TAO Sparse4D Training Skill
Guides agents through training, evaluation, export, quantization, and inference for multi-camera temporal 3D detection and tracking.
The skill declares Read and Bash as allowed tools and documents containers, data paths, checkpoints, and AutoML routing. No malware, credential theft, or covert exfiltration is evident. However, Bash is broad, per-run user confirmation and least-privilege boundaries are not specified, sensitive-data handling and rollback are absent, and W&B is enabled by default and may create external data flow; points were deducted.
SKILL.md, skill_info.yaml, templates, and error-pattern guidance cover conversion, training, evaluation, export, inference, and quantization with artifact checks and failure advice. The files provide no executable tests or CI reproducing the key paths; documented legacy-image failures and export-compatibility constraints add uncertainty, so the static ceiling and deductions apply.
Trigger phrases, target scenario, action scope, input format, hardware requirements, and several non-fit conditions are reasonably clear for multi-camera temporal 3D detection and tracking. Chinese-language support is not described, and the core depends on Docker, NVIDIA Container Toolkit, GPUs, specific TAO images, and potentially S3 data; mainland-China network and image reachability are unaddressed, so points were deducted.
The skill includes a main guide, reference documents, templates, metadata, a benchmark report, and a version number; the skill declares Apache-2.0 and publisher/maintenance context is partially clear. Recommended Instructions and Examples sections are missing, the author format is incomplete, and changelog, dependency-version, maintenance-owner, and unified installation/troubleshooting guidance are limited. The benchmark also records hierarchy and extra-schema findings.
The skill covers dataset conversion, training, evaluation, export, inference, and quantization, with data-source mappings, parameter templates, checkpoint inference, AutoML policy, and troubleshooting. This should reduce manual configuration substantially, but static review cannot confirm model results or end-to-end artifacts; the benchmark contains only one task and does not establish directly usable outputs for every action, so points were deducted.
Structured skill_info.yaml, auditable templates, inference mappings, a benchmark report, and a pinned revision provide limited traceability. There is only one evaluation task and no independent reproduction record, key-path test suite, CI coverage, or multi-source corroboration; the reported PASS is not execution verification for this static review, so full marks are not justified.
- Before execution, confirm Bash operations, data mounts, W&B external communication, and AutoML routing, and verify permissions for images, GPUs, data sources, and checkpoints.
- Quantization and export have documented image, camera-count, anchor-count, and checkpoint-compatibility limitations; perform a small validation run and retain recoverable outputs.
- Chinese support and mainland-China network reachability are not documented; TAO containers and remote data sources may not be directly accessible.
What does this skill do, and when should you use it?
This skill targets the NVIDIA TAO Toolkit Sparse4D model for multi-camera temporal 3D object detection and tracking. It covers dataset conversion, training, evaluation, export, quantization, and inference, including instance-bank handling, temporal reasoning, and checkpoint handoff. AutoML is enabled by default for training but can be disabled for an individual run. Docker, nvidia-container-toolkit, and NVIDIA GPUs are required, with 40GB+ of VRAM per GPU recommended.
Reads model metadata, packaged JSON schemas, specification templates, and reference documentation; requires dataset_convert to convert AICity data into OVPKL annotations and generate anchor_init.npy; builds the required spec_overrides and data paths for train, evaluate, export, inference, and quantize; routes eligible training requests through tao-skill-bank:tao-run-automl; invokes the relevant containers and commands; and documents handling for checkpoints, H5 depth paths, quantization failures, and temporal out-of-memory errors.
- TAO users training Sparse4D on multi-camera data for 3D detection.
- Vision teams evaluating detection or tracking metrics on validation or test splits.
- Engineers exporting a trained model or running inference from an exact checkpoint.
- Deployment teams quantizing a Sparse4D checkpoint with TorchAO.
- Researchers running AutoML training or a plain-training experiment with HPO disabled for that run.
What are this skill's strengths and limitations?
- Covers Sparse4D from dataset conversion through inference.
- Includes AutoML routing, dataset dependency mappings, and checkpoint-resolution rules.
- Documents AICity-to-OVPKL conversion, H5 depth-path normalization, and temporal memory troubleshooting.
- Records multi-GPU, multi-node, and important model parameters.
- Requires dataset_convert before the main workflow, with sparse4d/ovpkl data artifacts.
- High hardware requirements: minimum 2 GPUs, 8 recommended, with 40GB+ VRAM per GPU.
- The supplied material references schemas, templates, and model containers whose complete contents are not provided.
- The supplied material gives no standalone test suite, platform-specific validation results, or quantization success rate.
How do you install this skill?
Install the individual skill with the NVIDIA skills CLI:
npx skills add nvidia/skills --skill tao-train-sparse4d --yes
The README does not document the exact local destination for each platform; the CLI handles the installation target.
How do you use this skill?
After installation, prompt the agent with phrases such as “train Sparse4D,” “multi-camera 3D detection,” “temporal 3D tracker,” or “sparse query 3D perception.” Prepare compatible conversion artifacts and run dataset_convert first to produce OVPKL annotations and anchor_init.npy. For training, explicitly use automl_policy: on or automl_policy: off when needed. Other supported actions include evaluate, export, inference, and quantize.