---

name: mmyolo description: "Use MMYOLO for OpenMMLab YOLO object detection configs, datasets, training/testing, inference, model APIs, and deployment conversion." disable-model-invocation: true metadata: disco-role: operating license: GPL 3.0

MMYOLO repo skill

Use this skill when the task involves MMYOLO, OpenMMLab YOLO-family object detection, MMEngine configs, MIM train/test commands, detection datasets, MMYOLO inference/visualization, MMYOLO registries/model components, or MMYOLO deployment/export/conversion.

MMYOLO v0.6.0 is an OpenMMLab toolbox for YOLO-style object detection and related detection workflows. It builds on PyTorch, MMCV, MMEngine, and MMDetection.

First checks

1. Read repository provenance if you need to decide whether this skill matches a checkout or package version.
2. Read installation notes before changing an environment or diagnosing imports.
3. Run scripts/check_mmyolo_environment.py when you need a safe import/version/config parse check.
4. Use cross-cutting troubleshooting for import/version/backend/MIM failures before entering a specific workflow.

Minimal import check:

python scripts/check_mmyolo_environment.py --json

Config parse check:

python scripts/check_mmyolo_environment.py --config CONFIG.py

Route by user task

User task	Read
Choose a YOLOv5/6/7/8, YOLOX, RTMDet, or PPYOLOE baseline; edit config inheritance, class counts, dataloaders, evaluators, hooks, TTA, or --cfg-options.	config-customization
Prepare or validate COCO/YOLO/VOC/DOTA/LabelMe data, convert labels, inspect annotation JSON, plan anchor optimization, or wire data paths into a config.	data-tools
Construct safe train/resume/AMP/test/evaluation/prediction-output commands, plan distributed/Slurm launch, analyze logs/schedulers/confusion matrices, or troubleshoot training/evaluation.	training-evaluation
Plan image/folder/URL/video inference, save visualized detections, export LabelMe annotations, inspect DetDataSample, use SAHI large-image recipes, or plan feature-map/CAM visualization.	inference-visualization
Inspect registries, public model/dataset components, backbones/necks/heads/losses/assigners, custom project patterns, plugins, or switch_to_deploy.	model-api
Convert upstream YOLO checkpoints, plan ONNXRuntime/TensorRT/RKNN/MMDeploy export, check optional deployment dependencies, or reason about backend artifacts.	deployment-conversion

Common workflow paths

Fine-tune on a custom detection dataset

1. Use data-tools to validate annotation schema, category order, image paths, and dataset layout.
2. Use config-customization to create a child config with correct metainfo, num_classes, dataloaders, evaluators, epochs, hooks, and pretrained initialization.
3. Use training-evaluation to build a safe mim train mmyolo command and decide GPU/AMP/resume settings.
4. Return to training-evaluation for mim test mmyolo, prediction dumps, painted outputs, and metric troubleshooting.

Run prediction and export labels

1. Use config-customization if the config/class metadata is uncertain.
2. Use inference-visualization for CPU/GPU image/folder/URL inference, score/class filters, and LabelMe output.
3. Use data-tools only when the produced labels need validation or conversion for a later training dataset.

Prepare deployment

1. Use training-evaluation or inference-visualization to confirm the checkpoint/config pair works before export.
2. Use deployment-conversion to select ONNXRuntime, TensorRT, RKNN, MMDeploy, or checkpoint-conversion paths.
3. Run the deployment dependency checker before proposing vendor-specific commands.

Boundaries

• Do not launch training, evaluation, inference, downloads, exports, or vendor builds unless the caller explicitly asks for execution and the side effects are acceptable.
• Do not treat CPU import/config success as proof of CUDA, TensorRT, RKNN, or DeepStream readiness.
• Do not invent dataset paths, Slurm partitions, WandB credentials, checkpoint URLs, or vendor toolchain availability.
• Prefer package-level OpenMIM commands and bundled helpers over checkout-local source script paths.
• If the task is generic PyTorch, generic computer vision, or another OpenMMLab package without MMYOLO-specific configs/APIs, route to a more specific skill.