function w = dbnc(dsTrain, varargin)
prtrace(mfilename);
% Settings for the different training algorithms.
if(~exist('nnet','dir'))
error('Neural network toolbox not found')
end
cInput_types = {'gaussian', 'logsigm'};
%argument definition
p = inputParser; % Create instance of inputParser class.
p.addRequired('dsTrain', @(x)( isdataset(x)));
p.addParamValue('input_type', 'logsigm', @(x)( any(strcmpi(x,cInput_types))) );
p.addParamValue('num_hidden', 3, @(x)( isnumeric(x) && x > 0 ));
p.addParamValue('size_hidden', [], @(x)( (all(isnumeric(x)) && all(x > 0)) || isempty(x) ) );
p.addParamValue('Initialization_dataset', [], @(x)( isdataset(x) ) );
p.addParamValue('Validation_dataset', [], @(x)( isdataset(x) ) );
% Check arguments
try
p.parse( dsTrain, varargin{:} );
catch MyError
rethrow(MyError);
end
dsTrain = p.Results.dsTrain;
dsInitialization = p.Results.Initialization_dataset;
dsValidation = p.Results.Validation_dataset;
num_hidden = p.Results.num_hidden;
size_hidden = p.Results.size_hidden;
input_type = p.Results.input_type;
% Dont know why this variable uses a lot of bytes to store at disk.
clear p
%Check the datasets
islabtype(dsTrain,'crisp');
isvaldfile(dsTrain,1,2); % At least 1 object per class, 2 classes
dsTrain = testdatasize(dsTrain);
%force domain to be in [0-1]
w_scale_domain = scalem(dsTrain, 'domain');
dsTrain = dsTrain * w_scale_domain;
if( isempty(dsValidation) )
dsValidation = dsTrain;
else
dsValidation = testdatasize(dsValidation);
iscomdset(dsTrain, dsValidation); % Check whether training and tuning set match
dsValidation = dsValidation * w_scale_domain;
end
if( isempty(dsInitialization) )
dsInitialization = dsTrain;
else
dsInitialization = dsInitialization * w_scale_domain;
end
[train_m, train_k, train_c] = getsize(dsTrain);
[ train_labels, train_lablist] = getnlab(dsTrain);
if( isempty(size_hidden) )
% num_hidden layers each of train_k size by default
size_hidden = repmat(train_k, num_hidden, 1);
end
%train the first RBM on data
top = +dsInitialization;
%train all other RBM's on top of each other
for ii=1:num_hidden
[model{ii} , top ] = rbmBB(top, size_hidden(ii));
end
clear top
model{num_hidden}.Wc = 0.1*randn(train_c, size(model{num_hidden}.W,2) );
model{num_hidden}.cc = 0.1*randn(1, train_c);
model{num_hidden}.labels = 1:train_c;
%fine-tuning of the DBN
targets = zeros(train_m, train_c);
for ii = 1:train_c
targets(train_labels == ii, ii) = 1;
end
for epoch = 1:20
%%%%%%%%%%%%%%% PERFORM CONJUGATE GRADIENT WITH 3 LINESEARCHES %%%%%%%%%%%%%%%%%%%%%%%%%%%%%
max_iter=3;
if epoch<6 % First update top-level weights holding other weights fixed.
top = +dsTrain;
for ii=1:num_hidden
top = rbmVtoH(model{ii}, top);
end
VV = colvec([ model{num_hidden}.Wc'; model{num_hidden}.cc ]);
Dim = [ size_hidden(end); train_c];
XX = minimize(VV,'CG_CLASSIFY_INIT',max_iter,Dim,top,targets);
aux = reshape(XX, size_hidden(end)+1, train_c);
model{num_hidden}.Wc = aux(1:end-1,:)';
model{num_hidden}.cc = aux(end,:);
else
VV = [];
for ii=1:num_hidden
VV = [VV; colvec([model{ii}.W; model{ii}.b ]) ];
end
VV = [VV; colvec([model{ii}.Wc'; model{ii}.cc ])];
Dim = [ train_k; colvec(size_hidden); train_c];
XX = minimize(VV,'CG_CLASSIFY',max_iter,Dim, +dsTrain,targets);
xxx = 0;
for ii=1:num_hidden
aux = reshape(XX(xxx+1:xxx+(Dim(ii)+1)*Dim(ii+1)),Dim(ii)+1,Dim(ii+1));
model{ii}.W = aux(1:end-1,:);
model{ii}.b = aux(end,:);
xxx = xxx+(Dim(ii)+1)*Dim(ii+1);
end
aux = reshape(XX(xxx+1:xxx+(Dim(ii+1)+1)*Dim(ii+2)),Dim(ii+1)+1,Dim(ii+2));
model{ii}.Wc = aux(1:end-1,:)';
model{ii}.cc = aux(end,:);
end
end
% Create mapping.
w = w_scale_domain*mapping('dbnPredict', 'trained', model, train_lablist, train_k, train_c);
w = setname(w, 'DBN classifier');
% w = setcost(w,a);
return