Usage of the BDTG standalone C++ class from TMVA

TMVA
1

Hello TMVA experts,

I am trying to use the BDTG standalone c++ class and having some trouble with it.
The BDTG response is completely different when I compare it from the TMVAGUI and the autogenerated C++ class implemented in a homemade program
Here, the TMVAGUI output on the left and the c++ class output on the right :

bdtg_output_tmva_vs_homemade
bdtg_output_tmva_vs_homemade.png1239×441 74.3 KB

I don’t understand the differences even so the input events are exactly the same in TMVA and my homemade program, I also tried with a Reader but I get the same result.

To investigate this, I have tried the same procedure with MLPBNN and it works fine.

Maybe is there differences in the usage of the BDTG class compared with MLPBNN one ?

Thank you for your help.

Cheers,
Maxime

2

Hi,

Just to clarify; Is the output consistent between the TMVA::Reader and the stand-alone c++ class?

This would hint at a mismatch between the setups because the output when running TrainAllMethods and TestAllMethods is actually generated in a way very similar to how the Reader does it (recreating the method from the xml-file).

Notice that your plot does not include negative outputs. Could you regenerate your plot using the range -1 to 1?
Also notice that the TMVA graph output is normalised so that the histogram sums to 1 for (i.e signal is normalised to sum to 1, and background is normalised to sum to 1.).

Cheers,
Kim

3

Hi,

Thank you very much, you pointed out my silly mistake, indeed my neural net output range is [0,1] but the BDT is in the range [-1,1].
Now that I have normalized in the good range these output are consistent.

cheers,
Maxime

4

Great Maxime!

Please mark the topic as solved if indeed it is, and good luck with your machine learning :slight_smile:

Cheers,
Kim