NN testing code
in [Matlab]
|
Prev: sum rows by attributes
Next: Data points from line
From: Dalmar on 4 Jun 2010 05:39 I have written this NN to classify different cases. Simulation is excellent but I could not get the required target during the testing. Please let me know your suggestion. I am desperate to solve this problem due to submission of my postgraduate final year. Regards This is the Matlab code Inline Attachment Follows: matrixdatatraining2.m %% Neural Network Classification % training Matrix data load P0;load P1;load P2;load P3;load P4; % training data (Inputs) size(Pi=909x3) [r0 c0] = size(P0); [r1 c1] = size(P1); [r2 c2] = size(P2); [r3 c3] = size(P3); [r4 c4] = size(P4); t0=zeros(r0,1); t1=ones(r1,1)*-2; t2=ones(r2,1)*2 ;t3=ones(r3,1)*-5; t4=ones(r4,1)*5; %% Taking Input P=[P0; P1; P2; P3; P4]; T = [zeros(r0,1); ones(r1,1)*-2; ones(r2,1)*2 ;ones(r3,1)*-5; ones(r4,1)*5]; %training data(Tatgets) % %T=[t0, t1, t2, t3, t4]; % % T = [zeros(r0,c0,1); ones(r1,c1,1);ones(r2,c2,1)*2; ones(r3,c3,1)*3; ones(r4,c4,1)*4]; % % T=T(1,:); P = transpose(P); T = transpose(T); % [rP cP] = size(P); % 4x1058 % [rT cT] = size(T); % 1x1058 load P2a;%load P2a;load P3a;load P4a; % Testing data size(Pi=909x3) [Pn,minp,maxp,Tn,mint,maxt]=premnmx(P,T); [an,mina,maxa,sn,mins,maxs]=premnmx(P1a',t2');% normalization %% Creating and training of the Neural Network %We would like to create two layers neural net with 5 nodes in hidden %layer. %net = newff(minmax(P),[15 1],{'tansig','purelin'},'traingdm'); net = newff(minmax(Pn),[8,12,8,1],{'tansig','purelin','tansig', 'purelin'},'trainbr');%'trainbfg');%trainbr', trainlm); %net = newff(P,T,[20 1],{'tansig' 'purelin'},'traingdx'); %net = newff(minmax(P),[15 1],{'tansig' 'logsig'},'trainbr');%scg %net=newff(minmax(p),[15 1],{'tansig' 'purelin'},traingdm); net.trainParam.show = 100%inf; net.trainParam.goal = 1e-5; net.trainParam.lr = 0.01; net.trainparam.lr_inc = 1.05; net.trainparam.lr_dec = 0.7; net.trainparam.max_perfect_inc = 1.04; net.trainparam.mc = 0.9; net.trainparam.min_grad = 1e-10; net.trainParam.epochs = 500; [net,tr] = train(net,Pn,Tn); %% Save the Network save matrixdatatraining, net; %% simulatation %Y1 = sim(net,Pn); s=t1; y = sim(net,an); %un-ormalization %[a] = postmnmx(Y1,mint,maxt); %[a1] = postmnmx(y,mins,maxs); %a1=a1'; tt=postmnmx(y',mins,maxs); % if you like plot the output figure(1) plot(tt,'r') hold %Current plot held plot(t2) title('Comparison between actual targets and predictions') % [m,b,r] = postreg(a,T); %% Plotting the training %Here the network is simulated and its output plotted against % the targets. figure(2) plot(s') hold on plot(T', 'r') title('Comparison between actual targets and predictions') hold off xlabel('Epoches') % Here the network is trained for 50 epochs. Again the network's % output is plotted. % figure(2) % % net.trainParam.epochs = 50; % % net = train(net,P,T); % % Y = sim(net,P); % %plot(P,T,P,Y,'o') % plot(P,Y,'r') % %plot(p,t,'o',p,y1,'x',p,y2,'*') %% %% Testing the Network %load getmatrixdata Ptesti; load Ptest3 PP=Ptest3; PP = transpose(PP); load matrixdatatraining, net; ytest=sim(net,PP); figure(3) err=y-t2; plot(err) title('eroor between actual targets and predictions') xlabel('Epoches') figure(4) x=ytest(200:900); plot(x') h=mean(x) pi=mean(ytest) %% Trials % [net,tr]=train(net,Pn,Tn,P1,[],[],test); % plot(tr.epoch,tr.perf,tr.epoch,tr.vperf,tr.epoch,tr.tperf) % legend('Training','Validation','Test',-1); % ylabel('Squared Error'); xlabel('Epoch'); % % an = sim(net,ptrans); % a = poststd(an,meant,stdt); % for i=1:3 % figure(i) % [m(i),b(i),r(i)] = postreg(a(i,:),t(i,:)); % end % p = [-0.92 0.73 -0.47 0.74 0.29; -0.08 0.86 -0.67 -0.52 0.93]; % t = [-0.08 3.4 -0.82 0.69 3.1]; % [pn,minp,maxp,tn,mint,maxt] = premnmx(p,t); % net = newff(minmax(pn),[5 1],{'tansig' 'purelin'},'trainlm'); % net = train(net,pn,tn); % an = sim(net,pn); % [a] = postmnmx(an,mint,maxt); % [m,b,r] = postreg(a,t);
From: Dalmar on 4 Jun 2010 05:43 On 4 June, 10:39, Dalmar <ybul...(a)googlemail.com> wrote: > I have written this NN to classify different cases. Simulation is > excellent but I could not get the required target during the testing. > Please let me know your suggestion. I am desperate to solve this > problem due to submission of my postgraduate final year. > Regards > This is the Matlab code > Inline Attachment Follows: matrixdatatraining2.m > > %% Neural Network Classification > % training Matrix data > load P0;load P1;load P2;load P3;load P4; % training data (Inputs) > size(Pi=909x3) > [r0 c0] = size(P0); > [r1 c1] = size(P1); > [r2 c2] = size(P2); > [r3 c3] = size(P3); > [r4 c4] = size(P4); > t0=zeros(r0,1); t1=ones(r1,1)*-2; t2=ones(r2,1)*2 ;t3=ones(r3,1)*-5; > t4=ones(r4,1)*5; > > %% Taking Input > P=[P0; P1; P2; P3; P4]; > T = [zeros(r0,1); ones(r1,1)*-2; ones(r2,1)*2 ;ones(r3,1)*-5; > ones(r4,1)*5]; %training data(Tatgets) > > % %T=[t0, t1, t2, t3, t4]; > % % T = [zeros(r0,c0,1); ones(r1,c1,1);ones(r2,c2,1)*2; > ones(r3,c3,1)*3; ones(r4,c4,1)*4]; > % % T=T(1,:); > P = transpose(P); > T = transpose(T); > % [rP cP] = size(P); % 4x1058 > > % [rT cT] = size(T); % 1x1058 > load P2a;%load P2a;load P3a;load P4a; % Testing data size(Pi=909x3) > > [Pn,minp,maxp,Tn,mint,maxt]=premnmx(P,T); > [an,mina,maxa,sn,mins,maxs]=premnmx(P1a',t2');% normalization > > %% Creating and training of the Neural Network > %We would like to create two layers neural net with 5 nodes in hidden > %layer. > %net = newff(minmax(P),[15 1],{'tansig','purelin'},'traingdm'); > > net = newff(minmax(Pn),[8,12,8,1],{'tansig','purelin','tansig', > 'purelin'},'trainbr');%'trainbfg');%trainbr', trainlm); > %net = newff(P,T,[20 1],{'tansig' 'purelin'},'traingdx'); > %net = newff(minmax(P),[15 1],{'tansig' 'logsig'},'trainbr');%scg > %net=newff(minmax(p),[15 1],{'tansig' 'purelin'},traingdm); > > net.trainParam.show = 100%inf; > net.trainParam.goal = 1e-5; > net.trainParam.lr = 0.01; > net.trainparam.lr_inc = 1.05; > net.trainparam.lr_dec = 0.7; > net.trainparam.max_perfect_inc = 1.04; > net.trainparam.mc = 0.9; > net.trainparam.min_grad = 1e-10; > net.trainParam.epochs = 500; > > [net,tr] = train(net,Pn,Tn); > %% Save the Network > > save matrixdatatraining, net; > %% simulatation > %Y1 = sim(net,Pn); > s=t1; > y = sim(net,an); > %un-ormalization > %[a] = postmnmx(Y1,mint,maxt); > %[a1] = postmnmx(y,mins,maxs); > %a1=a1'; > tt=postmnmx(y',mins,maxs); > % if you like plot the output > figure(1) > plot(tt,'r') > hold > %Current plot held > plot(t2) > title('Comparison between actual targets and predictions') > > % [m,b,r] = postreg(a,T); > > %% Plotting the training > %Here the network is simulated and its output plotted against > % the targets. > figure(2) > > plot(s') > hold