Event format

Every dataset returns an event as a flat dict of NumPy arrays with one row per point, and every model reads a packed batch of such events.

One event

Point-aligned fields have one row per point; scalar quantities keep a trailing axis of length 1. A PILArNet-M event from PILArNetH5Dataset contains:

Key	Shape	Content
coord	(N, 3) float32	point position
energy	(N, 1) float32	deposited energy
segment_motif	(N, 1)	semantic class
segment_pid	(N, 1)	particle class
instance_particle	(N, 1)	particle id within the event
instance_interaction	(N, 1)	interaction id within the event
segment_interaction	(N, 1)	background flag
momentum	(N, 1)	momentum of the point’s particle (GeV)
vertex	(N, 3)	interaction vertex of the point’s particle
is_primary	(N, 1)	primary-particle flag
name, split, revision	strings	bookkeeping

Code	segment_motif	segment_pid
0	shower	photon
1	track	electron
2	Michel	muon
3	delta	pion
4	low-energy deposit	proton
5	—	none (low-energy deposit)

-1 marks points to ignore. PILArNet-M coordinates lie in a cube of side 768; recipes rescale them, as described in Transforms. Other readers emit the same core keys (coord, energy, segment, instance) plus their own; see Supported detectors.

See one

PILArNet-M-mini test event 0 after the Panda transforms: 1,175 points, colored by their true labels. Drag to rotate, scroll to zoom, and switch between semantic and particle classes.

A batch

After the transforms, Collect keeps the keys you name and joins feat_keys into feat. collate_fn concatenates the events and builds offset:

Key	Shape	Content
coord	(N_total, 3)	transformed positions
grid_coord	(N_total, 3) int	grid cells, when the pipeline grid-samples
feat	(N_total, C)	features in feat_keys order
segment	(N_total,)	training target, when present
offset	(B,)	cumulative end index of each event; offset[-1] == N_total

Inside a model, Point(batch) wraps the dict and derives the per-point event index from offset.