Skip to content

two track MVA

Niklas Stefan Nolte requested to merge NN_NN into master

This implements the two-track neural network architecture as presented in https://indico.cern.ch/event/1065920/#3-hlt1-inclusive-lines-tuning

It is done similarly to the catboost lines: Parameters are read from a json file during init and model is given to the algorithm via constants.

The json file also has the nominal cut to achieve 660 kHz, but its not used for now.
Should it be a property to be configured, maybe defaulting to whatever is in the json? How would I do that?

Also, while there is some preselection on the tracks that go into secondary vertex creation, the mva is still evaluated way more often than needed, because it does not consider the preselection that is done in the line. Maybe that does not matter all that much, but its still kind of annoying.

Edited by Niklas Stefan Nolte

Merge request reports