second-pytorch
Guide historical SECOND and PointPillars LiDAR 3D detection workflows for KITTI and NuScenes, including data contracts, configuration, geometry, evaluation, guarded training/inference, and the web viewer.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- second-pytorch
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Guide historical SECOND and PointPillars LiDAR 3D detection workflows for KITTI and NuScenes, including data contracts, configuration, geometry, evaluation, guarded training/inference, and the web viewer.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/compatibility.md3.2 KB
- 参考文档references/repo-provenance.md2.1 KB
- 参考文档references/repo-routing-metadata.json337 B
- 参考文档references/troubleshooting.md4.7 KB
03 SKILL.md 正文
SECOND PyTorch
Use this repo skill for source-grounded work involving the deprecated
traveller59/second.pytorch 1.6.0-alpha-era codebase: KITTI/NuScenes LiDAR
3D detection, VoxelNet/SECOND, PointPillars, voxel and box geometry, text
protobuf configs, evaluation, and the historical KITTI web viewer.
First decision: historical code or maintained replacement?
The repository is explicitly deprecated. For new detector development, prefer OpenPCDet or MMDetection3D. Use this skill when a historical config, checkpoint, experiment, data artifact, or compatibility question requires this code's contracts. Read repo-provenance.md before assuming this skill matches a checkout, and read compatibility.md before importing the detector.
Runtime boundary
- The checkout has no
pyproject.toml,setup.py, requirements file, or supported modern package install. Use a private isolated environment and an explicitly supplied checkout only when the historical runtime is truly needed. - The executable, verified portion of this graph is CPU-safe/static: dataset layout validation, config/schema reasoning, NumPy geometry, and safe helper probes.
- Full sparse detector construction, GPU NMS, training, evaluation, checkpoint
restoration, and viewer checkpoint inference are legacy-backend guarded.
Current
spconv2.x APIs are not interchangeable with the source's old symbols. A successful Torch CUDA smoke orspconvimport is not detector verification. - Do not download datasets/checkpoints, start a service, launch the Qt viewer, run full training/evaluation, or mutate data as a skill smoke test.
Read troubleshooting.md for cross-cutting failures before changing dependencies, paths, or config files.
Public environment setup
There is no supported package installer for this source tree. For the verified
CPU-safe/static routes, create an isolated Python 3.10+ environment and install
only the public scientific/runtime dependencies needed by the selected helper,
for example numpy, scipy, numba, scikit-image, Pillow, protobuf<4,
fire, and optional nuscenes-devkit or flask-cors. The bundled geometry and
layout helpers use standard Python/NumPy and do not require the detector stack.
For model work, install a separately reproducible legacy Torch/spconv/Numba
combination and run the backend probe. Do not install the newest spconv and
assume compatibility; the source expects old symbols that current 2.x wheels
may not expose. Never mutate an existing environment or use a CUDA import as a
substitute for the gate in compatibility.md.
Route by task
| Request signal | Read next |
|---|---|
| KITTI/NuScenes root, infos, reduced clouds, ground-truth database, sweeps, custom dataset, voxel/preprocess layout | data-preparation |
| train/evaluate, VoxelNet, PointPillars, checkpoint, config selection, fp16, multi-GPU, spconv/Numba compatibility | training-and-inference |
| boxes, corners, yaw, camera/lidar conversion, anchors, encoding, IoU/NMS, KITTI AP, NuScenes results | geometry-and-evaluation |
| KITTI web viewer, Flask/CORS, browser/backend URLs, point-cloud or detection payloads, buildNet/inference UI | visualization-and-serving |
Cross-route requests should start with data preparation, then configuration and training/inference, while geometry/evaluation owns box conventions and metrics. The viewer is a guarded consumer of the same data and checkpoint contracts.
Minimal safe checks
Run the bundled helpers from their linked sub-skills before any user-controlled workflow. They are self-contained and do not require the original checkout:
python sub-skills/data-preparation/scripts/validate_dataset_layout.py --help
python sub-skills/geometry-and-evaluation/scripts/geometry_smoke.py
python sub-skills/training-and-inference/scripts/check_legacy_backend.py
python sub-skills/visualization-and-serving/scripts/check_viewer_deps.py --json
The geometry helper should finish with geometry smoke: PASS. The other probes
are diagnostic; missing optional components or legacy symbols must be reported,
not silently substituted.
Public contract reminders
- Internal LiDAR boxes normally use
[x, y, z, w, l, h, yaw]; KITTI camera boxes use a different dimension/order convention. Never swap dimensions by visual intuition. - KITTI preparation expects
training/andtesting/trees with matching image, calibration, point-cloud, and (for training) label stems. NuScenes preparation depends on version metadata, samples, sweeps, and a matching velocity/non-velocity dataset class. - Historical config field names such as
kitti_root_pathandkitti_info_pathare used by both dataset families. Keep dataset class, info files, feature width, box dimensionality, and class order consistent.
Handoff
A useful result reports the selected route, source/skill provenance, validated layout/config/geometry observations, exact generated output paths, backend gate status, and any unverified legacy limitation. Do not claim a metric, detector run, or viewer inference result from a static or CPU-safe check alone.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).