1 | /* ======================================================================== *\
|
---|
2 | ! $Name: not supported by cvs2svn $:$Id: MRanForest.cc,v 1.26 2006-11-02 08:57:00 tbretz Exp $
|
---|
3 | ! --------------------------------------------------------------------------
|
---|
4 | !
|
---|
5 | ! *
|
---|
6 | ! * This file is part of MARS, the MAGIC Analysis and Reconstruction
|
---|
7 | ! * Software. It is distributed to you in the hope that it can be a useful
|
---|
8 | ! * and timesaving tool in analysing Data of imaging Cerenkov telescopes.
|
---|
9 | ! * It is distributed WITHOUT ANY WARRANTY.
|
---|
10 | ! *
|
---|
11 | ! * Permission to use, copy, modify and distribute this software and its
|
---|
12 | ! * documentation for any purpose is hereby granted without fee,
|
---|
13 | ! * provided that the above copyright notice appear in all copies and
|
---|
14 | ! * that both that copyright notice and this permission notice appear
|
---|
15 | ! * in supporting documentation. It is provided "as is" without express
|
---|
16 | ! * or implied warranty.
|
---|
17 | ! *
|
---|
18 | !
|
---|
19 | !
|
---|
20 | ! Author(s): Thomas Hengstebeck 3/2003 <mailto:hengsteb@physik.hu-berlin.de>
|
---|
21 | ! Author(s): Thomas Bretz <mailto:tbretz@astro.uni-wuerzburg.de>
|
---|
22 | !
|
---|
23 | ! Copyright: MAGIC Software Development, 2000-2006
|
---|
24 | !
|
---|
25 | !
|
---|
26 | \* ======================================================================== */
|
---|
27 |
|
---|
28 | /////////////////////////////////////////////////////////////////////////////
|
---|
29 | //
|
---|
30 | // MRanForest
|
---|
31 | //
|
---|
32 | // ParameterContainer for Forest structure
|
---|
33 | //
|
---|
34 | // A random forest can be grown by calling GrowForest.
|
---|
35 | // In advance SetupGrow must be called in order to initialize arrays and
|
---|
36 | // do some preprocessing.
|
---|
37 | // GrowForest() provides the training data for a single tree (bootstrap
|
---|
38 | // aggregate procedure)
|
---|
39 | //
|
---|
40 | // Essentially two random elements serve to provide a "random" forest,
|
---|
41 | // namely bootstrap aggregating (which is done in GrowForest()) and random
|
---|
42 | // split selection (which is subject to MRanTree::GrowTree())
|
---|
43 | //
|
---|
44 | /////////////////////////////////////////////////////////////////////////////
|
---|
45 | #include "MRanForest.h"
|
---|
46 |
|
---|
47 | #include <TRandom.h>
|
---|
48 |
|
---|
49 | #include "MHMatrix.h"
|
---|
50 | #include "MRanTree.h"
|
---|
51 | #include "MData.h"
|
---|
52 | #include "MDataArray.h"
|
---|
53 | #include "MParList.h"
|
---|
54 |
|
---|
55 | #include "MArrayI.h"
|
---|
56 | #include "MArrayF.h"
|
---|
57 |
|
---|
58 | #include "MLog.h"
|
---|
59 | #include "MLogManip.h"
|
---|
60 |
|
---|
61 | ClassImp(MRanForest);
|
---|
62 |
|
---|
63 | using namespace std;
|
---|
64 |
|
---|
65 | // --------------------------------------------------------------------------
|
---|
66 | //
|
---|
67 | // Default constructor.
