executing-plans
executing-plans
Use when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available
Explorar Skills
Productividad · SKILL.md
training-evaluation
training-evaluationGuides VAD training, checkpoint loading, safe configuration overrides, single-GPU evaluation, result formatting, and diagnosis of legacy-stack runtime failures.
Independent GitHub snapshot · Automated classification · Original documentation governs
Documentación de la Skill
GitHubname: training-evaluation description: "Guides VAD training, checkpoint loading, safe configuration overrides, single-GPU evaluation, result formatting, and diagnosis of legacy-stack runtime failures." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0
VAD training and evaluation
Use this route for launch commands, work directories, resume/override behavior, checkpoint evaluation, metrics, or result formatting.
Route
- Prepare and validate data with data-preparation.
- Select and parse a VAD config with architecture-configuration and its safe checker.
- Read workflow-reference.md and cli-reference.md.
- Run
python scripts/check_training_contract.py CONFIGbefore an expensive command. - Train with the selected single/distributed launch recipe; evaluate with one non-distributed GPU unless you have independently verified a compatible alternative. The repository warns that distributed evaluation can produce inaccurate results.
Full training/evaluation was not run during construction: it needs external nuScenes data, checkpoints, and CUDA/native operators.
Scope boundaries
- Raw data and temporal PKLs: data-preparation.
- Model graph and config semantics: architecture-configuration.
- Rendering result artifacts: visualization.
Use troubleshooting.md for import, launcher, normalization, and output failures.