Models¶

warpconvnet.models is a small zoo of reference 3D point-cloud and sparse-voxel networks. Every model is an importable nn.Module that consumes Points (or Voxels) and returns a Geometry or tensor.

from warpconvnet.models import (
    DGCNN, DGCNNEncoder,
    FIGConvNet, FIGConvNetDrivAer,
    MaskFormer, MaskTransformer,
    MinkUNet18, MinkUNet34, MinkUNet50, MinkUNet101,
    PointMinkUNet18, PointMinkUNet34,
    PointNet,
    PointTransformerV3,
    SpaCeFormer,
)

Model index¶

Model	Input	Task	Paper
PointNet	Points	Classification / Segmentation	Qi et al. 2017
DGCNN	Points	Classification / Segmentation	Wang et al. 2019
PointTransformerV3	Points (serialized)	Segmentation	Wu et al. 2024
MinkUNet18/34/50/101	Voxels	Semantic segmentation	Choy et al. 2019
PointMinkUNet18/34	Points → Voxels → Points	Per-point segmentation	—
FIGConvNet	Points	Regression / Dense prediction	Choy et al. 2024
MaskFormer	Points	Instance / mask prediction	Cheng et al. 2021 (3D variant)
SpaCeFormer	Voxels	Hierarchical sparse-voxel U-Net	Choy et al. 2026

Importing in Hydra¶

Hydra _target_ strings resolve to the public path, so any of the above classes work as a drop-in target:

model:
  _target_: warpconvnet.models.MinkUNet34
  in_channels: 3
  out_channels: 20

python examples/train/scannet.py model._target_=warpconvnet.models.MinkUNet34