/** 
 * @file GMM.h
 * @brief Gaussian Mixture Model based on   
  article{Calinon07SMC,
  title="On Learning, Representing and Generalizing a Task in a Humanoid 
  Robot",
  author="S. Calinon and F. Guenter and A. Billard",
  journal="IEEE Transactions on Systems, Man and Cybernetics, Part B. 
  Special issue on robot learning by observation, demonstration and 
  imitation",
  year="2007",
  volume="37",
  number="2",
  pages="286--298"
}
 * @author Björn Fröhlich
 * @date 05/14/2009

 */
#ifndef GMMINCLUDE
#define GMMINCLUDE

#include "core/vector/VectorT.h"
#include "core/vector/MatrixT.h"
 
#include "vislearning/cbaselib/MultiDataset.h"
#include "vislearning/cbaselib/LocalizationResult.h"
#include "vislearning/cbaselib/CachedExample.h"
#include "vislearning/cbaselib/Example.h"

#include "vislearning/math/cluster/ClusterAlgorithm.h"
#include "core/vector/VVector.h"

namespace OBJREC {

class GMM : public ClusterAlgorithm
{
	protected:
		
		//! number of gaussians
		int gaussians;
		
		//! dimension of each feature
		int dim;

		//! mean vectors
		NICE::VVector mu;
	
		//! sparse sigma vectors
		NICE::VVector sparse_sigma;
			
		//! sparse inverse sigma vectors (if usediag)
		NICE::VVector sparse_inv_sigma;
		
		//! save det_sigma for fast computing
		std::vector<double> log_det_sigma;
				
		//! the configfile
		const NICE::Config *conf;
		
		//! the weight for each gaussian
		std::vector<double> priors;
		
		//! parameters for other GM to compare
		NICE::VVector mu2;
		NICE::VVector sparse_sigma2;
		std::vector<double> priors2;
		bool comp;
		
		//! maximum number of iterations for EM
		int maxiter;
		
		//! how many features to use, use 0 for alle input features
		int featsperclass;
				
		//! for faster computing cdimval = dim*2*PI
		double cdimval;
		
		//! parameter for the map estimation
		double tau;
		
		//! whether to use pyramid initialisation or not
		bool pyramid;
	public:
		/**
	 	 * simplest constructor
		 */
		GMM();
		
		/**
		 * simple constructor
		 * @param _no_classes 
		 */
		GMM(int _no_classes);
		
		/**
		 * standard constructor
		 * @param conf a Configfile
		 * @param _no_classes number of gaussian
		 */
		GMM(const NICE::Config *conf, int _no_classes = -1);
		
		/**
		 * standard destructor
		 */
		~GMM(){std::cerr << "dadada" << std::endl;};
		
		/**
		 * computes the mixture
		 * @param examples the input features
		 */
		void computeMixture(Examples examples);
		
		/**
		 * computes the mixture
		 * @param DataSet the input features
		 */
		void computeMixture(const NICE::VVector &DataSet);
		
		/**
		 * returns the probabilities for each gaussian in a sparse vector
		 * @param vin input vector
		 * @param probs BoV output vector
		 */
		void getProbs(const NICE::Vector &vin, SparseVector &probs);
		
		/**
		 * returns the probabilities for each gaussian
		 * @param vin input vector
		 * @param probs BoV output vector
		 */
		void getProbs(const NICE::Vector &vin, NICE::Vector &probs);
		
		/**
		 * returns the fisher score for the gmm
		 * @param vin input vector
		 * @param probs Fisher score output vector
		 */
		void getFisher(const NICE::Vector &vin, SparseVector &probs);
		
		/**
		 *   init the GaussianMixture by selecting randomized mean vectors and using the coovariance of all features
		 * @param DataSet input Matrix
		 */
		void initEM(const NICE::VVector &DataSet);
				
		/**
		 *   alternative for initEM: init the GaussianMixture with a K-Means clustering
		 * @param DataSet input Matrix
		 */
		void initEMkMeans(const NICE::VVector &DataSet);
		
		/**
		 * performs Expectation Maximization on the Dataset, 	in order to obtain a nState GMM Dataset is a Matrix(no_classes,nDimensions)
		 * @param DataSet input Matrix
		 * @param gaussians number gaussians to use
		 * @return number of iterations
		 */
		int doEM(const NICE::VVector &DataSet, int nbgaussians);
			
		/**
		 * Compute log probabilty of vector v for the given state. *
		 * @param Vin 
		 * @param state 
		 * @return 
				 */
		double logpdfState(const NICE::Vector &Vin,int state);
		
		/**
		 * determine the best mixture for the input feature
		 * @param v input feature
		 * @param bprob probability of the best mixture
		 * @return numer of the best mixture
		 */
		int getBestClass(const NICE::Vector &v, double *bprob = NULL);
		
		/**
		 * Cluster a given Set of features and return the labels for each feature
		 * @param features input features
		 * @param prototypes mean of the best gaussian
		 * @param weights weight of the best gaussian
		 * @param assignment number of the best gaussian
		 */
		void cluster ( const NICE::VVector & features, NICE::VVector & prototypes, std::vector<double> & weights, std::vector<int> & assignment );
		
		/**
		 * save GMM data
		 * @param filename filename
		 */
		void saveData(const std::string filename);
		
        /**
		 * load GMM data
		 * @param filename filename
		 * @return true if everything works fine
	 	 */
		bool loadData(const std::string filename);
		
		/**
		 * return the parameter of the mixture
		 * @param mu 
		 * @param sSigma 
		 * @param p 
		 */
		void getParams(NICE::VVector &mean, NICE::VVector &sSigma, std::vector<double> &p);
		
		/**
		 * Set the parameters of an other mixture for comparing with this one
		 * @param mean mean vectors
		 * @param sSigma diagonal covariance Matrixs
		 * @param p weights
		 */
		void setCompareGM(NICE::VVector mean, NICE::VVector sSigma, std::vector<double> p);
		
		/**
		 * probability product kernel
		 * @param sigma1 
		 * @param sigma2 
		 * @param mu1 
		 * @param mu2 
		 * @param p 
		 * @return 
		 */
		double kPPK(NICE::Vector sigma1, NICE::Vector sigma2, NICE::Vector mu1, NICE::Vector mu2, double p);
		
		/**
		 * starts a comparison between this Mixture and a other one seted bei "setComparGM"
		 */
		double compare();
		
		/**
		 * whether to compare or not
		 * @param c 
		 */
		void comparing(bool c = true);
		
		int getSize(){return gaussians;}
};


} // namespace

#endif