|
---|
68 | //
|
---|
69 | MRanForest::MRanForest(const char *name, const char *title)
|
---|
70 | : fClassify(kTRUE), fNumTrees(100), fNumTry(0), fNdSize(1),
|
---|
71 | fRanTree(NULL), fRules(NULL), fMatrix(NULL), fUserVal(-1)
|
---|
72 | {
|
---|
73 | fName = name ? name : "MRanForest";
|
---|
74 | fTitle = title ? title : "Storage container for Random Forest";
|
---|
75 |
|
---|
76 | fForest=new TObjArray();
|
---|
77 | fForest->SetOwner(kTRUE);
|
---|
78 | }
|
---|
79 |
|
---|
80 | MRanForest::MRanForest(const MRanForest &rf)
|
---|
81 | {
|
---|
82 | // Copy constructor
|
---|
83 | fName = rf.fName;
|
---|
84 | fTitle = rf.fTitle;
|
---|
85 |
|
---|
86 | fClassify = rf.fClassify;
|
---|
87 | fNumTrees = rf.fNumTrees;
|
---|
88 | fNumTry = rf.fNumTry;
|
---|
89 | fNdSize = rf.fNdSize;
|
---|
90 | fTreeNo = rf.fTreeNo;
|
---|
91 | fRanTree = NULL;
|
---|
92 |
|
---|
93 | fRules=new MDataArray();
|
---|
94 | fRules->Reset();
|
---|
95 |
|
---|
96 | MDataArray *newrules=rf.fRules;
|
---|
97 |
|
---|
98 | for(Int_t i=0;i<newrules->GetNumEntries();i++)
|
---|
99 | {
|
---|
100 | MData &data=(*newrules)[i];
|
---|
101 | fRules->AddEntry(data.GetRule());
|
---|
102 | }
|
---|
103 |
|
---|
104 | // trees
|
---|
105 | fForest=new TObjArray();
|
---|
106 | fForest->SetOwner(kTRUE);
|
---|
107 |
|
---|
108 | TObjArray *newforest=rf.fForest;
|
---|
109 | for(Int_t i=0;i<newforest->GetEntries();i++)
|
---|
110 | {
|
---|
111 | MRanTree *rantree=(MRanTree*)newforest->At(i);
|
---|
112 |
|
---|
113 | MRanTree *newtree=new MRanTree(*rantree);
|
---|
114 | fForest->Add(newtree);
|
---|
115 | }
|
---|
116 |
|
---|
117 | fHadTrue = rf.fHadTrue;
|
---|
118 | fHadEst = rf.fHadEst;
|
---|
119 | fDataSort = rf.fDataSort;
|
---|
120 | fDataRang = rf.fDataRang;
|
---|
121 | fClassPop = rf.fClassPop;
|
---|
122 | fWeight = rf.fWeight;
|
---|
123 | fTreeHad = rf.fTreeHad;
|
---|
124 |
|
---|
125 | fNTimesOutBag = rf.fNTimesOutBag;
|
---|
126 |
|
---|
127 | }
|
---|
128 |
|
---|
129 | // --------------------------------------------------------------------------
|
---|
130 | // Destructor.
|
---|
131 | MRanForest::~MRanForest()
|
---|
132 | {
|
---|
133 | delete fForest;
|
---|
134 | if (fMatrix)
|
---|
135 | delete fMatrix;
|
---|
136 | if (fRules)
|
---|
137 | delete fRules;
|
---|
138 | }
|
---|
139 |
|
---|
140 | void MRanForest::Print(Option_t *o) const
|
---|
141 | {
|
---|
142 | *fLog << inf << GetDescriptor() << ": " << endl;
|
---|
143 | MRanTree *t = GetTree(0);
|
---|
144 | if (t)
|
---|
145 | {
|
---|
146 | *fLog << "Setting up RF for training on target:" << endl;
|
---|
147 | *fLog << " " << t->GetTitle() << endl;
|
---|
148 | }
|
---|
149 | if (fRules)
|
---|
150 | {
|
---|
151 | *fLog << "Following rules are used as input to RF:" << endl;
|
---|
152 | for (Int_t i=0;i<fRules->GetNumEntries();i++)
|
---|
153 | *fLog << " " << i << ") " << (*fRules)[i].GetRule() << endl;
|
---|
154 | }
|
---|
