[TMVA] Pre-training initialization of Neural-net weights

TMVA
1

Dear TMVA-experts,

Is it possible to initialize the connection-weights of a neural network to some arbitrary values prior to beginning training? My understanding is the the TMVA::Reader class is able to read the .xml files produced by trained models in order to recreate said models, but I don’t know that the TMVA::Factory class can do this.

Thanks!
Brian

2

Hi,

sorry, while in principle that would indeed a fairly easy thing to do, it’s not implemented for the MLP. The new DNN implemented by Peter had in his stand alone version some ‘resume-training’ from a given weightfile implemented, but it seems to be commented out for the TMVA version - I guess because it simply doesn’t ‘fit’ into the structure :frowning:

Helge

3

Hi Helge,

Thank you for the quick reply :slight_smile:.

That is unfortunate that no such feature exists :frowning: (or rather, that it is not implemented in the TMVA version of the current ML methods). It seems that there are already a lot of new features in the pipeline for development in TMVA; nevertheless, is there a good person to contact to request this feature as well?

Cheers,
Brian