nnfw/src/liboutputfunctions.cpp

/********************************************************************************
 *  Neural Network Framework.                                                   *
 *  Copyright (C) 2005-2009 Gianluca Massera <emmegian@yahoo.it>                *
 *                                                                              *
 *  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 2 of the License, or           *
 *  (at your option) any later version.                                         *
 *                                                                              *
 *  This program is distributed in the hope that it will be useful,             *
 *  but WITHOUT ANY WARRANTY; without even the implied warranty of              *
 *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the               *
 *  GNU General Public License for more details.                                *
 *                                                                              *
 *  You should have received a copy of the GNU General Public License           *
 *  along with this program; if not, write to the Free Software                 *
 *  Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA  *
 ********************************************************************************/

#include "liboutputfunctions.h"
#include "cluster.h"
#include <QStringList>
#include <QRegExp>
#include <cmath>

namespace farsa {

IdentityFunction::IdentityFunction()
    : OutputFunction() {
}

void IdentityFunction::apply( DoubleVector& inputs, DoubleVector& outputs ) {
    outputs.copyValues( inputs );
}

bool IdentityFunction::derivate( const DoubleVector&, const DoubleVector&, DoubleVector& derivates ) const {
    derivates.setAll( 1.0 );
    return true;
}

void IdentityFunction::configure(ConfigurationParameters& params, QString prefix)
{
    ((void)params);
    ((void)prefix);
    // Nothing to configure
}

void IdentityFunction::save(ConfigurationParameters& params, QString prefix)
{
    // Just telling our type to ConfigurationParameters
    params.startObjectParameters(prefix, "IdentityFunction", this);
}

ScaleFunction::ScaleFunction( double rate )
    : OutputFunction() {
    this->rate = rate;
}

void ScaleFunction::apply( DoubleVector& inputs, DoubleVector& outputs ) {
    mul( outputs, rate, inputs );
}

bool ScaleFunction::derivate( const DoubleVector&, const DoubleVector&, DoubleVector& derivates ) const {
    derivates.setAll( rate );
    return true;
}

void ScaleFunction::configure(ConfigurationParameters& params, QString prefix)
{
    rate = 1.0;
    QString str = params.getValue(prefix + "rate");
    if (!str.isEmpty()) {
        bool ok;
        rate = str.toDouble(&ok);
        if (!ok) {
            rate = 1.0;
        }
    }
}

void ScaleFunction::save(ConfigurationParameters& params, QString prefix)
{
    params.startObjectParameters(prefix, "ScaleFunction", this);
    params.createParameter(prefix, "rate", QString::number(rate));
}

GainFunction::GainFunction( double gain )
    : OutputFunction() {
    gainv = gain;
}

void GainFunction::apply( DoubleVector& inputs, DoubleVector& outputs ) {
    outputs.copyValues( inputs );
    outputs += gainv;
}

bool GainFunction::derivate( const DoubleVector&, const DoubleVector&, DoubleVector& derivates ) const {
    derivates.setAll( 1.0 );
    return true;
}

void GainFunction::configure(ConfigurationParameters& params, QString prefix)
{
    gainv = 1.0;
    QString str = params.getValue(prefix + "gain");
    if (!str.isEmpty()) {
        bool ok;
        gainv = str.toDouble(&ok);
        if (!ok) {
            gainv = 1.0;
        }
    }
}

void GainFunction::save(ConfigurationParameters& params, QString prefix)
{
    params.startObjectParameters(prefix, "GainFunction", this);
    params.createParameter(prefix, "gain", QString::number(gainv));
}

