Re: TMVA and efficiency scan with rectangular optimisation

From: Peter Speckmayer <peter.speckmayer_at_cern.ch>
Date: Fri, 26 Jun 2009 10:40:16 +0200


Hi Andrea,

You can change the number of signal efficiency bins by changing the value of fNbinsXOfROCCurve in the file Config.cxx (don't forget to compile src/ afterwards)

   fVariablePlotting.fNbinsXOfROCCurve = 100;

--> change this value to 1000 for instance.

(Unfortunately you cannot change the binning in a certain region only.)

For getting the cuts printed out for the finer binning add the following lines after factory->EvaluateAllMethods() in TMVAClassification.cxx

   TMVA::MethodCuts* cuts = dynamic_cast<TMVA::MethodCuts*>(factory->GetMethod("Cuts"));    Double_t epsilon = 0.0001;
   for (Double_t eff=0.05; eff<0.95; eff += 0.005) {

      cuts->PrintCuts( eff+epsilon );
   }

Where you change the string "Cuts" in GetMethod for the name of the booked cut-method  (e.g. CutsGA). Unfortunately this does not work with the macro TMVAClassification.C but only with the compiled TMVAClassification.cxx.

let me know if this works for you,
Peter

> Betreff:
> [ROOT] TMVA and efficiency scan with rectangular optimisation
> Von:
> Andrea Massironi <andrea.massironi_at_cern.ch>
> Datum:
> Thu, 25 Jun 2009 21:41:05 +0200
> An:
> <roottalk_at_cern.ch>
>
> An:
> <roottalk_at_cern.ch>
>
>
> Dear Rooters,
> I'm using the TMVA package for a rectangular optimization.
> Following the example
>
> TMVAnalysis.C
>
> I create TMVA::Factory *factory
> and then I train it with
>
> factory->BookMethod( TMVA::Types::kCuts, "Cuts",
>
> "!H:!V:FitMethod=MC:EffSel:SampleSize=200000:VarProp=FSmart" );
>
> then
> factory->TrainAllMethods();
>
> In this way I get a file
> weights/TMVAnalysis_Cuts.weights.txt
> with the optimised selections.
>
> However the optimisation has been performed with steps in signal
> efficiency of 0.01.
> Is it possible to have an efficiency scan with a different step (i.e.
> 0.005) and in a fixed efficiency range (for example with signal
> efficiency between 0. and 0.1)?
>
> Thanks, Andrea
Received on Fri Jun 26 2009 - 10:40:20 CEST

This archive was generated by hypermail 2.2.0 : Fri Jun 26 2009 - 11:50:02 CEST