PunchThroughClassifier.cxx

Go to the documentation of this file.

/*
  Copyright (C) 2002-2024 CERN for the benefit of the ATLAS collaboration
*/
 
#include "PunchThroughClassifier.h"
 
#include <fstream>
#include <memory>
 
// PathResolver
#include "PathResolver/PathResolver.h"
 
//LWTNN
#include "lwtnn/parse_json.hh"
 
//libXML
#include <libxml/xmlmemory.h>
#include <libxml/parser.h>
#include <libxml/tree.h>
#include <libxml/xmlreader.h>
#include <libxml/xpath.h>
#include <libxml/xpathInternals.h>
 
#include "CxxUtils/libxml2Helper.h"
 
using namespace CxxUtils;
 
ISF::PunchThroughClassifier::PunchThroughClassifier(const std::string& type, const std::string& name, const IInterface*  parent)
    : base_class(type, name, parent) {
 
    declareProperty( "ScalerConfigFileName",          m_scalerConfigFileName   );
    declareProperty( "NetworkConfigFileName",          m_networkConfigFileName   );
    declareProperty( "CalibratorConfigFileName",          m_calibratorConfigFileName   );
}
 
double ISF::PunchThroughClassifier::computePunchThroughProbability(const ISF::ISFParticle &isfp, const TFCSSimulationState& simulstate) const {
 
    std::map<std::string, std::map<std::string, double> > networkInputs = computeInputs(isfp, simulstate); //compute inputs
 
    networkInputs = scaleInputs(networkInputs); //scale inputs
 
    std::map<std::string, double> networkOutputs = m_graph->compute(networkInputs); //call neural network on inputs
 
    double calibratedOutput = calibrateOutput(networkOutputs["out_0"]); //calibrate neural network output
 
    return calibratedOutput;
}
 
 
StatusCode ISF::PunchThroughClassifier::initialize(){
 
    ATH_MSG_VERBOSE( "[ punchthroughclassifier ] initialize()" );
 
    std::string resolvedScalerFileName = PathResolverFindCalibFile (m_scalerConfigFileName);
    if ( initializeScaler(resolvedScalerFileName) != StatusCode::SUCCESS)
    {
        ATH_MSG_ERROR("[ punchthroughclassifier ] unable to load punchthroughclassifier input scaler");
    }
 
    std::string resolvedNetworkFileName = PathResolverFindCalibFile (m_networkConfigFileName);
    if ( initializeNetwork(resolvedNetworkFileName) != StatusCode::SUCCESS)
    {
        ATH_MSG_ERROR("[ punchthroughclassifier ] unable to load punchthroughclassifier network");
    }
 
    std::string resolvedCalibratorFileName = PathResolverFindCalibFile (m_calibratorConfigFileName);
    if ( initializeCalibrator(resolvedCalibratorFileName) != StatusCode::SUCCESS)
    {
        ATH_MSG_ERROR("[ punchthroughclassifier ] unable to load punchthroughclassifier calibrator");
    }
 
    return StatusCode::SUCCESS;
}
 
StatusCode ISF::PunchThroughClassifier::initializeScaler(const std::string & scalerConfigFile){
 
    //parse xml that contains config for MinMaxScaler for each of the network inputs
 
    xmlDocPtr doc = xmlParseFile( scalerConfigFile.c_str() );
 
    ATH_MSG_INFO( "[ punchthroughclassifier ] Loading scaler: " << scalerConfigFile);
 
    for( xmlNodePtr nodeRoot = doc->children; nodeRoot != nullptr; nodeRoot = nodeRoot->next) {
 
        if (xmlStrEqual( nodeRoot->name, BAD_CAST "Transformations" )) {
            for( xmlNodePtr nodeTransform = nodeRoot->children; nodeTransform != nullptr; nodeTransform = nodeTransform->next ) {
 
                //Get min and max values that we normalise values to
                if (xmlStrEqual( nodeTransform->name, BAD_CAST "ScalerValues" )) {
                    m_scalerMin = GetXmlAttr<double>( nodeTransform, "min" );
                    m_scalerMax = GetXmlAttr<double>( nodeTransform, "max" );
                }
 
                //Get values necessary to normalise each input variable
                if (xmlStrEqual( nodeTransform->name, BAD_CAST "VarScales" )) {
                    std::string name = GetXmlAttr<std::string>( nodeTransform, "name" );
                    double min = GetXmlAttr<double>( nodeTransform, "min" );
                    double max = GetXmlAttr<double>( nodeTransform, "max" );
                    m_scalerMinMap.emplace ( name, min );
                    m_scalerMaxMap.emplace ( name, max );
                }
            }
        }
    }
 
    return StatusCode::SUCCESS;
}
 
StatusCode ISF::PunchThroughClassifier::initializeNetwork(const std::string & networkConfigFile){
 
