paddle-detection
Guides PaddleDetection object detection, segmentation, keypoint, multi-object tracking, configuration, training, evaluation, inference, export, deployment, and PP-Human/PP-Vehicle pipeline workflows.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- paddle-detection
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Guides PaddleDetection object detection, segmentation, keypoint, multi-object tracking, configuration, training, evaluation, inference, export, deployment, and PP-Human/PP-Vehicle pipeline workflows.
- disable-model-invocation
- true
02 资源文件 6
- 参考文档references/install-and-environment.md3.7 KB
- 参考文档references/repo-provenance.md1.7 KB
- 参考文档references/repo-routing-metadata.json505 B
- 参考文档references/troubleshooting.md4.7 KB
- 参考文档references/verification-candidates.md1.5 KB
- 脚本scripts/check_paddledet_environment.py4.5 KB
03 SKILL.md 正文
PaddleDetection
Use this skill when a task names PaddleDetection, paddledet, PP-YOLO/PP-YOLOE/PicoDet, PaddleDetection YAML configs, tools/train.py, tools/infer.py, model export, PP-Human, PP-Vehicle, PP-Tracking, or PaddleDetection deployment.
This is an operating router, not a copy of the repository. Work against the user's PaddleDetection checkout or installed package, and use the bundled references/scripts in this skill for decisions and preflight checks. Do not assume the creation checkout or its private environment still exists.
First route
- Identify the task family and choose exactly one primary route below.
- Establish the target checkout/package version, PaddlePaddle device, dataset layout, config path, weights source, and whether network/GPU/vendor runtimes are allowed.
- Run the bundled environment preflight before importing deep modules:
scripts/check_paddledet_environment.py. - Keep
use_gpu=falseor--device=CPUfor a deterministic smoke test unless CUDA/TensorRT is explicitly required and available. - Read
references/troubleshooting.mdwhenever imports, weights, config paths, optional dependencies, or device checks fail.
Routes
- Choose a model, inspect or override YAML, use the registry, or list model families: read
sub-skills/model-zoo-and-configuration/SKILL.md. - Prepare COCO/VOC/WIDER/MOT/keypoint/custom data, convert annotations, validate paths, split semi-supervised data, or slice small-object data: read
sub-skills/data-preparation/SKILL.md. - Train, fine-tune, resume, evaluate, run repository inference, use distributed launch, AMP, VisualDL/W&B, or troubleshoot
tools/*.py: readsub-skills/training-evaluation-inference/SKILL.md. - Export a trained model or deploy with Paddle Inference, Serving, Lite, ONNX, FastDeploy, or TensorRT: read
sub-skills/deployment-and-export/SKILL.md. - Run PP-Human, PP-Vehicle, PP-Tracking, ReID, plate/attribute/action/violation analysis, or image/video/RTSP pipelines: read
sub-skills/industrial-pipelines/SKILL.md.
Package baseline
The distribution is paddledet and the primary import is ppdet. The repository's baseline requirements include PaddleDetection utilities plus NumPy below 2, OpenCV, PyYAML, Shapely, SciPy, pycocotools, VisualDL, MOT dependencies, and image augmentation utilities. Install a compatible PaddlePaddle CPU or GPU build first, then install the repository requirements and the package in editable mode when working from source.
Minimal import check:
python -c "import paddle, ppdet; print(paddle.__version__, ppdet.__version__)"
python -c "import paddle; paddle.utils.run_check()"
The source checkout currently advertises package version 0.0.0 even when its Git tag is newer. Treat ppdet.model_zoo.get_config_file() as a network/cache operation and prefer a checked-out config path for source workflows; see the model-zoo route for the observed 404 failure mode.
Shared operating rules
- Configs are YAML and may recursively merge
_BASE_files. Use-o key=valueonly for deliberate, logged overrides; verify paths andnum_classesafter merging. - Weights may be local files or remote URLs. Never download weights or datasets without explicit network approval; record the exact source and checksum when reproducibility matters.
- CPU checks validate parsing/import/data contracts only. They do not prove CUDA, TensorRT, Paddle Lite, Serving, FastDeploy, or vendor accelerator behavior.
- Treat training, benchmark, full pipeline, and TIPC commands as expensive unless a tiny fixture and explicit runtime budget are available.
- Keep generated outputs outside the source tree when possible: checkpoints under a chosen
save_dir, inference exports under a separateoutput_inference-style directory, and visualizations/results under a separate output directory. - Before changing a config, inspect the dataset reader, metric, input shape, class count, device flags, and pretrained/weights fields together.
Read references/install-and-environment.md for supported dependency variants and backend boundaries, references/verification-candidates.md for safe checks, and references/repo-provenance.md before deciding whether this skill is stale.
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).