本文へ移動
cccskills
無料GitHub で公開

yolo-master-agent

Use when the user wants to run a YOLO-Master task (train/val/predict/track/export/benchmark) or use the Agent Skill dispatcher.

Tencent/YOLO-Master
インストール方法を見る

含まれるファイル（93）

  • SKILL.md3.0 KB
  • assets/autotrain_cases.json35.0 KB
  • assets/autotrain_cases/01_yolo_benchmark.json733 B
  • assets/autotrain_cases/02_yolo_export.json696 B
  • assets/autotrain_cases/03_yolo_lora_train.json1.5 KB
  • assets/autotrain_cases/04_yolo_model_inspect.json751 B
  • assets/autotrain_cases/05_yolo_moe_diagnose.json728 B
  • assets/autotrain_cases/06_yolo_moe_prune.json738 B
  • assets/autotrain_cases/07_yolo_multimodal_evaluate.json3.7 KB
  • assets/autotrain_cases/08_yolo_multimodal_infer.json14.1 KB
  • assets/autotrain_cases/09_yolo_pipeline_experiment.json1.9 KB
  • assets/autotrain_cases/10_yolo_predict.json2.9 KB
  • assets/autotrain_cases/11_yolo_solutions_run.json669 B
  • assets/autotrain_cases/12_yolo_system.json2.6 KB
  • assets/autotrain_cases/13_yolo_track.json904 B
  • assets/autotrain_cases/14_yolo_train.json3.0 KB
  • assets/autotrain_cases/15_yolo_tune.json631 B
  • assets/autotrain_cases/16_yolo_ui_launch.json499 B
  • assets/autotrain_cases/17_yolo_val.json1.6 KB
  • assets/autotrain_cases/18_yolo_lora_diagnose.json780 B
  • assets/autotrain_cases/19_yolo_eval_peft_compare.json1.2 KB
  • assets/autotrain_cases/20_stability_check.json7.0 KB
  • assets/autotrain_cases/21_yolo_eval_sparse_sahi_compare.json736 B
  • assets/autotrain_cases/21_yolo_release_audit.json506 B
  • assets/autotrain_cases/22_yolo_job.json1.1 KB
  • assets/mini-detect/images/train/bus.jpg134.2 KB
  • assets/mini-detect/images/val/bus.jpg134.2 KB
  • assets/mini-detect/labels/train/bus.txt38 B
  • assets/mini-detect/labels/val/bus.txt38 B
  • assets/mini-detect/mini_detect.yaml83 B
  • assets/open_world_label_aliases.json603 B
  • assets/open-world-taxonomy/lvis_1203_classes.json73.6 KB
  • assets/open-world-taxonomy/lvis_1203_classes.txt19.3 KB
  • assets/open-world-taxonomy/sources.json780 B
  • assets/open-world-taxonomy/v3det_13204_classes.json814.3 KB
  • assets/open-world-taxonomy/v3det_13204_classes.txt205.9 KB
  • assets/prompts/vlm_coco_multitask.md3.9 KB
  • assets/prompts/vlm_open_world_caption_misses_compact.md1.4 KB
  • assets/prompts/vlm_open_world_detect_classify_compact.md1.6 KB
  • assets/prompts/vlm_open_world_detection_compact.md1.7 KB
  • assets/prompts/vlm_open_world_detection.md4.0 KB
  • metadata/openai.yaml3.1 KB
  • README.md7.5 KB
  • references/autotrain.md1.4 KB
  • references/manual-probes.md1.6 KB
  • references/multimodal-evaluation.md2.3 KB
  • references/multimodal-inference.md4.8 KB
  • references/pipeline-peft.md2.3 KB
  • references/skill-architecture.md42.9 KB
  • references/thinking-with-image.md9.1 KB
  • runtime/__init__.py60 B
  • runtime/cli/__init__.py73 B
  • runtime/cli/async_jobs.py7.5 KB
  • runtime/cli/compare_open_world_profiles.py9.5 KB
  • runtime/cli/contract.py13.3 KB
  • runtime/cli/core_handlers.py26.3 KB
  • runtime/cli/dataset.py9.6 KB
  • runtime/cli/device.py10.6 KB
  • runtime/cli/dispatcher.py13.3 KB
  • runtime/cli/executor.py16.8 KB
  • runtime/cli/job_handlers.py1.3 KB
  • runtime/cli/launcher_handlers.py6.0 KB
  • runtime/cli/lora_tools.py8.0 KB
  • runtime/cli/model_handlers.py1.8 KB
  • runtime/cli/moe_tools.py9.0 KB
  • runtime/cli/multimodal_handlers.py49.5 KB
  • runtime/cli/normalize.py10.4 KB
  • runtime/cli/peft_compare.py7.8 KB
  • runtime/cli/pipeline.py9.8 KB
  • runtime/cli/progress.py889 B
  • runtime/cli/regenerate_open_world_report.py8.5 KB
  • runtime/cli/release.py2.6 KB
  • runtime/cli/sahi_compare.py7.5 KB
  • runtime/cli/snapshot.py13.4 KB
  • runtime/cli/stability.py23.5 KB
  • runtime/cli/system_handlers.py8.3 KB
  • runtime/cli/validate.py41.7 KB
  • runtime/evaluation/__init__.py55 B
  • runtime/evaluation/metrics.py50.0 KB
  • runtime/multimodal/__init__.py64 B
  • runtime/multimodal/fusion.py29.6 KB
  • runtime/multimodal/providers/dashscope.yaml283 B
  • runtime/multimodal/providers/openai.yaml242 B
  • runtime/multimodal/runtime.py30.2 KB
  • runtime/multimodal/visual.py5.3 KB
  • runtime/open_world/__init__.py53 B
  • runtime/open_world/taxonomy.py31.4 KB
  • runtime/release.py24.9 KB
  • scripts/compare_open_world_profiles.py338 B
  • scripts/regenerate_open_world_report.py339 B
  • scripts/run_async_job.py1.1 KB
  • scripts/run_yolo_master_skill.py321 B
  • scripts/validate_yolo_master_skill.py319 B

