The prediction probability given by TMVA DNN is all 0

TMVA
1

Dear tmva users,
I have trained DNN model using TMVA with ROOT6.20version.The model can application normally with ROOT6.18. But when I run it with ROOT6.20,the output is all zeros! I am extremely perplexed as to why changing the ROOT version would result in the same program being unable to produce the correct output.
This my model and Application code
model.zip (1.1 MB)
pion.C (6.2 KB)
pionTrain.root (62.1 KB)

Thanks a lot!

2

Hello,

Sorry for the late reply. I can confirm the problem and investigate it. It might be that we have deprecated the DNN method in TMVA in favour of the DL method, and this caused a problem evaluating old model. I will investigate this further.

I would suggest to trying using the DL method as shown in the tutorial TMVAClassification.C (see ROOT: tutorials/tmva/TMVAClassification.C File Reference)

Best regards

Lorenzo

3

I have updated your weight file to work with the DL method. I attached here, in case you cannot re-train the method. This produces the same result as using 6.18

Lorenzo

pion.C (6.3 KB)
weights.tar.gz (1.1 MB)

4

Dear moneta:
Thank you very much for your help, the model trained by DL can be used correctly!

Thanks a lot!