基于Matlab实现的ANN、DNN网络对MNIST数据分类源码+MNIST数据集+项目说明.zip

function [Yn On Yt Ot] = ann2159347(ID, N_ep, lr, bp, u, v, w, cf) % Where, for the MNIST data sets, % - ID is your student ID (e.g. 2216789) % - N_ep the number of epochs, lr is the learning rate % - bp = 1, heuristic unscaled, 0 calculus based bp % - u, v, w the number of nodes on the three hidden layers % - cf = 1 use total squared error; 2 use cross-entropy % - Yn are the exact training labels/targets, Ot the training predictions % - Yt are the exact testing labels/targets, On the testing predictions % The MNIST CSV files are not to be altered in any way, and should be % in the same folder as this matlab code. % hidden layer=1 % ann2159347(2216789, 100, 0.008, bp, 0, 0, 0, cf); % ann2159347(2216789, 100, 0.008, 1, 0, 0, 0, 1); % ann2159347(2216789, 100, 0.008, 1, 0, 0, 0, 2); % ann2159347(2216789, 100, 0.008, 0, 0, 0, 0, 1); % ann2159347(2216789, 100, 0.008, 0, 0, 0, 0, 2); % hidden layer=3 % ann2159347(2216789, 100, 0.008, bp, 26, 28, 27, cf); % ann2159347(2216789, 100, 0.008, 1, 26, 28, 27, 1); % ann2159347(2216789, 100, 0.008, 1, 26, 28, 27, 2); % ann2159347(2216789, 100, 0.008, 0, 26, 28, 27, 1); % ann2159347(2216789, 100, 0.008, 0, 26, 28, 27, 2); close all %% set some useful defaults set(0,'DefaultLineLineWidth', 2); set(0,'DefaultLineMarkerSize', 10); % As a professional touch we should test the validity of our input if or(N_ep <= 0, lr <= 0) error('N_ep and/or lr are not valid') end if ~ismember(bp,[1,0]) error('back prop choice is not valid') end if ~ismember(cf,[1,2]) error('performance index choice is not valid') end %% obtain the raw training data A = readmatrix('mnist_train_1000.csv'); % convert and NORMALIZE it into training inputs and target outputs X_train = A(:,2:end)'/255; % beware - transpose, data is in columns! N_train = size(X_train,2); % size(X_train,1/2) gives number of rows/columns Y_train = zeros(10,N_train); % set up the one-hot encoding - note that we have to increment by 1 for i=1:N_train Y_train(1+A(i,1),i) = 1; end %% if hidden layer=3 then if u~=0&&v~=0&&w~=0 % default variables Ni = 784; % number of input nodes N1 = u; % number of nodes on first hidden layer N2 = v; % number of nodes on second hidden layer N3 = w; % number of nodes on third hidden layer No = 10; % number of output nodes % set up weights and biases W2 = 0.5-rand(Ni,N1); b2 = zeros(N1,1); W3 = 0.5-rand(N1, N2); b3 = zeros(N2,1); W4 = 0.5-rand(N2, N3); b4 = zeros(N3,1); W5 = 0.5-rand(N3, No); b5 = zeros(No,1); % set up a sigmoid activation function for layers 2 and 3 sig2 = @(x) 1./(1+exp(-x)); dsig2 = @(x) exp(-x)./((1+exp(-x)).^2); sig3 = @(x) 1./(1+exp(-x)); dsig3 = @(x) exp(-x)./((1+exp(-x)).^2); sig4 = @(x) 1./(1+exp(-x)); dsig4 = @(x) exp(-x)./((1+exp(-x)).^2); % output layer activation depends on cf if cf == 1 sig5 = @(x) 1./(1+exp(-x)); dsig5 = @(x) exp(-x)./((1+exp(-x)).