KLInferenceMethod.h

Go to the documentation of this file.

/*
 * Copyright (c) The Shogun Machine Learning Toolbox
 * Written (w) 2014 Wu Lin
 * All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions are met:
 *
 * 1. Redistributions of source code must retain the above copyright notice, this
 *    list of conditions and the following disclaimer.
 * 2. Redistributions in binary form must reproduce the above copyright notice,
 *    this list of conditions and the following disclaimer in the documentation
 *    and/or other materials provided with the distribution.
 *
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
 * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
 * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
 * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *
 * The views and conclusions contained in the software and documentation are those
 * of the authors and should not be interpreted as representing official policies,
 * either expressed or implied, of the Shogun Development Team.
 *
 * Code adapted from
 * http://hannes.nickisch.org/code/approxXX.tar.gz
 * and Gaussian Process Machine Learning Toolbox
 * http://www.gaussianprocess.org/gpml/code/matlab/doc/
 * and the reference paper is
 * Nickisch, Hannes, and Carl Edward Rasmussen.
 * "Approximations for Binary Gaussian Process Classification."
 * Journal of Machine Learning Research 9.10 (2008).
 *
 */

#ifndef _KLINFERENCEMETHOD_H_
#define _KLINFERENCEMETHOD_H_

#include <shogun/lib/config.h>

#ifdef HAVE_EIGEN3
#include <shogun/machine/gp/InferenceMethod.h>
#include <shogun/optimization/lbfgs/lbfgs.h>
#include <shogun/machine/gp/VariationalGaussianLikelihood.h>

namespace Eigen
{
    template <class, int, int, int, int, int> class Matrix;
    template <class, int> class LDLT;

    typedef Matrix<float64_t,-1,-1,0,-1,-1> MatrixXd;
}

namespace shogun
{

class CKLInferenceMethod: public CInferenceMethod
{
public:
    CKLInferenceMethod();

    CKLInferenceMethod(CKernel* kernel, CFeatures* features,
            CMeanFunction* mean, CLabels* labels, CLikelihoodModel* model);

    virtual ~CKLInferenceMethod();

    virtual EInferenceType get_inference_type() const { return INF_KL; }

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

    virtual float64_t get_negative_log_marginal_likelihood();

    virtual SGVector<float64_t> get_posterior_mean();

    virtual SGMatrix<float64_t> get_posterior_covariance();

    virtual bool supports_regression() const
    {
        check_members();
        return m_model->supports_regression();
    }

    virtual bool supports_binary() const
    {
        check_members();
        return m_model->supports_binary();
    }

    virtual void set_model(CLikelihoodModel* mod);

    virtual void update();

    /* set L-BFGS parameters
     * For details please see shogun/optimization/lbfgs/lbfgs.h
     * @param m The number of corrections to approximate the inverse hessian matrix.
     * Default value is 100.
     * @param max_linesearch The maximum number of trials to do line search for each L-BFGS update.
     * Default value is 1000.
     * @param linesearch The line search algorithm.
     * Default value is using the backtracking with the strong Wolfe condition line search
     * @param max_iterations The maximum number of iterations for L-BFGS update.
     * Default value is 1000.
     * @param delta Delta for convergence test based on the change of function value.
     * Default value is 0.
     * @param past Distance for delta-based convergence test.
     * Default value is 0.
     * @param epsilon Epsilon for convergence test based on the change of gradient.
     * Default value is 1e-5
     * @param min_step The minimum step of the line search.
     * The default value is 1e-20
     * @param max_step The maximum step of the line search.
     * The default value is 1e+20
     * @param ftol A parameter used in Armijo condition.
     * Default value is 1e-4
     * @param wolfe A parameter used in curvature condition.
     * Default value is 0.9
     * @param gtol A parameter used in Morethuente linesearch to control the accuracy.
     * Default value is 0.9
     * @param xtol The machine precision for floating-point values.
     * Default value is 1e-16.
     * @param orthantwise_c Coeefficient for the L1 norm of variables.
     * This parameter should be set to zero for standard minimization problems.
     * Setting this parameter to a positive value activates
     * Orthant-Wise Limited-memory Quasi-Newton (OWL-QN) method. Default value is 0.
     * @param orthantwise_start Start index for computing L1 norm of the variables.
     * This parameter is valid only for OWL-QN method. Default value is 0.
     * @param orthantwise_end End index for computing L1 norm of the variables.
     * Default value is 1.
     */
    virtual void set_lbfgs_parameters(int m = 100,
            int max_linesearch = 1000,
            int linesearch = LBFGS_LINESEARCH_BACKTRACKING_STRONG_WOLFE,
            int max_iterations = 1000,
            float64_t delta = 0.0,
            int past = 0,
            float64_t epsilon = 1e-5,
            float64_t min_step = 1e-20,
            float64_t max_step = 1e+20,
            float64_t ftol = 1e-4,
            float64_t wolfe = 0.9,
            float64_t gtol = 0.9,
            float64_t xtol = 1e-16,
            float64_t orthantwise_c = 0.0,
            int orthantwise_start = 0,
            int orthantwise_end = 1);