SigmoidFunction::SigmoidFunction( double l ) : OutputFunction() {
    lambda = l;
}

void SigmoidFunction::apply( DoubleVector& inputs, DoubleVector& outputs ) {
    // ____________1_________________
    //   exp( -lamba*inputs ) + 1
    inv( exp( mul( outputs, -lambda, inputs ) ) += 1.0 );
}

bool SigmoidFunction::derivate( const DoubleVector&, const DoubleVector& outputs, DoubleVector& derivates ) const {
    // derivates <- lambda * out * (1.0-out)
    subtract( derivates, 1.0, outputs );
    derivates *= outputs;
    derivates *= lambda;
    return true;
}

void SigmoidFunction::configure(ConfigurationParameters& params, QString prefix)
{
    lambda = 1.0;
    QString str = params.getValue(prefix + "lambda");
    if (!str.isEmpty()) {
        bool ok;
        lambda = str.toDouble(&ok);
        if (!ok) {
            lambda = 1.0;
        }
    }
}

void SigmoidFunction::save(ConfigurationParameters& params, QString prefix)
{
    params.startObjectParameters(prefix, "SigmoidFunction", this);
    params.createParameter(prefix, "lambda", QString::number(lambda));
}

FakeSigmoidFunction::FakeSigmoidFunction( double l )
    : OutputFunction() {
    lambda = l;
}

void FakeSigmoidFunction::apply( DoubleVector& inputs, DoubleVector& outputs ) {
    unsigned int size = inputs.size();
    double x;
    double x0 = 6. + 2./3.;
    double zero = 0.5;
    for ( unsigned int i = 0; i<size; i++ ) {
        x = inputs[i];
        x *= lambda;
        x -= (.5 - zero) / (.075 + zero);
        if ( x <= -x0 ) {
            outputs[i] = 0.0;
        } else {
            if ( x < x0 ) {
                outputs[i] = .5 + .575 * x / ( 1.0 + fabs(x) );
            } else {
                outputs[i] = 1.0;
            }
        }
    }
}

bool FakeSigmoidFunction::derivate( const DoubleVector&, const DoubleVector& outputs, DoubleVector& derivates ) const {
    // derivates <- lambda * out * (1.0-out)
    subtract( derivates, 1.0, outputs );
    derivates *= outputs;
    derivates *= lambda;
    return true;
}

void FakeSigmoidFunction::configure(ConfigurationParameters& params, QString prefix)
{
    lambda = 1.0;
    QString str = params.getValue(prefix + "lambda");
    if (!str.isEmpty()) {
        bool ok;
        lambda = str.toDouble(&ok);
        if (!ok) {
            lambda = 1.0;
        }
    }
}

void FakeSigmoidFunction::save(ConfigurationParameters& params, QString prefix)
{
    params.startObjectParameters(prefix, "FakeSigmoidFunction", this);
    params.createParameter(prefix, "lambda", QString::number(lambda));
}

ScaledSigmoidFunction::ScaledSigmoidFunction( double l, double min, double max )
    : OutputFunction() {
    lambda = l;
    this->min = min;
    this->max = max;
}

void ScaledSigmoidFunction::apply( DoubleVector& inputs, DoubleVector& outputs ) {
    //--- compute the sigmoid
    // ____________1_________________
    //   exp( -lamba*inputs ) + 1
    inv( exp( mul( outputs, -lambda, inputs ) ) += 1.0 );
    //--- and scale it
    outputs *= max-min;
    outputs += min;
}

bool ScaledSigmoidFunction::derivate( const DoubleVector&, const DoubleVector& outputs, DoubleVector& derivates ) const {
    // derivates <- lambda * out * (1.0-out)
    subtract( derivates, 1.0, outputs );
    derivates *= outputs;
    derivates *= lambda;
    return true;
}

void ScaledSigmoidFunction::configure(ConfigurationParameters& params, QString prefix)
{
    lambda = 1.0;
    QString str = params.getValue(prefix + "lambda");
    if (!str.isEmpty()) {
        bool ok;
        lambda = str.toDouble(&ok);
        if (!ok) {
            lambda = 1.0;
        }
    }

    min = -1.0;
    str = params.getValue(prefix + "min");
    if (!str.isEmpty()) {
        bool ok;
        min = str.toDouble(&ok);
        if (!ok) {
            min = -1.0;
        }
    }

    max = 1.0;
    str = params.getValue(prefix + "max");
    if (!str.isEmpty()) {
        bool ok;
        max = str.toDouble(&ok);
        if (!ok) {
            max = 1.0;
        }
    }
}

void ScaledSigmoidFunction::save(ConfigurationParameters& params, QString prefix)
{
    params.startObjectParameters(prefix, "ScaledSigmoidFunction", this);
    params.createParameter(prefix, "lambda", QString::number(lambda));
    params.createParameter(prefix, "min", QString::number(min));
    params.createParameter(prefix, "max", QString::number(max));
}

