How erikwijmans/pointnet2_pytorch works
Project type and purpose
• This is a research-oriented PyTorch library/CLI for PointNet++ (3D point-cloud neural network) training and inference.
• The README confirms it implements PointNet2 in PyTorch, wraps custom CUDA ops, supports multi-GPU via PyTorch DataParallel, uses Hydra for configuration and PyTorch Lightning for training orchestration.
File structure
• Top-level:
• setup.py, pyproject.toml, requirements.txt: packaging and dependencies
• pointnet2/: the main Python package
• pointnet2_ops_lib/: the custom CUDA extension library
• tests/: unit tests for classification and semantic segmentation
• pointnet2/ subdirectories:
• config/: all Hydra YAML configs (global config.yaml, subfolders model/ task/ task_model/)
• data/: Python DataLoader classes and utilities
• models/: Python model definitions for SSG/MSG and cls/semseg variants
• train.py: the entry-point training script
• utils/: miscellaneous Python helpers
• pointnet2_ops_lib/:
• pointnet2_ops/: Python bindings and helpers
• _ext-src/: C++/CUDA source for sampling, grouping, interpolation, ball-query kernels
• setup.py: build/install script for the extension
README for additional insights
• Project uses:
• PyTorch ≥1.0
• CUDA for custom ops
• Hydra for experiment/config management
• PyTorch Lightning for training loops & multi-GPU
• Usage pattern:
• “pip install -e .” to install Python package
• “pip install pointnet2_ops_lib/” to build CUDA kernels
• “python pointnet2/train.py task=…” to launch experiments
Detail without over-fragmentation
• Test suite as a validation overlay or side component that imports & runs key modules.