source: trunk/MagicSoft/Mars/mranforest/MRanForest.cc@ 7684

Last change on this file since 7684 was 7684, checked in by tbretz, 18 years ago
*** empty log message ***
File size: 17.8 KB
Line 
1/* ======================================================================== *\
2!
3! *
4! * This file is part of MARS, the MAGIC Analysis and Reconstruction
5! * Software. It is distributed to you in the hope that it can be a useful
6! * and timesaving tool in analysing Data of imaging Cerenkov telescopes.
7! * It is distributed WITHOUT ANY WARRANTY.
8! *
9! * Permission to use, copy, modify and distribute this software and its
10! * documentation for any purpose is hereby granted without fee,
11! * provided that the above copyright notice appear in all copies and
12! * that both that copyright notice and this permission notice appear
13! * in supporting documentation. It is provided "as is" without express
14! * or implied warranty.
15! *
16!
17!
18! Author(s): Thomas Hengstebeck 3/2003 <mailto:hengsteb@physik.hu-berlin.de>
19!
20! Copyright: MAGIC Software Development, 2000-2005
21!
22!
23\* ======================================================================== */
24
25/////////////////////////////////////////////////////////////////////////////
26//
27// MRanForest
28//
29// ParameterContainer for Forest structure
30//
31// A random forest can be grown by calling GrowForest.
32// In advance SetupGrow must be called in order to initialize arrays and
33// do some preprocessing.
34// GrowForest() provides the training data for a single tree (bootstrap
35// aggregate procedure)
36//
37// Essentially two random elements serve to provide a "random" forest,
38// namely bootstrap aggregating (which is done in GrowForest()) and random
39// split selection (which is subject to MRanTree::GrowTree())
40//
41/////////////////////////////////////////////////////////////////////////////
42#include "MRanForest.h"
43
44#include <TVector.h>
45#include <TRandom.h>
46
47#include "MHMatrix.h"
48#include "MRanTree.h"
49#include "MData.h"
50#include "MDataArray.h"
51#include "MParList.h"
52
53#include "MLog.h"
54#include "MLogManip.h"
55
56ClassImp(MRanForest);
57
58using namespace std;
59
60// --------------------------------------------------------------------------
61//
62// Default constructor.
63//
64MRanForest::MRanForest(const char *name, const char *title)
65 : fClassify(kTRUE), fNumTrees(100), fNumTry(0), fNdSize(1),
66 fRanTree(NULL), fRules(NULL), fMatrix(NULL), fUserVal(-1)
67{
68 fName = name ? name : "MRanForest";
69 fTitle = title ? title : "Storage container for Random Forest";
70
71 fForest=new TObjArray();
72 fForest->SetOwner(kTRUE);
73}
74
75MRanForest::MRanForest(const MRanForest &rf)
76{
77 // Copy constructor
78 fName = rf.fName;
79 fTitle = rf.fTitle;
80
81 fClassify = rf.fClassify;
82 fNumTrees = rf.fNumTrees;
83 fNumTry = rf.fNumTry;
84 fNdSize = rf.fNdSize;
85 fTreeNo = rf.fTreeNo;
86 fRanTree = NULL;
87
88 fRules=new MDataArray();
89 fRules->Reset();
90
91 MDataArray *newrules=rf.fRules;
92
93 for(Int_t i=0;i<newrules->GetNumEntries();i++)
94 {
95 MData &data=(*newrules)[i];
96 fRules->AddEntry(data.GetRule());
97 }
98
99 // trees
100 fForest=new TObjArray();
101 fForest->SetOwner(kTRUE);
102
103 TObjArray *newforest=rf.fForest;
104 for(Int_t i=0;i<newforest->GetEntries();i++)
105 {
106 MRanTree *rantree=(MRanTree*)newforest->At(i);
107
108 MRanTree *newtree=new MRanTree(*rantree);
109 fForest->Add(newtree);
110 }
111
112 fHadTrue = rf.fHadTrue;
113 fHadEst = rf.fHadEst;
114 fDataSort = rf.fDataSort;
115 fDataRang = rf.fDataRang;
116 fClassPop = rf.fClassPop;
117 fWeight = rf.fWeight;
118 fTreeHad = rf.fTreeHad;
119
120 fNTimesOutBag = rf.fNTimesOutBag;
121
122}
123
124// --------------------------------------------------------------------------
125// Destructor.