155 | *fLog << "Random forest parameters:" << endl;
|
---|
156 | if (t)
|
---|
157 | {
|
---|
158 | *fLog << " - " << (t->IsClassify()?"classification":"regression") << " tree" << endl;
|
---|
159 | *fLog << " - Number of trys: " << t->GetNumTry() << endl;
|
---|
160 | *fLog << " - Node size: " << t->GetNdSize() << endl;
|
---|
161 | }
|
---|
162 | *fLog << " - Number of trees: " << fNumTrees << endl;
|
---|
163 | *fLog << " - User value: " << fUserVal << endl;
|
---|
164 | *fLog << endl;
|
---|
165 | }
|
---|
166 |
|
---|
167 | void MRanForest::SetNumTrees(Int_t n)
|
---|
168 | {
|
---|
169 | //at least 1 tree
|
---|
170 | fNumTrees=TMath::Max(n,1);
|
---|
171 | }
|
---|
172 |
|
---|
173 | void MRanForest::SetNumTry(Int_t n)
|
---|
174 | {
|
---|
175 | fNumTry=TMath::Max(n,0);
|
---|
176 | }
|
---|
177 |
|
---|
178 | void MRanForest::SetNdSize(Int_t n)
|
---|
179 | {
|
---|
180 | fNdSize=TMath::Max(n,1);
|
---|
181 | }
|
---|
182 |
|
---|
183 | void MRanForest::SetWeights(const TArrayF &weights)
|
---|
184 | {
|
---|
185 | fWeight=weights;
|
---|
186 | }
|
---|
187 |
|
---|
188 | void MRanForest::SetGrid(const TArrayD &grid)
|
---|
189 | {
|
---|
190 | const int n=grid.GetSize();
|
---|
191 |
|
---|
192 | for(int i=0;i<n-1;i++)
|
---|
193 | if(grid[i]>=grid[i+1])
|
---|
194 | {
|
---|
195 | *fLog<<warn<<"Grid points must be in increasing order! Ignoring grid."<<endl;
|
---|
196 | return;
|
---|
197 | }
|
---|
198 |
|
---|
199 | fGrid=grid;
|
---|
200 |
|
---|
201 | //*fLog<<inf<<"Following "<<n<<" grid points are used:"<<endl;
|
---|
202 | //for(int i=0;i<n;i++)
|
---|
203 | // *fLog<<inf<<" "<<i<<") "<<fGrid[i]<<endl;
|
---|
204 | }
|
---|
205 |
|
---|
206 | MRanTree *MRanForest::GetTree(Int_t i) const
|
---|
207 | {
|
---|
208 | return static_cast<MRanTree*>(fForest->UncheckedAt(i));
|
---|
209 | }
|
---|
210 |
|
---|
211 | Int_t MRanForest::GetNumDim() const
|
---|
212 | {
|
---|
213 | return fMatrix ? fMatrix->GetNcols() : 0;
|
---|
214 | }
|
---|
215 |
|
---|
216 | Int_t MRanForest::GetNumData() const
|
---|
217 | {
|
---|
218 | return fMatrix ? fMatrix->GetNrows() : 0;
|
---|
219 | }
|
---|
220 |
|
---|
221 | Int_t MRanForest::GetNclass() const
|
---|
222 | {
|
---|
223 | int maxidx = TMath::LocMax(fClass.GetSize(),fClass.GetArray());
|
---|
224 |
|
---|
225 | return int(fClass[maxidx])+1;
|
---|
226 | }
|
---|
227 |
|
---|
228 | void MRanForest::PrepareClasses()
|
---|
229 | {
|
---|
230 | const int numdata=fHadTrue.GetSize();
|
---|
231 |
|
---|
232 | if(fGrid.GetSize()>0)
|
---|
233 | {
|
---|
234 | // classes given by grid
|
---|
235 | const int ngrid=fGrid.GetSize();
|
---|
236 |
|
---|
237 | for(int j=0;j<numdata;j++)
|
---|
238 | {
|
---|
239 | // Array is supposed to be sorted prior to this call.
|
---|
240 | // If match is found, function returns position of element.
|
---|
241 | // If no match found, function gives nearest element smaller
|
---|
242 | // than value.
