---

name: customization-extension description: "Extend MMDetection 3.3.0 with custom registries, models, losses, datasets, transforms, hooks, optimizers, project plugins, structures, and 2.x-to-3.x migration fixes." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Customization Extension

Use this sub-skill when an agent must add, register, debug, or migrate custom MMDetection components. It covers MMEngine registry usage, custom_imports, project plugin layout, custom model/data/runtime classes, and DetDataSample/box/mask structure expectations.

Route First

• For config inheritance, model zoo selection, _base_, or cfg-options, use ../configuration-model-zoo/SKILL.md.
• For dataset file layout, COCO conversion, annotation schemas, or metrics, use ../datasets-evaluation/SKILL.md.
• For train/test launch commands, resume, distributed jobs, or result dumping, use ../training-testing/SKILL.md.
• For inference API calls or visualization outputs, use ../inference-visualization/SKILL.md.

Core References

• references/registries-and-apis.md: registry nodes, decorators, import/scope rules, structures, and migration API changes.
• references/extension-workflows.md: workflows for custom models, losses, datasets, transforms, hooks, optimizers, projects plugins, and structures.
• references/troubleshooting.md: diagnosis tables for registration, imports, default scope, class counts, data samples, transforms, optimizers, and 2.x migration failures.
• scripts/registry_probe.py: inspect MMDetection registry contents and verify import paths without training.

Fast Workflow

1. Identify the component boundary: model/loss/head/backbone uses MODELS; dataset uses DATASETS; transform uses TRANSFORMS; hooks use HOOKS; optimizers and constructors use optimizer registries.
2. Register the Python class with the correct decorator from mmdet.registry, then ensure its module is imported by package __init__.py or by config-level custom_imports.
3. Set default_scope='mmdet' or initialize the mmdet scope before registry builds when running standalone checks.
4. Validate the component in isolation with scripts/registry_probe.py and a minimal Registry.build() or transform call before launching training.
5. When modifying data flow, confirm pipeline keys before and after each transform and ensure final packed samples expose the fields consumed by the model.

Hard Usability Cases

• A new bbox head appears in config as type='MyBBoxHead' but build fails with KeyError; debug registry node, import side effects, and default scope before editing model logic.
• A custom transform returns missing or renamed keys; trace expected input/output keys through the pipeline and confirm packed DetDataSample fields before training.