ProcMatrix.cc

Go to the documentation of this file.

#include <cstring>
#include <vector>
#include <memory>

#include <xercesc/dom/DOM.hpp>

#include <TMatrixD.h>
#include <TMatrixF.h>
#include <TH2.h>

#include "FWCore/Utilities/interface/Exception.h"

#include "PhysicsTools/MVAComputer/interface/AtomicId.h"

#include "PhysicsTools/MVATrainer/interface/XMLDocument.h"
#include "PhysicsTools/MVATrainer/interface/MVATrainer.h"
#include "PhysicsTools/MVATrainer/interface/TrainProcessor.h"
#include "PhysicsTools/MVATrainer/interface/LeastSquares.h"

XERCES_CPP_NAMESPACE_USE

using namespace PhysicsTools;

namespace { // anonymous

class ProcMatrix : public TrainProcessor {
    public:
    typedef TrainProcessor::Registry<ProcMatrix>::Type Registry;

    ProcMatrix(const char *name, const AtomicId *id,
               MVATrainer *trainer);
    virtual ~ProcMatrix();

    virtual void configure(DOMElement *elem) override;
    virtual Calibration::VarProcessor *getCalibration() const override;

    virtual void trainBegin() override;
    virtual void trainData(const std::vector<double> *values,
                           bool target, double weight) override;
    virtual void trainEnd() override;

    virtual bool load() override;
    virtual void save() override;

    protected:
    virtual void *requestObject(const std::string &name) const override;

    private:
    enum Iteration {
        ITER_FILL,
        ITER_DONE
    } iteration;

    typedef std::pair<unsigned int, double> Rank;

    std::vector<Rank> ranking() const;

    std::unique_ptr<LeastSquares>   lsSignal, lsBackground;
    std::unique_ptr<LeastSquares>   ls;
    std::vector<double>     vars;
    bool                fillSignal;
    bool                fillBackground;
    bool                doNormalization;
    bool                doRanking;
};

static ProcMatrix::Registry registry("ProcMatrix");

ProcMatrix::ProcMatrix(const char *name, const AtomicId *id,
                             MVATrainer *trainer) :
    TrainProcessor(name, id, trainer),
    iteration(ITER_FILL), fillSignal(true), fillBackground(true),
    doRanking(false)
{
}

ProcMatrix::~ProcMatrix()
{
}

void ProcMatrix::configure(DOMElement *elem)
{
    ls.reset(new LeastSquares(getInputs().size()));

    DOMNode *node = elem->getFirstChild();
    while(node && node->getNodeType() != DOMNode::ELEMENT_NODE)
        node = node->getNextSibling();

    if (!node)
        return;

    if (std::strcmp(XMLSimpleStr(node->getNodeName()), "fill") != 0)
        throw cms::Exception("ProcMatrix")
                << "Expected fill tag in config section."
                << std::endl;

    elem = static_cast<DOMElement*>(node);

    fillSignal =
        XMLDocument::readAttribute<bool>(elem, "signal", false);
    fillBackground =
        XMLDocument::readAttribute<bool>(elem, "background", false);
    doNormalization =
        XMLDocument::readAttribute<bool>(elem, "normalize", false);

    doRanking = XMLDocument::readAttribute<bool>(elem, "ranking", false);
    if (doRanking)
        fillSignal = fillBackground = doNormalization = true;

    if (doNormalization && fillSignal && fillBackground) {
        lsSignal.reset(new LeastSquares(getInputs().size()));
        lsBackground.reset(new LeastSquares(getInputs().size()));
    }

    node = node->getNextSibling();
    while(node && node->getNodeType() != DOMNode::ELEMENT_NODE)
        node = node->getNextSibling();

    if (node)
        throw cms::Exception("ProcMatrix")
            << "Superfluous tags in config section."
            << std::endl;

    if (!fillSignal && !fillBackground)
        throw cms::Exception("ProcMatrix")
            << "Filling neither background nor signal in config."
            << std::endl;
}

Calibration::VarProcessor *ProcMatrix::getCalibration() const
{
    if (doRanking)
        return 0;