126MRanForest::~MRanForest()
127{
128 delete fForest;
129 if (fMatrix)
130 delete fMatrix;
131 if (fRules)
132 delete fRules;
133}
134
135void MRanForest::Print(Option_t *o) const
136{
137 *fLog << inf << GetDescriptor() << ": " << endl;
138 MRanTree *t = GetTree(0);
139 if (t)
140 {
141 *fLog << "Setting up RF for training on target:" << endl;
142 *fLog << " " << t->GetTitle() << endl;
143 }
144 if (fRules)
145 {
146 *fLog << "Following rules are used as input to RF:" << endl;
147 for (Int_t i=0;i<fRules->GetNumEntries();i++)
148 *fLog << " " << i << ") " << (*fRules)[i].GetRule() << endl;
149 }
150 *fLog << "Random forest parameters:" << endl;
151 if (t)
152 {
153 *fLog << " - " << (t->IsClassify()?"classification":"regression") << " tree" << endl;
154 *fLog << " - Number of trys: " << t->GetNumTry() << endl;
155 *fLog << " - Node size: " << t->GetNdSize() << endl;
156 }
157 *fLog << " - Number of trees: " << fNumTrees << endl;
158 *fLog << " - User value: " << fUserVal << endl;
159 *fLog << endl;
160}
161
162void MRanForest::SetNumTrees(Int_t n)
163{
164 //at least 1 tree
165 fNumTrees=TMath::Max(n,1);
166}
167
168void MRanForest::SetNumTry(Int_t n)
169{
170 fNumTry=TMath::Max(n,0);
171}
172
173void MRanForest::SetNdSize(Int_t n)
174{
175 fNdSize=TMath::Max(n,1);
176}
177
178void MRanForest::SetWeights(const TArrayF &weights)
179{
180 fWeight=weights;
181}
182
183void MRanForest::SetGrid(const TArrayD &grid)
184{
185 const int n=grid.GetSize();
186
187 for(int i=0;i<n-1;i++)
188 if(grid[i]>=grid[i+1])
189 {
190 *fLog<<warn<<"Grid points must be in increasing order! Ignoring grid."<<endl;
191 return;
192 }
193
194 fGrid=grid;
195
196 //*fLog<<inf<<"Following "<<n<<" grid points are used:"<<endl;
197 //for(int i=0;i<n;i++)
198 // *fLog<<inf<<" "<<i<<") "<<fGrid[i]<<endl;
199}
200
201MRanTree *MRanForest::GetTree(Int_t i) const
202{
203 return static_cast<MRanTree*>(fForest->UncheckedAt(i));
204}
205
206Int_t MRanForest::GetNumDim() const
207{
208 return fMatrix ? fMatrix->GetNcols() : 0;
209}
210
211Int_t MRanForest::GetNumData() const
212{
213 return fMatrix ? fMatrix->GetNrows() : 0;
214}
215
216Int_t MRanForest::GetNclass() const
217{
218 int maxidx = TMath::LocMax(fClass.GetSize(),fClass.GetArray());
219
220 return int(fClass[maxidx])+1;
221}
222
223void MRanForest::PrepareClasses()
224{
225 const int numdata=fHadTrue.GetSize();
226
227 if(fGrid.GetSize()>0)
228 {
229 // classes given by grid
230 const int ngrid=fGrid.GetSize();
231
232 for(int j=0;j<numdata;j++)
233 {
234 // Array is supposed to be sorted prior to this call.
235 // If match is found, function returns position of element.
236 // If no match found, function gives nearest element smaller
237 // than value.