|
---|
243 | int k=TMath::BinarySearch(ngrid, fGrid.GetArray(), fHadTrue[j]);
|
---|
244 |
|
---|
245 | fClass[j] = k;
|
---|
246 | }
|
---|
247 |
|
---|
248 | int minidx = TMath::LocMin(fClass.GetSize(),fClass.GetArray());
|
---|
249 | int min = fClass[minidx];
|
---|
250 | if(min!=0) for(int j=0;j<numdata;j++)fClass[j]-=min;
|
---|
251 |
|
---|
252 | }else{
|
---|
253 | // classes directly given
|
---|
254 | for (Int_t j=0;j<numdata;j++)
|
---|
255 | fClass[j] = TMath::Nint(fHadTrue[j]);
|
---|
256 | }
|
---|
257 | }
|
---|
258 |
|
---|
259 | Double_t MRanForest::CalcHadroness()
|
---|
260 | {
|
---|
261 | TVector event;
|
---|
262 | *fRules >> event;
|
---|
263 |
|
---|
264 | return CalcHadroness(event);
|
---|
265 | }
|
---|
266 |
|
---|
267 | Double_t MRanForest::CalcHadroness(const TVector &event)
|
---|
268 | {
|
---|
269 | fTreeHad.Set(fNumTrees);
|
---|
270 |
|
---|
271 | Double_t hadroness=0;
|
---|
272 | Int_t ntree =0;
|
---|
273 |
|
---|
274 | TIter Next(fForest);
|
---|
275 |
|
---|
276 | MRanTree *tree;
|
---|
277 | while ((tree=(MRanTree*)Next()))
|
---|
278 | hadroness += (fTreeHad[ntree++]=tree->TreeHad(event));
|
---|
279 |
|
---|
280 | return hadroness/ntree;
|
---|
281 | }
|
---|
282 |
|
---|
283 | Bool_t MRanForest::AddTree(MRanTree *rantree=NULL)
|
---|
284 | {
|
---|
285 | fRanTree = rantree ? rantree : fRanTree;
|
---|
286 |
|
---|
287 | if (!fRanTree) return kFALSE;
|
---|
288 |
|
---|
289 | MRanTree *newtree=new MRanTree(*fRanTree);
|
---|
290 | fForest->Add(newtree);
|
---|
291 |
|
---|
292 | return kTRUE;
|
---|
293 | }
|
---|
294 |
|
---|
295 | Bool_t MRanForest::SetupGrow(MHMatrix *mat,MParList *plist)
|
---|
296 | {
|
---|
297 | //-------------------------------------------------------------------
|
---|
298 | // access matrix, copy last column (target) preliminarily
|
---|
299 | // into fHadTrue
|
---|
300 | if (fMatrix)
|
---|
301 | delete fMatrix;
|
---|
302 | fMatrix = new TMatrix(mat->GetM());
|
---|
303 |
|
---|
304 | int dim = fMatrix->GetNcols()-1;
|
---|
305 | int numdata = fMatrix->GetNrows();
|
---|
306 |
|
---|
307 | fHadTrue.Set(numdata);
|
---|
308 | fHadTrue.Reset();
|
---|
309 |
|
---|
310 | for (Int_t j=0;j<numdata;j++)
|
---|
311 | fHadTrue[j] = (*fMatrix)(j,dim);
|
---|
312 |
|
---|
313 | // remove last col
|
---|
314 | fMatrix->ResizeTo(numdata,dim);
|
---|
315 |
|
---|
316 | //-------------------------------------------------------------------
|
---|
317 | // setup labels for classification/regression
|
---|
318 | fClass.Set(numdata);
|
---|
319 | fClass.Reset();
|
---|
320 |
|
---|
321 | if (fClassify)
|
---|
322 | PrepareClasses();
|
---|
323 |
|
---|
324 | //-------------------------------------------------------------------
|
---|
325 | // allocating and initializing arrays
|
---|
326 | fHadEst.Set(numdata);
|
---|
327 | fHadEst.Reset();
|
---|
328 |
|
---|
329 | fNTimesOutBag.Set(numdata);
|
---|
330 | fNTimesOutBag.Reset();
|
---|
331 |
|
---|
332 | fDataSort.Set(dim*numdata);
|
---|
333 | fDataSort.Reset();
|
---|
334 |
|
---|
335 | fDataRang.Set(dim*numdata);
|
---|
336 | fDataRang.Reset();
|
---|
337 |
|
---|
338 | Bool_t useweights = fWeight.GetSize()==numdata;
|
---|
339 | if (!useweights)
|
---|
340 | {
|
---|
341 | fWeight.Set(numdata);
|
---|
342 | fWeight.Reset(1.);
|
---|
343 | *fLog << inf <<"Setting weights to 1 (no weighting)"<< endl;
|
---|
344 | }
|
---|
345 |
|
---|
346 | //-------------------------------------------------------------------
