DomainAdaptationSVMLinear.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.
 *
 * Written (W) 2007-2011 Christian Widmer
 * Copyright (C) 2007-2011 Max-Planck-Society
 */

#include <shogun/lib/config.h>

#ifdef HAVE_LAPACK

#include <shogun/transfer/domain_adaptation/DomainAdaptationSVMLinear.h>
#include <shogun/io/SGIO.h>
#include <shogun/base/Parameter.h>
#include <shogun/labels/Labels.h>
#include <shogun/labels/BinaryLabels.h>
#include <shogun/labels/RegressionLabels.h>
#include <iostream>
#include <vector>

using namespace shogun;


CDomainAdaptationSVMLinear::CDomainAdaptationSVMLinear() : CLibLinear(L2R_L1LOSS_SVC_DUAL)
{
    init(NULL, 0.0);
}


CDomainAdaptationSVMLinear::CDomainAdaptationSVMLinear(float64_t C, CDotFeatures* f, CLabels* lab, CLinearMachine* pre_svm, float64_t B_param) : CLibLinear(C, f, lab)
{
    init(pre_svm, B_param);

}


CDomainAdaptationSVMLinear::~CDomainAdaptationSVMLinear()
{

    SG_UNREF(presvm);
    SG_DEBUG("deleting DomainAdaptationSVMLinear\n");
}


void CDomainAdaptationSVMLinear::init(CLinearMachine* pre_svm, float64_t B_param)
{

    if (pre_svm)
    {
        // increase reference counts
        SG_REF(pre_svm);

        // set bias of parent svm to zero
        pre_svm->set_bias(0.0);
    }

    this->presvm = pre_svm;
    this->B = B_param;
    this->train_factor = 1.0;

    set_liblinear_solver_type(L2R_L1LOSS_SVC_DUAL);

    // invoke sanity check
    is_presvm_sane();

    // serialization code
    m_parameters->add((CSGObject**) &presvm, "presvm", "SVM to regularize against");
    m_parameters->add(&B, "B",  "Regularization strenth B.");
    m_parameters->add(&train_factor, "train_factor",  "train_factor");

}


bool CDomainAdaptationSVMLinear::is_presvm_sane()
{

    if (!presvm) {

        SG_WARNING("presvm is null");

    } else {

        if (presvm->get_bias() != 0) {
            SG_ERROR("presvm bias not set to zero");
        }

        if (presvm->get_features()->get_feature_type() != this->get_features()->get_feature_type()) {
            SG_ERROR("feature types do not agree");
        }
    }

    return true;

}


bool CDomainAdaptationSVMLinear::train_machine(CFeatures* train_data)
{

    CDotFeatures* tmp_data;

    if (m_labels->get_label_type() != LT_BINARY)
        SG_ERROR("DomainAdaptationSVMLinear requires binary labels\n");

    if (train_data)
    {
        if (!train_data->has_property(FP_DOT))
            SG_ERROR("DotFeatures expected\n");

        if (((CBinaryLabels*) m_labels)->get_num_labels() != train_data->get_num_vectors())
            SG_ERROR("Number of training vectors does not match number of labels\n");

        tmp_data = (CDotFeatures*) train_data;
    }
    else
    {
        tmp_data = features;
    }

    CBinaryLabels* labels = (CBinaryLabels*) get_labels();
    int32_t num_training_points = labels->get_num_labels();

    std::vector<float64_t> lin_term = std::vector<float64_t>(num_training_points);

    if (presvm)
    {
        ASSERT(presvm->get_bias() == 0.0);

        // bias of parent SVM was set to zero in constructor, already contains B
        CBinaryLabels* parent_svm_out = presvm->apply_binary(tmp_data);

        SG_DEBUG("pre-computing linear term from presvm\n");

        // pre-compute linear term
        for (int32_t i=0; i!=num_training_points; i++)
        {
            lin_term[i] = train_factor * B * labels->get_value(i) * parent_svm_out->get_value(i) - 1.0;
        }

        // set linear term for QP
        this->set_linear_term(
                SGVector<float64_t>(&lin_term[0], lin_term.size()));

    }

    SG_UNREF(labels);

    /*
    // warm-start liblinear
    //TODO test this code, measure speed-ups
    //presvm w stored in presvm
    float64_t* tmp_w;
    presvm->get_w(tmp_w, w_dim);

    //copy vector
    float64_t* tmp_w_copy = SG_MALLOC(float64_t, w_dim);
    std::copy(tmp_w, tmp_w + w_dim, tmp_w_copy);

    for (int32_t i=0; i!=w_dim; i++)
    {
        tmp_w_copy[i] = B * tmp_w_copy[i];
    }

    //set w (copied in setter)
    set_w(tmp_w_copy, w_dim);
    SG_FREE(tmp_w_copy);
    */

    bool success = false;

    //train SVM
    if (train_data)
    {
        success = CLibLinear::train_machine(train_data);
    } else {
        success = CLibLinear::train_machine();
    }

    //ASSERT(presvm)

    return success;

}


CLinearMachine* CDomainAdaptationSVMLinear::get_presvm()
{
    return presvm;
}


float64_t CDomainAdaptationSVMLinear::get_B()
{
    return B;
}


float64_t CDomainAdaptationSVMLinear::get_train_factor()
{
    return train_factor;
}


void CDomainAdaptationSVMLinear::set_train_factor(float64_t factor)
{
    train_factor = factor;
}


CBinaryLabels* CDomainAdaptationSVMLinear::apply_binary(CFeatures* data)
{
    ASSERT(presvm->get_bias()==0.0);

    int32_t num_examples = data->get_num_vectors();

    CBinaryLabels* out_current = CLibLinear::apply_binary(data);

    SGVector<float64_t> out_combined(num_examples);
    if (presvm)
    {
        // recursive call if used on DomainAdaptationSVM object
        CBinaryLabels* out_presvm = presvm->apply_binary(data);


        // combine outputs
        for (int32_t i=0; i!=num_examples; i++)
            out_combined[i] = out_current->get_value(i) + B*out_presvm->get_value(i);

        SG_UNREF(out_presvm);
    }

    SG_UNREF(out_current);

    return new CBinaryLabels(out_combined);
}

#endif //HAVE_LAPACK