238 int k=TMath::BinarySearch(ngrid, fGrid.GetArray(), fHadTrue[j]);
239
240 fClass[j] = k;
241 }
242
243 int minidx = TMath::LocMin(fClass.GetSize(),fClass.GetArray());
244 int min = fClass[minidx];
245 if(min!=0) for(int j=0;j<numdata;j++)fClass[j]-=min;
246
247 }else{
248 // classes directly given
249 for (Int_t j=0;j<numdata;j++)
250 fClass[j] = TMath::Nint(fHadTrue[j]);
251 }
252}
253
254Double_t MRanForest::CalcHadroness()
255{
256 TVector event;
257 *fRules >> event;
258
259 return CalcHadroness(event);
260}
261
262Double_t MRanForest::CalcHadroness(const TVector &event)
263{
264 fTreeHad.Set(fNumTrees);
265
266 Double_t hadroness=0;
267 Int_t ntree =0;
268
269 TIter Next(fForest);
270
271 MRanTree *tree;
272 while ((tree=(MRanTree*)Next()))
273 hadroness += (fTreeHad[ntree++]=tree->TreeHad(event));
274
275 return hadroness/ntree;
276}
277
278Bool_t MRanForest::AddTree(MRanTree *rantree=NULL)
279{
280 fRanTree = rantree ? rantree : fRanTree;
281
282 if (!fRanTree) return kFALSE;
283
284 MRanTree *newtree=new MRanTree(*fRanTree);
285 fForest->Add(newtree);
286
287 return kTRUE;
288}
289
290Bool_t MRanForest::SetupGrow(MHMatrix *mat,MParList *plist)
291{
292 //-------------------------------------------------------------------
293 // access matrix, copy last column (target) preliminarily
294 // into fHadTrue
295 if (fMatrix)
296 delete fMatrix;
297 fMatrix = new TMatrix(mat->GetM());
298
299 int dim = fMatrix->GetNcols()-1;
300 int numdata = fMatrix->GetNrows();
301
302 fHadTrue.Set(numdata);
303 fHadTrue.Reset(0);
304
305 for (Int_t j=0;j<numdata;j++)
306 fHadTrue[j] = (*fMatrix)(j,dim);
307
308 // remove last col
309 fMatrix->ResizeTo(numdata,dim);
310
311 //-------------------------------------------------------------------
312 // setup labels for classification/regression
313 fClass.Set(numdata);
314 fClass.Reset(0);
315
316 if (fClassify)
317 PrepareClasses();
318
319 //-------------------------------------------------------------------
320 // allocating and initializing arrays
321 fHadEst.Set(numdata);
322 fHadEst.Reset(0);
323
324 fNTimesOutBag.Set(numdata);
325 fNTimesOutBag.Reset(0);
326
327 fDataSort.Set(dim*numdata);
328 fDataSort.Reset(0);
329
330 fDataRang.Set(dim*numdata);
331 fDataRang.Reset(0);
332
333 if(fWeight.GetSize()!=numdata)
334 {
335 fWeight.Set(numdata);
336 fWeight.Reset(1.);
337 *fLog << inf <<"Setting weights to 1 (no weighting)"<< endl;
338 }
339
340 //-------------------------------------------------------------------
341 // setup rules to be used for classification/regression
342 const MDataArray *allrules=(MDataArray*)mat->GetColumns();
343 if (allrules==NULL)
344 {
345 *fLog << err <<"Rules of matrix == null, exiting"<< endl;
346 return kFALSE;
347 }
348
349 if (allrules->GetNumEntries()!=dim+1)
350 {
351 *fLog << err <<"Rules of matrix " << allrules->GetNumEntries() << " mismatch dimension+1 " << dim+1 << "...exiting."<< endl;
352 return kFALSE;
353 }
354
355 if (fRules)
356 delete fRules;
357
358 fRules = new MDataArray();
359 fRules->Reset();
360
361 const TString target_rule = (*allrules)[dim].GetRule();
362 for (Int_t i=0;i<dim;i++)
363 fRules->AddEntry((*allrules)[i].GetRule());
364
365 *fLog << inf << endl;
366 *fLog << "Setting up RF for training on target:" << endl;
367 *fLog << " " << target_rule.Data() << endl;
368 *fLog << "Following rules are used as input to RF:" << endl;
369 for (Int_t i=0;i<dim;i++)
370 *fLog << " " << i << ") " << (*fRules)[i].GetRule() << endl;
371 *fLog << endl;
372
373 //-------------------------------------------------------------------
374 // prepare (sort) data for fast optimization algorithm
375 if (!CreateDataSort())
376 return kFALSE;
377
378 //-------------------------------------------------------------------
379 // access and init tree container
380 fRanTree = (MRanTree*)plist->FindCreateObj("MRanTree");
381 if(!fRanTree)
382 {
383 *fLog << err << dbginf << "MRanForest, fRanTree not initialized... aborting." << endl;
384 return kFALSE;
385 }
386 //fRanTree->SetName(target_rule); // Is not stored anyhow
387
388 const Int_t tryest = TMath::Nint(TMath::Sqrt(dim));
389
390 *fLog << inf << endl;
391 *fLog << "Following input for the tree growing are used:"<<endl;
392 *fLog << " Number of Trees : "<<fNumTrees<<endl;
393 *fLog << " Number of Trials: "<<(fNumTry==0?tryest:fNumTry)<<(fNumTry==0?" (auto)":"")<<endl;
394 *fLog << " Final Node size : "<<fNdSize<<endl;