    ATH_MSG_INFO( "[ punchthroughclassifier ] Loading classifier: " << networkConfigFile);
 
    std::ifstream input(networkConfigFile);
    if(!input){
        ATH_MSG_ERROR("Could not find json file " << networkConfigFile );
        return StatusCode::FAILURE;
    }
 
    m_graph  = std::make_unique<lwt::LightweightGraph>(lwt::parse_json_graph(input));
    if(!m_graph){
        ATH_MSG_ERROR("Could not parse graph json file " << networkConfigFile );
        return StatusCode::FAILURE;
    }
 
 
    return StatusCode::SUCCESS;
}
 
 
StatusCode ISF::PunchThroughClassifier::initializeCalibrator(const std::string & calibratorConfigFile){
 
    //parse xml that contains config for isotonic regressor used to calibrate the network output
    ATH_MSG_INFO( "[ punchthroughclassifier ] Loading calibrator: " << calibratorConfigFile);
 
    xmlDocPtr doc = xmlParseFile( calibratorConfigFile.c_str() );
 
    for( xmlNodePtr nodeRoot = doc->children; nodeRoot != nullptr; nodeRoot = nodeRoot->next) {
 
        if (xmlStrEqual( nodeRoot->name, BAD_CAST "Transformations" )) {
            for( xmlNodePtr nodeTransform = nodeRoot->children; nodeTransform != nullptr; nodeTransform = nodeTransform->next ) {
 
                //get lower and upper bounds of isotonic regressor
                if (xmlStrEqual( nodeTransform->name, BAD_CAST "LimitValues" )) {
                    m_calibrationMin = GetXmlAttr<double>( nodeTransform, "min" );
                    m_calibrationMax = GetXmlAttr<double>( nodeTransform, "max" );
                }
 
                //get defined points where isotonic regressor knows transform
                if (xmlStrEqual( nodeTransform->name, BAD_CAST "LinearNorm" )) {
                    double orig = GetXmlAttr<double>( nodeTransform, "orig" );
                    double norm = GetXmlAttr<double>( nodeTransform, "norm" );
                    m_calibrationMap.emplace ( orig, norm );
                }
            }
        }
    }
 
    return StatusCode::SUCCESS;
}
 
std::map<std::string, std::map<std::string, double> > ISF::PunchThroughClassifier::computeInputs(const ISF::ISFParticle &isfp, const TFCSSimulationState& simulstate) {
 
    //calculate inputs for NN
 
    std::map<std::string, std::map<std::string, double> > networkInputs;
 
    //add initial particle and total energy variables
    networkInputs["node_0"] = {
        {"variable_0", isfp.momentum().mag() },
        {"variable_1", std::abs(isfp.position().eta()) },
        {"variable_2", isfp.position().phi() },
        {"variable_3", simulstate.E()},
    };
 
    //add energy fraction variables
    for (unsigned int i = 0; i < 24; i++){
        networkInputs["node_0"].insert({"variable_" + std::to_string(i + 4), simulstate.Efrac(i)});
    }
 
    return networkInputs;
}
 
std::map<std::string, std::map<std::string, double> > ISF::PunchThroughClassifier::scaleInputs(std::map<std::string, std::map<std::string, double> >& inputs) const{
 
    //apply MinMaxScaler to network inputs
 
    for (auto& var : inputs["node_0"]) {
 
        double x_std;
        if(m_scalerMaxMap.at(var.first) != m_scalerMinMap.at(var.first)){
            x_std = (var.second - m_scalerMinMap.at(var.first)) / (m_scalerMaxMap.at(var.first) - m_scalerMinMap.at(var.first));
        }
        else{
            x_std = (var.second - m_scalerMinMap.at(var.first));
        }
        var.second = x_std * (m_scalerMax - m_scalerMin) + m_scalerMin;
    }
 
    return inputs;
}
 
double ISF::PunchThroughClassifier::calibrateOutput(double& networkOutput) const {
 
    //calibrate output of network using isotonic regressor model
 
    //if network output is outside of the range of isotonic regressor then return min and max values
    if (networkOutput < m_calibrationMin){
        return m_calibrationMin;
    }
    else if (networkOutput > m_calibrationMax){
        return m_calibrationMax;
    }
 
    //otherwise find neighbouring points in isotonic regressor
    auto upper = m_calibrationMap.upper_bound(networkOutput);
    auto lower = upper--;
 
    //Perform linear interpolation between points
    double m = (upper->second - lower->second)/(upper->first - lower->first);
    double c = lower->second - m * lower->first;
    double calibrated = m * networkOutput + c;
 
    return calibrated;
}