RampFunction::RampFunction()
    : OutputFunction() {
    min_x = 0.0;
    max_x = 0.0;
    min_y = 0.0;
    max_y = 0.0;
}

RampFunction::RampFunction( double minX, double maxX, double minY, double maxY )
    : OutputFunction() {
    min_x = minX;
    max_x = maxX;
    min_y = minY;
    max_y = maxY;
}

void RampFunction::apply( DoubleVector& inputs, DoubleVector& outputs ) {
    unsigned int size = inputs.size();
    double m = ( max_y-min_y )/( max_x-min_x );
    double q = min_y - m*min_x;
    for ( unsigned int i = 0; i<size; i++ ) {
        double ret = m*(inputs[i]) + q;
        if (ret < min_y) {
            outputs[i] = min_y;
        } else if (ret > max_y) {
            outputs[i] = max_y;
        } else {
            outputs[i] = ret;
        }
    }
}

bool RampFunction::derivate( const DoubleVector& inputs, const DoubleVector&, DoubleVector& derivates ) const {
    for( unsigned int i=0; i<inputs.size(); i++ ) {
        if ( inputs[i] >= min_x && inputs[i] <= max_x ) {
            derivates[i] = ( max_y-min_y )/( max_x-min_x );
        } else {
            double y;
            y = 1.0/( 1.0 + std::exp( -inputs[i] ) );
            derivates[i] = y * ( 1.0 - y );
        }
    }
    return true;
}

void RampFunction::configure(ConfigurationParameters& params, QString prefix)
{
    min_x = 0.0;
    QString str = params.getValue(prefix + "minX");
    if (!str.isEmpty()) {
        bool ok;
        min_x = str.toDouble(&ok);
        if (!ok) {
            min_x = 0.0;
        }
    }

    max_x = 0.0;
    str = params.getValue(prefix + "maxX");
    if (!str.isEmpty()) {
        bool ok;
        max_x = str.toDouble(&ok);
        if (!ok) {
            max_x = 0.0;
        }
    }

    min_y = 0.0;
    str = params.getValue(prefix + "minY");
    if (!str.isEmpty()) {
        bool ok;
        min_y = str.toDouble(&ok);
        if (!ok) {
            min_y = 0.0;
        }
    }

    max_y = 0.0;
    str = params.getValue(prefix + "maxY");
    if (!str.isEmpty()) {
        bool ok;
        max_y = str.toDouble(&ok);
        if (!ok) {
            max_y = 0.0;
        }
    }
}

void RampFunction::save(ConfigurationParameters& params, QString prefix)
{
    params.startObjectParameters(prefix, "RampFunction", this);
    params.createParameter(prefix, "minX", QString::number(min_x));
    params.createParameter(prefix, "maxX", QString::number(max_x));
    params.createParameter(prefix, "minY", QString::number(min_y));
    params.createParameter(prefix, "maxY", QString::number(max_y));
}

LinearFunction::LinearFunction()
    : OutputFunction() {
    m = 0.0;
    b = 0.0;
}