395 *fLog << " Using Grid : "<<(fGrid.GetSize()>0?"Yes":"No")<<endl;
396 *fLog << " Using Weights : "<<(fWeight.GetSize()>0?"Yes":"No")<<endl;
397 *fLog << " Number of Events: "<<numdata<<endl;
398 *fLog << " Number of Params: "<<dim<<endl;
399
400 if(fNumTry==0)
401 {
402 fNumTry=tryest;
403 *fLog << inf << endl;
404 *fLog << "Set no. of trials to the recommended value of round(";
405 *fLog << TMath::Sqrt(dim) << ") = " << fNumTry << endl;
406 }
407
408 fRanTree->SetNumTry(fNumTry);
409 fRanTree->SetClassify(fClassify);
410 fRanTree->SetNdSize(fNdSize);
411
412 fTreeNo=0;
413
414 return kTRUE;
415}
416
417Bool_t MRanForest::GrowForest()
418{
419 if(!gRandom)
420 {
421 *fLog << err << dbginf << "gRandom not initialized... aborting." << endl;
422 return kFALSE;
423 }
424
425 fTreeNo++;
426
427 //-------------------------------------------------------------------
428 // initialize running output
429
430 float minfloat=fHadTrue[TMath::LocMin(fHadTrue.GetSize(),fHadTrue.GetArray())];
431 Bool_t calcResolution=(minfloat>0.001);
432
433 if (fTreeNo==1)
434 {
435 *fLog << inf << endl << underline;
436
437 if(calcResolution)
438 *fLog << "no. of tree no. of nodes resolution in % (from oob-data -> overest. of error)" << endl;
439 else
440 *fLog << "no. of tree no. of nodes rms in % (from oob-data -> overest. of error)" << endl;
441 // 12345678901234567890123456789012345678901234567890
442 }
443
444 const Int_t numdata = GetNumData();
445 const Int_t nclass = GetNclass();
446
447 //-------------------------------------------------------------------
448 // bootstrap aggregating (bagging) -> sampling with replacement:
449
450 TArrayF classpopw(nclass);
451 TArrayI jinbag(numdata); // Initialization includes filling with 0
452 TArrayF winbag(numdata); // Initialization includes filling with 0
453
454 float square=0;
455 float mean=0;
456
457 for (Int_t n=0; n<numdata; n++)
458 {
459 // The integer k is randomly (uniformly) chosen from the set
460 // {0,1,...,numdata-1}, which is the set of the index numbers of
461 // all events in the training sample
462
463 const Int_t k = Int_t(gRandom->Rndm()*numdata);
464
465 if(fClassify)
466 classpopw[fClass[k]]+=fWeight[k];
467 else
468 classpopw[0]+=fWeight[k];
469
470 mean +=fHadTrue[k]*fWeight[k];
471 square+=fHadTrue[k]*fHadTrue[k]*fWeight[k];
472
473 winbag[k]+=fWeight[k];
474 jinbag[k]=1;
475
476 }
477
478 //-------------------------------------------------------------------
479 // modifying sorted-data array for in-bag data:
480
481 // In bagging procedure ca. 2/3 of all elements in the original
482 // training sample are used to build the in-bag data
483 TArrayI datsortinbag=fDataSort;
484 Int_t ninbag=0;
485
486 ModifyDataSort(datsortinbag, ninbag, jinbag);
487
488 fRanTree->GrowTree(fMatrix,fHadTrue,fClass,datsortinbag,fDataRang,classpopw,mean,square,
489 jinbag,winbag,nclass);
490
491 //-------------------------------------------------------------------
492 // error-estimates from out-of-bag data (oob data):
493 //
494 // For a single tree the events not(!) contained in the bootstrap sample of
495 // this tree can be used to obtain estimates for the classification error of
496 // this tree.
497 // If you take a certain event, it is contained in the oob-data of 1/3 of
498 // the trees (see comment to ModifyData). This means that the classification error
499 // determined from oob-data is underestimated, but can still be taken as upper limit.
500
501 for (Int_t ievt=0;ievt<numdata;ievt++)
502 {
503 if (jinbag[ievt]>0)
504 continue;
505
506 fHadEst[ievt] +=fRanTree->TreeHad((*fMatrix), ievt);
507 fNTimesOutBag[ievt]++;
508
509 }
510
511 Int_t n=0;
512 Float_t ferr=0;
513
514 for (Int_t ievt=0;ievt<numdata;ievt++)
515 {
516 if(fNTimesOutBag[ievt]!=0)
517 {
518 float val = fHadEst[ievt]/float(fNTimesOutBag[ievt])-fHadTrue[ievt];
519 if(calcResolution) val/=fHadTrue[ievt];
520
521 ferr += val*val;
522 n++;
523 }
524 }
525 ferr = TMath::Sqrt(ferr/n);
526
527 //-------------------------------------------------------------------
528 // give running output
529 *fLog << setw(5) << fTreeNo;
530 *fLog << setw(18) << fRanTree->GetNumEndNodes();
531 *fLog << Form("%18.2f", ferr*100.);
532 *fLog << endl;
533
534 fRanTree->SetError(ferr);
535
536 // adding tree to forest
537 AddTree();
538
539 return fTreeNo<fNumTrees;
540}
541
542Bool_t MRanForest::CreateDataSort()
543{
544 // fDataSort(m,n) is the event number in which fMatrix(m,n) occurs.