|
---|
347 | // setup rules to be used for classification/regression
|
---|
348 | const MDataArray *allrules=(MDataArray*)mat->GetColumns();
|
---|
349 | if (allrules==NULL)
|
---|
350 | {
|
---|
351 | *fLog << err <<"Rules of matrix == null, exiting"<< endl;
|
---|
352 | return kFALSE;
|
---|
353 | }
|
---|
354 |
|
---|
355 | if (allrules->GetNumEntries()!=dim+1)
|
---|
356 | {
|
---|
357 | *fLog << err <<"Rules of matrix " << allrules->GetNumEntries() << " mismatch dimension+1 " << dim+1 << "...exiting."<< endl;
|
---|
358 | return kFALSE;
|
---|
359 | }
|
---|
360 |
|
---|
361 | if (fRules)
|
---|
362 | delete fRules;
|
---|
363 |
|
---|
364 | fRules = new MDataArray();
|
---|
365 | fRules->Reset();
|
---|
366 |
|
---|
367 | const TString target_rule = (*allrules)[dim].GetRule();
|
---|
368 | for (Int_t i=0;i<dim;i++)
|
---|
369 | fRules->AddEntry((*allrules)[i].GetRule());
|
---|
370 |
|
---|
371 | *fLog << inf << endl;
|
---|
372 | *fLog << "Setting up RF for training on target:" << endl;
|
---|
373 | *fLog << " " << target_rule.Data() << endl;
|
---|
374 | *fLog << "Following rules are used as input to RF:" << endl;
|
---|
375 | for (Int_t i=0;i<dim;i++)
|
---|
376 | *fLog << " " << i << ") " << (*fRules)[i].GetRule() << endl;
|
---|
377 | *fLog << endl;
|
---|
378 |
|
---|
379 | //-------------------------------------------------------------------
|
---|
380 | // prepare (sort) data for fast optimization algorithm
|
---|
381 | if (!CreateDataSort())
|
---|
382 | return kFALSE;
|
---|
383 |
|
---|
384 | //-------------------------------------------------------------------
|
---|
385 | // access and init tree container
|
---|
386 | fRanTree = (MRanTree*)plist->FindCreateObj("MRanTree");
|
---|
387 | if(!fRanTree)
|
---|
388 | {
|
---|
389 | *fLog << err << dbginf << "MRanForest, fRanTree not initialized... aborting." << endl;
|
---|
390 | return kFALSE;
|
---|
391 | }
|
---|
392 | //fRanTree->SetName(target_rule); // Is not stored anyhow
|
---|
393 |
|
---|
394 | const Int_t tryest = TMath::Nint(TMath::Sqrt(dim));
|
---|
395 |
|
---|
396 | *fLog << inf << endl;
|
---|
397 | *fLog << "Following input for the tree growing are used:"<<endl;
|
---|
398 | *fLog << " Forest type : "<<(fClassify?"classification":"regression")<<endl;
|
---|
399 | *fLog << " Number of Trees : "<<fNumTrees<<endl;
|
---|
400 | *fLog << " Number of Trials: "<<(fNumTry==0?tryest:fNumTry)<<(fNumTry==0?" (auto)":"")<<endl;
|
---|
401 | *fLog << " Final Node size : "<<fNdSize<<endl;
|
---|
402 | *fLog << " Using Grid : "<<(fGrid.GetSize()>0?"Yes":"No")<<endl;
|
---|
403 | *fLog << " Using Weights : "<<(useweights?"Yes":"No")<<endl;
|
---|
404 | *fLog << " Number of Events: "<<numdata<<endl;
|
---|
405 | *fLog << " Number of Params: "<<dim<<endl;
|
---|
406 |
|
---|
407 | if(fNumTry==0)
|
---|
408 | {
|
---|
409 | fNumTry=tryest;
|
---|
410 | *fLog << inf << endl;
|
---|
411 | *fLog << "Set no. of trials to the recommended value of round(";
|
---|
412 | *fLog << TMath::Sqrt(dim) << ") = " << fNumTry << endl;
|
---|
413 | }
|
---|
414 |
|
---|
415 | fRanTree->SetNumTry(fNumTry);
|
---|
416 | fRanTree->SetClassify(fClassify);
|
---|
417 | fRanTree->SetNdSize(fNdSize);
|
---|
418 |
|
---|
419 | fTreeNo=0;
|
---|
420 |
|
---|
421 | return kTRUE;
|
---|
422 | }
|
---|
423 |
|
---|
424 | Bool_t MRanForest::GrowForest()
|
---|
425 | {
|
---|
426 | if(!gRandom)
|
---|
427 | {
|
---|
