TMVA SplitMode=Block/Alternate gives 0 testing error

TMVA
1

Hi,

I’m trying to train a Deep Neural Network with TMVA and want to use SplitMode = Block or Alternate but whenever I set this option, every epoch gives me 0 testing error (as seen in the standard output to terminal). Weirdly enough, I do get a non-zero roughly-expected value for training error. If I switch to using SplitMode=Random, the testing error goes back to a non-zero value and everything seems to function fine. How exactly do I get SplitMode=Block or Alternate working?

For reference, my TMVA::DataLoader call is…

dataLoader->PrepareTrainingAndTestTree("", 13621, 25727, 13621, 25727, “SplitMode=block”);

Cheers,
James

2

Hi,

Could you try lowering the batch size of your DNN, does that change your reported test score? (Try e.g. 10 just for the sake of it.)

Otherwise, could you post a complete example that reproduces the error?

Cheers,
Kim