trainnet.m

细胞生长结构可视化工具箱-MATLAB Toolbox1999.zip

% Version where classes are replicated to all have same size. Samples are then presented once each in a random order

% General purpose Growing Cell Structure Visualisation and Classification
%%
function TrainNet(GcsProj)

GcsProj.Gcs.Status = 'Training';
GcsProj.Gcs.StatusColour = [1 0 0];
UpdateGUI;

%seed=sum(100*clock);
%rand('seed',seed);
rand('seed',1)

vers;	% Store Matlab version 

% Set background colour
	whitebg('white');
   
   % If network is untrained then perform necessary setup - normalisation, splitting data into training and test sets.
if strcmp(GcsProj.Gcs.Trained,'No')   
   
	% Might want to move this to separate m file and offer as option	
	% Normalise data 
	if (GcsProj.Gcs.Normalise == 1) & isempty(GcsProj.Gcs.NormF1)	% normalisation selected and has not already been performed
		for i=1:GcsProj.Gcs.n 
			if std(GcsProj.dataAll(:,i)) ~=0	 % (~( max(train(:,i))==1 & min(train(:,i))==0 ) & (std(train(:,i)) ~=0) )	%
	         disp(['normalising column ',int2str(i)]);
	         eval(['GcsProj.Gcs.NormF' int2str(i) ' = std(GcsProj.dataAll(:,i));']);
	         eval(['GcsProj.dataAll(:,i) = GcsProj.dataAll(:,i)./GcsProj.Gcs.NormF' int2str(i) ';']);
	       %  eval(['GcsProj.test(:,i) = GcsProj.test(:,i)./GcsProj.Gcs.NormF' int2str(i) ';']);	% Both sets must be normalised by original calculated value
	         
			end
	   end
	end
   
   % Create list of indeces for samples prior to extracting test samples
   GcsProj.Gcs.index = [1:size(GcsProj.dataAll,1)];
	% Split data into training and test sets
	GcsProj.train = GcsProj.dataAll;
   GcsProj.test = [];
   testindex =[];
	for i=1:GcsProj.NoTest
	   index = floor(rand(1,1)*(size(GcsProj.train,1)-1)+1);	% Randomly select a sample
	   GcsProj.test = [GcsProj.test; GcsProj.train(index,1:GcsProj.Gcs.n+1)];	% copy sample to test set
      GcsProj.train(index,:) = [];		% Remove test sample fromm training set
      testindex = [testindex; GcsProj.Gcs.index(index)];	% Add sample No to list of indeces
      GcsProj.Gcs.index(index) = [];	% Remove index from index array
	end
	GcsProj.Gcs.index = testindex;		% Store test sample indeces in GCS structure

	if isempty(GcsProj.Gcs.w)
		% Set inital node weights to three random training samples
		Ip1=GcsProj.train(floor((rand(1,1)*(size(GcsProj.train,1)-1)+1)),1:GcsProj.Gcs.n);
		Ip2=GcsProj.train(floor((rand(1,1)*(size(GcsProj.train,1)-1)+1)),1:GcsProj.Gcs.n);
		Ip3=GcsProj.train(floor((rand(1,1)*(size(GcsProj.train,1)-1)+1)),1:GcsProj.Gcs.n);
		% Node weight array
  		GcsProj.Gcs.w = [Ip1;Ip2;Ip3;];
	end


	% Split data into different classes ready for equal prior training
	for cindex =1:GcsProj.Gcs.NoClasses
		eval(['GcsProj.data' int2str(cindex) ' =[];']);	% init matrix for current class
		eval(['i = find(GcsProj.train(:,GcsProj.Gcs.n+1)==' int2str(cindex) ');']);
		eval(['GcsProj.data' int2str(cindex) ' = GcsProj.train(i,:);']);
		eval(['GcsProj.Gcs.Prior' int2str(cindex) ' = size(GcsProj.data' int2str(cindex) ',1)./size(GcsProj.train,1);']);
	end

	if GcsProj.Gcs.EqualPriors
	   GcsProj=equalsze(GcsProj);
	end

	GcsProj.train=[];	% Complete data set
	for cindex=1:GcsProj.Gcs.NoClasses
		eval(['GcsProj.train = [GcsProj.train; GcsProj.data' int2str(cindex) '];']);
	end
   
   % Save Training Data so that it can be reloaded at a latter date
   trdata = GcsProj;
   trdata.Gcs=[];
   filebase = GcsProj.Gcs.Name;
   filename = [filebase,'.mat'];
   GcsProj.Gcs.Trainfilename =filename;	% Store name of file in GCS structure
	eval(['save ' filename ' trdata;']);
end
   
%eval(['save ' filebase '_train train test dataAll']);

GcsProj.lamda = GcsProj.epochspernode*size(GcsProj.train,1);

