Task #19936
Change FD CNN to use TF
Start date:
05/13/2018
Due date:
% Done:
0%
Estimated time:
Description
We would like to change the FD CNN model to use the TF to improve speed and to keep up with the latest development.
I think there are two steps:
1. Apply the ProtoDUNE TF model to the FD and evaluate its performance.
2. Retrain a TF model for the FD.
History
#1 Updated by Tingjun Yang over 2 years ago
Here is information from Leigh:
if ((fNNetModelFilePath.length() > 5) && (fNNetModelFilePath.compare(fNNetModelFilePath.length() - 5, 5, ".nnet") == 0)) { fNNet = new nnet::KerasModelInterface(fNNetModelFilePath.c_str()); } else if ((fNNetModelFilePath.length() > 3) && (fNNetModelFilePath.compare(fNNetModelFilePath.length() - 3, 3, ".pb") == 0)) { fNNet = new nnet::TfModelInterface(fNNetModelFilePath.c_str()); } else { mf::LogError("PointIdAlg") << "File name extension not supported."; }
If you give it a .nnet file it runs Keras, if you give it a .pb file it runs TF
An FD version of the definition in here would be appropriate: dunetpc/fcl/common/featurelabelingmodules.fcl
The only definition there at the moment is the protoDUNE one
I am going to give it a try.