WarpConvNet is a high-performance library for 3D deep learning, built on NVIDIA's Warp framework. It provides efficient implementations of:
- Point cloud processing
- Sparse voxel convolutions
- Attention mechanisms for 3D data
- Geometric operations and transformations
# Specify the CUDA version
export CUDA=cu121 # For CUDA 12.1
pip install torch==2.5.1 torchvision==0.20.1 --index-url https://download.pytorch.org/whl/${CUDA}
pip install torch-scatter -f https://data.pyg.org/whl/torch-2.5.0+${CUDA}.html
pip install warpconvnet
Available optional dependency groups:
warpconvnet[dev]
: Development tools (pytest, coverage, pre-commit)warpconvnet[docs]
: Documentation building toolswarpconvnet[models]
: Additional dependencies for model training (wandb, hydra, etc.)
./
├── docker/ # Docker build files
│ ├── build.sh
│ └── Dockerfile
├── examples/ # Example applications
│ ├── mnist.py
│ └── modelnet.py
├── scripts/ # Development utilities
│ ├── build_docs.py
│ ├── dir_struct.sh
│ └── serve_docs.py
├── tests/ # Test suite
│ ├── base/ # Core functionality tests
│ ├── coords/ # Coordinate operation tests
│ ├── features/ # Feature processing tests
│ ├── nn/ # Neural network tests
│ └── types/ # Geometry type tests
└── warpconvnet/ # Main package
├── geometry/ # Geometric operations
│ ├── base/ # Core definitions
│ ├── coords/ # Coordinate operations
│ ├── features/ # Feature operations
│ └── types/ # Geometry types
├── nn/ # Neural networks
│ ├── functional/ # Neural network functions
│ └── modules/ # Neural network modules
├── ops/ # Basic operations
└── utils/ # Utility functions
For complete directory structure, run bash scripts/dir_struct.sh
.
python examples/modelnet.py
pip install warpconvnet[models]
cd warpconvnet/models
python examples/scannet.py train.batch_size=12 model=mink_unet
Build and run with GPU support:
# Build container
cd warpconvnet/docker
docker build -t warpconvnet .
# Run container
docker run --gpus all \
--shm-size=32g \
-it \
-v "/home/${USER}:/root" \
-v "$(pwd):/workspace" \
warpconvnet:latest
# Run all tests
pytest tests/
# Run specific test suite
pytest tests/nn/
pytest tests/coords/
# Run with benchmarks
pytest tests/ --benchmark-only
# Build docs
python scripts/build_docs.py
# Serve locally
python scripts/serve_docs.py
Apache 2.0
If you use this code in your research, please cite:
@misc{warpconvnet2025,
author = {Chris Choy and NVIDIA Research},
title = {WarpConvNet: High-Performance 3D Deep Learning Library},
year = {2025},
publisher = {NVIDIA Corporation},
howpublished = {\url{https://github.com/NVlabs/warpconvnet}}
}