FeatureBlockLogisticRegression.cpp

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 Sergey Lisitsyn
 */

#include <shogun/classifier/FeatureBlockLogisticRegression.h>
#include <shogun/lib/slep/slep_solver.h>
#include <shogun/lib/slep/slep_options.h>

#include <shogun/lib/IndexBlockGroup.h>
#include <shogun/lib/IndexBlockTree.h>

namespace shogun
{

CFeatureBlockLogisticRegression::CFeatureBlockLogisticRegression() :
    CLinearMachine(), 
    m_feature_relation(NULL), m_z(0.0)
{
    register_parameters();
}

CFeatureBlockLogisticRegression::CFeatureBlockLogisticRegression(
     float64_t z, CDotFeatures* train_features, 
     CBinaryLabels* train_labels, CIndexBlockRelation* feature_relation) :
    CLinearMachine(), 
    m_feature_relation(NULL)
{
    set_feature_relation(feature_relation);
    set_z(z);
    set_q(2.0);
    set_features(train_features);
    set_labels(train_labels);
    set_termination(0);
    set_regularization(0);
    set_tolerance(1e-3);
    set_max_iter(1000);
    register_parameters();
}

CFeatureBlockLogisticRegression::~CFeatureBlockLogisticRegression()
{
    SG_UNREF(m_feature_relation);
}

void CFeatureBlockLogisticRegression::register_parameters()
{
    SG_ADD((CSGObject**)&m_feature_relation, "feature_relation", "feature relation", MS_NOT_AVAILABLE);
    SG_ADD(&m_z, "z", "regularization coefficient", MS_AVAILABLE);
    SG_ADD(&m_q, "q", "q of L1/Lq", MS_AVAILABLE);
    SG_ADD(&m_termination, "termination", "termination", MS_NOT_AVAILABLE);
    SG_ADD(&m_regularization, "regularization", "regularization", MS_NOT_AVAILABLE);
    SG_ADD(&m_tolerance, "tolerance", "tolerance", MS_NOT_AVAILABLE);
    SG_ADD(&m_max_iter, "max_iter", "maximum number of iterations", MS_NOT_AVAILABLE);
}

CIndexBlockRelation* CFeatureBlockLogisticRegression::get_feature_relation() const
{
    SG_REF(m_feature_relation);
    return m_feature_relation;
}

void CFeatureBlockLogisticRegression::set_feature_relation(CIndexBlockRelation* feature_relation)
{
    SG_UNREF(m_feature_relation);
    SG_REF(feature_relation);
    m_feature_relation = feature_relation;
}

int32_t CFeatureBlockLogisticRegression::get_max_iter() const
{
    return m_max_iter;
}

int32_t CFeatureBlockLogisticRegression::get_regularization() const
{
    return m_regularization;
}

int32_t CFeatureBlockLogisticRegression::get_termination() const
{
    return m_termination;
}

float64_t CFeatureBlockLogisticRegression::get_tolerance() const
{
    return m_tolerance;
}

float64_t CFeatureBlockLogisticRegression::get_z() const
{
    return m_z;
}

float64_t CFeatureBlockLogisticRegression::get_q() const
{
    return m_q;
}

void CFeatureBlockLogisticRegression::set_max_iter(int32_t max_iter)
{
    ASSERT(max_iter>=0);
    m_max_iter = max_iter;
}

void CFeatureBlockLogisticRegression::set_regularization(int32_t regularization)
{
    ASSERT(regularization==0 || regularization==1);
    m_regularization = regularization;
}

void CFeatureBlockLogisticRegression::set_termination(int32_t termination)
{
    ASSERT(termination>=0 && termination<=4);
    m_termination = termination;
}

void CFeatureBlockLogisticRegression::set_tolerance(float64_t tolerance)
{
    ASSERT(tolerance>0.0);
    m_tolerance = tolerance;
}

void CFeatureBlockLogisticRegression::set_z(float64_t z)
{
    m_z = z;
}

void CFeatureBlockLogisticRegression::set_q(float64_t q)
{
    m_q = q;
}

bool CFeatureBlockLogisticRegression::train_machine(CFeatures* data)
{
    if (data && (CDotFeatures*)data)
        set_features((CDotFeatures*)data);

