GetSeparation error in evaluation stage of training

TMVA
1

Hello,
I’m training a BDT on some large nTuples and I have a stubborn error in the EvaluateAllMethods stage. It reads:
<GetSeparation> signal and background histograms have different or invalid dimensions
and I suspect it has something to do with the calculated RMS of a certain variable going to -nan in the table of variables (line 433 in the included log nohup.txt). This value seems to be reasonable in the other tables that TFHandler_BDT generates.
The separation on this variable isn’t great, but I would like to include it in the analysis if possible. Any idea what might be causing this error?

Thanks,
Gannon

Files:
Command log: nohup.txt (56.6 KB)
Macro: performAnalysis.py (4.9 KB)
(Let me know if you want to see the dependencies for the macro or the ntuples themselves. They’re fairly large so I’m not sure how to share them here.)

4

Hi,
I think this is caused by some BDT output values being a NaN for some events, so a check in Tools.cxx:152 fails and ends up in a FATAL.
I think the issue should be solved if one understand why you are getting NaN values. Are you sure all inputs are OK and not NaN.
I would need access to your input files to understand more about the problem

Lorenzo

5

Hi Lorenzo,
I’ve checked all the events and none of the inputs are NaN. I’m going to try running the BDT on the input tree in segments to see if I can find which event(s) lead to a NaN BDT output. If you want to take a look at the input files, I can share them. What’s the easiest way to transfer them?
Gannon

6

Hi,
You can share via cernbox for example. Please also share your C++ code running the BDT.

Cheers

Lorenzo

7

Hi,
Here’s a link to the files I was analyzing (WW as signal, ttbar as background.) I am using a set of python files instead of C++ to run the BDT, which I included. The performAnalysis.py file in run is what i actually run, while src contains middleware used to load the data files, declare variables etc.
Thanks,
Gannon