This repository contains code for training and evaluating a sparse implemenation of ResNet on Summit at ORNL. The expected input is Richardson-Lucy deconvolved 2νββ and background events from the NEXT detector.
Jobs are submitted through the job submission script scn_hv.lsf with the command
bsub scn_hv.lsf
To run the training code, change the line in scn_hv.lsf to execute run_training.py:
jsrun -n 24 -a 1 -c 2 -g 1 python -m run_training.py
To run the evaluation code, change the line in scn_hv.lsf to execute run_score_new_events.py:
jsrun -n 1 -a 1 -c 2 -g 1 python -m run_score_new_events.py
The -n flag specifies the number of GPUs to use. It should be equal to six times the number of compute nodes (which is specified above by -nnodes).
The input training and testing files are defined in larcvconfig_train_lr.txt and larcvconfig_test_lr.txt. All other parameters can be changed within the dnn_larcv.py script (learning rate, number of epochs, output file paths, etc.)