www.gusucode.com > 有监督的 CNN 网络完成对MNIST 数字的识别 > 有监督的 CNN 网络完成对MNIST 数字的识别/CNN—卷积神经网络数字识别/@cnn/cutrain.m
function [out_cnet] = cutrain(cnet,Ip,labels,I_testp, labels_test)
%TRAIN train convolutional neural network using stochastic Levenberg-Marquardt
%
% Syntax
%
% [cnet, perf_plot] = train(cnet,Ip,labels,I_testp, labtst)
%
% Description
% Input:
% cnet - Convolutional neural network class object
% Ip - cell array, containing preprocessed images of handwriten digits
% labels - cell array of labels, corresponding to images
% I_testp - cell array, containing preprocessed images of handwriten
% digits of test set
% labtst - cell array of labels, corresponding to images of test set
% Output:
% cnet - trained convolutional neural network
% perf_plot - performance data
%Initialize GUI
h_gui = cnn_gui();
%Progress bars
h_HessPatch = findobj(h_gui,'Tag','HessPatch');
h_HessEdit = findobj(h_gui,'Tag','HessPrEdit');
h_TrainPatch = findobj(h_gui,'Tag','TrainPatch');
h_TrainEdit = findobj(h_gui,'Tag','TrainPrEdit');
%Axes
h_MCRaxes = findobj(h_gui,'Tag','MCRaxes');
h_RMSEaxes = findobj(h_gui,'Tag','RMSEaxes');
%Info textboxes
h_EpEdit = findobj(h_gui,'Tag','EpEdit');
h_ItEdit = findobj(h_gui,'Tag','ItEdit');
h_RMSEedit = findobj(h_gui,'Tag','RMSEedit');
h_MCRedit = findobj(h_gui,'Tag','MCRedit');
h_TetaEdit = findobj(h_gui,'Tag','TetaEdit');
%Buttons
h_AbortButton = findobj(h_gui,'Tag','AbortButton');
tic; %Fix the start time
perf_plot = []; %Array for storing performance data
%Init the cuda cnn
singnet = cnn2singlestruct(cnet);
cudacnn('init',singnet);
%Initial MCR calculation
mcr(1)=cucalcMCR(I_testp, labels_test, 1:50);
plot(h_MCRaxes,mcr);
SetText(h_MCRedit,mcr(end));
numPats = length(Ip);
if(cnet.HcalcMode == 1)
for i=1:cnet.HrecalcSamplesNum
%Setting the right output to 1, others to -1
d = -ones(1,10);
d(labels(i)+1) = 1;
%Simulating
[out, cnet] = sim(cnet,Ip{i});
%Calculate the error
e = out-d;
%Calculate Jacobian times error, or in other words calculate
%gradient
[cnet,je] = calcje(cnet,e);
[cnet,hx] = calchx(cnet);
% jj = jj+diag(sparse(hx));
SetHessianProgress(h_HessPatch,h_HessEdit,i/cnet.HrecalcSamplesNum);
end
% %Averaging
% jj = jj/cnet.HrecalcSamplesNum;
end
%For all epochs
for t=1:cnet.epochs
SetText(h_EpEdit,t);
SetTextHP(h_TetaEdit,cnet.teta);
%For all patterns
for n=1:numPats
%Setting the right output to 1, others to -1
d = -ones(1,10);
d(labels(n)+1) = 1;
%Simulating
out = cudacnn('sim',single(Ip{n}))';
%Calculate the error
e = out-d;
%Calculate Jacobian times error, or in other words calculate
%gradient
cudacnn('adapt',single(e));
%Calculate Hessian diagonal approximation
% if(cnet.HcalcMode == 0)
% [cnet,hx] = calchx(cnet);
% %Calculate the running estimate of Hessian diagonal approximation
% jj = gamma*diag(sparse(hx))+sparse((1-gamma)*jj);
% else
% if(mod(t*numPats+n,cnet.Hrecalc)==0) %If it is time to recalculate Hessian
% if(n+cnet.HrecalcSamplesNum>numPats)
% stInd = numPats-cnet.HrecalcSamplesNum;
% else
% stInd = n;
% end
% for i=stInd:stInd+cnet.HrecalcSamplesNum
% %Setting the right output to 1, others to -1
% d = -ones(1,10);
% d(labels(i)+1) = 1;
% %Simulating
% [out, cnet] = sim(cnet,Ip{i});
% %Calculate the error
% e = out-d;
% %Calculate Jacobian times error, or in other words calculate
% %gradient
% [cnet,je] = calcje(cnet,e);
% [cnet,hx] = calchx(cnet);
% jj = jj+diag(sparse(hx));
%
% SetHessianProgress(h_HessPatch,h_HessEdit,(i-stInd)/cnet.HrecalcSamplesNum);
% end
% %Averaging
% jj = jj/cnet.HrecalcSamplesNum;
% end
% end
%The following is usefull for debugging.