LinearFunction::LinearFunction( double m, double b )
    : OutputFunction() {
    this->m = m;
    this->b = b;
}

void LinearFunction::apply( DoubleVector& inputs, DoubleVector& outputs ) {
    mul( outputs, m, inputs ) += b;
}

bool LinearFunction::derivate( const DoubleVector& , const DoubleVector&, DoubleVector& derivates ) const {
    derivates.setAll( m );
    return true;
}

void LinearFunction::configure(ConfigurationParameters& params, QString prefix)
{
    m = 0.0;
    QString str = params.getValue(prefix + "m");
    if (!str.isEmpty()) {
        bool ok;
        m = str.toDouble(&ok);
        if (!ok) {
            m = 0.0;
        }
    }

    b = 0.0;
    str = params.getValue(prefix + "b");
    if (!str.isEmpty()) {
        bool ok;
        b = str.toDouble(&ok);
        if (!ok) {
            b = 0.0;
        }
    }
}

void LinearFunction::save(ConfigurationParameters& params, QString prefix)
{
    params.startObjectParameters(prefix, "LinearFunction", this);
    params.createParameter(prefix, "m", QString::number(m));
    params.createParameter(prefix, "b", QString::number(b));
}

StepFunction::StepFunction( double min, double max, double threshold )
    : OutputFunction() {
    this->min = min;
    this->max = max;
    this->threshold = threshold;
}

void StepFunction::apply( DoubleVector& inputs, DoubleVector& outputs ) {
    unsigned int size = inputs.size();
    for ( unsigned int i = 0; i<size; i++ ) {
        ( inputs[i] > threshold ) ? outputs[i] = max : outputs[i] = min;
    }
}

bool StepFunction::derivate( const DoubleVector& inputs, const DoubleVector&, DoubleVector& derivates ) const {
    //--- return the same as if it is a sigmoid with lambda = 1
    for( unsigned int i=0; i<inputs.size(); i++ ) {
        double y;
        y = 1.0/( 1.0 + std::exp( -inputs[i] ) );
        derivates[i] = y * ( 1.0 - y );
    }
    return true;
}

void StepFunction::configure(ConfigurationParameters& params, QString prefix)
{
    min = 0.0;
    QString str = params.getValue(prefix + "min");
    if (!str.isEmpty()) {
        bool ok;
        min = str.toDouble(&ok);
        if (!ok) {
            min = 0.0;
        }
    }

    max = 1.0;
    str = params.getValue(prefix + "max");
    if (!str.isEmpty()) {
        bool ok;
        max = str.toDouble(&ok);
        if (!ok) {
            max = 1.0;
        }
    }

    threshold = 0.0;
    str = params.getValue(prefix + "threshold");
    if (!str.isEmpty()) {
        bool ok;
        threshold = str.toDouble(&ok);
        if (!ok) {
            threshold = 0.0;
        }
    }
}

void StepFunction::save(ConfigurationParameters& params, QString prefix)
{
    params.startObjectParameters(prefix, "StepFunction", this);
    params.createParameter(prefix, "min", QString::number(min));
    params.createParameter(prefix, "max", QString::number(max));
    params.createParameter(prefix, "threshold", QString::number(threshold));
}

LeakyIntegratorFunction::LeakyIntegratorFunction()
    : OutputFunction(), delta(), outprev()
{
}

LeakyIntegratorFunction::LeakyIntegratorFunction( const DoubleVector& d )
    : OutputFunction(), delta(d.size()), outprev(d.size()) {
    delta.copyValues( d );
    outprev.zeroing();
}

void LeakyIntegratorFunction::apply( DoubleVector& inputs, DoubleVector& outputs ) {
    //--- y <- delta*y(t-1) + (1.0-delta)*inputs
    //---  its equivalent to
    //--- y <- delta*( y(t-1) - inputs ) + inputs
    outputs = subtract( outputs, outprev, inputs );
    outputs *= delta;
    outputs += inputs;
    outprev.copyValues( outputs );
}

void LeakyIntegratorFunction::zeroingStatus() {
    outprev.zeroing();
}

void LeakyIntegratorFunction::clusterSetted() {
    if ( clusterv->numNeurons() != delta.size() ) {
        delta.resize( clusterv->numNeurons() );
        outprev.resize( clusterv->numNeurons() );
    }
}

void LeakyIntegratorFunction::configure(ConfigurationParameters& params, QString prefix)
{
    // Delta is a vector, that is a list of space-separated values
    QString str = params.getValue(prefix + "delta");
    if (!str.isEmpty()) {
        QStringList list = str.split(QRegExp("\\s+"), QString::SkipEmptyParts);
        delta.resize(list.size());
        for( int i = 0; i < list.size(); i++) {
            bool ok;
            delta[i] = list[i].toDouble(&ok);
            if (!ok) {
                delta[i] = 0.0;
            }
        }
    }

