End-to-end solutions

[marcorossi5]

Inference mode in Larsoft

Input

root file from dunetpc, containing raw::RawDigit and sim::SimChannel products.

Output

NN output containing roi and dn outputs as larsoft products inside ROOT files

Interface

A C++ interface is required to be able to export PyTorch model into an appropriate format for larsoft. Like this one: WaveformRoiFinder_module.cc

Possible solutions:

• ONNX: save the model trained in PyTorch and load it back in Tensorflow since larsoft contains tf (link here)
• TMVA: save the model from PyTorch and load it via TMVA. TMVA operates directly on the ROOT format. Not sure if this conforms to larsoft standards.
  Useful links: Class Reference, Pytorch tutorial