source: trunk/MagicSoft/Mars/mranforest/MRanForest.cc@ 7396

/* ======================================================================== *\
!
! *
! * This file is part of MARS, the MAGIC Analysis and Reconstruction
! * Software. It is distributed to you in the hope that it can be a useful
! * and timesaving tool in analysing Data of imaging Cerenkov telescopes.
! * It is distributed WITHOUT ANY WARRANTY.
! *
! * Permission to use, copy, modify and distribute this software and its
! * documentation for any purpose is hereby granted without fee,
! * provided that the above copyright notice appear in all copies and
! * that both that copyright notice and this permission notice appear
! * in supporting documentation. It is provided "as is" without express
! * or implied warranty.
! *
!
!
!   Author(s): Thomas Hengstebeck 3/2003 <mailto:hengsteb@physik.hu-berlin.de>
!
!   Copyright: MAGIC Software Development, 2000-2005
!
!
\* ======================================================================== */

/////////////////////////////////////////////////////////////////////////////
//
// MRanForest
//
// ParameterContainer for Forest structure
//
// A random forest can be grown by calling GrowForest.
// In advance SetupGrow must be called in order to initialize arrays and
// do some preprocessing.
// GrowForest() provides the training data for a single tree (bootstrap
// aggregate procedure)
//
// Essentially two random elements serve to provide a "random" forest,
// namely bootstrap aggregating (which is done in GrowForest()) and random
// split selection (which is subject to MRanTree::GrowTree())
//
/////////////////////////////////////////////////////////////////////////////
#include "MRanForest.h"

#include <TVector.h>
#include <TRandom.h>

#include "MHMatrix.h"
#include "MRanTree.h"
#include "MData.h"
#include "MDataArray.h"
#include "MParList.h"

#include "MLog.h"
#include "MLogManip.h"

ClassImp(MRanForest);

using namespace std;

// --------------------------------------------------------------------------
//
// Default constructor.
//
MRanForest::MRanForest(const char *name, const char *title) : fClassify(1), fNumTrees(100), fNumTry(0), fNdSize(1), fRanTree(NULL), fUserVal(-1)
{
    fName  = name  ? name  : "MRanForest";
    fTitle = title ? title : "Storage container for Random Forest";

    fForest=new TObjArray();
    fForest->SetOwner(kTRUE);
}

MRanForest::MRanForest(const MRanForest &rf)
{
    // Copy constructor
    fName  = rf.fName;
    fTitle = rf.fTitle;

    fClassify = rf.fClassify;
    fNumTrees = rf.fNumTrees;
    fNumTry   = rf.fNumTry;
    fNdSize   = rf.fNdSize;
    fTreeNo   = rf.fTreeNo;
    fRanTree  = NULL;

    fRules=new MDataArray();
    fRules->Reset();

    MDataArray *newrules=rf.fRules;

    for(Int_t i=0;i<newrules->GetNumEntries();i++)
    {
        MData &data=(*newrules)[i];
        fRules->AddEntry(data.GetRule());
    }

    // trees
    fForest=new TObjArray();
    fForest->SetOwner(kTRUE);

    TObjArray *newforest=rf.fForest;
    for(Int_t i=0;i<newforest->GetEntries();i++)
    {
        MRanTree *rantree=(MRanTree*)newforest->At(i);

        MRanTree *newtree=new MRanTree(*rantree);
        fForest->Add(newtree);
    }

    fHadTrue  = rf.fHadTrue;
    fHadEst   = rf.fHadEst;
    fDataSort = rf.fDataSort;
    fDataRang = rf.fDataRang;
    fClassPop = rf.fClassPop;
    fWeight   = rf.fWeight;
    fTreeHad  = rf.fTreeHad;

    fNTimesOutBag = rf.fNTimesOutBag;

}

// --------------------------------------------------------------------------
// Destructor. 
MRanForest::~MRanForest()
{
    delete fForest;
}