    // Also reloading outprev (resizing it to match delta length)
    str = params.getValue(prefix + "outprev");
    if (!str.isEmpty()) {
        QStringList list = str.split(QRegExp("\\s+"), QString::SkipEmptyParts);
        outprev.resize(list.size());
        for( int i = 0; i < list.size(); i++) {
            bool ok;
            outprev[i] = list[i].toDouble(&ok);
            if (!ok) {
                outprev[i] = 0.0;
            }
        }
    }
    outprev.resize(delta.size());
}

void LeakyIntegratorFunction::save(ConfigurationParameters& params, QString prefix)
{
    params.startObjectParameters(prefix, "LeakyIntegratorFunction", this);

    // First creating a string list, then transforming to a single string
    QStringList list;
    for (unsigned int i = 0; i < delta.size(); i++) {
        list.push_back(QString::number(delta[i]));
    }
    params.createParameter(prefix, "delta", list.join(" "));

    // Saving in the same way also outprev
    list.clear();
    for (unsigned int i = 0; i < delta.size(); i++) {
        list.push_back(QString::number(outprev[i]));
    }
    params.createParameter(prefix, "outprev", list.join(" "));
}

LogLikeFunction::LogLikeFunction( double A, double B )
    : OutputFunction() {
    this->A = A;
    this->B = B;
}

void LogLikeFunction::apply( DoubleVector& inputs, DoubleVector& outputs ) {
    //--- y <- x / ( 1+A*x+b )
    outputs = inv( mul( outputs, A, inputs ) += (1.0+B) );
    outputs *= inputs;
}

void LogLikeFunction::configure(ConfigurationParameters& params, QString prefix)
{
    A = 1.0;
    QString str = params.getValue(prefix + "A");
    if (!str.isEmpty()) {
        bool ok;
        A = str.toDouble(&ok);
        if (!ok) {
            A = 1.0;
        }
    }

    B = 5.0;
    str = params.getValue(prefix + "B");
    if (!str.isEmpty()) {
        bool ok;
        B = str.toDouble(&ok);
        if (!ok) {
            B = 5.0;
        }
    }
}

void LogLikeFunction::save(ConfigurationParameters& params, QString prefix)
{
    params.startObjectParameters(prefix, "LogLikeFunction", this);
    params.createParameter(prefix, "A", QString::number(A));
    params.createParameter(prefix, "B", QString::number(B));
}

CompositeFunction::CompositeFunction()
    : OutputFunction(), first(), second(), mid() {
}

CompositeFunction::CompositeFunction( OutputFunction *f, OutputFunction *g )
    : OutputFunction(), first(f), second(g), mid() {
}

CompositeFunction::~CompositeFunction() {
    // auto_ptr will release memory for us
}

void CompositeFunction::apply( DoubleVector& inputs, DoubleVector& outputs ) {
    first->apply( inputs, mid );
    second->apply( mid, outputs );
}

bool CompositeFunction::setFirstFunction( OutputFunction *f ) {
    first.reset(f);
    first->setCluster( clusterv );
    return true;
}

OutputFunction* CompositeFunction::getFirstFunction() {
    return first.get();
}

bool CompositeFunction::setSecondFunction( OutputFunction *g ) {
    second.reset(g);
    second->setCluster( clusterv );
    return true;
}