GcsProj.Gcs.cmap1 = [];	% Clear so that plot routines will correctly recalculate colourmaps after further training
% Train the network
train = GcsProj.train;	% Copy to working set from which used samples are removed
iter=0;	%current iteration / training sample no
hwait = waitbar(0,'Training ...');
	while (iter<=GcsProj.lamda*(GcsProj.NoNewNodes+1))	% +1 so that last node is added and trained before exiting
      iter=iter+1;	% update sample number
      figure(hwait);
      waitbar(iter/(GcsProj.lamda*(GcsProj.NoNewNodes+1)))
		if (rem(iter,GcsProj.lamda)==1)
			disp(['Training epoch ',int2str(ceil(iter/GcsProj.lamda)),' ...']);
		end
	% 1: Select input
		index = ceil(size(train,1)*rand(1)); %index = ceil(size(GcsProj.train,1)*rand(1));
		Ip = train(index,1:GcsProj.Gcs.n);	% Select next input
		train(index,:) = [];
		if isempty(train)	% If have used all samples, replenish data
			train = GcsProj.train;
		end
	% Use random samplling without removal with equal priors
	%	sampleclass = ceil(NoClasses*rand(1));
	%	eval(['sampleindex = ceil(size(data' int2str(cindex) ',1)*rand(1));']);
	%	eval(['Ip = data' int2str(sampleclass) '(sampleindex,1:n);']);

	% 2: Find winning node
		d=[];
		for i=1:size(GcsProj.Gcs.w,1)
		 	d(i) = norm(Ip-GcsProj.Gcs.w(i,:),GcsProj.Gcs.metric)^2;	% Find squared error
		end
		[minval,s] = min(d);

	% 3: Update error of winning node
		GcsProj.Gcs.E(s) = GcsProj.Gcs.E(s) + minval; % Sum squared error
	
	% 4: Adapt ref vectors
		GcsProj.Gcs.w(s,:) = GcsProj.Gcs.w(s,:) + GcsProj.Gcs.eb*(Ip-GcsProj.Gcs.w(s,:));	% Update winning node
	
		for i=1:size(GcsProj.Gcs.w,1)			% find neighbours and update them
			if GcsProj.Gcs.C(s,i) == 1
				GcsProj.Gcs.w(i,:) = GcsProj.Gcs.w(i,:) + GcsProj.Gcs.en*(Ip-GcsProj.Gcs.w(i,:));
			end
		end
	% Supervised learning	
		if rem(iter,GcsProj.lamda) == 0
		%  : Update the node counts	
			GcsProj.Gcs = count(GcsProj.Gcs,GcsProj.train);
		% Now form the bayes classifier and test
			GcsProj.Gcs = baycopt(GcsProj.Gcs,GcsProj.test);	% bayes classifier using optimal threshold	%bayclass;
         % Store performance results and plot history
         
         % Set Trained state to Yes
         GcsProj.Gcs.Trained = 'Yes';
         UpdateGUI;
		end

	% 6: Save data to file
		if rem(iter,GcsProj.lamda) == 0
			% Base filename
%			filebase = ['gcs'];
%         filename = [filebase,'_',int2str(size(GcsProj.Gcs.C,1))];
         SaveGcs(GcsProj.Gcs);
 %        eval(['save ' filename ' Gcs;']);	% Save only Gcs network - trainng set already saved, no not implemented yet!
 %        clear Gcs
			%if GcsProj.Gcs.NoClasses == 2
				% PLot ROC Curve
			%	figure(2);
				%areav = [areav, roc_parw(filename)];
				%eval(['save ' filename ]);	% redo to add areav to saved data
         %end
         
		end

	% 7: Insert new node if Necessary No of iterations have passed
   if rem(iter,GcsProj.lamda) == 0	
      if iter ~= GcsProj.lamda*(GcsProj.NoNewNodes+1)
         GcsProj.Gcs = AddNode(GcsProj.Gcs);
      end
   end

	% 8: Decrease error of all units
		%E = E*(1-beta);
		%Freq = (1-beta)*Freq;

end

msg = [ 'Run complete. The network was saved after each epoch in the files gcs_x,  ';
	'where x is the number of nodes in the network.                            ';
	'                                                                          ';
	'Use the m file factor to perform factor analysis on the current network.  ';
	'Use replot to redisplay all plots.                                        ';
	'Use plotparw to redisplay the frequency visualisation plot.               ';
	'Use post to redisplay the posterior probability estimate.                 ';
	'USe roc_parw to redisplay the receiver operating characterisitc curve.    '];

%disp(msg);
close(hwait)

GcsProj.Gcs.Status = 'Idle';
GcsProj.Gcs.StatusColour = 'default';
UpdateGUI;