function [W1,W2]=wrescale(W1,W2,Uscale,Yscale,NN)
% WRESCALE
% --------
%         Rescale weights of a trained network so that the network
%         can work on unscaled data.
%
% Calling [W1,W2]=wrescale(W1,W2,Uscale,Yscale,NN) rescales the weights
% for networks with LINEAR OUTPUT UNITS. The function works for
% plain feed-forward networks as well as for input-output models of dynamic
% systems (i.e. NNAR(X), NNARMA(X) and NNOE type models). Notice that when
% the function is used on a pruned network it will reintrouce the biases removed
% during the pruning session.
%
%
% NN is left out for static models
%
% INPUTS:
%      W1     - Input-to-hidden of network trained on scaled data
%      W2     - Hidden-to-output weights
%      Uscale - Matrix containing sample mean and standard deviation for
%               each input. For time series an empty matrix, [], is passed. 
%      Yscale - Matrix containing mean and std's for each output.
%      NN     - Vector containing lag spaces (see nnarx, nnarmax, nnoe ..)
%
% OUTPUTS:
%      W1, W2 - Scaled weight matrices.
%
% NB! See function 'dscale' on how to scale the data before training.
%
%
% Programmed by Magnus Norgaard, IAU/IMM/EI Technical University of Denmark
% LastEditDate: July 19, 1995

[hidden,netinputs] = size(W1);
[inputs,dummy]  = size(Uscale);
[outputs,dummy] = size(Yscale);

% ----- Rescale input-to-hidden weights -----
% -- Static networks --
if nargin<5
  meanvec = Uscale(:,1);
  stdvec  = Uscale(:,2);
  
% -- Dynamic models --
else
  if (2+2*inputs) >= netinputs,
    error('You have fucked up in the call of function "wrscale"')
  end
  meanvec = Yscale(1)*ones(NN(1),1);
  stdvec  = Yscale(2)*ones(NN(1),1);
  if length(NN)==2+2*inputs,   % NARMAX model
    nc = NN(inputs+2);
    nab=sum(NN(1:inputs+1));
    if nab==netinputs-1,       % NARMAX model, linear noise filter
      nc = 0;
    end
  end
  for k=1:inputs,
    meanvec = [meanvec;Uscale(k,1)*ones(NN(k+1),1)];
    stdvec  = [stdvec;Uscale(k,2)*ones(NN(k+1),1)];
  end
  if nc>0,         % NARMAX models
    meanvec = [meanvec;zeros(nc,1)];
    stdvec  = [stdvec;Yscale(2)*ones(nc,1)];
  end
end

for k=1:netinputs-1,
  W1(:,k) = W1(:,k)/stdvec(k);
  W1(:,netinputs) = W1(:,netinputs) - W1(:,k)*meanvec(k);
end


% ----- Rescale hidden-to-output weights -----
meanvec = Yscale(:,1);
stdvec  = Yscale(:,2);
for k=1:outputs,
  W2(k,:) = W2(k,:)*stdvec(k);
  W2(k,hidden+1) = W2(k,hidden+1)+meanvec(k);
end