand(1,set_node(i+1));
end%% Learning coeff = 0.7 & Iteration = 10% 141108, Success rate = 0.725
%lrn_rate = 0.3;
%max_iter = 100;lrn_rate = 0.3;
max_iter = 50;%最大迭代次数，越大训练时间越长精度越高ticAct=cell(num_layer,1);
Err=cell(num_layer-1,1);err_trace=[];for index_inter= 1:max_iterif mod(index_inter,50) ==0index_interendAct_trace=[];Train_trace = [];for j= 1:num_trainP = randperm(num_train);train_input = input{P(j)};train_output = output{P(j)};% Forward Propagation[Act] = FP(train_input,Act,W,B,num_layer);% Backward Propagation & Template update[W,B,Err] = BP(train_output,Act,W,B,num_layer,lrn_rate,Err);% Debug[row,col]=find(Act{end}==max(Act{end}));Act_trace(end+1)=col-1;[row,col]=find(train_output==max(train_output));Train_trace(end+1)=col-1;end All_arr(index_inter).act = Act_trace;All_arr(index_inter).train=Train_trace;All_arr(index_inter).err = Act_trace-Train_trace;
endtocsave
disp('Training Ends')if (strcmp(Mode,'XOR'))grid = [0:0.01:1];Z=-1*ones(length(grid),length(grid));for i=1:length(grid)for j=1:length(grid)test = [grid(i) grid(j)];Act_new = FP(test,Act,W,B,num_layer);Z(i,j) = Act_new{3};endend[X,Y] = meshgrid(grid);mesh(X,Y,Z)elseif (strcmp(Mode,'MNIST'))Guess_arr = [];for i=1:length(test_input)[guess_result] = FP(test_input{i},Act,W,B,num_layer);[row,col]=find(guess_result{end}==max(guess_result{end}));Guess_arr(end+1)=col-1;end
endZ=zeros(10,10);for i =1:length(test_labels)Z(Guess_arr(i)+1,test_labels(i)+1)=Z(Guess_arr(i)+1,test_labels(i)+1)+1;
endAbs_err = Guess_arr-double(test_labels)';
success_rate = sum(Guess_arr-double(test_labels)'==0)/1000
%plot(abs(Guess_arr-double(test_labels)'))%figure(1);scatter((Guess_arr*10),test_labels)
%figure(2);plot(err_trace);运行结果：
【源码下载地址】 https://download.csdn.net/download/FL1623863129/88600486
