Skip to content
Back to skills

Config And Model Architecture

ASecurity

"Explain UniAD config variants, plugin registration, public model

  • 247 stars
  • 0 votes
  • 0 copies
  • 2 views
  • Added September 9, 2026
businessapi

Works with

  • api

Security analysis

A100/100

Pro scans all 6 files and shows the line behind each finding

Scanned September 9, 2026

npx -y skills add VectorSpaceLab/AREX-Skill --skill config-and-model-architecture --agent claude-code

Installs 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 with every re-scan.

Security grade badge for Config And Model Architecture
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/vectorspacelab-config-and-model-architecture/badge)](https://www.skillsdirectory.com/skills/vectorspacelab-config-and-model-architecture)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
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.

Files in this skill

  • SKILL.md2.2 KB
  • references/api-reference.md7.3 KB
  • references/config-reference.md4.3 KB
  • references/model-architecture.md5 KB
  • references/troubleshooting.md4 KB
  • scripts/summarize_uniad_config.py18.8 KB

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…