How to set test error limit in TMVA DNN method?

As the TMVAUserGuide.pdf said, the Convergence criterion of DNN method is setting the number of
training epochs without improvement in test error to be performed before ending the training phase. But how can we determine whether the test error has improved,and how can we control it? Can we set a number such as 0.01, when test error is reduced by more than 0.01, we think it has improved.

@couet Could you help me?Thank you!

As it is a TMVA question the best person to help you is @moneta.

Thank you very much! :smiley:

In the TMVA DNN you can control the convergence with two criteria,

  • the maximum number of training epochs, MaxEpochs
  • the number of epochs that the test error is not reduced, ConvergenceSteps.
    For example if you set ConvergeSteps=5 the training will stop after seeing 5 training epochs where the test error is not reduced, i.e. is larger than the minimum test error found so far.

Unfortunately we don’t have another parameter controlling the actual reduction of the test error. For example if it is not reduced by more than 0.01. We could eventually implement this if it is needed



Hi, thanks for your reply. :grinning: I have understood the convergence criteria. But it seems that there is not “MaxEpochs” in the options of training strategy. Where can I search the whole options.


Which version of the TMVA Users guide are you having ?
If you download it from here, you will have MaxEpochs documented in Table 23.


1 Like

Thank you very much!My users guide version is 4.2,it is a little old :grinning:


This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.