^2); elseif cf == 2 sig5 = @(x) exp(x)/sum(exp(x)); else error('cf has improper value') end % we now train by looping N_ep times through the training set for epoch = 1:N_ep mixup = randperm(N_train); for j = 1:N_train i = mixup(j); % get X_train(:,i) as an input to the network a1 = X_train(:,i); % forward prop to the next layer, activate it, repeat n2 = W2'*a1 + b2; a2 = sig2(n2); n3 = W3'*a2 + b3; a3 = sig3(n3); n4 = W4'*a3 + b4; a4 = sig4(n4); n5 = W5'*a4 + b5; a5 = sig5(n5); % this is then the output y = a5; % calculate A, the diagonal matrices of activation derivatives A4 = diag(dsig4(n4)); A3 = diag(dsig3(n3)); A2 = diag(dsig2(n2)); % we calculate the error in this output, and get the S3 vector e5 = Y_train(:,i) - y; if cf == 1 A5 = diag(dsig5(n5)); S5 = -2*A5*e5; elseif cf == 2 S5 = -e5; end % back prop the error if bp == 1 e4 = W5*e5; e3 = W4*e4; e2 = W3*e3; % from these we compute the S vectors S4 = -2*A4*e4; S3 = -2*A3*e3; S2 = -2*A2*e2; elseif bp == 0 S4 = A4*W5*S5; S3 = A3*W4*S4; S2 = A2*W3*S3; end % and use a learning rate to update weights and biases W5 = W5 - lr * a4*S5'; b5 = b5 - lr * S5; W4 = W4 - lr * a3*S4'; b4 = b4 - lr * S4; W3 = W3 - lr * a2*S3'; b3 = b3 - lr * S3; W2 = W2 - lr * a1*S2'; b2 = b2 - lr * S2; end end % loop through the training set and evaluate accuracy of prediction wins = 0; y_pred = zeros(No,N_train); last_train=zeros(2,N_train); for i = 1:N_train y = sig3(W3'*sig2(W2'*X_train(:,i) + b2) + b3); y_pred(:,i) = sig5(W5'*sig4(W4'*y + b4) + b5); [~, indx1] = max(y_pred(:,i)); [~, indx2] = max(Y_train(:,i)); last_train(1,i)=indx1-1; last_train(2,i)=indx2-1; if indx1 == indx2; wins = wins+1; end end %% if hidden layer=1 then else % default variables Ni = 784; % number of input nodes Nh = 10; % number of hidden nodes No = 10; % number of output nodes % set up weights and biases W2 = 0.5-rand(Ni,Nh); b2 = zeros(Nh,1); W3 = 0.5-rand(Nh, No); b3 = zeros(No,1); % set up a sigmoid activation function for layers 2 and 3 sig2 = @(x) 1./(1+exp(-x)); dsig2 = @(x) exp(-x)./((1+exp(-x)).^2); if cf == 1 sig3 = @(x) 1./(1+exp(-x)); dsig3 = @(x) exp(-x)./((1+exp(-x)).^2); elseif cf == 2 sig3 = @(x) exp(x)/sum(exp(x)); else error('cf has improper value') end % we now train by looping N_ep times through the training set for epoch = 1:N_ep % Make the training order random randperm (n) return a row vector， % which contains a random arrangement of integers from 1 to n without repeating elements. mixup = randperm(N_train); for j = 1:N_train i = mixup(j); % get X_train(:,i) as an input to the network a1 = X_train(:,i); % forward prop to the next layer, activate it, repeat n2 = W2'*a1 + b2; a2 = sig2(n2); n3 = W3'*a2 + b3; a3 = sig3(n3); % this is then the output y = a3; % calculate A, the diagonal matrices of activation derivatives A2 = diag(dsig2(n2)); % we calculate the error in this output, and get the S3 vector e3 = Y_train(:,i) - y; if cf == 1 A3 = diag(dsig3(n3)); S3 = -2*A3*e3; elseif cf == 2 S3 = -e3; end % back prop the error if bp == 1; e2 = W3*e3; S2 = -2*A2*e2; elseif bp == 0; S2 = A2*W3*S3; end % and use a learning rate to update weights and biases W3 = W3 - lr * a2*S3'; b3 = b3 - lr * S3; W2 = W2 - lr * a1*S2'; b2 = b2 - lr * S2; end end % loop through the training set and evaluate accuracy of prediction wins = 0; y_pred = zeros(10,N_train);%y_pred Each column stores the likelihood of each number last_train=zeros(2,N_train); for i = 1:N_train y_pred(:,i) = sig3(W3'*sig2(W2'*X_train(:,i) + b2) + b3); [~, indx1] = max(y_pred(:,i)); [~, indx2] = max(Y_train(:,i)); last_train(1,i)=indx1-1;%last_train The first line stores the predicted numbers last_train(2,i)=indx2-1;%last_train The second line stores real numbers if indx1 == indx2 wins = wins+1; end end end fprintf('recognize number training set wins = %d/%d, %f%%\n',wins,