MRanTree *MRanForest::GetTree(Int_t i)
{
    return (MRanTree*)(fForest->At(i));
}

void MRanForest::SetNumTrees(Int_t n)
{
    //at least 1 tree
    fNumTrees=TMath::Max(n,1);
    fTreeHad.Set(fNumTrees);
    fTreeHad.Reset();
}

void MRanForest::SetNumTry(Int_t n)
{
    fNumTry=TMath::Max(n,0);
}

void MRanForest::SetNdSize(Int_t n)
{
    fNdSize=TMath::Max(n,1);
}

void MRanForest::SetWeights(const TArrayF &weights)
{
    const int n=weights.GetSize();
    fWeight.Set(n);
    fWeight=weights;

    return;
}

void MRanForest::SetGrid(const TArrayD &grid)
{
    const int n=grid.GetSize();

    for(int i=0;i<n-1;i++)
        if(grid[i]>=grid[i+1])
        {
            *fLog<<inf<<"Grid points must be in increasing order! Ignoring grid."<<endl;
            return;
        }

    fGrid=grid;

    //*fLog<<inf<<"Following "<<n<<" grid points are used:"<<endl;
    //for(int i=0;i<n;i++)
    //    *fLog<<inf<<" "<<i<<") "<<fGrid[i]<<endl;

    return;
}

Int_t MRanForest::GetNumDim() const
{
    return fMatrix ? fMatrix->GetNcols() : 0;
}

Int_t MRanForest::GetNumData() const
{
    return fMatrix ? fMatrix->GetNrows() : 0;
}

Int_t MRanForest::GetNclass() const
{
    int maxidx = TMath::LocMax(fClass.GetSize(),fClass.GetArray());

    return int(fClass[maxidx])+1;
}

void MRanForest::PrepareClasses()
{
    const int numdata=fHadTrue.GetSize();

    if(fGrid.GetSize()>0)
    {
        // classes given by grid
        const int ngrid=fGrid.GetSize();

        for(int j=0;j<numdata;j++)
        {
            // Array is supposed  to be sorted prior to this call.
            // If match is found, function returns position of element.
            // If no match found, function gives nearest element smaller
            // than value.
            int k=TMath::BinarySearch(ngrid, fGrid.GetArray(), fHadTrue[j]);

            fClass[j]   = k;
        }

        int minidx = TMath::LocMin(fClass.GetSize(),fClass.GetArray());
        int min = fClass[minidx];
        if(min!=0) for(int j=0;j<numdata;j++)fClass[j]-=min;

    }else{
        // classes directly given
        for (Int_t j=0;j<numdata;j++)
            fClass[j] = int(fHadTrue[j]+0.5);
    }

    return;
}

/*
Bool_t MRanForest::PreProcess(MParList *plist)
{
    if (!fRules)
    {
        *fLog << err << dbginf << "MDataArray with rules not initialized... aborting." << endl;
        return kFALSE;
    }

    if (!fRules->PreProcess(plist))
    {
        *fLog << err << dbginf << "PreProcessing of MDataArray failed... aborting." << endl;
        return kFALSE;
    }

    return kTRUE;
}
*/

Double_t MRanForest::CalcHadroness()
{
    TVector event;
    *fRules >> event;

    return CalcHadroness(event);
}

Double_t MRanForest::CalcHadroness(const TVector &event)
{
    Double_t hadroness=0;
    Int_t ntree=0;

    TIter Next(fForest);

    MRanTree *tree;
    while ((tree=(MRanTree*)Next()))
    {
        hadroness+=(fTreeHad[ntree]=tree->TreeHad(event));
        ntree++;
    }
    return hadroness/ntree;
}

Bool_t MRanForest::AddTree(MRanTree *rantree=NULL)
{
    fRanTree = rantree ? rantree:fRanTree;

    if (!fRanTree) return kFALSE;

