This MR is implementing the improved tau ID based on a deep set NN (ATLTAU-1798), which is meant to replace the current default RNN tau ID. It is largely reusing the RNN implementation. The NN uses a bunch of extra variables compared to the RNN, namely the RNN track classification scores (from the RNN tau track classifier, 4 probabilities per track). Support has been added for both RecExCommon and CA configurations.
The calibration files (3 json files for the NN and 3 ROOT files for the WPs, for 1p/2p/3p taus) are on cvmfs.
The goal is to run this new tau ID in AODtoDAOD via a wrapper in DerivationFrameworkTau, in a MR yet to come (implementation ready but requires this MR and !59677 (merged) to be merged), so that the deep set ID goes into the next DAOD_PHYS production to be launched imminently. The new tau ID is implemented by not deployed, this MR is not FT0V.
Tagging @dta , @sineadf , @mmlynari , @ghamity , @mhodgkin , @christos , @nstyles , @jcatmore , @calpigia , and @jcardena as the main developer of the deep set tau ID. Juan, please check the MR to make sure it looks as expected ;-)