This package is a work-in-progress for generating inferred signals in dead regions of a DUNE-ND-like liquid argon time projection chamber (LArTPC). Right now, the targeted architecture for this model is a masked autoencoder, which is to be adapted to sparse 3D images.
The framework is being developed with the DUNE-ND 2x2 prototype in mind, and development is utilizing existing 2x2 simulation starting from the larnd2supera stage. Small subrun samples can be found on SLAC's SDF computing system in /sdf/group/neutrino/cyifan/larnd2supera/larcv_output/output_00679-larcv.root (thank you to Yifan and others for providing these samples!)
To pre-process the larnd2supera simulation into patched images, the dataprep.py script is provided. This script should be called like
python dataprep.py [-h] inputRoot [inputRoot ...] preppedOutput
This will iterate through 2x2 images, find small images (30x30x30 voxels, this is configurable), and then apply a patching scheme (6x6x6 patches per image, so that each patch is 5x5x5 voxels, this is also subject to optimization in the future). The resulting patches are saved in a sparse representation along with a record of the patching scheme to an hdf5 file for faster reading by the train-time data loader.
A simple data loader is defined in dataloader.py which will read from this prepped hdf5 file and apply a run-time mask to the patches. The probability to keep or mask a given patch is given as an argument to the dataloader at initialization. A very simple example of iterating through batches with this dataloader is included with the definition. Constructing sparse tensors from these sparse patches is still an open question, but a utility function is included.