Inferers¶

Inferers are used to process point clouds at inference time. One of their main responsibilities is to split large point clouds into smaller chunks for efficient infererence and avoiding memory issues. This is also the place where test-time augmentation (TTA) is applied, if desired.

from torch_pointcloud.inferers import SlidingWindowInferer, TTAInferer
from torch_pointcloud.transforms import Compose, RandomFlip, RandomRotate

inferer = TTAInferer(
    base=SlidingWindowInferer(block_size=6.0),
    transforms=Compose([
        RandomRotate(keys="pos", angle_range=(-180.0, 180.0), axis=2, p=1.0),
        RandomFlip(keys="pos", axes=[0, 1], p=0.5),
    ]),
    num_passes=4,
    aggregate="mean",
)

data = {"pos": torch.randn(1000, 3), "x": torch.randn(1000, 3), "batch": torch.zeros(1000)}
probs = inferer(data, predictor=lambda d: model(d["x"], d["pos"], d["batch"]))

Which one to use¶

Inferer	Runs the predictor on	Reproduces
`SimpleInferer`	the whole scene, once	single-pass evaluation, the Lightning default
`SlidingWindowInferer`	cubic blocks on a regular grid	block-based S3DIS / ScanNet protocols
`VoxelPartitionInferer`	whole-extent downsamples, one point per voxel	the fragment protocol of sparse and point transformers
`KNNWindowInferer`	fixed-budget kNN crops around the least-covered point	possibility-driven crop voting (RandLA-Net)
`PotentialSphereInferer`	radius spheres drawn from a potential grid	potential sphere voting (KPConv)
`TTAInferer`	any base inferer, under several views	test-time augmentation and voting
`PartRefinementInferer`	any base inferer, then a neighbor vote	part-segmentation label refinement

The last two wrap another inferer rather than replacing it.