wolffd@0: function [initState, transmat, mu, Nproto, pick] = mhmmParzen_train_observed(obsData, hiddenData, ...
wolffd@0: 						  nstates, maxNproto, varargin)
wolffd@0: % mhmmParzentrain_observed  with mixture of Gaussian outputs from fully observed sequences
wolffd@0: % function [initState, transmat, mu, Nproto] = mhmm_train_observed_parzen(obsData, hiddenData, ...
wolffd@0: %						  nstates, maxNproto)
wolffd@0: %
wolffd@0: %
wolffd@0: % INPUT
wolffd@0: % If all sequences have the same length
wolffd@0: % obsData(:,t,ex) 
wolffd@0: % hiddenData(ex,t)  - must be ROW vector if only one sequence
wolffd@0: % If sequences have different lengths, we use cell arrays
wolffd@0: % obsData{ex}(:,t) 
wolffd@0: % hiddenData{ex}(t)
wolffd@0: %
wolffd@0: % Optional argumnets
wolffd@0: % dirichletPriorWeight - for smoothing transition matrix counts
wolffd@0: % mkSymmetric
wolffd@0: %
wolffd@0: % Output
wolffd@0: % mu(:,q)
wolffd@0: % Nproto(q) is the number of prototypes (mixture components) chosen for state q
wolffd@0: 
wolffd@0: [transmat, initState] = transmat_train_observed(...
wolffd@0:     hiddenData, nstates, varargin{:});
wolffd@0: 
wolffd@0: % convert to obsData(:,t*nex)
wolffd@0: if ~iscell(obsData)
wolffd@0:   [D T Nex] = size(obsData);
wolffd@0:   obsData = reshape(obsData, D, T*Nex);
wolffd@0: else
wolffd@0:   obsData = cat(2, obsData{:});
wolffd@0:   hiddenData = cat(2, hiddenData{:});
wolffd@0: end
wolffd@0: [mu, Nproto, pick] = parzen_fit_select_unif(obsData, hiddenData(:), maxNproto);
wolffd@0: 
wolffd@0: