'std::bad_alloc' after using TMVA::Factory ::BookMethod

willemv · July 19, 2017, 7:35pm

I am trying to use TMVA for the first time, hoping to train a BDT discriminator. At the moment I was trying to train a simple discriminator using 2 variables, and to add signal and background training and test events “by hand”. The relevant parts of my code could be summarized as:

TString outfileName("TMVA.root");
TMVA::Factory *factory = new TMVA::Factory( "TMVAClassification", outputFile, "!V:!Silent:Color:DrawProgressBar:Transformations=I;D;P;G,D:AnalysisType=Classification" );
TMVA::DataLoader *dataloader=new TMVA::DataLoader("dataset");
dataloader->AddVariable( "a", 'F' );
dataloader->AddVariable( "b", 'F' );

for(Long64_t it = 0; it < treeSignal->GetEntries();; ++it){
	treeSignal->GetEntry(it);
	std::vector<Double_t> vars(2);
	vars[0] = a;
	vars[1] = b;
	dataloader->AddSignalTrainingEvent( vars, 1. );
}

for(Long64_t it = 0; it < treeBkg->GetEntries();; ++it){
	treeBkg->GetEntry(it);
	std::vector<Double_t> vars(2);
	vars[0] = a;
	vars[1] = b;
	dataloader->AddBackgroundTrainingEvent( vars, 1. );
}
factory->BookMethod( dataloader, TMVA::Types::kBDT, "BDT", "!H:!V:NTrees=850:MinNodeSize=2.5%:MaxDepth=3:BoostType=AdaBoost:AdaBoostBeta=0.5:UseBaggedBoost:BaggedSampleFraction=0.5:SeparationType=GiniIndex:nCuts=20");
factory->TrainAllMethods();

However, the moment I try “factory->BookMethod(…)” the code crashes with the error message “terminate called after throwing an instance of ‘std::bad_alloc’ what(): std::bad_alloc Aborted (core dumped)”, and I can not seem to figure out what the problem is. What am I missing here that makes the code crash? (extra note : For the moment my code and trees are organized in such a way that I can only add training events by hand, and can not use “->AddSignalTree” etc.) Thanks in advance for the help!

kialbert · July 19, 2017, 8:13pm

Strange. It’s hard to say what the problem is with only this information, but you should definitely at least get a sensible error message in this situation.

If you could

verify that a and b have the values that you expect when adding the event.
try another method, e.g. Fisher just to see if this is specific to the BDT.
Provide a minimal reproducer (a file containing as little code as possible while still reproducing the error). Perhaps using TMVA::Classification as base to get some example trees.

And get back to us, that would help us debug the problem

Cheers,
Kim

willemv · July 20, 2017, 6:47am

Hello Kim,

Thanks for the reply! I checked the variables a and b, and they were always sensible. Upon trying a Fisher discriminant however, the error chages. Now the code manages to book the method in fact, but crashes on the training stage, giving the error

Factory                  : Booking method: Fisher                                                    ]
                     : 
Factory                  : Train all methods
<FATAL>                         : No input data for the training provided!
***> abort program execution
terminate called after throwing an instance of 'std::runtime_error'
what():  FATAL error
Aborted (core dumped)

Does this help you in localizing what the issue might be? So the difference is basically that for a BDT I can not “book” the method while for a Fisher discriminant I can book the method, but not train it. I do not understand where the “No input data for the training provided” comes from as my code clearly calls “>AddSignalTrainingEvent(…)” and the same for the background, and the variables given to this function were all sensible.