    MRanTree *newtree=new MRanTree(*fRanTree);
    fForest->Add(newtree);

    return kTRUE;
}

Bool_t MRanForest::SetupGrow(MHMatrix *mat,MParList *plist)
{
    //-------------------------------------------------------------------
    // access matrix, copy last column (target) preliminarily
    // into fHadTrue
    TMatrix mat_tmp = mat->GetM();
    int dim         = mat_tmp.GetNcols();
    int numdata     = mat_tmp.GetNrows();

    fMatrix=new TMatrix(mat_tmp);

    fHadTrue.Set(numdata);
    fHadTrue.Reset(0);

    for (Int_t j=0;j<numdata;j++)
        fHadTrue[j] = (*fMatrix)(j,dim-1);

    // remove last col
    fMatrix->ResizeTo(numdata,dim-1);
    dim=fMatrix->GetNcols();

    //-------------------------------------------------------------------
    // setup labels for classification/regression
    fClass.Set(numdata);
    fClass.Reset(0);

    if(fClassify) PrepareClasses();

    //-------------------------------------------------------------------
    // allocating and initializing arrays
    fHadEst.Set(numdata);       fHadEst.Reset(0);
    fNTimesOutBag.Set(numdata); fNTimesOutBag.Reset(0);
    fDataSort.Set(dim*numdata); fDataSort.Reset(0);
    fDataRang.Set(dim*numdata); fDataRang.Reset(0);

    if(fWeight.GetSize()!=numdata)
    {
        fWeight.Set(numdata);
        fWeight.Reset(1.);
        *fLog << inf <<"Setting weights to 1 (no weighting)"<< endl;
    }

    //-------------------------------------------------------------------
    // setup rules to be used for classification/regression
    MDataArray *allrules=(MDataArray*)mat->GetColumns();
    if(allrules==NULL)
    {
        *fLog << err <<"Rules of matrix == null, exiting"<< endl;
        return kFALSE;
    }

    fRules=new MDataArray(); fRules->Reset();
    TString target_rule;

    for(Int_t i=0;i<dim+1;i++)
    {
        MData &data=(*allrules)[i];
        if(i<dim)
            fRules->AddEntry(data.GetRule());
        else
            target_rule=data.GetRule();
    }

    *fLog << inf <<endl;
    *fLog << inf <<"Setting up RF for training on target:"<<endl<<" "<<target_rule.Data()<<endl;
    *fLog << inf <<"Following rules are used as input to RF:"<<endl;

    for(Int_t i=0;i<dim;i++)
    {
        MData &data=(*fRules)[i];
        *fLog<<inf<<" "<<i<<") "<<data.GetRule()<<endl<<flush;
    }

    *fLog << inf <<endl;

    //-------------------------------------------------------------------
    // prepare (sort) data for fast optimization algorithm
    if(!CreateDataSort()) return kFALSE;

    //-------------------------------------------------------------------
    // access and init tree container
    fRanTree = (MRanTree*)plist->FindCreateObj("MRanTree");

    if(!fRanTree)
    {
        *fLog << err << dbginf << "MRanForest, fRanTree not initialized... aborting." << endl;
        return kFALSE;
    }

    fRanTree->SetClassify(fClassify);
    fRanTree->SetNdSize(fNdSize);

    if(fNumTry==0)
    {
        double ddim = double(dim);

        fNumTry=int(sqrt(ddim)+0.5);
        *fLog<<inf<<endl;
        *fLog<<inf<<"Set no. of trials to the recommended value of round("<<sqrt(ddim)<<") = ";
        *fLog<<inf<<fNumTry<<endl;

    }
    fRanTree->SetNumTry(fNumTry);