545 // fDataRang(m,n) is the rang of fMatrix(m,n), i.e. if rang = r:
546 // fMatrix(m,n) is the r-th highest value of all fMatrix(m,.).
547 //
548 // There may be more then 1 event with rang r (due to bagging).
549
550 const Int_t numdata = GetNumData();
551 const Int_t dim = GetNumDim();
552
553 TArrayF v(numdata);
554 TArrayI isort(numdata);
555
556
557 for (Int_t mvar=0;mvar<dim;mvar++)
558 {
559
560 for(Int_t n=0;n<numdata;n++)
561 {
562 v[n]=(*fMatrix)(n,mvar);
563 isort[n]=n;
564
565 if(TMath::IsNaN(v[n]))
566 {
567 *fLog << err <<"Event no. "<<n<<", matrix column no. "<<mvar;
568 *fLog << err <<" has the value NaN."<<endl;
569 return kFALSE;
570 }
571 }
572
573 TMath::Sort(numdata,v.GetArray(),isort.GetArray(),kFALSE);
574
575 // this sorts the v[n] in ascending order. isort[n] is the event number
576 // of that v[n], which is the n-th from the lowest (assume the original
577 // event numbers are 0,1,...).
578
579 // control sorting
580 for(int n=1;n<numdata;n++)
581 if(v[isort[n-1]]>v[isort[n]])
582 {
583 *fLog << err <<"Event no. "<<n<<", matrix column no. "<<mvar;
584 *fLog << err <<" not at correct sorting position."<<endl;
585 return kFALSE;
586 }
587
588 for(Int_t n=0;n<numdata-1;n++)
589 {
590 const Int_t n1=isort[n];
591 const Int_t n2=isort[n+1];
592
593 fDataSort[mvar*numdata+n]=n1;
594 if(n==0) fDataRang[mvar*numdata+n1]=0;
595 if(v[n1]<v[n2])
596 {
597 fDataRang[mvar*numdata+n2]=fDataRang[mvar*numdata+n1]+1;
598 }else{
599 fDataRang[mvar*numdata+n2]=fDataRang[mvar*numdata+n1];
600 }
601 }
602 fDataSort[(mvar+1)*numdata-1]=isort[numdata-1];
603 }
604 return kTRUE;
605}
606
607void MRanForest::ModifyDataSort(TArrayI &datsortinbag, Int_t ninbag, const TArrayI &jinbag)
608{
609 const Int_t numdim=GetNumDim();
610 const Int_t numdata=GetNumData();
611
612 ninbag=0;
613 for (Int_t n=0;n<numdata;n++)
614 if(jinbag[n]==1) ninbag++;
615
616 for(Int_t m=0;m<numdim;m++)
617 {
618 Int_t k=0;
619 Int_t nt=0;
620 for(Int_t n=0;n<numdata;n++)
621 {
622 if(jinbag[datsortinbag[m*numdata+k]]==1)
623 {
624 datsortinbag[m*numdata+nt]=datsortinbag[m*numdata+k];
625 k++;
626 }else{
627 for(Int_t j=1;j<numdata-k;j++)
628 {
629 if(jinbag[datsortinbag[m*numdata+k+j]]==1)
630 {
631 datsortinbag[m*numdata+nt]=datsortinbag[m*numdata+k+j];
632 k+=j+1;
633 break;
634 }
635 }
636 }
637 nt++;
638 if(nt>=ninbag) break;
639 }
640 }
641}
642
643Bool_t MRanForest::AsciiWrite(ostream &out) const
644{
645 Int_t n=0;
646 MRanTree *tree;
647 TIter forest(fForest);
648
649 while ((tree=(MRanTree*)forest.Next()))
650 {
651 tree->AsciiWrite(out);
652 n++;
653 }
654
655 return n==fNumTrees;
656}
Note: See TracBrowser for help on using the repository browser.