training-and-data
Design and validate YOLOX training, data, Exp, caching, logging, and evaluation experiments.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- training-and-data
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Design and validate YOLOX training, data, Exp, caching, logging, and evaluation experiments.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/data-and-exp-reference.md4.9 KB
- 参考文档references/training-workflows.md5.0 KB
- 参考文档references/troubleshooting.md3.4 KB
- 脚本scripts/inspect_yolox_exp.py9.1 KB
03 SKILL.md 正文
YOLOX Training And Data
Use this sub-skill when a task involves YOLOX training or evaluation setup, dataset layout, custom Exp files, config overrides, caching, image-size behavior, freezing, assignment visualization, or training loggers. Full training and evaluation are dataset-, checkpoint-, and accelerator-dependent; this sub-skill helps plan and validate them before expensive runs.
Route PyTorch demo inference and postprocess/vis details to ../inference-and-api/SKILL.md. Route ONNX, TorchScript, TensorRT, or deployment exports to ../export-and-deployment/SKILL.md.
Read or run these first
- Read references/training-workflows.md for train/eval commands, distributed flags, image-size controls, freezing, assignment visualization, and TensorBoard/MLflow/W&B logging.
- Read references/data-and-exp-reference.md when preparing COCO/VOC/custom datasets, writing an
Exp, usingYOLOX_DATADIR, applyingopts, or reasoning about caching/default model sizes. - Read references/troubleshooting.md for annotation, image, class-count, checkpoint, cache, distributed, logger, resume, or evaluator failures.
- Run scripts/inspect_yolox_exp.py before expensive training to load an
Exp, inspect fields, and optionally check COCO/VOC paths without starting dataloaders or training.
Fast operating rules
- Prefer installed module commands such as
python -m yolox.tools.train,python -m yolox.tools.eval, andpython -m yolox.tools.visualize_assign. - Specify exactly one experiment source:
-n yolox-sfor a packaged default or-f path/to/exp.pyfor custom data/classes/architecture. If both are supplied toget_exp, the file wins. - Keep dataset roots explicit. Set
YOLOX_DATADIRfor built-in COCO/VOC-style experiments, or setself.data_dirin theExpfor custom COCO layouts. - Validate
num_classes,input_size,test_size, annotations, image directories, and checkpoint compatibility before multi-GPU or cached runs. - Treat
--cacheas a resource decision: RAM cache is fast but memory-heavy; disk cache needs writable dataset-root storage. - Put non-default architecture, dataset, evaluator, augmentation, image-size, freezing, optimizer, or scheduler behavior in an
Expsubclass. Use CLIoptsonly for small typed overrides of existing fields.
Safe Exp inspection
From this sub-skill directory:
python scripts/inspect_yolox_exp.py --name yolox-s --expected-format none
python scripts/inspect_yolox_exp.py --exp-file path/to/exp.py --check-data --expected-format coco
The helper never starts training/evaluation. Missing dataset paths under --check-data are reported as preflight failures so the user can fix paths before launching YOLOX.
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).