Training data

The converters transform the raw ROOT sources into event-level graph datasets for PyTorch.

Locations

Dataset	Format	JLab path or status
VTP-reconstructed graphs	One LibTorch .pt file per event	/expphy/volatile/hallc/c-kaonlt/ckin/nps-data/reco_vtp/*.pt
Geant4+SIMC overlap graphs	One LibTorch .pt file per combined event	/lustre24/expphy/volatile/hallc/c-kaonlt/ckin/nps-data/geant4_overlaps/*.pt
VTP/JANA2 event arrays	One directory of .npy arrays per event	users/ckin/nps-data/vtp_cluster_data in the work area; temporary

The volatile and work-area datasets are not archival. Confirm that a path still exists before submitting jobs, and avoid treating it as the only copy.

Common LibTorch format

The tensor fields follow the graph-data conventions described in PyTorch Geometric's Creating Graph Datasets guide, with node features, connectivity, targets, and positions stored for each event.

Each .pt event serializes five tensors in this order:

(x, edge_index, edge_attr, y, pos)

Tensor	Shape	Meaning
x	[num_nodes, num_features]	Waveform or pulse features
edge_index	[2, num_edges]	Optional intra-cluster graph connectivity
edge_attr	[num_edges, 0]	Empty edge-feature tensor in the current converters
y	[num_nodes, 1]	Cluster identifier for each node
pos	[num_nodes, 2]	Detector (column, row)

Block IDs are used while constructing a graph but are not saved as a separate tensor. See Datasets and data loaders for loading both the .pt and .npy layouts.

VTP-reconstructed graphs

reco_vtp.exe reads the same replay waveforms, emulates the fADC250 and VTP clustering logic, and matches reconstructed clusters to recorded VTP clusters. It saves the same five-tensor layout, with waveform features, reconstructed cluster IDs, detector positions, and optional intra-cluster edges.

Geant4+SIMC overlap graphs

sim_data.exe accumulates --overlaps consecutive simulation events and emits one combined graph:

• Node: a calorimeter block occurrence belonging to a reconstructed cluster.
• Node feature x: all pulse (energy, time) pairs for that physical block, padded with zeros to 4 * overlaps values. The default overlaps=5 produces 20 features per node.
• Target y: a graph-local cluster ID starting at 1 across the combined input events.
• Position pos: detector (column, row).
• Edges: optional intra-cluster connectivity.

Only complete groups of overlaps input events are saved. The current implementation accepts --dt but does not apply that time-gap value when building pulse features.

Access and conversion

Use JLab VDI or another host with access to the listed filesystems. To produce a new dataset, follow the ROOT-to-PyTorch converter tutorial.