    ASSERT(features);
    ASSERT(m_labels);

    int32_t n_vecs = m_labels->get_num_labels();
    SGVector<float64_t> y(n_vecs);
    for (int32_t i=0; i<n_vecs; i++)
        y[i] = ((CBinaryLabels*)m_labels)->get_label(i);
    
    slep_options options = slep_options::default_options();
    options.q = m_q;
    options.regularization = m_regularization;
    options.termination = m_termination;
    options.tolerance = m_tolerance;
    options.max_iter = m_max_iter;
    options.loss = LOGISTIC;

    EIndexBlockRelationType relation_type = m_feature_relation->get_relation_type();
    switch (relation_type)
    {
        case GROUP:
        {
            CIndexBlockGroup* feature_group = (CIndexBlockGroup*)m_feature_relation;
            SGVector<index_t> ind = feature_group->get_SLEP_ind();
            options.ind = ind.vector;
            options.n_feature_blocks = ind.vlen-1;
            if (ind[ind.vlen-1] > features->get_num_vectors())
                SG_ERROR("Group of features covers more vectors than available\n");
        
            options.gWeight = SG_MALLOC(double, options.n_feature_blocks);
            for (int32_t i=0; i<options.n_feature_blocks; i++)
                options.gWeight[i] = 1.0;
            options.mode = FEATURE_GROUP;
            options.loss = LOGISTIC;
            options.n_nodes = 0;
            slep_result_t result = slep_solver(features, y.vector, m_z, options);

            SG_FREE(options.gWeight);
            int32_t n_feats = features->get_dim_feature_space();
            SGVector<float64_t> new_w(n_feats);
            for (int i=0; i<n_feats; i++)
                new_w[i] = result.w[i];
            set_bias(result.c[0]);

            w = new_w;
        }
        break;
        case TREE: 
        {
            CIndexBlockTree* feature_tree = (CIndexBlockTree*)m_feature_relation;

            SGVector<float64_t> ind_t = feature_tree->get_SLEP_ind_t();
            SGVector<float64_t> G;
            if (feature_tree->is_general())
            {
                G = feature_tree->get_SLEP_G();
                options.general = true;
            }
            options.ind_t = ind_t.vector;
            options.G = G.vector;
            options.n_nodes = ind_t.vlen/3;
            options.n_feature_blocks = ind_t.vlen/3;
            options.mode = FEATURE_TREE;
            options.loss = LOGISTIC;

            slep_result_t result = slep_solver(features, y.vector, m_z, options);
            
            int32_t n_feats = features->get_dim_feature_space();
            SGVector<float64_t> new_w(n_feats);
            for (int i=0; i<n_feats; i++)
                new_w[i] = result.w[i];

            set_bias(result.c[0]);

            w = new_w;
        }
        break;
        default: 
            SG_ERROR("Not supported feature relation type\n");
    }

    return true;
}

float64_t CFeatureBlockLogisticRegression::apply_one(int32_t vec_idx)
{
    return CMath::exp(-(features->dense_dot(vec_idx, w.vector, w.vlen) + bias));
}

SGVector<float64_t> CFeatureBlockLogisticRegression::apply_get_outputs(CFeatures* data)
{
    if (data)
    {
        if (!data->has_property(FP_DOT))
            SG_ERROR("Specified features are not of type CDotFeatures\n");

        set_features((CDotFeatures*) data);
    }

    if (!features)
        return SGVector<float64_t>();

    int32_t num=features->get_num_vectors();
    ASSERT(num>0);
    ASSERT(w.vlen==features->get_dim_feature_space());

    float64_t* out=SG_MALLOC(float64_t, num);
    features->dense_dot_range(out, 0, num, NULL, w.vector, w.vlen, bias);
    for (int32_t i=0; i<num; i++)
        out[i] = 2.0/(1.0+CMath::exp(-out[i])) - 1.0;
    return SGVector<float64_t>(out,num);
}

}