AdaDeltaUpdater.cpp

Go to the documentation of this file.

/*
 * Copyright (c) The Shogun Machine Learning Toolbox
 * Written (w) 2015 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.
 *
 */

#include <shogun/optimization/AdaDeltaUpdater.h>
#include <shogun/lib/config.h>
using namespace shogun;

AdaDeltaUpdater::AdaDeltaUpdater()
    :DescendUpdaterWithCorrection()
{
    init();
}

AdaDeltaUpdater::AdaDeltaUpdater(float64_t learning_rate,float64_t epsilon,float64_t decay_factor)
    :DescendUpdaterWithCorrection()
{
    init();
    set_learning_rate(learning_rate);
    set_epsilon(epsilon);
    set_decay_factor(decay_factor);
}

void AdaDeltaUpdater::set_learning_rate(float64_t learning_rate)
{
    REQUIRE(learning_rate>0,"Learning_rate (%f) must be positive\n",
        learning_rate);
    m_build_in_learning_rate=learning_rate;
}

void AdaDeltaUpdater::set_epsilon(float64_t epsilon)
{
    REQUIRE(epsilon>=0,"Epsilon (%f) must be non-negative\n",
        epsilon);
    m_epsilon=epsilon;
}

void AdaDeltaUpdater::set_decay_factor(float64_t decay_factor)
{
    REQUIRE(decay_factor>=0.0 && decay_factor<1.0,
        "Decay factor (%f) must in [0,1)\n",
        decay_factor);
    m_decay_factor=decay_factor;
}

AdaDeltaUpdater::~AdaDeltaUpdater()
{
}

void AdaDeltaUpdater::init()
{
    m_decay_factor=0.9;
    m_epsilon=1e-6;
    m_build_in_learning_rate=1.0;
    m_gradient_accuracy=SGVector<float64_t>();
    m_gradient_delta_accuracy=SGVector<float64_t>();
}

void AdaDeltaUpdater::update_context(CMinimizerContext* context)
{
    DescendUpdaterWithCorrection::update_context(context);
    REQUIRE(context, "Context must set\n");

    SGVector<float64_t> value(m_gradient_accuracy.vlen);
    std::copy(m_gradient_accuracy.vector,
        m_gradient_accuracy.vector+m_gradient_accuracy.vlen,
        value.vector);
    std::string key="AdaDeltaUpdater::m_gradient_accuracy";
    context->save_data(key, value);

    value=SGVector<float64_t>(m_gradient_delta_accuracy.vlen);
    std::copy(m_gradient_delta_accuracy.vector,
        m_gradient_delta_accuracy.vector+m_gradient_delta_accuracy.vlen,
        value.vector);
    key="AdaDeltaUpdater::m_gradient_delta_accuracy";
    context->save_data(key, value);
}

void AdaDeltaUpdater::load_from_context(CMinimizerContext* context)
{
    DescendUpdaterWithCorrection::load_from_context(context);
    REQUIRE(context, "context must set\n");

    std::string key="AdaDeltaUpdater::m_gradient_accuracy";
    SGVector<float64_t> value=context->get_data_sgvector_float64(key);
    m_gradient_accuracy=SGVector<float64_t>(value.vlen);
    std::copy(value.vector, value.vector+value.vlen,
        m_gradient_accuracy.vector);

    key="AdaDeltaUpdater::m_gradient_delta_accuracy";
    value=context->get_data_sgvector_float64(key);
    m_gradient_delta_accuracy=SGVector<float64_t>(value.vlen);
    std::copy(value.vector, value.vector+value.vlen,
        m_gradient_delta_accuracy.vector);
}

float64_t AdaDeltaUpdater::get_negative_descend_direction(float64_t variable,
    float64_t gradient, index_t idx, float64_t learning_rate)
{
    REQUIRE(idx>=0 && idx<m_gradient_accuracy.vlen,
        "Index (%d) is invalid\n", idx);
    REQUIRE(idx>=0 && idx<m_gradient_delta_accuracy.vlen,
        "Index (%d) is invalid\n", idx);
    float64_t scale=m_decay_factor*m_gradient_accuracy[idx]+
        (1.0-m_decay_factor)*gradient*gradient;
    m_gradient_accuracy[idx]=scale;
    float64_t res=m_build_in_learning_rate*gradient*CMath::sqrt(m_gradient_delta_accuracy[idx]+m_epsilon)/CMath::sqrt(scale+m_epsilon);
    m_gradient_delta_accuracy[idx]=m_decay_factor*m_gradient_delta_accuracy[idx]+(1.0-m_decay_factor)*res*res;
    return res;

}

void AdaDeltaUpdater::update_variable(SGVector<float64_t> variable_reference,
    SGVector<float64_t> raw_negative_descend_direction, float64_t learning_rate)
{
    REQUIRE(variable_reference.vlen>0,"variable_reference must set\n");
    REQUIRE(variable_reference.vlen==raw_negative_descend_direction.vlen,
        "The length of variable_reference (%d) and the length of gradient (%d) do not match\n",
        variable_reference.vlen,raw_negative_descend_direction.vlen);
    if(m_gradient_accuracy.vlen==0)
    {
        m_gradient_accuracy=SGVector<float64_t>(variable_reference.vlen);
        m_gradient_accuracy.set_const(0.0);

        m_gradient_delta_accuracy=SGVector<float64_t>(m_gradient_accuracy.vlen);
        m_gradient_delta_accuracy.set_const(0.0);
    }
    if(m_correction)
    {
        MomentumCorrection* momentum_correction=dynamic_cast<MomentumCorrection *>(m_correction);
        if(momentum_correction)
        {
            if(!momentum_correction->is_initialized())
                momentum_correction->initialize_previous_direction(variable_reference.vlen);
        }

        for(index_t idx=0; idx<variable_reference.vlen; idx++)
        {
            float64_t neg_des_dir=get_negative_descend_direction(
                variable_reference[idx], raw_negative_descend_direction[idx], idx, learning_rate);

            DescendPair pair=m_correction->get_corrected_descend_direction(
                neg_des_dir, idx);
            float64_t delta=pair.delta;
            variable_reference[idx]+=pair.descend_direction;
            m_gradient_delta_accuracy[idx]+=
                (1.0-m_decay_factor)*(delta*delta-neg_des_dir*neg_des_dir);
        }
    }
    else
    {
        DescendUpdaterWithCorrection::update_variable(variable_reference, raw_negative_descend_direction, learning_rate);
    }
}