---

name: customization-extensions description: "Customize MMDetection3D datasets, data pipelines, model components, runtime settings, and optional project extensions." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

customization-extensions

Use this sub-skill when the user needs to add or adapt extension code that must be registered and wired into a config.

Start here

• references/customization.md for registry, dataset, pipeline, model, and runtime patterns.
• references/projects.md for optional projects/ packages and compiled extras.
• references/troubleshooting.md for import, registry, config, and shape pitfalls.
• scripts/scaffold_custom_component.py to generate a starter tree in a caller-provided output directory.

Typical targets

• new datasets or dataset subclasses
• new point or image transforms
• new backbones, necks, heads, losses, assigners, samplers, or hooks
• custom optimizers, optimizer constructors, schedulers, or runtime hooks
• optional project modules that stay outside the core package

Routing rules

• If the request is mainly config selection or model zoo lookup, route to the configuration sub-skill.
• If the request is only training, testing, or evaluation command construction, route to the training sub-skill.
• If the request is raw dataset conversion or info-file generation, route to the data-preparation sub-skill.
• If the work needs project-specific extras or optional compilation, keep it here and consult references/projects.md.

Operating rule

Prefer custom_imports and generated starter files over editing core package code unless the user explicitly wants an in-tree implementation.