nnfw/src/backpropagationalgo.cpp

/********************************************************************************
 *  Neural Network Framework.                                                   *
 *  Copyright (C) 2005-2011 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 "neuralnet.h"
#include "matrixlinker.h"
#include "biasedcluster.h"
#include "backpropagationalgo.h"

namespace farsa {

BackPropagationAlgo::BackPropagationAlgo( NeuralNet *n_n, UpdatableList up_order, double l_r )
    : LearningAlgorithm(n_n), learn_rate(l_r), update_order(up_order) {
    useMomentum = false;
    momentumv = 0.0f;
    neuralNetChanged();
}

BackPropagationAlgo::BackPropagationAlgo()
    : LearningAlgorithm(), learn_rate(0.0) {
    useMomentum = false;
    momentumv = 0.0f;
    neuralNetChanged();
}

BackPropagationAlgo::~BackPropagationAlgo( ) {
    /* nothing to do ?!?! */
}

void BackPropagationAlgo::neuralNetChanged() {
    NeuralNet* net = neuralNet();
    if ( !net ) return;
    //--- clear all data
    mapIndex.clear();
    cluster_deltas_vec.clear();
    //--- insert new data for the new NeuralNet
    Cluster *cluster_temp;
    // pushing the info for output cluster
    ClusterList outs = net->outputClusters();
    for( int i=0; i<(int)outs.size(); i++ ) {
        addCluster( outs[i], true );
    }
    // --- generate information for backpropagation of deltas
    for( int i=0; i<(int)update_order.size(); i++ ) {
        cluster_temp = dynamic_cast<Cluster*>(update_order[i]);
        if ( cluster_temp ) {
            addCluster( cluster_temp, false );
            continue;
        }
        MatrixLinker* linker_temp = dynamic_cast<MatrixLinker*>(update_order[i]);
        if ( linker_temp ) {    //Dot linker subclass of Matrixlinker
            addLinker( linker_temp );
        }
    }
}

void BackPropagationAlgo::setUpdateOrder( const UpdatableList& update_order ) {
    this->update_order = update_order;
    this->neuralNetChanged();
}

void BackPropagationAlgo::setTeachingInput( Cluster* output, const DoubleVector& ti ) {
    if ( mapIndex.count( output ) == 0 ) { 
        return;
    }
    int index = mapIndex[ output ];
    subtract( cluster_deltas_vec[index].deltas_outputs, output->outputs(), ti );
    return;
}

DoubleVector BackPropagationAlgo::getError( Cluster* cl ) {
    if ( mapIndex.count( cl ) == 0 ) {
        qWarning() << "Cluster not present in BackPropagationAlgo";
        return DoubleVector();
    }
    int index = mapIndex[ cl ];
    return cluster_deltas_vec[index].deltas_outputs;
}

void BackPropagationAlgo::enableMomentum() {
    for ( int i=0; i<cluster_deltas_vec.size(); ++i ) {
        for ( int j=0;  j<cluster_deltas_vec[i].incoming_linkers_vec.size(); ++j ) {
            // --- zeroing data
            cluster_deltas_vec[i].incoming_last_outputs[j].zeroing();
            cluster_deltas_vec[i].last_deltas_inputs.zeroing();
        }
    }
    useMomentum = true;
}

void BackPropagationAlgo::propagDeltas() {
    DoubleVector diff_vec;
    for( int i=0; i<(int)cluster_deltas_vec.size(); i++ ) {
        cluster_deltas_vec[i].incoming_linkers_vec;
        // --- propagate DeltaOutput to DeltaInputs
        cluster_deltas_vec[i].deltas_inputs.copyValues( cluster_deltas_vec[i].deltas_outputs );
        Cluster* cl = cluster_deltas_vec[i].cluster;
        OutputFunction* func = cl->outFunction();
        diff_vec.resize( cluster_deltas_vec[i].deltas_inputs.size() );
        if ( func->derivate( cl->inputs(), cl->outputs(), diff_vec ) ) {
            cluster_deltas_vec[i].deltas_inputs *= diff_vec;
        }
        // --- propagate DeltaInputs to DeltaOutput through MatrixLinker
        for( int k=0; k<cluster_deltas_vec[i].incoming_linkers_vec.size( ); ++k ) {
            MatrixLinker* link = dynamic_cast<MatrixLinker*>(cluster_deltas_vec[i].incoming_linkers_vec[k]);
            if ( mapIndex.count(link->from()) == 0 ) {
                // --- the from() cluster is not in Learning
                continue;
            }
            int from_index = mapIndex[ link->from() ];
            amul( cluster_deltas_vec[from_index].deltas_outputs, link->matrix(), cluster_deltas_vec[i].deltas_inputs );
        }
    }
    return;
}

void BackPropagationAlgo::learn() {
    // --- zeroing previous step delta information
    for ( int i=0; i<cluster_deltas_vec.size(); ++i ) {
        if ( cluster_deltas_vec[i].isOutput ) continue;
        cluster_deltas_vec[i].deltas_outputs.zeroing();
    }
    // --- propagating the error through the net
    propagDeltas();
    // --- make the learn !!
    for ( int i=0; i<cluster_deltas_vec.size(); ++i ) {
        if ( cluster_deltas_vec[i].cluster != NULL) {
            DoubleVector minus_ones( cluster_deltas_vec[i].cluster->outputs().size( ), -1.0f );
            deltarule( cluster_deltas_vec[i].cluster->biases(), -learn_rate, minus_ones, cluster_deltas_vec[i].deltas_inputs );
        }

        for ( int j=0;  j<cluster_deltas_vec[i].incoming_linkers_vec.size(); ++j ) {
            if ( cluster_deltas_vec[i].incoming_linkers_vec[j] != NULL ) {
                deltarule(
                    cluster_deltas_vec[i].incoming_linkers_vec[j]->matrix(),
                    -learn_rate,
                    cluster_deltas_vec[i].incoming_linkers_vec[j]->from()->outputs(),
                    cluster_deltas_vec[i].deltas_inputs
                );
                if ( !useMomentum ) continue;
                // --- add the momentum
                deltarule(
                    cluster_deltas_vec[i].incoming_linkers_vec[j]->matrix(),
                    -learn_rate*momentumv,
                    cluster_deltas_vec[i].incoming_last_outputs[j],
                    cluster_deltas_vec[i].last_deltas_inputs
                );
                // --- save datas for momentum on the next step
                cluster_deltas_vec[i].incoming_last_outputs[j].copyValues( cluster_deltas_vec[i].incoming_linkers_vec[j]->from()->outputs() );
                cluster_deltas_vec[i].last_deltas_inputs.copyValues( cluster_deltas_vec[i].deltas_inputs );
            }
        }
    }
    return;
}

void BackPropagationAlgo::learn( const Pattern& pat ) {
    // --- set the inputs of the net
    ClusterList clins = neuralNet()->inputClusters();
    for( int i=0; i<clins.size(); i++ ) {
        clins[i]->inputs().copyValues( pat.inputsOf( clins[i] ) );
    }
    // --- spread the net
    neuralNet()->step();
    // --- set the teaching input
    ClusterList clout = neuralNet()->outputClusters();
    for( int i=0; i<clout.size(); i++ ) {
        setTeachingInput( clout[i], pat.outputsOf( clout[i] ) );
    }
    learn();
}

double BackPropagationAlgo::calculateMSE( const Pattern& pat ) {
    // --- set the inputs of the net
    ClusterList clins = neuralNet()->inputClusters();
    for( int i=0; i<clins.size(); i++ ) {
        clins[i]->inputs().copyValues( pat.inputsOf( clins[i] ) );
    }
    // --- spread the net
    neuralNet()->step();
    // --- calculate the MSE
    ClusterList clout = neuralNet()->outputClusters();
    double mseacc = 0.0;
    int dim = (int)clout.size();
    for( int i=0; i<dim; i++ ) {
        mseacc += mse( clout[i]->outputs(), pat.outputsOf( clout[i] ) );
    }
    return mseacc/dim;
}