OutputFunction* CompositeFunction::getSecondFunction() {
    return second.get();
}

void CompositeFunction::clusterSetted() {
    mid.resize( clusterv->numNeurons() );
    first->setCluster( clusterv );
    second->setCluster( clusterv );
}

void CompositeFunction::configure(ConfigurationParameters& params, QString prefix)
{
    // We don't need configured component functions here (and they will be
    // configured after exiting from this function)
    first.reset(params.getObjectFromParameter<OutputFunction>(prefix + "first", false, false));
    second.reset(params.getObjectFromParameter<OutputFunction>(prefix + "second", false, false));

    // We don't need to reload a reference to the cluster as he calls our setCluster
    // function after our creation
}

void CompositeFunction::save(ConfigurationParameters& params, QString prefix)
{
    params.startObjectParameters(prefix, "CompositeFunction", this);
    params.createParameter(prefix, "first", first.get());
    params.createParameter(prefix, "second", second.get());

    // We don't need to save the reference to the cluster as he calls our setCluster
    // function after our creation
}

LinearComboFunction::LinearComboFunction()
    : OutputFunction(), first(), second(), mid()
{
    this->w1 = 0.0;
    this->w2 = 0.0;
}

LinearComboFunction::LinearComboFunction( double w1, OutputFunction *f, double w2, OutputFunction *g )
    : OutputFunction(), first(f), second(g), mid() {
    this->w1 = w1;
    this->w2 = w2;
}

LinearComboFunction::~LinearComboFunction() {
    // auto_ptr will release memory for us
}

void LinearComboFunction::apply( DoubleVector& inputs, DoubleVector& outputs ) {
    first->apply( inputs, mid );
    mid *= w1;
    second->apply( inputs, outputs );
    outputs *= w2;
    outputs += mid;
}

bool LinearComboFunction::setFirstFunction( OutputFunction *f ) {
    first.reset(f);
    first->setCluster( clusterv );
    return true;
}

OutputFunction* LinearComboFunction::getFirstFunction() {
    return first.get();
}

bool LinearComboFunction::setFirstWeight( double v ) {
    w1 = v;
    return true;
}

double LinearComboFunction::getFirstWeight() {
    return w1;
}

bool LinearComboFunction::setSecondFunction( OutputFunction *g ) {
    second.reset(g);
    second->setCluster( clusterv );
    return true;
}

OutputFunction* LinearComboFunction::getSecondFunction() {
    return second.get();
}

bool LinearComboFunction::setSecondWeight( double v ) {
    w2 = v;
    return true;
}

double LinearComboFunction::getSecondWeight() {
    return w2;
}

void LinearComboFunction::clusterSetted() {
    mid.resize( clusterv->numNeurons() );
    first->setCluster( clusterv );
    second->setCluster( clusterv );
}

void LinearComboFunction::configure(ConfigurationParameters& params, QString prefix)
{
    // We don't need configured component functions here (and they will be
    // configured after exiting from this function)
    first.reset(params.getObjectFromParameter<OutputFunction>(prefix + "first", false, false));

    w1 = 0.0;
    QString str = params.getValue(prefix + "w1");
    if (!str.isEmpty()) {
        bool ok;
        w1 = str.toDouble(&ok);
        if (!ok) {
            w1 = 0.0;
        }
    }

    second.reset(params.getObjectFromParameter<OutputFunction>(prefix + "second", false, false));

    w2 = 0.0;
    str = params.getValue(prefix + "w2");
    if (!str.isEmpty()) {
        bool ok;
        w2 = str.toDouble(&ok);
        if (!ok) {
            w2 = 0.0;
        }
    }

    // We don't need to reload a reference to the cluster as he calls our setCluster
    // function after our creation
}

void LinearComboFunction::save(ConfigurationParameters& params, QString prefix)
{
    params.startObjectParameters(prefix, "LinearComboFunction", this);
    params.createParameter(prefix, "first", first.get());
    params.createParameter(prefix, "w1", QString::number(w1));
    params.createParameter(prefix, "second", second.get());
    params.createParameter(prefix, "w2", QString::number(w2));

    // We don't need to save the reference to the cluster as he calls our setCluster
    // function after our creation
}

}