Multiclass inference with TMVA SOFIE

TMVA
,
1

Dear experts,

I’m trying to use TMVA SOFIE to make inference starting from a keras (.h5) model.
My model is a multiclass DNN, and the output that I expect from the inference is a vector of dimension N for each event, where N is the number of classes. What I get when I call sofie_functor is a single number, that I think, looking at the line of the code TMVA/SOFIEHelpers.hxx#L49 is the first element of the vector. Is there a way to get all the numbers, or am I missing something here? Thanks a lot.

Technical details:
I’m using ROOT version 6.28, and TMVA SOFIE with python and RDataFrame.
To write the code I followed the tutorial root.cern/doc/master/TMVA__SOFIE__RDataFrame_8py.html without any relevant modifications.

Let me know if you need any additional information.
Thanks a lot,
Sofia

2

Hi Sofia,

Thanks for reaching out and welcome to the ROOT Community!

Let me add in the loop @moneta , who will be able to address this inquiry.

Cheers,
Danilo

3

Hello,

It is correct the corrent SOFIR functor implemented in TMVA/SOFIEHelper.hxx returns a single value. If you have a multiclass model and you need to return a vector of values, you can easly change it to return an std::vector by changing the signature of the operator(), function as following:

std::vector<T>  operator()(unsigned slot, AlwaysT<N>... args) {
      fInput[slot] = {args...};
      auto y =  fSessions[slot].infer(fInput[slot].data());
      return y;
   }

Please let me know if you have any issue using this modifies functor.

Best regards

Lorenzo

4

Hi,
it works, thank you!
Best regards,
Sofia