基于DBN-ELM深度信念网络的在线序贯极限学习机的数据分类matlab仿真,包含仿真操作录像,代码注释

function [net]=DBN_OSELM(In,TsIn,Tar,TsTar,Options) % DBN-ELM : deep belief network based Online Sequential Extreme Learning Machine activF=Options.activF; % Activation function net_Architecture=Options.net_Architecture; % Network Architecture mini_batch=Options.mini_batch; % minibatch size % devide into mini-batches [Trinputs,Trtargets]=devide_blocks(In,Tar,mini_batch); % Start training % Initialization phase % ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ start_time_train=cputime; for i=1:length(net_Architecture) % load initial mini-batches if i==1 P = Trinputs{1}; else P=Hn{1}'; end % generate input weights input_weights=randn(net_Architecture(i),size(P,2)); % calculating the temporal hidden layer tempH=input_weights*P'; % Activation function switch lower(activF) case {'sig','sigmoid'} %%%%%%%% Sigmoid H = 1 ./ (1 + exp(-tempH)); case {'radbas'} %%%%%%%% Radial basis function H = radbas(tempH); case {'hardlim'} %%%%%%%% hardlim function H = hardlim(tempH); case {'hardlims'} %%%%%%%% hardlims function H = hardlims(tempH); case {'sin'} %%%%%%%% sin function H = sin(tempH); case {'relu'} %%%%%%%% sin function H = max(tempH,0); %%%%%% More activation functions can be added here end % unsupervised learning Hn{1}=H; % save the new features representations B=pinv(H') * P ; % calculate B of the encoder M= pinv(H * H'); % calculate the covariance matrix E{1}=P-(H'*B); % calculate reconstruction error % supervised learning if i==length(net_Architecture) T=Trtargets{1}; % load targets beta=pinv(H') * T ; % determine initial beta Esup{1}=T-(Hn{1}' * beta); % calculate prediction error end % end of initialization phase % ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ % start sequential Learning if numel(Trinputs)>1% for t=2:numel(Trinputs)-1 % load new mini-batches if i==1 Pnew = Trinputs{t}; else Pnew =Hn{t}'; end % temporal hidden layer; We no longer need the old input weighting matrix, we will use pinv (B ') instead. Hn{t}=pinv(B')*Pnew'; % no need to the use of activation function ; activation beta is already calculated after activation %% unsupervised learning E{t}=Pnew-(Hn{t}' * B); K =(M * Hn{t}) *((eye(size(Pnew,1))+(Hn{t})' * M * Hn{t})^(-1)); M = (M -( K * (Hn{t})' * M)); B = B + M * Hn{t} * (E{t}); %% supervised learning if i== length(net_Architecture) Tnew=Trtargets{t}; Esup{t}=Tnew-(Hn{t}' * beta); beta = beta + M * Hn{t} * (Esup{t}); end end end %save updated beta (unsupervised) weights{i}=pinv(B'); end end_time_train=cputime; Tr_Time=end_time_train-start_time_train; % ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ %%%%%% training accuracy % encoding input=In; for i=1:length(net_Architecture) input=(weights{i}*input')'; end mapping_TRAIN=input; ytr_hat= smooth(S(mapping_TRAIN*beta,min(Tar),max(Tar)));% Scaling + removing outliers trainingAccuracy=sqrt(mse(ytr_hat-Tar)); %~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ %%%%%% testing accuracy start_time_test=cputime; % encoding input=TsIn; for i=1:length(net_Architecture) input=(weights{i}*input')'; end mapping_TEST=input; yts_hat= smooth(S(mapping_TEST*beta,min(TsTar),max(TsTar))); % Scaling + removing outliers Ts_Time=cputime-start_time_test; testingAccuracy=sqrt(mse(yts_hat-TsTar)); %~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ % Save results net.Tr_Time=Tr_Time; net.Ts_Time=Ts_Time; net.Tr_acc=trainingAccuracy; net.Ts_acc=testingAccuracy; net.ytr_hat=ytr_hat; net.yts_hat=yts_hat; end