As requested I also made a small example that gives me the error messages in question:

    #include "TMVA/Factory.h"
#include "TMVA/DataLoader.h"
#include "TMVA/Tools.h"


void Loop(){
	const TString fileName = "test.root";
	//Read Trees from ROOT files
	TFile* file =  new TFile("../data_april17/" + fileName);
	file->cd("FakeElectrons");
	TTree* tree = (TTree*) (file->Get("FakeElectrons/fakeTree"));
	Double_t _lPt[20];
	Double_t _lEta[20];
	TBranch *b__lPt;
	TBranch *b__lEta;
	tree->SetBranchAddress("_lPt", _lPt, &b__lPt);
	tree->SetBranchAddress("_lEta", _lEta, &b__lEta);


	TMVA::Tools::Instance();

	TString outfileName("TMVA.root");
   	TFile* outputFile = TFile::Open(outfileName, "RECREATE" );
	TMVA::Factory *factory = new TMVA::Factory( "TMVAClassification", outputFile,
		                                           "!V:!Silent:Color:DrawProgressBar:Transformations=I;D;P;G,D:AnalysisType=Classification" );
	TMVA::DataLoader *dataloader=new TMVA::DataLoader("dataset");
	dataloader->AddVariable( "a", 'F' );
   	dataloader->AddVariable( "b", 'F' );

	//Loop over sample
	for(Long64_t it = 0; it < tree->GetEntries(); ++it){
		tree->GetEntry(it);
		std::vector<Double_t> vars = {_lPt[0], _lEta[0]};
		dataloader->AddSignalTrainingEvent( vars, 1. );
		vars[1] = _lPt[0];
		vars[0] = _lEta[0];
		dataloader->AddBackgroundTrainingEvent( vars, 1. );
	}
	factory->BookMethod( dataloader, TMVA::Types::kBDT, "BDT");//, "!H:!V:NTrees=850:MinNodeSize=2.5%:MaxDepth=3:BoostType=AdaBoost:AdaBoostBeta=0.5:UseBaggedBoost:BaggedSampleFraction=0.5:SeparationType=GiniIndex:nCuts=20");
	//factory->BookMethod( dataloader, TMVA::Types::kFisher, "Fisher", "H:!V:Fisher:VarTransform=None:CreateMVAPdfs:PDFInterpolMVAPdf=Spline2:NbinsMVAPdf=50:NsmoothMVAPdf=10" );
	factory->TrainAllMethods();
	outputFile->Close();
	delete factory;
	delete dataloader;
}

I read two variables out of my tree “test.root” , being “_lPt”, and “_lEta”, which are arrays of 1 to 20 elements, and atleast the first entry was always initialized. Running this small example code, crashes on “bookmethod” when using a BDT discriminator and during the training stage for a Fisher discriminant.

(The ROOT version I am using is 6.08.06)

kialbert · July 20, 2017, 8:26am

Thanks, that is very helpful. I’ll look into it as soon as possible.

Cheers,
Kim

moneta · July 20, 2017, 8:59am

HI,

It looks to me you had a memory problem. How many events are you using and how many variables ?

Cheers

Lorenzo

willemv · July 20, 2017, 9:04am

Hello Lorenzo,

I used 2 variables as a first try. The sample I have contains a few million events, but regardless whether I run over a tiny fraction of it or the whole sample I get the same result. The small snippet of code I provided produces all the errors (atleast in my case), and shows what variables I use etc.

regards,

Willem

willemv · July 27, 2017, 8:29am

Is there anyone still looking into this problem? It still has not been solved.

omazapa · July 27, 2017, 11:58am

Hi Willem

please try calling the method dataloader->PrepareTrainingAndTestTree
If not, can you provide the root file to reproduce the problem?

Cheers
Omar.

kialbert · August 2, 2017, 10:08am

Hi,

Any update on giving access to the file “test.root” or trying out adding the dataloder->PrepareTrainingAndTestTree?

Cheers,
Kim

fiorini · September 23, 2022, 9:44am

Hi,
I stumbled on a similar problem myself and I would like to report the solution my case for future reference.
In my case, the issue causing the error was that the
“dataloder->PrepareTrainingAndTestTree” command was before the “AddXXXEvent” part.

The error is trivial to solve, but the error message doesn’t help much to solve it (especially in the case of the BDT).