Dear experts,
I have a vector branch in my root file and I parse it to the TMVA classification (1). The code run fine and the vector variable looks fine (2). If this makes sense, I wonder how can I parse this “vector variable” to the reader? using (3) does not work with the “vector” type.
Regards
I run the same code with the vector of pt (the leading pt is the 1st value in the vector) and the leading pt variable. I checked that the vector[0] always correspond to the leading pt, but TMVA give me different output, here are the log file:
you can see that the ROC is 0.727 for the vector and 0.719 for the leading pt variable. Depending on how I configure the DNN these ROC are significantly different.
Here are the response and input variable:
3) vector:
TMVA will include concatenate all entries in an array variable, keeping all other non-array variables constant. Having two variables, one scalar (x) and one vector (y) with a length of two would create the following TMVA events
Dear Kialbert,
my problem is that I do not know what the TMVA code does, I wonder if an expert know what the code does if I use: dataloader->AddVariable( “vector_pT”, “vector_pT”, “MeV”, ‘F’ );
I mean I’m just running the tools. When I used “vector_pT” I see much better separation than if I use " vector_pT[0]", so I wonder what can explain this difference? does it use all the values in the vector, or only one of them…?
Regards
This is what I’m trying, and unfortunately failing, to explain. Let me try with a different approach.
lead_jet_pT is a property of an event while vec_jet_pT contains data for individual jets. TMVA tries to be helpful when you add a vector variable and assumes that you want to do classification on the individual jets (as opposed to the whole physics event). So it creates new TMVA::Event's for each entry in the vector; A vector with length 5 would, instead of 1 TMVA::Event, generate 5 TMVA::Event's.
The TMVA::Reader is not directly aware of this behaviour, so you would have to replicate it manually to have consistent results between training/testing (TMVA::Factory) and application (TMVA::Reader).
Something along the lines of the code snippet below would be necessary.
int x = 0;
float vec_pT = 0.;
reader.AddVariable("x", &x);
reader.AddVariable("vec_pT", &vec_pT);
tree = GetTree();
tree.SetBranchAddress("x", &x);
// Note we are not using `AddVariable("vec_pT")` here.
// We'll be loading that value manually later.
for (ievent = 0; i < tree->GetEntries(); ++ievent) {
// This will get the correct value of `x`
tree->GetEntry(ievent);
for (ijet = 0; i < numJets; ++ijet) {
// This would get, in turn, each of the jet pt's and feed them into tmva
// Warning: there are probably better ways of doing this, and I'm not even sure
// it works as is written here, but I hope you understand the idea.
vec_pT = *(static_cast<float *>(tree->FindLeaf("vec_pT")->GetValuePointer()) + ijet);
float prediction = reader->EvaluateMVA("MyMVA");
std::cout << "My prediction is: " << prediction << " for event " << ievent*numJets + ijet << std::endl;
}
}
in TMVAClassification.C decided to used (1). To parse that to the reader, I did (2), but I got the error message (3). Do you see what is wrong? You can see the entire code here (4).
Regards
In your code the variable jet_pT is declared as a pointer to vector of float vector<float> *jet_pT;. Changing your assignment from r_jet_pT_0 = jet_pT[0]; to r_jet_pT_0 = (*jet_pT)[0]; should do the trick.
(That is instead of setting r_jet_pT_0 to the first vector of floats in jet_pT, set it to the first element of the pointed-to vector.)
