How liu-feng/pct-tensorflow works
Project type and purpose
• This is a machine-learning training framework in TensorFlow implementing PCT (Point Cloud Transformer) for 3D point cloud classification.
• It uses ModelNet40 as its dataset and provides both training (train.py) and evaluation (test.py) scripts.
• Custom high-performance TensorFlow ops (grouping, sampling, interpolation) are written in C++/CUDA under tf_ops.
File structure
• Top level
• README.md – project overview and usage
• model.py – definition of the PCT network (backbone + transformer layers)
• train.py / test.py – entry points for training and testing
• utils/ – helper modules (data provider, PointNet utilities, TF utilities)
• tf_ops/ – custom TF kernels, each under its own subdirectory:
– grouping/ (tf_grouping.cpp, CUDA sources, Python wrapper, tests)
– sampling/ (tf_sampling.cpp, CUDA sources, Python wrapper, compile scripts)
– interpolation/ (tf_interpolate.cpp, Python wrapper, tests, visualization script)
• Build/deploy
• Each tf_op has a *_compile.sh to build the shared library
• Test harnesses under each op’s test/ directory
README for insights
• Describes classification task on ModelNet40
• Directs user to build tf_ops, place dataset in datasets/modelnet40_ply_hdf5_2048/
• Summarizes workflow: compile ops → train.py → test.py