    *fLog<<inf<<endl;
    *fLog<<inf<<"Following settings for the tree growing are used:"<<endl;
    *fLog<<inf<<" Number of Trees : "<<fNumTrees<<endl;
    *fLog<<inf<<" Number of Trials: "<<fNumTry<<endl;
    *fLog<<inf<<" Final Node size : "<<fNdSize<<endl;

    fTreeNo=0;

    return kTRUE;
}

Bool_t MRanForest::GrowForest()
{
    if(!gRandom)
    {
        *fLog << err << dbginf << "gRandom not initialized... aborting." << endl;
        return kFALSE;
    }

    fTreeNo++;

    //-------------------------------------------------------------------
    // initialize running output

    float minfloat=fHadTrue[TMath::LocMin(fHadTrue.GetSize(),fHadTrue.GetArray())];
    Bool_t calcResolution=(minfloat>0.001);

    if (fTreeNo==1)
    {
        *fLog << inf << endl;
        *fLog << underline;

        if(calcResolution)
            *fLog << "no. of tree    no. of nodes    resolution in % (from oob-data -> overest. of error)" << endl;
        else
            *fLog << "no. of tree    no. of nodes    rms in % (from oob-data -> overest. of error)" << endl;
                     //        12345678901234567890123456789012345678901234567890
    }

    const Int_t numdata = GetNumData();
    const Int_t nclass  = GetNclass();

    //-------------------------------------------------------------------
    // bootstrap aggregating (bagging) -> sampling with replacement:

    TArrayF classpopw(nclass);
    TArrayI jinbag(numdata); // Initialization includes filling with 0
    TArrayF winbag(numdata); // Initialization includes filling with 0

    float square=0; float mean=0;

    for (Int_t n=0; n<numdata; n++)
    {
        // The integer k is randomly (uniformly) chosen from the set
        // {0,1,...,numdata-1}, which is the set of the index numbers of
        // all events in the training sample
  
        const Int_t k = Int_t(gRandom->Rndm()*numdata);

        if(fClassify)
            classpopw[fClass[k]]+=fWeight[k];
        else
            classpopw[0]+=fWeight[k];

        mean  +=fHadTrue[k]*fWeight[k];
        square+=fHadTrue[k]*fHadTrue[k]*fWeight[k];
  
        winbag[k]+=fWeight[k];
        jinbag[k]=1;

    }

    //-------------------------------------------------------------------
    // modifying sorted-data array for in-bag data:

    // In bagging procedure ca. 2/3 of all elements in the original
    // training sample are used to build the in-bag data
    TArrayI datsortinbag=fDataSort;
    Int_t ninbag=0;

    ModifyDataSort(datsortinbag, ninbag, jinbag);

    fRanTree->GrowTree(fMatrix,fHadTrue,fClass,datsortinbag,fDataRang,classpopw,mean,square,
                       jinbag,winbag,nclass);

    //-------------------------------------------------------------------
    // error-estimates from out-of-bag data (oob data):
    //
    // For a single tree the events not(!) contained in the bootstrap sample of
    // this tree can be used to obtain estimates for the classification error of
    // this tree.
    // If you take a certain event, it is contained in the oob-data of 1/3 of
    // the trees (see comment to ModifyData). This means that the classification error
    // determined from oob-data is underestimated, but can still be taken as upper limit.

    for (Int_t ievt=0;ievt<numdata;ievt++)
    {
        if (jinbag[ievt]>0) continue;

        fHadEst[ievt] +=fRanTree->TreeHad((*fMatrix), ievt);
        fNTimesOutBag[ievt]++;

    }

    Int_t n=0;
    double ferr=0;

    for (Int_t ievt=0;ievt<numdata;ievt++)
    {
        if(fNTimesOutBag[ievt]!=0)
        {
            float val = fHadEst[ievt]/float(fNTimesOutBag[ievt])-fHadTrue[ievt];
            if(calcResolution) val/=fHadTrue[ievt];

            ferr += val*val;
            n++;
        }
    }
    ferr = TMath::Sqrt(ferr/n);

