function [trans,field,lrec]=gentransnet(tokens,c,thresh)

% c = insensitivity factor
% thresh = cutoff for similarity

%% Create 'field'
start=10000*ones(size(tokens{1},1),1);    % 'start' state
stop=start*2;                             % 'stop' state
field=[];
for i=1:length(tokens)
   field=[field,tokens{i}];               % iteratively concatenate
   lrec(i)=size(tokens{i},2);
end
field=[start,field,stop];                 % circumfix start/stop states

F=size(field,2);
% v=var(field');v(v==0)=inf;              % this has to do with weighting
                                          % the features, ignore for now
% Create 'trans'
% trans=spalloc(F,F,3e7);                 % it would be nice to keep it
                                          % sparse to reduce memory, but
                                          % Matlab is bad at accessing
                                          % chunks of sparse matrices

r=min(size(tokens{1},1),13);

%% Initializations
trans=zeros(F,F);
starts=[0,cumsum(lrec(1:end-1))]+2;
stops=cumsum(lrec)+1;

%% Main loop
for i=1:length(tokens)
   for j=i:length(tokens)
%        A=rowkron(1./v',tokens{i});      % feature weighting, ignore
%        B=rowkron(1./v',tokens{j});
%       A=tokens{i}(1:13,:);B=tokens{j}(1:13,:);
      A=tokens{i}(1:r,:);B=tokens{j}(1:r,:);
      [r,cA]=size(A);[cB]=size(B,2);
      A=A(:,:,ones(cB,1));
      B=reshape(B,r,1,cB);B=B(:,ones(cA,1),:);
      s=exp(-c*sqrt(squeeze(sum((A-B).^2))));   % similarity
      s=(s>thresh).*s;                          % zero out if below threshold
      % patch the token matrices into a big quilt
      trans(starts(i):stops(i),starts(j)+1:stops(j)+1)=sparse(s);
      trans(starts(j):stops(j),starts(i)+1:stops(i)+1)=sparse(s');     
   end
   trans(:,starts(i))=([1;zeros(F-1,1)]);
   trans(stops(i),:)=([zeros(1,F-1),1]);
end
trans=sparse(trans);