%===========DEBUG
% tmp(1)=check_finit_dif(cnet,1,Ip{n},d,1);
%===========DEBUG
perf(n) = mse(double(e)); %Store the error
%Uncoment this if you want a gradient descent
%dW = cnet.teta(t)*je;
%Actually Levenberg-Marquardt
%dW = (jj+cnet.mu*ii)\(cnet.teta*je);
%Apply calculated weight updates
%cnet = adapt_dw(cnet,dW);
%Plot mean of performance for every 10 patterns
% if(n>500)
if(n>1)
if(~mod(n-1,400))
mcr = [mcr cucalcMCR(I_testp, labels_test, 1:200)];
plot(h_MCRaxes,mcr);
SetText(h_MCRedit,mcr(end));
end
if(~mod(n-1,10))
perf_plot = [perf_plot,mean(sqrt(perf(n-10:n)))];
plot(h_RMSEaxes,perf_plot);
SetText(h_RMSEedit,perf_plot(end));
end
end
% end;
%It looks smoother, but not suit to large datasets
% perf_plot = [perf_plot,mean(perf)];
% plot(perf_plot);
% grid on;
SetTrainingProgress(h_TrainPatch,h_TrainEdit,(n+(t-1)*numPats)/(numPats*cnet.epochs));
SetText(h_ItEdit,n);
drawnow;
if(~isempty(get(h_AbortButton,'UserData')))
fprintf('Training aborted \n');
out_cnet = singlestruct2cnn(cudacnn('save',singnet));
return;
end
end
cnet.teta = cnet.teta*cnet.teta_dec;
cudacnn('set_data',0,'teta',single(cnet.teta));
end
%display('Train is finished');
%display('Time was ');
toc
%display('Perfomance is ');
%mean(perf)
%Save new weights back to cnet
out_cnet = singlestruct2cnn(cudacnn('save',singnet));
%Sets Hessian progress
%hp - handle of patch
%hs - handle of editbox
%pr - value from 0 to 1
function SetHessianProgress(hp,hs,pr)
xpatch = [0 pr*100 pr*100 0];
set(hp,'XData',xpatch);
set(hs,'String',[num2str(pr*100,'%5.2f'),'%']);
drawnow;
%Sets Training progress
%hp - handle of patch
%hs - handle of editbox
%pr - value from 0 to 1
function SetTrainingProgress(hp,hs,pr)
xpatch = [0 pr*100 pr*100 0];
set(hp,'XData',xpatch);
set(hs,'String',[num2str(pr*100,'%5.2f'),'%']);
%Set numeric text in the specified edit box
%hs - handle of textbox
%num - number to convert and set
function SetText(hs,num)
set(hs,'String',num2str(num,'%5.2f'));
%Set numeric text in the specified edit box with high preceition
%hs - handle of textbox
%num - number to convert and set
function SetTextHP(hs,num)
set(hs,'String',num2str(num,'%5.3e'));