training
Guides YOLOP training, validation, evaluation metrics, staged multitask flags, checkpoints, auto-anchor behavior, and training smoke diagnostics.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- training
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Guides YOLOP training, validation, evaluation metrics, staged multitask flags, checkpoints, auto-anchor behavior, and training smoke diagnostics.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/evaluation.md2.7 KB
- 参考文档references/troubleshooting.md3.1 KB
- 参考文档references/workflows.md4.1 KB
- 脚本scripts/train_smoke.py4.2 KB
03 SKILL.md 正文
YOLOP Training and Evaluation
Use this sub-skill when the task asks to train YOLOP, run validation/evaluation, change end-to-end versus staged/single-task modes, diagnose tools/train.py or tools/test.py, reason about MultiHeadLoss, configure auto-anchor, or interpret YOLOP detection/segmentation metrics.
Do not use this sub-skill for only preparing data roots (use data-preparation), running demo inference (use inference), or exporting ONNX/TensorRT artifacts (use export).
Read first
- references/workflows.md gives the source training and evaluation command shapes, config fields, checkpoint behavior, and staged/single-task modes.
- references/evaluation.md explains validation outputs, detection metrics, segmentation metrics, and visualization side effects.
- references/troubleshooting.md covers data/config errors, torch-version loss issues, GPU/DDP pitfalls, batch size problems, and auto-anchor failures.
- Run scripts/train_smoke.py before expensive training to verify imports, model construction, forward shapes, and optionally expose loss-version incompatibilities on a tiny tensor.
Minimal workflow
- Validate BDD100K roots and masks with the
data-preparationsub-skill. - Edit or patch
lib.config.cfgvalues for dataset roots, image size, batch sizes, workers, training mode flags, pretrained checkpoint paths, andLOG_DIR. - Run a CPU or CUDA model smoke:
python sub-skills/training/scripts/train_smoke.py --repo-root /path/to/YOLOP --device cpu --image-size 128
- Start training from the YOLOP checkout root:
PYTHONPATH=. python tools/train.py
- For distributed training, only after a compatible CUDA torch/torchvision stack is installed and dataset/batch settings are ready:
PYTHONPATH=. python -m torch.distributed.launch --nproc_per_node=N tools/train.py
- Evaluate a checkpoint:
PYTHONPATH=. python tools/test.py --weights weights/End-to-end.pth
Important decisions
- All staged/single-task flags false means end-to-end multitask training.
TRAIN.SEG_ONLY,DET_ONLY,ENC_SEG_ONLY,ENC_DET_ONLY,DRIVABLE_ONLY, andLANE_ONLYchange both parameter freezing and loss components.NEED_AUTOANCHOR=Trueis not a cheap smoke check; it runs k-means/evolution over the training labels.tools/train.pyandtools/test.pyuse the yacs config more than their parser suggests; several parsed arguments are not applied byupdate_configin the current source.- CPU smokes validate code shape, but full training and reported paper/README metrics require real BDD100K data and practical GPU runtime.
Cross-links
- If dataset construction fails before the model starts, switch to ../data-preparation/SKILL.md.
- If the trained checkpoint should be used for a demo, switch to ../inference/SKILL.md.
- If the checkpoint should become ONNX or TensorRT artifacts, switch to ../export/SKILL.md.
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).