    Calibration::ProcMatrix *calib = new Calibration::ProcMatrix;

    unsigned int n = ls->getSize();
    const TMatrixD &rotation = ls->getRotation();

    calib->matrix.rows = n;
    calib->matrix.columns = n;

    for(unsigned int i = 0; i < n; i++)
        for(unsigned int j = 0; j < n; j++)
            calib->matrix.elements.push_back(rotation(j, i));

    return calib;
}

void ProcMatrix::trainBegin()
{
    if (iteration == ITER_FILL)
        vars.resize(ls->getSize());
}

void ProcMatrix::trainData(const std::vector<double> *values,
                           bool target, double weight)
{
    if (iteration != ITER_FILL)
        return;

    if (!(target ? fillSignal : fillBackground))
        return;

    LeastSquares *ls = target ? lsSignal.get() : lsBackground.get();
    if (!ls)
        ls = this->ls.get();

    for(unsigned int i = 0; i < ls->getSize(); i++, values++) {
        if (values->empty())
            throw cms::Exception("ProcMatrix")
                << "Variable \""
                << (const char*)getInputs().get()[i]->getName()
                << "\" is not set in ProcMatrix trainer."
                << std::endl;
        vars[i] = values->front();
    }

    ls->add(vars, target, weight);
}

void ProcMatrix::trainEnd()
{
    switch(iteration) {
        case ITER_FILL:
        vars.clear();
        if (lsSignal.get()) {
            unsigned int n = ls->getSize();
            double weight = lsSignal->getCoefficients()
                                (n + 1, n + 1);
            if (weight > 1.0e-9)
                ls->add(*lsSignal, 1.0 / weight);
            lsSignal.reset();
        }
        if (lsBackground.get()) {
            unsigned int n = ls->getSize();
            double weight = lsBackground->getCoefficients()
                                (n + 1, n + 1);
            if (weight > 1.0e-9)
                ls->add(*lsBackground, 1.0 / weight);
            lsBackground.reset();
        }
        ls->calculate();

        iteration = ITER_DONE;
        trained = true;
        break;

        default:
        /* shut up */;
    }

    if (iteration == ITER_DONE && monitoring) {
        TMatrixF matrix(ls->getCorrelations());
        TH2F *histo = monitoring->book<TH2F>("CorrMatrix", matrix);
        histo->SetNameTitle("CorrMatrix",
            (fillSignal && fillBackground)
            ? "correlation matrix (signal + background)"
            : (fillSignal ? "correlation matrix (signal)"
                          : "correlation matrix (background)"));

        std::vector<SourceVariable*> inputs = getInputs().get();
        for(std::vector<SourceVariable*>::const_iterator iter =
            inputs.begin(); iter != inputs.end(); ++iter) {

            unsigned int idx = iter - inputs.begin();
            SourceVariable *var = *iter;
            std::string name =
                (const char*)var->getSource()->getName()
                + std::string("_")
                + (const char*)var->getName();

            histo->GetXaxis()->SetBinLabel(idx + 1, name.c_str());
            histo->GetYaxis()->SetBinLabel(idx + 1, name.c_str());
            histo->GetXaxis()->SetBinLabel(inputs.size() + 1,
                                           "target");
            histo->GetYaxis()->SetBinLabel(inputs.size() + 1,
                                           "target");
        }
        histo->LabelsOption("d");
        histo->SetMinimum(-1.0);
        histo->SetMaximum(+1.0);

        if (!doRanking)
            return;

        std::vector<Rank> ranks = ranking();
        TVectorD rankVector(ranks.size());
        for(unsigned int i = 0; i < ranks.size(); i++)
            rankVector[i] = ranks[i].second;
        TH1F *rank = monitoring->book<TH1F>("Ranking", rankVector);
        rank->SetNameTitle("Ranking", "variable ranking");
        rank->SetYTitle("correlation to target");
        for(unsigned int i = 0; i < ranks.size(); i++) {
            unsigned int v = ranks[i].first;
            std::string name;
            SourceVariable *var = inputs[v];
            name = (const char*)var->getSource()->getName()
                   + std::string("_")
                   + (const char*)var->getName();
            rank->GetXaxis()->SetBinLabel(i + 1, name.c_str());
        }
    }
}

void *ProcMatrix::requestObject(const std::string &name) const
{
    if (name == "linearAnalyzer")
        return static_cast<void*>(ls.get());

    return 0;
}

bool ProcMatrix::load()
{
    std::string filename = trainer->trainFileName(this, "xml");
    if (!exists(filename))
        return false;