    //-------------------------------------------------------------------
    // give running output
    *fLog << inf << setw(5)  << fTreeNo;
    *fLog << inf << setw(20) << fRanTree->GetNumEndNodes();
    *fLog << inf << Form("%20.2f", ferr*100.);
    *fLog << inf << endl;

    // adding tree to forest
    AddTree();

    return fTreeNo<fNumTrees;
}

Bool_t MRanForest::CreateDataSort()
{
    // fDataSort(m,n) is the event number in which fMatrix(m,n) occurs.
    // fDataRang(m,n) is the rang of fMatrix(m,n), i.e. if rang = r:
    //   fMatrix(m,n) is the r-th highest value of all fMatrix(m,.).
    //
    // There may be more then 1 event with rang r (due to bagging).

    const Int_t numdata = GetNumData();
    const Int_t dim = GetNumDim();

    TArrayF v(numdata);
    TArrayI isort(numdata);


    for (Int_t mvar=0;mvar<dim;mvar++)
    {

        for(Int_t n=0;n<numdata;n++)
        {
            v[n]=(*fMatrix)(n,mvar);
            isort[n]=n;

            if(TMath::IsNaN(v[n]))
            {
                *fLog << err <<"Event no. "<<n<<", matrix column no. "<<mvar;
                *fLog << err <<" has the value NaN."<<endl;
                return kFALSE;
            }
        }

        TMath::Sort(numdata,v.GetArray(),isort.GetArray(),kFALSE);

        // this sorts the v[n] in ascending order. isort[n] is the event number
        // of that v[n], which is the n-th from the lowest (assume the original
        // event numbers are 0,1,...).

        // control sorting
        for(int n=1;n<numdata;n++)
            if(v[isort[n-1]]>v[isort[n]])
            {
                *fLog << err <<"Event no. "<<n<<", matrix column no. "<<mvar;
                *fLog << err <<" not at correct sorting position."<<endl;
                return kFALSE;
            }

        for(Int_t n=0;n<numdata-1;n++)
        {
            const Int_t n1=isort[n];
            const Int_t n2=isort[n+1];

            fDataSort[mvar*numdata+n]=n1;
            if(n==0) fDataRang[mvar*numdata+n1]=0;
            if(v[n1]<v[n2])
            {
                fDataRang[mvar*numdata+n2]=fDataRang[mvar*numdata+n1]+1;
            }else{
                fDataRang[mvar*numdata+n2]=fDataRang[mvar*numdata+n1];
            }
        }
        fDataSort[(mvar+1)*numdata-1]=isort[numdata-1];
    }
    return kTRUE;
}

void MRanForest::ModifyDataSort(TArrayI &datsortinbag, Int_t ninbag, const TArrayI &jinbag)
{
    const Int_t numdim=GetNumDim();
    const Int_t numdata=GetNumData();

    ninbag=0;
    for (Int_t n=0;n<numdata;n++)
        if(jinbag[n]==1) ninbag++;

    for(Int_t m=0;m<numdim;m++)
    {
        Int_t k=0;
        Int_t nt=0;
        for(Int_t n=0;n<numdata;n++)
        {
            if(jinbag[datsortinbag[m*numdata+k]]==1)
            {
                datsortinbag[m*numdata+nt]=datsortinbag[m*numdata+k];
                k++;
            }else{
                for(Int_t j=1;j<numdata-k;j++)
                {
                    if(jinbag[datsortinbag[m*numdata+k+j]]==1)
                    {
                        datsortinbag[m*numdata+nt]=datsortinbag[m*numdata+k+j];
                        k+=j+1;
                        break;
                    }
                }
            }
            nt++;
            if(nt>=ninbag) break;
        }
    }
}

Bool_t MRanForest::AsciiWrite(ostream &out) const
{
    Int_t n=0;
    MRanTree *tree;
    TIter forest(fForest);

    while ((tree=(MRanTree*)forest.Next()))
    {
        tree->AsciiWrite(out);
        n++;
    }

    return n==fNumTrees;
}