Generalized MVA/ANN fast inference library and model storage
With ever growing TMVA auto-generated inefficient code being dumped in the LHCb code base, it is more and more needed to have a generalized MVA/ANN fast inference library and (independent) model storage (like in ParamFiles
). Something like onnx
file format with onnxruntime
as inference library is suggested. This is already being used in e.g. ATLAS.
First task it add a compile time SIMD custom version on in the inference side of NNs, as ONNXRuntime
with first tests was not fast enough, see e.g. LHCb!4305 (merged).
Leave the more general infrastructure with e.g. ONNXRuntime
to a later date / issue