Home | News | Documentation | Download

TMVAGui() not working for multiple methods of BDT in file

TMVA
#1

This relates to another post where I didn’t receive a satisfactory answer.

I create the following file MVAAllUpdate.C (14.0 KB) containing the results of training training and multiple TMVA methods. When I then do TMVA::TMVAGui("AllMVA40K.root"), the buttons labeled (10) Decision Trees (BDT) and (11) Decision Tree Control Plots (BDT) in the window titled “TMVA Plotting Macros for Classification” return the error
*** Error in macro "BDT.C": cannot find directory "Method_BDT" in file: AllMVA40K.root
Listing the contents of the file AllMVA40K.root, I get

TFile**		AllMVA40K.root	
 TFile*		AllMVA40K.root	
  TDirectoryFile*		AllResults	AllResults
   KEY: TH2F	CorrelationMatrixS;1	Correlation Matrix (signal)
   KEY: TH2F	CorrelationMatrixB;1	Correlation Matrix (background)
   KEY: TDirectoryFile	InputVariables_Id;1	InputVariables_Id
   KEY: TDirectoryFile	Method_hPolIdealFisher;1	Directory for all Fisher methods
   KEY: TDirectoryFile	Method_vPolIdealFisher;1	Directory for all Fisher methods
   KEY: TDirectoryFile	Method_hPolFisher;1	Directory for all Fisher methods
   KEY: TDirectoryFile	Method_vPolFisher;1	Directory for all Fisher methods
   KEY: TDirectoryFile	Method_hPolBDT;1	Directory for all BDT methods
   KEY: TDirectoryFile	Method_vPolBDT;1	Directory for all BDT methods
   KEY: TDirectoryFile	Method_hPolCorrBDT;1	Directory for all BDT methods
   KEY: TDirectoryFile	Method_vPolCorrBDT;1	Directory for all BDT methods
   KEY: TTree	TestTree;2	TestTree
   KEY: TTree	TestTree;1	TestTree
   KEY: TTree	TrainTree;2	TrainTree
   KEY: TTree	TrainTree;1	TrainTree
  KEY: TDirectoryFile	AllResults;1	AllResults

I think this problem would be resolved if all of the BDT methods were saved under a single Method_BDT TDirectoryFile. For example, rather than having directory paths Method_hPolBDT/hPolBDT and Method_vPolBDT/vPolBDT, how would I go about instead getting Method_BDT/hPolBDT and Method_BDT/vPolBDT?

#2

Hi,

This indeed seems like a bug. A fix has been submitted. You can follow the development here: https://github.com/root-project/root/pull/3253.

Working around this issue in the current version is a bit involved but it can be done using this function:

void BDT(TString dataset, Int_t itree, TString wfile,
         TString methName = "BDT",
         Bool_t useTMVAStyle = kTRUE);

Calling it looks e.g. like this directly in your terminal:

root -l -e 'TMVA::BDT("dataset", 0, "dataset/weights/TMVAClassification_BDT.weights.xml", "BDT")'

Edit: And for the control plots you need something akin to:

root -l -e 'auto f = TFile::Open("TMVA.root"); TMVA::bdtcontrolplots("dataset", (TDirectory *)f->Get("dataset/Method_BDT/BDT"))'
With multiple signal and background types, handling one signal or background type divided over multiple trees using weights
#3

Hi Kim,

Thank you for submitting a fix to this bug. Will this fix apply to other methods as well? For example, if I train two Fisher methods, Class_1 and Class_2, will the training results be saved under the directory paths Method_Fisher/Class_1 and Method_Fisher/Class_2? And will this fix allow differing methods to have the same label? if I were to name two also BDT methods Class_1 and Class_2 as well, can I get a file with directory paths

Method_Fisher/Class_1
Method_Fisher/Class_2
Method_BDT/Class_1
Method_BDT/Class_2

I think because of the bug, BookMethod() doesn’t allow different booking if different methods with the same title.

#4

Hi,

Yes, this fix will apply to all methods. However, TMVA is designed such that the name of a method must be unique, so unfortunately it will, explicitly, not be possible to have both Method_Fisher/Class_1 and Method_BDT/Class_1.

Cheers,
Kim

#5

Thanks, Kim, this answers my questions.

#6

There is now a fix landed in lastest master. This means that this should be included in the v6.18 release!

So if you have two BDT trainings in a single run, they will now be placed in Method_BDT/BDT1 and Method_BDT/BDT2 in the output root file. This means that the GUI should now be able to generate the two plots for all trained BDT’s.

Cheers,
Kim

1 Like