---

name: torch-points3d description: "Use Torch Points3D for point-cloud deep learning APIs, datasets/transforms, Hydra training/evaluation, checkpoints, sparse backend decisions, and registration workflows." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Torch Points3D

Use this repo skill when a task involves Torch Points3D, the torch_points3d Python package, point-cloud segmentation/classification/object-detection/ panoptic/registration experiments, Hydra configs under conf/, high-level PointNet2/KPConv/RSConv/SparseConv3d APIs, or troubleshooting the old PyTorch/PyG/sparse-backend stack.

First Checks

• Read repo provenance before deciding whether this skill matches a local checkout or should be refreshed.
• Read installation and compatibility before installing or debugging dependency versions.
• Read cross-cutting troubleshooting for import failures, CUDA/sparse backend mismatches, Hydra config errors, dataset downloads, W&B/TensorBoard surprises, and checkpoint issues.
• Run torch_points3d_env_probe.py for a safe import/version/backend diagnostic.

Minimal Install Pattern

Torch Points3D is a legacy PyTorch/PyG stack. Install a compatible PyTorch build first, then the matching PyG extension wheels, then Torch Points3D and optional workflow packages:

python -m pip install "torch==1.8.*" "torchvision==0.9.*"
python -m pip install "torch-geometric==1.7.*" "torch-scatter==2.0.*" \
  "torch-sparse>=0.6.10,<0.6.13" "torch-cluster==1.5.*"
python -m pip install torch-points3d
python - <<'PY'
import torch, torch_points3d
print(torch.__version__)
print(torch_points3d.__name__)
PY

For modern CUDA, old wheels may not exist. Prefer a version-matched legacy Python environment for reproduction, or build the compiled extensions from source. Do not claim sparse models are runnable until MinkowskiEngine or torchsparse imports in the target environment.

Route by Task

• Use model-apis for high-level application constructors (PointNet2, KPConv, RSConv, SparseConv3d), standalone forward smokes, pretrained registry tags, and dense/partial-dense/sparse backend choices.
• Use datasets-transforms for dataset config resolution, Torch Geometric Data/Batch expectations, transform/filter YAML, collate rules, ScanNet/S3DIS/ShapeNet/ModelNet/SemanticKITTI layouts, and safe dataset preflight checks.
• Use training-evaluation for Hydra train.py/eval.py workflows, Trainer, checkpoint resume/eval, visualization/logging, experiment output discovery, and forward inference from checkpoints.
• Use registration-workflows for 3DMatch/KITTI/ETH/TUM/KAIST/ModelNet registration configs, pair/fragment data contracts, descriptor evaluation, FPS sampling, and registration-specific backend caveats.

Shared Decision Rules

• Start with a CPU import/config smoke before any training, dataset download, checkpoint download, or GPU/sparse backend build.
• Treat conf/ as the source of command composition. Training commands need task, models, data, and model_name overrides that resolve together.
• Dense PointNet2/RSConv examples use batched tensors shaped like [B, N, C] for features and [B, N, 3] for positions; KPConv and sparse workflows use PyG-style variable-size Data batches.
• SparseConv3d, Minkowski models, PointGroup, PVCNN/MS-SVConv, and some registration paths require optional sparse backends. Missing optional backend imports are not package-wide failures; route to the relevant troubleshooting page.
• Disable or explicitly configure W&B/TensorBoard/profiling for short local smoke runs unless the user wants logging artifacts.
• Dataset acquisition can involve external licenses, large downloads, or credentials. Prefer tiny fixtures, config composition, and layout validation before launching downloads.

Safe Shared Command

python scripts/torch_points3d_env_probe.py --json

Use --require-package, --require-pyg, or --require-sparse-backend when a task needs a hard pass/fail diagnostic instead of an informational report.