FITCInferenceMethod.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.
 *
 * Copyright (C) 2012 Jacob Walker
 *
 *  * Code adapted from Gaussian Process Machine Learning Toolbox
 * http://www.gaussianprocess.org/gpml/code/matlab/doc/
 *
 */

#ifndef CFITCINFERENCEMETHOD_H_
#define CFITCINFERENCEMETHOD_H_

#include <shogun/lib/config.h>
#ifdef HAVE_EIGEN3
#ifdef HAVE_LAPACK

#include <shogun/regression/gp/InferenceMethod.h>

namespace shogun
{

class CFITCInferenceMethod: public CInferenceMethod
{

public:

    /*Default Constructor*/
    CFITCInferenceMethod();

    /* Constructor
     * @param kernel covariance function
     * @param features features to use in inference
     * @param mean mean function
     * @param labels labels of the features
     * @param model Likelihood model to use
     * @param latent features to use
     */
    CFITCInferenceMethod(CKernel* kernel, CFeatures* features,
            CMeanFunction* mean, CLabels* labels, CLikelihoodModel* model,
            CFeatures* latent_features);

    /*Destructor*/
    virtual ~CFITCInferenceMethod();

    virtual float64_t get_negative_marginal_likelihood();

    virtual CMap<TParameter*, SGVector<float64_t> > get_marginal_likelihood_derivatives(
            CMap<TParameter*, CSGObject*>& para_dict);

    virtual SGVector<float64_t> get_alpha();


    virtual SGMatrix<float64_t> get_cholesky();

    virtual SGVector<float64_t> get_diagonal_vector();

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

    /*Get the gradient
     *
     * @return Map of gradient. Keys are names of parameters, values are
     * values of derivative with respect to that parameter.
     */
    virtual CMap<TParameter*, SGVector<float64_t> > get_gradient(
            CMap<TParameter*, CSGObject*>& para_dict)
    {
        return get_marginal_likelihood_derivatives(para_dict);
    }

    /*Get the function value
     *
     * @return Vector that represents the function value
     */
    virtual SGVector<float64_t> get_quantity()
    {
        SGVector<float64_t> result(1);
        result[0] = get_negative_marginal_likelihood();
        return result;
    }

protected:
    virtual void update_alpha();
    virtual void update_chol();
    virtual void update_train_kernel();
    virtual void update_all();

private:

    void init();

private:

    void check_members();

    /*Kernel matrix with noise*/
    SGMatrix<float64_t> m_kern_with_noise;

    /*noise of the latent variables*/
    float64_t m_ind_noise;

    /*Cholesky of Covariance of
     * latent features
     */
    SGMatrix<float64_t> m_chol_uu;

    /*Cholesky of Covariance of
     * latent features
     * and training features
     */
    SGMatrix<float64_t> m_chol_utr;

    /* Covariance matrix of latent
     * features
     */
    SGMatrix<float64_t> m_kuu;

    /* Covariance matrix of latent
     * features and training features
     */
    SGMatrix<float64_t> m_ktru;

    /* Diagonal of Training
     * kernel matrix + noise
     * - diagonal of the matrix
     * (m_chol_uu^{-1}*m_ktru)*
     * (m_chol_uu^(-1)*m_ktru)'
     * = V*V'
     */
    SGVector<float64_t> m_dg;

    /*Labels adjusted for
     * noise and means
     */
    SGVector<float64_t> m_r;

    /* Solves the equation
     * V*r = m_chol_utr
     */
    SGVector<float64_t> m_be;



};

}
#endif // HAVE_EIGEN3
#endif // HAVE_LAPACK

#endif /* CFITCInferenceMethod_H_ */