    virtual SGMatrix<float64_t> get_cholesky();

    virtual void set_noise_factor(float64_t noise_factor);

    virtual void set_max_attempt(index_t max_attempt);

    virtual void set_exp_factor(float64_t exp_factor);

    virtual void set_min_coeff_kernel(float64_t min_coeff_kernel);
protected:

    virtual void compute_gradient();

    float64_t m_min_coeff_kernel;

    float64_t m_noise_factor;

    float64_t m_exp_factor;

    index_t m_max_attempt;

    virtual void update_init();

    virtual Eigen::LDLT<Eigen::MatrixXd,0x1> update_init_helper();

    virtual CVariationalGaussianLikelihood* get_variational_likelihood() const;

    virtual void check_variational_likelihood(CLikelihoodModel* mod) const;

    virtual void update_approx_cov()=0;

    virtual float64_t get_derivative_related_cov(SGMatrix<float64_t> dK)=0;

    virtual float64_t lbfgs_optimization();

    virtual SGVector<float64_t> get_derivative_wrt_inference_method(
            const TParameter* param);

    virtual SGVector<float64_t> get_derivative_wrt_likelihood_model(
            const TParameter* param);

    virtual SGVector<float64_t> get_derivative_wrt_kernel(
            const TParameter* param);

    virtual SGVector<float64_t> get_derivative_wrt_mean(
            const TParameter* param);

    virtual float64_t get_negative_log_marginal_likelihood_helper()=0;

    virtual float64_t get_nlml_wrt_parameters();

    virtual void get_gradient_of_nlml_wrt_parameters(SGVector<float64_t> gradient)=0;

    virtual bool lbfgs_precompute()=0;

    SGVector<float64_t> m_mu;

    SGMatrix<float64_t> m_Sigma;

    SGVector<float64_t> m_s2;

    /* The number of corrections to approximate the inverse hessian matrix.*/
    int m_m;

    /* The maximum number of trials to do line search for each L-BFGS update.*/
    int m_max_linesearch;

    /* The line search algorithm.*/
    int m_linesearch;

    /* The maximum number of iterations for L-BFGS update.*/
    int m_max_iterations;

    /* Delta for convergence test based on the change of function value.*/
    float64_t m_delta;

    /* Distance for delta-based convergence test.*/
    int m_past;

    /* Epsilon for convergence test based on the change of gradient.*/
    float64_t m_epsilon;

    /* The minimum step of the line search.*/
    float64_t m_min_step;

    /* The maximum step of the line search.*/
    float64_t m_max_step;

    /* A parameter used in Armijo condition.*/
    float64_t m_ftol;

    /* A parameter used in curvature condition.*/
    float64_t m_wolfe;

    /* A parameter used in Morethuente linesearch to control the accuracy.*/
    float64_t m_gtol;

    /* The machine precision for floating-point values.*/
    float64_t m_xtol;

    /* Coeefficient for the L1 norm of variables.*/
    float64_t m_orthantwise_c;

    /* Start index for computing L1 norm of the variables.*/
    int m_orthantwise_start;

    /* End index for computing L1 norm of the variables.*/
    int m_orthantwise_end;

private:
    void init();

    static float64_t evaluate(void *obj,
            const float64_t *parameters,
            float64_t *gradient,
            const int dim,
            const float64_t step);

};
}
#endif /* HAVE_EIGEN3 */
#endif /* _KLINFERENCEMETHOD_H_ */