428 | *fLog << err << dbginf << "gRandom not initialized... aborting." << endl;
|
---|
429 | return kFALSE;
|
---|
430 | }
|
---|
431 |
|
---|
432 | fTreeNo++;
|
---|
433 |
|
---|
434 | //-------------------------------------------------------------------
|
---|
435 | // initialize running output
|
---|
436 |
|
---|
437 | float minfloat=fHadTrue[TMath::LocMin(fHadTrue.GetSize(),fHadTrue.GetArray())];
|
---|
438 | Bool_t calcResolution=(minfloat>0.001);
|
---|
439 |
|
---|
440 | if (fTreeNo==1)
|
---|
441 | {
|
---|
442 | *fLog << inf << endl << underline;
|
---|
443 |
|
---|
444 | if(calcResolution)
|
---|
445 | *fLog << "TreeNum BagSize NumNodes TestSize res/% (from oob-data -> overest. of error)" << endl;
|
---|
446 | else
|
---|
447 | *fLog << "TreeNum BagSize NumNodes TestSize rms/% (from oob-data -> overest. of error)" << endl;
|
---|
448 | // 12345678901234567890123456789012345678901234567890
|
---|
449 | }
|
---|
450 |
|
---|
451 | const Int_t numdata = GetNumData();
|
---|
452 | const Int_t nclass = GetNclass();
|
---|
453 |
|
---|
454 | //-------------------------------------------------------------------
|
---|
455 | // bootstrap aggregating (bagging) -> sampling with replacement:
|
---|
456 |
|
---|
457 | MArrayF classpopw(nclass);
|
---|
458 | MArrayI jinbag(numdata); // Initialization includes filling with 0
|
---|
459 | MArrayF winbag(numdata); // Initialization includes filling with 0
|
---|
460 |
|
---|
461 | float square=0;
|
---|
462 | float mean=0;
|
---|
463 |
|
---|
464 | for (Int_t n=0; n<numdata; n++)
|
---|
465 | {
|
---|
466 | // The integer k is randomly (uniformly) chosen from the set
|
---|
467 | // {0,1,...,numdata-1}, which is the set of the index numbers of
|
---|
468 | // all events in the training sample
|
---|
469 |
|
---|
470 | const Int_t k = gRandom->Integer(numdata);
|
---|
471 |
|
---|
472 | if(fClassify)
|
---|
473 | classpopw[fClass[k]]+=fWeight[k];
|
---|
474 | else
|
---|
475 | classpopw[0]+=fWeight[k];
|
---|
476 |
|
---|
477 | mean +=fHadTrue[k]*fWeight[k];
|
---|
478 | square+=fHadTrue[k]*fHadTrue[k]*fWeight[k];
|
---|
479 |
|
---|
480 | winbag[k]+=fWeight[k]; // Increase weight if chosen more than once
|
---|
481 | jinbag[k]=1;
|
---|
482 | }
|
---|
483 |
|
---|
484 | //-------------------------------------------------------------------
|
---|
485 | // modifying sorted-data array for in-bag data:
|
---|
486 |
|
---|
487 | // In bagging procedure ca. 2/3 of all elements in the original
|
---|
488 | // training sample are used to build the in-bag data
|
---|
489 | const MArrayF hadtrue(fHadTrue.GetSize(), fHadTrue.GetArray());
|
---|
490 | const MArrayI fclass(fClass.GetSize(), fClass.GetArray());
|
---|
491 | const MArrayI datarang(fDataRang.GetSize(), fDataRang.GetArray());
|
---|
492 |
|
---|
493 | MArrayI datsortinbag(fDataSort.GetSize(), fDataSort.GetArray());
|
---|
494 |
|
---|
495 | ModifyDataSort(datsortinbag, jinbag);
|
---|
496 |
|
---|
497 | fRanTree->GrowTree(fMatrix,hadtrue,fclass,datsortinbag,datarang,classpopw,mean,square,
|
---|
498 | jinbag,winbag,nclass);
|
---|
499 |
|
---|
500 | //-------------------------------------------------------------------
|
---|
501 | // error-estimates from out-of-bag data (oob data):
|
---|
502 | //
|
---|
503 | // For a single tree the events not(!) contained in the bootstrap
|
---|
504 | // sample of this tree can be used to obtain estimates for the
|
---|
505 | // classification error of this tree.
