This repository contains a LArSoft package for generating numl-format event HDF5 files from artroot files. These event HDF5 files contain low-level information (simulated particles, energy deposits, detector hits etc), which can then be used for efficient downstream production of machine learning inputs using the pynuml package.
The develop branch of this repository is kept up-to-date with the DUNE flavour of LArSoft, with tagged versions produced regularly. If you have a dunesw development area set up via MRB, you can install this package simply by running
cd $MRB_SOURCE
mrb g -t <larsoft version tag> https://github.com/vhewes/numl
mrbsetenv
mrb i -j4
Once the repository is built, you should be able to run
lar -c hdf5maker_dune.fcl <artroot file>
to generate an event HDF5 file for downstream processing.
The pynuml package is designed to operate on a single HDF5 file that contains the full dataset. However, since high-statistics MC generation is typically performed at scale via grid computing, the output will typically be a larger number of small event HDF5 files. These files can be merged into a single file using the ph5concat package. This package can be installed manually, but for those working on Linux architectures, it is recommended to install via Anaconda as follows:
conda install -c conda-forge ph5concat
This concatenation utility leverages MPI to merge files efficiently in parallel on High-Performance Computing (HPC) nodes. The user can run
# merge files
ph5_concat -n <N> -i files.txt -o out.evt.h5
# add sequencing metadata to files
add_key -k event_table/event_id -c -f out.evt.h5
where files.txt should be a textfile containing the full paths to the event HDF5 files to be merged, formatted with one file per line, out.evt.h5 is the name of the merged output file, and <N> is the number of MPI ranks requested. For an HPC CPU node, <N> should be set to the number of CPU cores requested; in general, running on a personal computer is not recommended, and <N> should typically be <= 4.