vad
Provides self-contained operating guidance for the VAD autonomous-driving repository, including legacy environment setup, nuScenes/CAN-bus data preparation, model/plugin configuration, training, single-GPU evaluation, and prediction visualization.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- vad
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Provides self-contained operating guidance for the VAD autonomous-driving repository, including legacy environment setup, nuScenes/CAN-bus data preparation, model/plugin configuration, training, single-GPU evaluation, and prediction visualization.
- disable-model-invocation
- true
02 资源文件 5
- 参考文档references/environment-and-verification.md2.4 KB
- 参考文档references/repo-provenance.md2.0 KB
- 参考文档references/repo-routing-metadata.json506 B
- 参考文档references/troubleshooting.md2.4 KB
- 脚本scripts/check_environment.py2.5 KB
03 SKILL.md 正文
VAD repository skill
Use this skill when working with VAD v1 or the bounded VADv2 additions: preparing nuScenes temporal annotations, understanding the VAD plugin/config graph, training or evaluating VAD, or inspecting rendered predictions.
Route by task
- Install prerequisites, prepare nuScenes/CAN-bus, generate temporal PKLs, or validate paths: data-preparation.
- Understand VAD/VADHead/transformers, custom registries, config families, stage-1/stage-2/e2e, or VADv2 boundaries: architecture-configuration.
- Train, resume, override configs, evaluate, format results, or reproduce released weights: training-evaluation.
- Inspect result artifacts or render camera/LiDAR/map/trajectory predictions: visualization.
Operating order
- Read environment-and-verification.md and troubleshooting.md.
- Keep the repository's legacy dependency family together: Python 3.8-era PyTorch/CUDA,
mmcv-full==1.4.0,mmdet==2.14.0,mmsegmentation==0.14.1, MMDetection3Dv0.17.1,timm, andnuscenes-devkit==1.1.9. Use compatible builds for the target host rather than mixing modern OpenMMLab packages into these configs. - Validate data and config contracts before any expensive command. The bundled scripts are read-only and do not download data, build native operators, train, evaluate, or render.
- Treat actual VAD model/plugin execution as CUDA and native-extension dependent. CPU imports or config parsing do not prove model-runtime support.
- Record the exact config, checkpoint, normalization, data root, and output artifact for every experiment. Read repo-provenance.md when checking whether this guidance matches a repository snapshot.
Safe checks
From the generated skill tree, the following are safe preflights:
python scripts/check_environment.py
python sub-skills/data-preparation/scripts/check_data_layout.py --data-root DATA_ROOT --canbus-root CANBUS_PARENT
python sub-skills/architecture-configuration/scripts/check_config_contract.py CONFIG --check-plugin
python sub-skills/training-evaluation/scripts/check_training_contract.py CONFIG
python sub-skills/visualization/scripts/inspect_result_artifact.py RESULTS.pkl
Use the workflow sub-skills for the actual repository commands. Do not start full training/evaluation, download nuScenes/checkpoints, or run a renderer merely to test installation.
Important caveats
- VAD uses custom temporal nuScenes info files and CAN-bus-derived fields; stock MMDetection3D info PKLs are not drop-in replacements.
- The project documentation recommends two-stage training and warns that distributed evaluation can produce inaccurate metrics; evaluate with one GPU and
--launcher none. - Released checkpoint reproduction requires the legacy
img_norm_cfg(mean=[103.530,116.280,123.675],std=[1,1,1],to_rgb=False), not the newer setting shown in the current configs. - Missing
ball_query_extor another*_extis an environment/native-build block, not a model-config typo. Preserve that distinction in reports.
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).