SKILL.md（原文）

インストールする前に、エージェントに与えられる指示の中身を確認できます。

YOLO-Master Agent Skill

Use This Skill

  • train, val, predict, track, export, benchmark, tune
  • LoRA diagnose, PEFT compare, MoE diagnose/prune
  • Multimodal visual inference and batch evaluation
  • End-to-end orchestration via yolo.pipeline.experiment

Execution Rule

If yolo version fails, run pip install -e . first. Prefer the yolo CLI over raw Python API for supported commands. Use the bundled dispatcher for deterministic, structured runs:

python agent/scripts/run_yolo_master_skill.py --json '{"skill":"yolo.train","inputs":{"model":"yolo11n.pt","data":"coco8.yaml"},"params":{"epochs":1,"imgsz":32}}'

On Apple Silicon, the dispatcher defaults heavy compute (train, val, benchmark, predict, track) to device=mps. Override with runtime.device or params.device. If MPS/CUDA fails, the dispatcher retries once on CPU and returns a recovery record.

Use policy.dry_run=true for cheap coverage before real runs. For long jobs, set policy.async=true on train/tune/pipeline to submit a subprocess job.

Workflow

  1. Ensure yolo CLI is available; install if missing.
  2. Dispatch the task via run_yolo_master_skill.py with appropriate skill, inputs, and params.
  3. Return structured artifacts, metrics, and next actions from the dispatcher response.

Validation

python agent/scripts/validate_yolo_master_skill.py --suite quick --pretty --summary-only

quick = fast-smoke + dry-run + contract. Use --suite all for the full regression pass.

References

  • Skill architecture — skill registry, request/response contract, execution logic
  • Multimodal inference — VLM/LLM parameters, prompt templates, fusion, open-world taxonomy
  • Multimodal evaluation — batch evaluation, fusion guardrails, metric guardrails
  • AutoTrain loop — validator suites, case pack, tiered coverage
  • Pipeline & PEFT — pipeline experiments, release audit, LoRA diagnose, PEFT compare
  • Manual probes — extended CLI suite, direct train/val, environment doctor
  • Thinking with image — VLM visual reasoning, marked-image prompting, crop/zoom search
  • Open-world taxonomy assets in assets/open-world-taxonomy — LVIS 1203 and V3Det 13204 category lists

Guardrails

  • New Ultralytics CLI args must go through params dict, not as top-level dispatcher keys.
  • Consume evaluation in addition to metrics when judging train/val runs.
  • Prefer the recovery field over raw stderr when a run auto-falls back from MPS/CUDA to CPU.
  • Treat UI launchers and research scripts as launcher-style skills, not plain sync functions.

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

このスキルの問題を報告する