on > plot(T', 'r') > title('Comparison between actual targets and predictions') > hold off > xlabel('Epoches') > % Here the network is trained for 50 epochs. Again the network's > % output is plotted. > % figure(2) > % % net.trainParam.epochs = 50; > % % net = train(net,P,T); > % % Y = sim(net,P); > % %plot(P,T,P,Y,'o') > % plot(P,Y,'r') > % %plot(p,t,'o',p,y1,'x',p,y2,'*') > %% > %% Testing the Network > %load getmatrixdata Ptesti; > load Ptest3 > PP=Ptest3; > PP = transpose(PP); > load matrixdatatraining, net; > ytest=sim(net,PP); > > figure(3) > err=y-t2; > plot(err) > title('eroor between actual targets and predictions') > xlabel('Epoches') > > figure(4) > > x=ytest(200:900); > plot(x') > > h=mean(x) > pi=mean(ytest) Sorry the program is until here.. Dalmar
From: Greg Heath on 12 Jun 2010 17:19 On Jun 10, 4:42 pm, Dalmar <ybul...(a)googlemail.com> wrote: > On Jun 5, 10:15 am, Greg Heath <he...(a)alumni.brown.edu> wrote: > > On Jun 4, 5:43 am,Dalmar<ybul...(a)googlemail.com> wrote: > > > > I have written this NN to classify different cases. Simulation is > > > excellent but I could not get the required target during the testing. > > > Please let me know your suggestion. I am desperate to solve this > > > problem due to submission of my postgraduate final year. > > > Regards > > > This is the Matlab code -----SNIP > > I should thank you Mr Greg for your continuous support. > This is the NN code > clc, clear all, close all; > % classification neural network > % Taking Input data(Pi=909x3) > > load P0;load P1;load P2;load P3;load P4; > [r0 c0] = size(P0); > [r1 c1] = size(P1); > [r2 c2] = size(P2); > [r3 c3] = size(P3); > [r4 c4] = size(P4); > t0=zeros(r0,1); t1=ones(r1,1)*-2; t2=ones(r2,1)*2 ;t3=ones(r3,1)*-5; > t4=ones(r4,1)*5; > > P=[P0; P1; P2; P3; P4]; > %training data(Targets) {909 x 1} > T = [zeros(r0,1); ones(r1,1)*-2; ones(r2,1)*2 ;ones(r3,1)*-5; > ones(r4,1)*5]; > t0=zeros(r0,1); t1=ones(r1,1)*-2; t2=ones(r2,1)*2 ;t3=ones(r3,1)*-5; > t4=ones(r4,1)*5; > > P = transpose(P); > T = transpose(T); If this is a classification network for 5 classes, the output should be 5 dimensional with targets that are columns of the 5-dimensional unit matrix eye(5). Use LOGSIG as the output activation function. Inputs will be assigned to the class corresponding to the maximum output. Output values are estimates of class posterior probabilities conditional on the input P(i|x), i=1:5. > [I N] = size(P) > [O N] = size(T) size(P) = [3 4545] size(T) = [1 4545] with the current format size(T) = [5 4545] with the standard classification format. > %You should always design constant and linear models to understand > %the results before designing a neural net with a hidden layer. > > % Constant Model > > y00 = repmat(mean(T,2),1,N) ;% output > e00 = T-y00;% error > MSE00 = mse(e00); % mean-squared -error > > %Linear Model > W = T/[ones(1,N);P] % weights > y0 = W*[ones(1,N);P]; % output > e0 = T-y0; % error > MSE0 = mse(e0); % mean-square error > NMSE0 = MSE0/MSE00; > R20 = 1-NMSE0; % 0.13618 > > %R20 = 1 - MSE0/MSE00; % R-squared statistic > %'The Linear Model only accounts for ~ 13.6% of the output' > %'variance!' > %' ==> try a Neural Net' > > %Creating and training of the Neural Network > > rand('state',0) > Hmax = 10 How did you come up with Hmax = 10? Neq = prod(size(T)) = 5*4545 = 22725 % Equations Nw = (I+1)*H+(H+1)*O = O+(I+O+1)*H % Unknowns for Neq > ~ 10*Nw Hmax = floor( (Neq/10 -O)/(I+O+1) ) = floor( 2267.5/9 ) = 251 Meaning 1: Your training set is so large you can afford to have H = 251 hidden nodes without worrying about overtraining an overfit net. However, since the input is only 3-D, that would be a ridiculous choice. Meaning 2: H = 1:1:10 is probably not a good search strategy. You can easily afford to investigate a much larger net. I would probably start out by only using 1/3 of the training set with Neq > ~20*Nw. Then Hmax = floor(377.92/9) = 41 leading to a H = 1: 5 : 41 INITIAL search strategy. Once a narrower range is identified you can step with dH = 1. > ntrial = 0 > for H = 1:Hmax > for weighttrial = 1:10 > ntrial = ntrial + 1 > net = newff(minmax(P),[H 1]); % help newff for a classifier I recommend { 'tansig' 'logsig'} > net.trainParam.goal = MSE00/100; > net.trainParam.show = inf; > [net,tr] = train(net,P,T); % help trainlm > y = sim(net,P); > > e = T-y; > MSE(ntrial) = mse(e); > NMSE(ntrial) = MSE(ntrial)/MSE00; > R2(ntrial) = 1-NMSE(ntrial); > end > end > figure(1) > plot(y); hold on; plot(T); hold off; What is the significance of plotting the output of the100th trial ? ? Why not plot R2 vs ntrial and/or R2 vs H ??? > ylabel('simulation of the neural network') > xlabel('Epochs') No, not epochs. You are plotting output of the 100th trial vs input vector index... which is probably not very informative. Find the H and weight values that yield the best performance. Then plot those output performances vs input vector index. > [ R2max ntrialopt] = max(R2) > > rawsummary = [(1:ntrial)' R2']; Plot this! > sortedsummary = [ (1:ntrial)' -sortrows(-rawsummary,2)] Plot this! > %'53% of the nets are better than the Linear Model' > %'15% of the nets have R2 >~ 85% ' Which 15 ntrial and H values? Now you can use the ntrial values to determine the corresponding state of the RNG for the random initial weights and recreate those nets for the purpose of testing. However, if those H values are near Hmax = 10, I suggest, at least, exploring much higher values. > %% Creating and training of the Neural Network > k=k+1,figure(k), hold on > plot((1:ntrial),R2,'.') > plot(100*R20*ones(1,ntrial),'r') > xlabel('Trial') > ylabel('Per Cent of Output Variance Modeled by the NNET') > ylim([0 1]) > title('Neural Network Performance') > return > % Testing the Neural Network > Yntrn = sim(net,Pn); > MSEntrn = mse(Tn-Yntrn) > R2ntrn = 1- MSEntrn/MSETn00 > Yntst = sim(net,an); > MSEtst = mse(sn-Yntst) > R2ntst = 1- MSEntst/MSEsn00 1. Use the standard classification format using columns of eye(5) as training targets. 2. Investigate higher values of H 3. Find the best group of trained nets to recreate and test. You have only tested the last of 100. 4. The ultimate performance criterion for a classifier is the percent error rate. with a 5-dimensional output find the class assignment corresponding to the maximum then determine if it is correct or not. Report a. The percent error rate for each class b. The percent error rate for the 5-class mixture. Unfortunately, there is no good way to train a MLP using error rate as a performance measure. Hope this helps. Greg P.S. You probably shoold also calculate the per cent error rates for the constant and linear classifiers.
|
Pages: 1 Prev: sum rows by attributes Next: Data points from line |