GMM.h

Go to the documentation of this file.

/*
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation; either version 3 of the License, or
 * (at your option) any later version.
 *
 * Written (W) 2011 Alesis Novik
 * Copyright (C) 2011 Berlin Institute of Technology and Max-Planck-Society
 */
#ifndef _GMM_H__
#define _GMM_H__

#include <shogun/lib/config.h>

#ifdef HAVE_LAPACK

#include <shogun/distributions/Distribution.h>
#include <shogun/distributions/Gaussian.h>
#include <shogun/lib/common.h>

namespace shogun
{
class CGMM : public CDistribution
{
    public:
        CGMM();
        CGMM(int32_t n, ECovType cov_type=FULL);
        CGMM(SGVector<CGaussian*> components, SGVector<float64_t> coefficients, bool copy=false);
        virtual ~CGMM();

        void cleanup();

        virtual bool train(CFeatures* data=NULL);

        float64_t train_em(float64_t min_cov=1e-9, int32_t max_iter=1000, float64_t min_change=1e-9);

        float64_t train_smem(int32_t max_iter=100, int32_t max_cand=5, float64_t min_cov=1e-9, int32_t max_em_iter=1000, float64_t min_change=1e-9);

        void max_likelihood(SGMatrix<float64_t> alpha, float64_t min_cov);

        virtual int32_t get_num_model_parameters();

        virtual float64_t get_log_model_parameter(int32_t num_param);

        virtual float64_t get_log_derivative(
            int32_t num_param, int32_t num_example);

        virtual float64_t get_log_likelihood_example(int32_t num_example);

        virtual float64_t get_likelihood_example(int32_t num_example)
        {
            return CMath::exp(get_log_likelihood_example(num_example));
        }

        virtual inline SGVector<float64_t> get_nth_mean(int32_t num)
        {
            ASSERT(num<m_components.vlen);
            return m_components.vector[num]->get_mean();
        }

        virtual inline void set_nth_mean(SGVector<float64_t> mean, int32_t num)
        {
            ASSERT(num<m_components.vlen);
            m_components.vector[num]->set_mean(mean);
        }

        virtual inline SGMatrix<float64_t> get_nth_cov(int32_t num)
        {
            ASSERT(num<m_components.vlen);
            return m_components.vector[num]->get_cov();
        }

        virtual inline void set_nth_cov(SGMatrix<float64_t> cov, int32_t num)
        {
            ASSERT(num<m_components.vlen);
            m_components.vector[num]->set_cov(cov);
        }

        virtual inline SGVector<float64_t> get_coef()
        {
            return m_coefficients;
        }

        virtual inline void set_coef(SGVector<float64_t> coefficients)
        {
            m_coefficients.destroy_vector();
            m_coefficients=coefficients;
        }

        virtual inline SGVector<CGaussian*> get_comp()
        {
            return m_components;
        }

        virtual inline void set_comp(SGVector<CGaussian*> components)
        {
            for (int32_t i=0; i<m_components.vlen; i++)
            {
                SG_UNREF(m_components.vector[i]);
            }

            m_components.destroy_vector();
            m_components=components;

            for (int32_t i=0; i<m_components.vlen; i++)
            {
                SG_REF(m_components.vector[i]);
            }
        }

        SGVector<float64_t> sample();

        SGVector<float64_t> cluster(SGVector<float64_t> point);

        inline virtual const char* get_name() const { return "GMM"; }

    private:
        SGMatrix<float64_t> alpha_init(SGMatrix<float64_t> init_means);

        void register_params();

        void partial_em(int32_t comp1, int32_t comp2, int32_t comp3, float64_t min_cov, int32_t max_em_iter, float64_t min_change);

    protected:
        SGVector<CGaussian*> m_components;
        SGVector<float64_t> m_coefficients;
};
}
#endif //HAVE_LAPACK
#endif //_GMM_H__