LSSVMLearner.cpp

Go to the documentation of this file.

/*
 * Copyright (C) 2007-2011 RobotCub Consortium, European Commission FP6 Project IST-004370
 * author:  Arjan Gijsberts
 * email:   arjan.gijsberts@iit.it
 * website: www.robotcub.org
 * Permission is granted to copy, distribute, and/or modify this program
 * under the terms of the GNU General Public License, version 2 or any
 * later version published by the Free Software Foundation.
 *
 * A copy of the license can be found at
 * http://www.robotcub.org/icub/license/gpl.txt
 *
 * 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
 */
 
#include <cassert>
#include <sstream>
#include <cmath>
 
#include <yarp/math/Math.h>
#include <yarp/math/SVD.h>
 
#include "iCub/learningMachine/LSSVMLearner.h"
#include "iCub/learningMachine/Serialization.h"
 
using namespace yarp::math;
using namespace iCub::learningmachine::serialization;
 
namespace iCub {
namespace learningmachine {
 
double RBFKernel::evaluate(const yarp::sig::Vector& v1, const yarp::sig::Vector& v2) {
    assert(v1.size() == v2.size());
    double result = 0.0;
    double diff;
 
    for(size_t i = 0; i < v1.size(); i++) {
        diff = v1(i) - v2(i);
        result += diff * diff;
    }
    result *= -1 * this->gamma;
    return std::exp(result);
}
 
 
LSSVMLearner::LSSVMLearner(unsigned int dom, unsigned int cod, double c) {
    this->setName("LSSVM");
    this->kernel = new RBFKernel();
    // make sure to not use initialization list to constructor of base for
    // domain and codomain size, as it will not use overloaded mutators
    this->setDomainSize(dom);
    this->setCoDomainSize(cod);
    this->setC(c);
}
 
LSSVMLearner::LSSVMLearner(const LSSVMLearner& other)
  : IFixedSizeLearner(other), inputs(other.inputs), outputs(other.outputs),
    alphas(other.alphas), bias(other.bias), LOO(other.LOO), C(other.C),
    kernel(new RBFKernel(*other.kernel)) {
 
}
 
LSSVMLearner::~LSSVMLearner() {
    delete this->kernel;
}
 
LSSVMLearner& LSSVMLearner::operator=(const LSSVMLearner& other) {
    if(this == &other) return *this; // handle self initialization
 
    this->IFixedSizeLearner::operator=(other);
    this->inputs = other.inputs;
    this->outputs = other.outputs;
    this->alphas = other.alphas;
    this->bias = other.bias;
    this->LOO = other.LOO;
    this->C = other.C;
    delete this->kernel;
    this->kernel = new RBFKernel(*other.kernel);
 
    return *this;
}
 
void LSSVMLearner::feedSample(const yarp::sig::Vector& input, const yarp::sig::Vector& output) {
    // call parent method to let it do some validation for us
    this->IFixedSizeLearner::feedSample(input, output);
 
    this->inputs.push_back(input);
    this->outputs.push_back(output);
}
 
void LSSVMLearner::train() {
    assert(this->inputs.size() == this->outputs.size());
 
    // save wasting some time
    if(inputs.size() == 0) {
        return;
    }
 
    // create kernel matrix
    yarp::sig::Matrix K(inputs.size() + 1, inputs.size() + 1);
    for(int r = 0; r < K.rows() - 1; r++) {
        // symmetric matrix
        for(int c = 0; c <= r; c++) {
            K(r, c) = K(c, r) = this->kernel->evaluate(this->inputs[r], this->inputs[c]);
            if(r == c) K(r, c) += (1.0 / this->C);
        }
    }
    for(int i = 0; i < K.rows() - 1; i++) {
        K(i, K.cols() - 1) = K(K.rows() - 1, i) = 1.;
    }
    K(K.rows() - 1, K.cols() - 1) = 0.;
 
    // invert kernel matrix
    yarp::sig::Matrix Kinv = luinv(K);
 
    // compute solution
    yarp::sig::Matrix Y = zeros(this->outputs.size() + 1, this->getCoDomainSize());
    for(int r = 0; r < Y.rows() - 1; r++) {
        for(int c = 0; c < Y.cols(); c++) {
            Y(r, c) = this->outputs[r](c);
        }
    }
 
    yarp::sig::Matrix result = Kinv * Y;
    this->alphas = result.submatrix(0, result.rows() - 2, 0, result.cols() - 1);
    this->bias = result.getRow(result.rows() - 1);
 
    // compute LOO
    this->LOO = zeros(this->getCoDomainSize());
 
    for(unsigned int i = 0; i < this->getCoDomainSize(); i++) {
        yarp::sig::Vector alphas_i = this->alphas.getCol(i);
        for(size_t j = 0; j < alphas_i.size(); j++) {
            double err = alphas_i(j) / Kinv(j, j);
            this->LOO(i) += err * err;
        }
        this->LOO(i) /= alphas_i.size();
    }
 