|
---|
506 | // If you take a certain event, it is contained in the oob-data of
|
---|
507 | // 1/3 of the trees (see comment to ModifyData). This means that the
|
---|
508 | // classification error determined from oob-data is underestimated,
|
---|
509 | // but can still be taken as upper limit.
|
---|
510 | Int_t ninbag = 0;
|
---|
511 | for (Int_t ievt=0;ievt<numdata;ievt++)
|
---|
512 | {
|
---|
513 | if (jinbag[ievt]>0)
|
---|
514 | {
|
---|
515 | ninbag++;
|
---|
516 | continue;
|
---|
517 | }
|
---|
518 |
|
---|
519 | fHadEst[ievt] +=fRanTree->TreeHad((*fMatrix), ievt);
|
---|
520 | fNTimesOutBag[ievt]++;
|
---|
521 | }
|
---|
522 |
|
---|
523 | Int_t n=0;
|
---|
524 | Float_t ferr=0;
|
---|
525 |
|
---|
526 | for (Int_t ievt=0;ievt<numdata;ievt++)
|
---|
527 | {
|
---|
528 | if(fNTimesOutBag[ievt]!=0)
|
---|
529 | {
|
---|
530 | float val = fHadEst[ievt]/float(fNTimesOutBag[ievt])-fHadTrue[ievt];
|
---|
531 | if(calcResolution) val/=fHadTrue[ievt];
|
---|
532 |
|
---|
533 | ferr += val*val;
|
---|
534 | n++;
|
---|
535 | }
|
---|
536 | }
|
---|
537 | ferr = TMath::Sqrt(ferr/n);
|
---|
538 |
|
---|
539 | //-------------------------------------------------------------------
|
---|
540 | // give running output
|
---|
541 | *fLog << setw(4) << fTreeNo;
|
---|
542 | *fLog << Form(" %8.1f", 100.*ninbag/numdata);
|
---|
543 | *fLog << setw(9) << fRanTree->GetNumEndNodes();
|
---|
544 | *fLog << Form(" %9.1f", 100.*n/numdata);
|
---|
545 | *fLog << Form("%18.2f", ferr*100.);
|
---|
546 | *fLog << endl;
|
---|
547 |
|
---|
548 | fRanTree->SetError(ferr);
|
---|
549 |
|
---|
550 | // adding tree to forest
|
---|
551 | AddTree();
|
---|
552 |
|
---|
553 | return fTreeNo<fNumTrees;
|
---|
554 | }
|
---|
555 |
|
---|
556 | Bool_t MRanForest::CreateDataSort()
|
---|
557 | {
|
---|
558 | // fDataSort(m,n) is the event number in which fMatrix(m,n) occurs.
|
---|
559 | // fDataRang(m,n) is the rang of fMatrix(m,n), i.e. if rang = r:
|
---|
560 | // fMatrix(m,n) is the r-th highest value of all fMatrix(m,.).
|
---|
561 | //
|
---|
562 | // There may be more then 1 event with rang r (due to bagging).
|
---|
563 |
|
---|
564 | const Int_t numdata = GetNumData();
|
---|
565 | const Int_t dim = GetNumDim();
|
---|
566 |
|
---|
567 | TArrayF v(numdata);
|
---|
568 | TArrayI isort(numdata);
|
---|
569 |
|
---|
570 |
|
---|
571 | for (Int_t mvar=0;mvar<dim;mvar++)
|
---|
572 | {
|
---|
573 |
|
---|
574 | for(Int_t n=0;n<numdata;n++)
|
---|
575 | {
|
---|
576 | v[n]=(*fMatrix)(n,mvar);
|
---|
577 | //isort[n]=n;
|
---|
578 |
|
---|
579 | if (!TMath::Finite(v[n]))
|
---|
580 | {
|
---|
581 | *fLog << err <<"Event no. "<<n<<", matrix column no. "<<mvar;
|
---|
582 | *fLog << err <<" has a non finite value (eg. NaN)."<<endl;
|
---|
583 | return kFALSE;
|
---|
584 | }
|
---|
585 | }
|
---|
586 |
|
---|
587 | TMath::Sort(numdata,v.GetArray(),isort.GetArray(),kFALSE);
|
---|
588 |
|
---|
589 | // this sorts the v[n] in ascending order. isort[n] is the
|
---|
590 | // event number of that v[n], which is the n-th from the
|
---|
591 | // lowest (assume the original event numbers are 0,1,...).
