point-cnnUse the legacy PointCNN TensorFlow 1.x repository to build X-Conv models, prepare point-cloud datasets, train classification or segmentation workflows, and validate evaluation artifacts.
Independent GitHub snapshot · Automated classification · Original documentation governs
Use this skill for the public yangyanli/PointCNN codebase: legacy TensorFlow
1.x point-cloud classification and segmentation built around X-Conv/X-DeConv.
This is a versioned operating guide, not a copy of the source checkout. Read
provenance before deciding whether a different
checkout is still covered.
For a task that crosses routes, prepare and validate data first, select a model/setting second, run a bounded workflow third, and inspect/merge artifacts last. Keep the owning sub-skill in control of its detailed contract.
This repository uses TensorFlow 1.x graph-mode APIs (tf.contrib, tf.layers,
placeholders, sessions, and tf.py_func). Start with the bundled read-only
probe:
python scripts/check_environment.py --help
python scripts/check_environment.py
Use a deliberately isolated legacy environment and the public dependencies listed in the repository's requirements, adapting versions to the target platform. Do not assume that a current TensorFlow release or eager execution will work. Classification requires a functioning TensorFlow 1.x graph stack.
Every supplied segmentation setting uses sampling = 'fps'. FPS,
GatherPoint, and related operators are registered only for a CUDA GPU and
must be built against a compatible TensorFlow framework ABI, CUDA toolkit, C++
toolchain, driver, and visible GPU. A CPU import, CPU graph build, or successful
HDF5 validation is not segmentation verification. Read the FPS diagnostics
and keep the required backend result visible; do not silently replace it with a
CPU fallback.
Use explicit input and output paths. Historical shell launchers may background jobs or assume repository-relative directories; adapt them to foreground commands and disposable output roots. Do not make dataset downloads, archive extraction, full conversions, Semantic3D acquisition/decompression, long training, or benchmark runs part of a smoke check. Validate a tiny fixture or CLI help first, and record checkpoint/setting/data provenance before treating any metric as meaningful.
Cross-cutting symptoms and recovery order are in troubleshooting. The generated skill is not imported or synchronized to another agent by this creation run.