}
 
Prediction LSSVMLearner::predict(const yarp::sig::Vector& input) {
    this->checkDomainSize(input);
 
    if(this->inputs.size() == 0) {
        return zeros(this->getCoDomainSize());
    }
 
    // compute kernel expansion
    yarp::sig::Vector k(this->inputs.size());
    for(size_t i = 0; i < k.size(); i++) {
        k(i) = this->kernel->evaluate(this->inputs[i], input);
    }
 
    return Prediction((this->alphas.transposed() * k) + this->bias);
}
 
void LSSVMLearner::reset() {
    this->inputs.clear();
    this->outputs.clear();
    this->alphas = yarp::sig::Matrix();
    this->LOO.clear();
    this->bias.clear();
}
 
LSSVMLearner* LSSVMLearner::clone() {
    return new LSSVMLearner(*this);
}
 
std::string LSSVMLearner::getInfo() {
    std::ostringstream buffer;
    buffer << this->IFixedSizeLearner::getInfo();
    buffer << "C: " << this->getC() << " | ";
    buffer << "Collected Samples: " << this->inputs.size() << " | ";
    buffer << "Training Samples: " << this->alphas.rows() << " | ";
    buffer << "Kernel: " << this->kernel->getInfo() << std::endl;
    buffer << "LOO: " << this->LOO.toString() << std::endl;
    return buffer.str();
}
 
std::string LSSVMLearner::getConfigHelp() {
    std::ostringstream buffer;
    buffer << this->IFixedSizeLearner::getConfigHelp();
    //buffer << "  kernel idx|all cfg    Kernel configuration" << std::endl;
    buffer << "  c val                 Tradeoff parameter C" << std::endl;
    buffer << this->kernel->getConfigHelp() << std::endl;
    return buffer.str();
}
 
void LSSVMLearner::writeBottle(yarp::os::Bottle& bot) {
    // write kernel gamma
    bot << this->kernel->getGamma() << this->getC() << this->bias
        << this->alphas;
 
    // write inputs
    for(unsigned int i = 0; i < this->inputs.size(); i++) {
        for(unsigned int d = 0; d < this->getDomainSize(); d++) {
            bot.addFloat64(this->inputs[i](d));
        }
    }
    bot.addInt32(this->inputs.size());
 
    // write outputs
    for(unsigned int i = 0; i < this->outputs.size(); i++) {
        for(unsigned int d = 0; d < this->getCoDomainSize(); d++) {
            bot.addFloat64(this->outputs[i](d));
        }
    }
    bot.addInt32(this->outputs.size());
 
    // make sure to call the superclass's method
    this->IFixedSizeLearner::writeBottle(bot);
}
 
void LSSVMLearner::readBottle(yarp::os::Bottle& bot) {
    // make sure to call the superclass's method
    this->IFixedSizeLearner::readBottle(bot);
 
    // read outputs
    this->outputs.resize(bot.pop().asInt32());
    for(int i = this->outputs.size() - 1; i >= 0; i--) {
        this->outputs[i].resize(this->getCoDomainSize());
        for(int d = this->getCoDomainSize() - 1; d >= 0; d--) {
            this->outputs[i](d) = bot.pop().asFloat64();
        }
    }
 
    // read inputs
    this->inputs.resize(bot.pop().asInt32());
    for(int i = this->inputs.size() - 1; i >= 0; i--) {
        this->inputs[i].resize(this->getDomainSize());
        for(int d = this->getDomainSize() - 1; d >= 0; d--) {
            this->inputs[i](d) = bot.pop().asFloat64();
        }
    }
 
    double c;
    double gamma;
    bot >> this->alphas >> this->bias >> c >> gamma;
    this->setC(c);
    this->kernel->setGamma(gamma);
}
 
void LSSVMLearner::setDomainSize(unsigned int size) {
    this->IFixedSizeLearner::setDomainSize(size);
}
 
void LSSVMLearner::setCoDomainSize(unsigned int size) {
    this->IFixedSizeLearner::setCoDomainSize(size);
    //this->initAll(this->getCoDomainSize());
}
 
bool LSSVMLearner::configure(yarp::os::Searchable& config) {
    bool success = this->IFixedSizeLearner::configure(config);
 
    // format: set c dbl
    if(config.find("c").isFloat64() || config.find("c").isInt32()) {
        double val = config.find("c").asFloat64();
        if(val > 0) {
            this->setC(config.find("c").asFloat64());
            success = true;
        }
    }
 
    success |= this->kernel->configure(config);
 
    return success;
}
 
} // learningmachine
} // iCub