Transforms#

A transform takes an event dict and returns it, changed. A config lists transforms in order, and Compose runs them on every event.

The Panda pipeline, step by step#

Step	Event after the step
raw event	`coord` (N, 3) in detector units, `energy` (N, 1)
`NormalizeCoord(center=[384, 384, 384], scale=768·√3/2)`	`coord` inside the unit ball
`LogTransform(min_val=0.01, max_val=20.0)`	`energy` on [-1, 1], log-scaled
`GridSample(grid_size=0.001, return_grid_coord=True)`	one point per occupied cell, M ≤ N points; integer `grid_coord` added
`Copy(keys_dict={"segment_motif": "segment"})`	training target under the name models expect
`ToTensor()`	every array is a torch tensor
`Collect(keys=("coord", "grid_coord", "segment"), feat_keys=("coord", "energy"))`	only the listed keys, plus `feat` (M, 4)

Copy appears in training recipes only; inference needs no target.

Common transforms#

Transform	What it does
`NormalizeCoord(center, scale)`	`coord = (coord - center) / scale`; uses the centroid and the largest radius when omitted
`LogTransform(min_val, max_val, keys=("energy",))`	log-scales the keys onto [-1, 1]
`GridSample(grid_size, mode, return_grid_coord)`	keeps one point per occupied cell and subsamples every point-aligned key
`Copy(keys_dict)`	copies keys under new names
`ToTensor()`	converts NumPy arrays to tensors
`Collect(keys, feat_keys)`	keeps `keys`, concatenates `feat_keys` into `feat`, records the point count for `offset`
`RandomRotate`, `RandomFlip`, `RandomScale`, `RandomJitter`	geometric augmentation
`RandomDropout`, `ShufflePoint`, `SphereCrop`	drop, reorder or crop points

There are 46 registered transforms in all; Python API lists each one.

Point-aligned keys#

Transforms that drop or reorder points, such as GridSample, crops and dropout, subsample every key named in the event’s index_valid_keys list. A new per-point key must be in that list before the first of those transforms, or it no longer lines up with coord.

Write a transform#

See Add a component.