void BackPropagationAlgo::addCluster( Cluster* cl, bool isOut ) {
    if( mapIndex.count( cl ) == 0 ) {
        cluster_deltas temp;
        int size = cl->numNeurons();
        temp.cluster = dynamic_cast<BiasedCluster*>(cl);
        temp.isOutput = isOut;
        temp.deltas_outputs.resize( size );
        temp.deltas_inputs.resize( size );
        temp.last_deltas_inputs.resize( size );
        cluster_deltas_vec.push_back( temp );
        mapIndex[cl] = cluster_deltas_vec.size()-1;
    }
}

void BackPropagationAlgo::addLinker( Linker* link ) {
    if ( mapIndex.count( link->to() ) == 0 ) {
        cluster_deltas temp;
        int size = link->to()->numNeurons();
        temp.cluster = dynamic_cast<BiasedCluster*>(link->to());
        temp.isOutput = false;
        temp.deltas_outputs.resize( size );
        temp.deltas_inputs.resize( size );
        temp.last_deltas_inputs.resize( size );
        temp.incoming_linkers_vec.push_back( dynamic_cast<MatrixLinker*>(link) );
        temp.incoming_last_outputs.push_back( DoubleVector( link->from()->numNeurons() ) );
        cluster_deltas_vec.push_back( temp );
        mapIndex[temp.cluster] = cluster_deltas_vec.size()-1;
    }
    else {
        int tmp = mapIndex[link->to()];
        cluster_deltas_vec[ tmp ].incoming_linkers_vec.push_back( dynamic_cast<MatrixLinker*>(link) );
        cluster_deltas_vec[ tmp ].incoming_last_outputs.push_back( DoubleVector( link->from()->numNeurons() ) );
    }
}

void BackPropagationAlgo::configure(ConfigurationParameters& params, QString prefix) {
    params.startRememberingGroupObjectAssociations();
    learn_rate = params.getValue( prefix + "rate" ).toDouble();
    momentumv = params.getValue( prefix + "momentum" ).toDouble();
    if ( momentumv == 0.0 ) {
        useMomentum = false;
    } else {
        useMomentum = true;
    }
    QString str = params.getValue( prefix + "order" );
    update_order.clear();
    if ( !str.isEmpty() ) {
        QStringList list = str.split(QRegExp("\\s+"), QString::SkipEmptyParts);
        foreach( QString upl, list ) {
            Updatable* up = params.getObjectFromGroup<Updatable>( upl, true );
            update_order << up;
        }
    }
    NeuralNet* net = params.getObjectFromParameter<NeuralNet>( prefix+"neuralnet", false, true );
    setNeuralNet( net );
    params.stopRememberingGroupObjectAssociations();
}

void BackPropagationAlgo::save(ConfigurationParameters& params, QString prefix) {
    params.startObjectParameters(prefix, "BackPropagationAlgo", this);
    params.createParameter( prefix, "neuralnet", neuralNet() );
    params.createParameter( prefix, "rate", QString::number(learn_rate) );
    if ( useMomentum ) {
        params.createParameter( prefix, "momentum", QString::number(momentumv) );
    }
    QStringList list;
    foreach( Updatable* up, update_order ) {
        list << up->name();
    }
    params.createParameter( prefix, "order", list.join(" ") );
    //--- save the neuralnet in the group corresponding to its name
    neuralNet()->save( params, neuralNet()->name() );
}

void BackPropagationAlgo::describe( QString type ) {
    Descriptor d = addTypeDescription( type, "Backpropagation Learning Algorithm" );
    d.describeObject( "neuralnet" ).type( "NeuralNet" ).props( IsMandatory ).help( "The neural network to learn by backpropagation" );
    d.describeReal( "rate" ).limits( 0.0, 1.0 ).def( 0.2 ).help( "The learning rate" );
    d.describeReal( "momentum" ).limits( 0.0, 1.0 ).help( "The momentum rate; if zero momentum will be disabled" );
}

}