GradientModelSelection.cpp

浏览该文件的文档.

/*
 * 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) 2013 Roman Votyakov
 * Copyright (C) 2012 Jacob Walker
 */

#include <shogun/modelselection/GradientModelSelection.h>

#ifdef HAVE_NLOPT

#include <shogun/evaluation/GradientResult.h>
#include <shogun/modelselection/ParameterCombination.h>
#include <shogun/modelselection/ModelSelectionParameters.h>
#include <shogun/machine/Machine.h>
#include <nlopt.h>

using namespace shogun;

#ifndef DOXYGEN_SHOULD_SKIP_THIS

struct nlopt_params
{
    CMachineEvaluation* machine_eval;

    CParameterCombination* current_combination;

    CMap<TParameter*, CSGObject*>* parameter_dictionary;

    bool print_state;
};

double nlopt_function(unsigned n, const double* x, double* grad, void* func_data)
{
    nlopt_params* params=(nlopt_params*)func_data;

    CMachineEvaluation* machine_eval=params->machine_eval;
    CParameterCombination* current_combination=params->current_combination;
    CMap<TParameter*, CSGObject*>* parameter_dictionary=params->parameter_dictionary;
    bool print_state=params->print_state;

    index_t offset=0;

    // set parameters from vector x
    for (index_t i=0; i<parameter_dictionary->get_num_elements(); i++)
    {
        CMapNode<TParameter*, CSGObject*>* node=parameter_dictionary->get_node_ptr(i);

        TParameter* param=node->key;
        CSGObject* parent=node->data;

        if (param->m_datatype.m_ctype==CT_VECTOR ||
                param->m_datatype.m_ctype==CT_SGVECTOR ||
                param->m_datatype.m_ctype==CT_SGMATRIX ||
                param->m_datatype.m_ctype==CT_MATRIX)
        {

            for (index_t j=0; j<param->m_datatype.get_num_elements(); j++)
            {

                bool result=current_combination->set_parameter(param->m_name,
                        (float64_t)x[offset++], parent, j);
                 REQUIRE(result, "Parameter %s not found in combination tree\n",
                         param->m_name)
            }
        }
        else
        {
            bool result=current_combination->set_parameter(param->m_name,
                    (float64_t)x[offset++], parent);
            REQUIRE(result, "Parameter %s not found in combination tree\n",
                    param->m_name)
        }
    }

    // apply current combination to the machine
    CMachine* machine=machine_eval->get_machine();
    current_combination->apply_to_machine(machine);
    if (print_state)
    {
        SG_SPRINT("Current combination\n");
        current_combination->print_tree();
    }
    SG_UNREF(machine);

    // evaluate the machine
    CEvaluationResult* evaluation_result=machine_eval->evaluate();
    CGradientResult* gradient_result=CGradientResult::obtain_from_generic(
            evaluation_result);
    SG_UNREF(evaluation_result);

    if (print_state)
    {
        SG_SPRINT("Current result\n");
        gradient_result->print_result();
    }

    // get value of the function, gradients and parameter dictionary
    SGVector<float64_t> value=gradient_result->get_value();
    CMap<TParameter*, SGVector<float64_t> >* gradient=gradient_result->get_gradient();
    CMap<TParameter*, CSGObject*>* gradient_dictionary=
        gradient_result->get_paramter_dictionary();
    SG_UNREF(gradient_result);

    offset=0;

    // set derivative for each parameter from parameter dictionary
    for (index_t i=0; i<parameter_dictionary->get_num_elements(); i++)
    {
        CMapNode<TParameter*, CSGObject*>* node=parameter_dictionary->get_node_ptr(i);

        SGVector<float64_t> derivative;

        for (index_t j=0; j<gradient_dictionary->get_num_elements(); j++)
        {
            CMapNode<TParameter*, CSGObject*>* gradient_node=
                gradient_dictionary->get_node_ptr(j);

            if (gradient_node->data==node->data &&
                    !strcmp(gradient_node->key->m_name, node->key->m_name))
            {
                derivative=gradient->get_element(gradient_node->key);
            }
        }

        REQUIRE(derivative.vlen, "Can't find gradient wrt %s parameter!\n",
                node->key->m_name);

        memcpy(grad+offset, derivative.vector, sizeof(double)*derivative.vlen);

        offset+=derivative.vlen;
    }

    SG_UNREF(gradient);
    SG_UNREF(gradient_dictionary);

    return (double)(SGVector<float64_t>::sum(value));
}

#endif /* DOXYGEN_SHOULD_SKIP_THIS */

CGradientModelSelection::CGradientModelSelection() : CModelSelection()
{
    init();
}

