%NEURC Automatic neural network classifier
%
% W = NEURC(A,UNITS)
% W = A*NEURC([],UNITS)
% W = A*NEURC(UNITS)
%
% INPUT
% A Dataset
% UNITS Number of units
% Default: 0.2 x size smallest class in A.
%
% OUTPUT
% W Trained feed-forward neural network mapping
%
% DESCRIPTION
% Automatically trained feed-forward neural network classifier with UNITS
% units in a single hidden layer. Training, by LMNC, is stopped when the
% performance on an artificially generated tuning set of 1000 samples per
% class (based on k-nearest neighbour interpolation) does not improve anymore.
%
% NEURC always tries three random initialisations, with fixed random seeds,
% and returns the best result according to the tuning set. This is done in
% order to obtain a reproducable result.
%
% If UNITS is NaN it is optimised by REGOPTC. This may take a long
% computing time and is often not significantly better than the default.
%
% Uses the Mathworks' neural network toolbox.
%
% SEE ALSO (<a href="http://37steps.com/prtools">PRTools Guide</a>)
% MAPPINGS, DATASETS, LMNC, BPXNC, GENDATK, REGOPTC
% Copyright: R.P.W. Duin, r.p.w.duin@37steps.com
% Faculty EWI, Delft University of Technology
% P.O. Box 5031, 2600 GA Delft, The Netherlands
% $Id: neurc.m,v 1.9 2008/07/03 09:11:44 duin Exp $
function argout = neurc (varargin)
checktoolbox('nnet');
mapname = 'AutoNeuralNet';
argin = shiftargin(varargin,'integer');
argin = setdefaults(argin,[],[]);
if mapping_task(argin,'definition')
argout = define_mapping(argin,'untrained',mapname);
elseif mapping_task(argin,'training') % Train a mapping.
[a,units] = deal(argin{:});
n_attempts = 3; % Try three different random initialisations.
[m,k] = size(a);
if isempty(units)
cs = classsizes(a);
units = ceil(0.2*min(cs));
end
if isnan(units) % optimize complexity parameter: number of neurons
defs = {[]};
parmin_max = [1,30];
w = regoptc(a,mfilename,{units},defs,[1],parmin_max,testc([],'soft'),0);
return
end
islabtype(a,'crisp');
isvaldfile(a,1,2); % at least 1 object per class, 2 classes
a = testdatasize(a);
a = setprior(a,getprior(a,0));
% train a network.
% Reproducability: always use same seeds.
randstate = randreset(1); opt_err = inf; opt_mapping = [];
% Try a number of random initialisations.
s = sprintf('%i neural network initializations: ',n_attempts);
prwaitbar(n_attempts,s);
for attempt = 1:n_attempts
prwaitbar(n_attempts,attempt,[s int2str(attempt)]);
prwarning(4,'training with initialisation %d of %d',attempt,n_attempts);
t = gendatk(a,1000,2,1); % Create tuning set based on training set.
w = lmnc(a,units,inf,[],t); % Find LMNC mapping.
e = t*w*testc; % Calculate classification error.
if (e < opt_err)
% If this is the best of the three repetitions, store it.
opt_mapping = w; opt_err = e;
end
end
prwaitbar(0);
randreset(randstate); % return original state
% Output is best network found.
argout = setname(opt_mapping,mapname);
else % Evaluation
[a,w] = deal(argin{1:2});
nodatafile(a);
data = getdata(w);
if (length(data) > 1)
% "Old" neural network - network is second parameter: unpack.
data = getdata(w); weights = data{1};
pars = data{2}; numlayers = length(pars);
output = a; % Output of first layer: dataset.
for j = 1:numlayers-1
% Number of inputs (n_in) and outputs (n_out) of neurons in layer J.
n_in = pars(j); n_out = pars(j+1);
% Calculate output of layer J+1. Note that WEIGHTS contains both
% weights (multiplied by previous layer's OUTPUT) and biases
% (multiplied by ONES).
this_weights = reshape(weights(1:(n_in+1)*n_out),n_in+1,n_out);
output = sigm([output,ones(m,1)]*this_weights);
% Remove weights of this layer.
weights(1:(n_in+1)*n_out) = [];
end
else
% "New" neural network: unpack and simulate using the toolbox.
net = data{1};
output = sim(net,+a')';
end;
% 2-class case, therefore 1 output: 2nd output is 1-1st output.
if (size(output,2) == 1)
output = [output (1-output)];
end
% Output is mapped dataset.
argout = setdat(a,output,w);
end
return