// @(#)root/tmva $Id: MethodANNBase.h 40005 2011-06-27 15:29:10Z stelzer $ // Author: Andreas Hoecker, Peter Speckmayer, Matt Jachowski, Jan Therhaag /********************************************************************************** * Project: TMVA - a Root-integrated toolkit for multivariate data analysis * * Package: TMVA * * Class : MethodANNBase * * Web : http://tmva.sourceforge.net * * * * Description: * * Artificial neural network base class for the discrimination of signal * * from background. * * * * Authors (alphabetical): * * Andreas Hoecker - CERN, Switzerland * * Matt Jachowski - Stanford University, USA * * Peter Speckmayer - CERN, Switzerland * * Joerg Stelzer - CERN, Switzerland * * Jan Therhaag - U of Bonn, Germany * * * * Small changes (regression): * * Krzysztof Danielowski - IFJ PAN & AGH, Poland * * Kamil Kraszewski - IFJ PAN & UJ , Poland * * Maciej Kruk - IFJ PAN & AGH, Poland * * * * Copyright (c) 2005-2011: * * CERN, Switzerland * * * * Redistribution and use in source and binary forms, with or without * * modification, are permitted according to the terms listed in LICENSE * * (http://tmva.sourceforge.net/LICENSE) * **********************************************************************************/ #ifndef ROOT_TMVA_MethodANNBase #define ROOT_TMVA_MethodANNBase ////////////////////////////////////////////////////////////////////////// // // // MethodANNBase // // // // Base class for all TMVA methods using artificial neural networks // // // ////////////////////////////////////////////////////////////////////////// #ifndef ROOT_TString #include "TString.h" #endif #include #ifndef ROOT_TTree #include "TTree.h" #endif #ifndef ROOT_TObjArray #include "TObjArray.h" #endif #ifndef ROOT_TRandom3 #include "TRandom3.h" #endif #ifndef ROOT_TMVA_MethodBase #include "TMVA/MethodBase.h" #endif #ifndef ROOT_TMVA_TActivation #include "TMVA/TActivation.h" #endif #ifndef ROOT_TMVA_TNeuron #include "TMVA/TNeuron.h" #endif #ifndef ROOT_TMVA_TNeuronInput #include "TMVA/TNeuronInput.h" #endif class TH1; class TH1F; namespace TMVA { class MethodANNBase : public MethodBase { public: // constructors dictated by subclassing off of MethodBase MethodANNBase( const TString& jobName, Types::EMVA methodType, const TString& methodTitle, DataSetInfo& theData, const TString& theOption, TDirectory* theTargetDir ); MethodANNBase( Types::EMVA methodType, DataSetInfo& theData, const TString& theWeightFile, TDirectory* theTargetDir ); virtual ~MethodANNBase(); // this does the real initialization work void InitANNBase(); // setters for subclasses void SetActivation(TActivation* activation) { if (fActivation != NULL) delete fActivation; fActivation = activation; } void SetNeuronInputCalculator(TNeuronInput* inputCalculator) { if (fInputCalculator != NULL) delete fInputCalculator; fInputCalculator = inputCalculator; } // this will have to be overridden by every subclass virtual void Train() = 0; // print network, for debugging virtual void PrintNetwork() const; using MethodBase::ReadWeightsFromStream; // write weights to file void AddWeightsXMLTo( void* parent ) const; void ReadWeightsFromXML( void* wghtnode ); // read weights from file virtual void ReadWeightsFromStream( istream& istr ); // calculate the MVA value virtual Double_t GetMvaValue( Double_t* err = 0, Double_t* errUpper = 0 ); virtual const std::vector &GetRegressionValues(); virtual const std::vector &GetMulticlassValues(); // write method specific histos to target file virtual void WriteMonitoringHistosToFile() const; // ranking of input variables const Ranking* CreateRanking(); // the option handling methods virtual void DeclareOptions(); virtual void ProcessOptions(); Bool_t Debug() const; enum EEstimator { kMSE=0,kCE}; protected: virtual void MakeClassSpecific( std::ostream&, const TString& ) const; std::vector* ParseLayoutString( TString layerSpec ); virtual void BuildNetwork( std::vector* layout, std::vector* weights=NULL, Bool_t fromFile = kFALSE ); void ForceNetworkInputs( const Event* ev, Int_t ignoreIndex = -1 ); Double_t GetNetworkOutput() { return GetOutputNeuron()->GetActivationValue(); } // debugging utilities void PrintMessage( TString message, Bool_t force = kFALSE ) const; void ForceNetworkCalculations(); void WaitForKeyboard(); // accessors Int_t NumCycles() { return fNcycles; } TNeuron* GetInputNeuron(Int_t index) { return (TNeuron*)fInputLayer->At(index); } TNeuron* GetOutputNeuron( Int_t index = 0) { return fOutputNeurons.at(index); } // protected variables TObjArray* fNetwork; // TObjArray of TObjArrays representing network TObjArray* fSynapses; // array of pointers to synapses, no structural data TActivation* fActivation; // activation function to be used for hidden layers TActivation* fOutput; // activation function to be used for output layers, depending on estimator TActivation* fIdentity; // activation for input and output layers TRandom3* frgen; // random number generator for various uses TNeuronInput* fInputCalculator; // input calculator for all neurons std::vector fRegulatorIdx; //index to different priors from every synapses std::vector fRegulators; //the priors as regulator EEstimator fEstimator; TString fEstimatorS; // monitoring histograms TH1F* fEstimatorHistTrain; // monitors convergence of training sample TH1F* fEstimatorHistTest; // monitors convergence of independent test sample // monitoring histograms (not available for regression) void CreateWeightMonitoringHists( const TString& bulkname, std::vector* hv = 0 ) const; std::vector fEpochMonHistS; // epoch monitoring hitograms for signal std::vector fEpochMonHistB; // epoch monitoring hitograms for background std::vector fEpochMonHistW; // epoch monitoring hitograms for weights // general TMatrixD fInvHessian; // zjh bool fUseRegulator; // zjh protected: Int_t fRandomSeed; // random seed for initial synapse weights private: // helper functions for building network void BuildLayers(std::vector* layout, Bool_t from_file = false); void BuildLayer(Int_t numNeurons, TObjArray* curLayer, TObjArray* prevLayer, Int_t layerIndex, Int_t numLayers, Bool_t from_file = false); void AddPreLinks(TNeuron* neuron, TObjArray* prevLayer); // helper functions for weight initialization void InitWeights(); void ForceWeights(std::vector* weights); // helper functions for deleting network void DeleteNetwork(); void DeleteNetworkLayer(TObjArray*& layer); // debugging utilities void PrintLayer(TObjArray* layer) const; void PrintNeuron(TNeuron* neuron) const; // private variables Int_t fNcycles; // number of epochs to train TString fNeuronType; // name of neuron activation function class TString fNeuronInputType; // name of neuron input calculator class TObjArray* fInputLayer; // cache this for fast access std::vector fOutputNeurons; // cache this for fast access TString fLayerSpec; // layout specification option // some static flags static const Bool_t fgDEBUG = kTRUE; // debug flag ClassDef(MethodANNBase,0) // Base class for TMVA ANNs }; } // namespace TMVA #endif