CGradientModelSelection::CGradientModelSelection(CMachineEvaluation* machine_eval,
        CModelSelectionParameters* model_parameters)
        : CModelSelection(machine_eval, model_parameters)
{
    init();
}

CGradientModelSelection::~CGradientModelSelection()
{
}

void CGradientModelSelection::init()
{
    m_max_evaluations=1000;
    m_grad_tolerance=1e-6;

    SG_ADD(&m_grad_tolerance, "gradient_tolerance", "Gradient tolerance",
            MS_NOT_AVAILABLE);
    SG_ADD(&m_max_evaluations, "max_evaluations", "Maximum number of evaluations",
            MS_NOT_AVAILABLE);
}

CParameterCombination* CGradientModelSelection::select_model(bool print_state)
{
    if (!m_model_parameters)
    {
        CMachine* machine=m_machine_eval->get_machine();

        CParameterCombination* current_combination=new CParameterCombination(machine);
        SG_REF(current_combination);

        if (print_state)
        {
            SG_PRINT("Initial combination:\n");
            current_combination->print_tree();
        }

        // get total length of variables
        index_t total_variables=current_combination->get_parameters_length();

        // build parameter->value map
        CMap<TParameter*, SGVector<float64_t> >* argument=
            new CMap<TParameter*, SGVector<float64_t> >();
        current_combination->build_parameter_values_map(argument);

        //  unroll current parameter combination into vector
        SGVector<double> x(total_variables);
        index_t offset=0;

        for (index_t i=0; i<argument->get_num_elements(); i++)
        {
            CMapNode<TParameter*, SGVector<float64_t> >* node=argument->get_node_ptr(i);
            memcpy(x.vector+offset, node->data.vector, sizeof(double)*node->data.vlen);
            offset+=node->data.vlen;
        }

        SG_UNREF(argument);

        // create nlopt object and choose MMA (Method of Moving Asymptotes)
        // optimization algorithm
        nlopt_opt opt=nlopt_create(NLOPT_LD_MMA, total_variables);

        // currently we assume all parameters are positive
        // (this is NOT true when inducing points and Full Matrix GaussianARDKernel are optimized)
        // create lower bound vector (lb=-inf)
        //SGVector<double> lower_bound(total_variables);
        //lower_bound.set_const(1e-6);

        // create upper bound vector (ub=inf)
        //SGVector<double> upper_bound(total_variables);
        //upper_bound.set_const(HUGE_VAL);

        // set upper and lower bound
        //nlopt_set_lower_bounds(opt, lower_bound.vector);
        //nlopt_set_upper_bounds(opt, upper_bound.vector);

        // set maximum number of evaluations
        nlopt_set_maxeval(opt, m_max_evaluations);

        // set absolute argument tolearance
        nlopt_set_xtol_abs1(opt, m_grad_tolerance);
        nlopt_set_ftol_abs(opt, m_grad_tolerance);

        // build parameter->sgobject map from current parameter combination
        CMap<TParameter*, CSGObject*>* parameter_dictionary=
            new CMap<TParameter*, CSGObject*>();
        current_combination->build_parameter_parent_map(parameter_dictionary);

        // nlopt parameters
        nlopt_params params;

        params.current_combination=current_combination;
        params.machine_eval=m_machine_eval;
        params.print_state=print_state;
        params.parameter_dictionary=parameter_dictionary;

        // choose evaluation direction (minimize or maximize objective function)
        if (m_machine_eval->get_evaluation_direction()==ED_MINIMIZE)
        {
            if (print_state)
                SG_PRINT("Minimizing objective function:\n");

            nlopt_set_min_objective(opt, nlopt_function, &params);
        }
        else
        {
            if (print_state)
                SG_PRINT("Maximizing objective function:\n");

            nlopt_set_max_objective(opt, nlopt_function, &params);
        }

        // the minimum objective value, upon return
        double minf;

        // optimize our function
        nlopt_result result=nlopt_optimize(opt, x.vector, &minf);

        REQUIRE(result>0, "NLopt failed while optimizing objective function!\n");

        if (print_state)
        {
            SG_PRINT("Best combination:\n");
            current_combination->print_tree();
        }

        // clean up
        nlopt_destroy(opt);
        SG_UNREF(machine);
        SG_UNREF(parameter_dictionary);

        return current_combination;
    }
    else
    {
        SG_NOTIMPLEMENTED
        return NULL;
    }
}

#endif /* HAVE_NLOPT */