"Explain UniAD config variants, plugin registration, public model
Scanned 9/9/2026
Install to Claude Code
npx -y skills add VectorSpaceLab/AREX-Skill --skill config-and-model-architecture --agent claude-codeInstalls into .claude/skills of the current project.
Are you the author of Config And Model Architecture?
Add the live security badge to your README — it updates automatically with every re-scan.
[](https://www.skillsdirectory.com/skills/vectorspacelab-config-and-model-architecture)More formats (shields.io, HTML) on the badges page.
---
name: config-and-model-architecture
description: "Explain UniAD config variants, plugin registration, public model
components, and task-head relationships."
disable-model-invocation: true
metadata:
disco-role: operating
license: Apache 2.0
---
# Config and Model Architecture
Use this sub-skill when the question is about choosing, editing, or comparing UniAD configs, understanding which model component owns a task, or diagnosing a BEV encoder swap or registry problem.
## Read first
- `references/config-reference.md` for the three public config families, queue lengths, checkpoints, and plugin fields.
- `references/model-architecture.md` for the detector/head ownership map and task flow.
- `references/api-reference.md` for verified constructor signatures and key call contracts.
- `references/troubleshooting.md` for import, memory, anchor, BEV swap, and planning-metric failures.
- `scripts/summarize_uniad_config.py` for a safe one-file config summary.
## Covers
- BEVFormer base, stage 1 track/map, and stage 2 end-to-end configs.
- `plugin = True` / `plugin_dir` and the `projects.mmdet3d_plugin` import root.
- Verified public classes: `UniAD`, `UniADTrack`, `BEVFormerHead`, `BEVFormerTrackHead`, `PansegformerHead`, `MotionHead`, `OccHead`, and `PlanningHeadSingleMode`.
- Queue length, `load_from`, freeze flags, and task-loss ownership.
- Planning metric ambiguity and the `planning_evaluation_strategy` switch.
- BEV encoder replacement rules: preserve the `bev_embed` / `bev_pos` contract.
## Do not cover
- Training or evaluation command construction → `training-evaluation`
- Dataset downloads or info generation → `data-preparation`
- Visual result interpretation or plotting → `visualization-and-results`
## Use this when
- You need to pick the right UniAD config for an experiment.
- You need to explain why a stage is missing a head or a loss.
- You are diagnosing plugin, registry, or import failures tied to model construction.
- You need to identify which module owns a reported loss or output.
- You want a safe summary of a config file before editing it.
## Output expectation
Lead with the minimal config or architecture answer, then cite the specific config field or class contract that justifies it.
Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.
No comments yet. Be the first to comment!