|
---|
592 |
|
---|
593 | // control sorting
|
---|
594 | /*
|
---|
595 | for(int n=0;n<numdata-1;n++)
|
---|
596 | if(v[isort[n]]>v[isort[n+1]])
|
---|
597 | {
|
---|
598 | *fLog << err <<"Event no. "<<n<<", matrix column no. "<<mvar;
|
---|
599 | *fLog << err <<" not at correct sorting position."<<endl;
|
---|
600 | return kFALSE;
|
---|
601 | }
|
---|
602 | */
|
---|
603 |
|
---|
604 | // DataRang is similar to the isort index starting from 0 it is
|
---|
605 | // increased by one for each event which is greater, but stays
|
---|
606 | // the same for the same value. (So to say it counts how many
|
---|
607 | // different values we have)
|
---|
608 | for(Int_t n=0;n<numdata-1;n++)
|
---|
609 | {
|
---|
610 | const Int_t n1=isort[n];
|
---|
611 | const Int_t n2=isort[n+1];
|
---|
612 |
|
---|
613 | // FIXME: Copying isort[n] to fDataSort[mvar*numdata]
|
---|
614 | // can be accelerated!
|
---|
615 | fDataSort[mvar*numdata+n]=n1;
|
---|
616 | if(n==0) fDataRang[mvar*numdata+n1]=0;
|
---|
617 | if(v[n1]<v[n2])
|
---|
618 | {
|
---|
619 | fDataRang[mvar*numdata+n2]=fDataRang[mvar*numdata+n1]+1;
|
---|
620 | }else{
|
---|
621 | fDataRang[mvar*numdata+n2]=fDataRang[mvar*numdata+n1];
|
---|
622 | }
|
---|
623 | }
|
---|
624 | fDataSort[(mvar+1)*numdata-1]=isort[numdata-1];
|
---|
625 | }
|
---|
626 | return kTRUE;
|
---|
627 | }
|
---|
628 |
|
---|
629 | // Reoves all indices which are not in the bag from the datsortinbag
|
---|
630 | void MRanForest::ModifyDataSort(MArrayI &datsortinbag, const MArrayI &jinbag)
|
---|
631 | {
|
---|
632 | const Int_t numdim=GetNumDim();
|
---|
633 | const Int_t numdata=GetNumData();
|
---|
634 |
|
---|
635 | Int_t ninbag=0;
|
---|
636 | for (Int_t n=0;n<numdata;n++)
|
---|
637 | if(jinbag[n]==1) ninbag++;
|
---|
638 |
|
---|
639 | for(Int_t m=0;m<numdim;m++)
|
---|
640 | {
|
---|
641 | Int_t *subsort = &datsortinbag[m*numdata];
|
---|
642 |
|
---|
643 | Int_t k=0;
|
---|
644 | for(Int_t n=0;n<ninbag;n++)
|
---|
645 | {
|
---|
646 | if(jinbag[subsort[k]]==1)
|
---|
647 | {
|
---|
648 | subsort[n] = subsort[k];
|
---|
649 | k++;
|
---|
650 | }else{
|
---|
651 | for(Int_t j=k+1;j<numdata;j++)
|
---|
652 | {
|
---|
653 | if(jinbag[subsort[j]]==1)
|
---|
654 | {
|
---|
655 | subsort[n] = subsort[j];
|
---|
656 | k = j+1;
|
---|
657 | break;
|
---|
658 | }
|
---|
659 | }
|
---|
660 | }
|
---|
661 | }
|
---|
662 | }
|
---|
663 | }
|
---|
664 |
|
---|
665 | Bool_t MRanForest::AsciiWrite(ostream &out) const
|
---|
666 | {
|
---|
667 | Int_t n=0;
|
---|
668 | MRanTree *tree;
|
---|
669 | TIter forest(fForest);
|
---|
670 |
|
---|
671 | while ((tree=(MRanTree*)forest.Next()))
|
---|
672 | {
|
---|
673 | tree->AsciiWrite(out);
|
---|
674 | n++;
|
---|
675 | }
|
---|
676 |
|
---|
677 | return n==fNumTrees;
|
---|
678 | }
|
---|