    XMLDocument xml(filename);
    DOMElement *elem = xml.getRootNode();
    if (std::strcmp(XMLSimpleStr(elem->getNodeName()), "ProcMatrix") != 0)
        throw cms::Exception("ProcMatrix")
            << "XML training data file has bad root node."
            << std::endl;

    DOMNode *node = elem->getFirstChild();
    while(node && node->getNodeType() != DOMNode::ELEMENT_NODE)
        node = node->getNextSibling();

    if (!node)
        throw cms::Exception("ProcMatrix")
            << "Train data file empty." << std::endl;

    ls->load(static_cast<DOMElement*>(node));

    node = elem->getNextSibling();
    while(node && node->getNodeType() != DOMNode::ELEMENT_NODE)
        node = node->getNextSibling();

    if (node)
        throw cms::Exception("ProcMatrix")
            << "Train data file contains superfluous tags."
            << std::endl;

    iteration = ITER_DONE;
    trained = true;
    return true;
}

void ProcMatrix::save()
{
    XMLDocument xml(trainer->trainFileName(this, "xml"), true);
    DOMDocument *doc = xml.createDocument("ProcMatrix");

    xml.getRootNode()->appendChild(ls->save(doc));
}

static void maskLine(TMatrixDSym &m, unsigned int line)
{
    unsigned int n = m.GetNrows();
    for(unsigned int i = 0; i < n; i++)
        m(i, line) = m(line, i) = 0.;
    m(line, line) = 1.;
}

static void restoreLine(TMatrixDSym &m, TMatrixDSym &o, unsigned int line)
{
    unsigned int n = m.GetNrows();
    for(unsigned int i = 0; i < n; i++) {
        m(i, line) = o(i, line);
        m(line, i) = o(line, i);
    }
}

static double targetCorrelation(const TMatrixDSym &coeffs,
                                const std::vector<bool> &use)
{
    unsigned int n = coeffs.GetNrows() - 2;
    
    TVectorD weights = LeastSquares::solveFisher(coeffs);
    weights.ResizeTo(n + 2);
    weights[n + 1] = weights[n];
    weights[n] = 0.;

    double v1 = 0.;
    double v2 = 0.;
    double v3 = coeffs(n, n);
    double N = coeffs(n + 1, n + 1);
    double M = 0.;
    for(unsigned int i = 0; i < n + 2; i++) {
        if (i < n && !use[n])
            continue;
        double w = weights[i];
        for(unsigned int j = 0; j < n + 2; j++) {
            if (i < n && !use[n])
                continue;
            v1 += w * weights[j] * coeffs(i, j);
        }
        v2 += w * coeffs(i, n);
        M += w * coeffs(i, n + 1);
    }

    double c1 = v1 * N - M * M;
    double c2 = v2 * N - M * coeffs(n + 1, n);
    double c3 = v3 * N - coeffs(n + 1, n) * coeffs(n + 1, n);

    double c = c1 * c3;
    return (c > 1.0e-9) ? c2 / std::sqrt(c) : 0.0;
}

std::vector<ProcMatrix::Rank> ProcMatrix::ranking() const
{
    TMatrixDSym coeffs = ls->getCoefficients();
    unsigned int n = coeffs.GetNrows() - 2;

    typedef std::pair<unsigned int, double> Rank;
    std::vector<Rank> ranking;
    std::vector<bool> use(n, true);

    double corr = targetCorrelation(coeffs, use);

    for(unsigned int nVars = n; nVars > 1; nVars--) {
        double bestCorr = -99999.0;
        unsigned int bestIdx = n;
        TMatrixDSym origCoeffs = coeffs;

        for(unsigned int i = 0; i < n; i++) {
            if (!use[i])
                continue;

            use[i] = false;
            maskLine(coeffs, i);
            double newCorr = targetCorrelation(coeffs, use);
            use[i] = true;
            restoreLine(coeffs, origCoeffs, i);

            if (newCorr > bestCorr) {
                bestCorr = newCorr;
                bestIdx = i;
            }
        }

        ranking.push_back(Rank(bestIdx, corr));
        corr = bestCorr;
        use[bestIdx] = false;
        maskLine(coeffs, bestIdx);
    }

    for(unsigned int i = 0; i < n; i++)
        if (use[i])
            ranking.push_back(Rank(i, corr));

    return ranking;
}

} // anonymous namespace