What does this skill do, and when should you use it?
This skill targets AOI, PCB, and quality-control data, using Siamese networks to compare image pairs for manufacturing defects, anomalies, and quality issues. It covers training, evaluation, inference, ONNX export, and TensorRT deployment for TAO Optical Inspection models. AutoML is enabled by default for training and can be disabled for an individual run. Users need Docker, nvidia-container-toolkit, and an NVIDIA GPU with at least 8GB of VRAM.
Reads and applies TAO Optical Inspection dataset, model, and deployment specifications; constructs training, validation, test, and inference paths from images.tar.gz and dataset.csv splits; runs train, evaluate, inference, and export actions; resolves checkpoints from upstream training or AutoML jobs; exports ONNX models; reads the deployment reference before TensorRT engine generation, TensorRT evaluation, or TensorRT inference; and diagnoses CSV format errors, extracted-image root mismatches, invalid training batch sizes, and trusted-checkpoint loading failures.
- A manufacturing engineer wants to train an AOI defect classifier from PCB image pairs.
- A computer-vision team needs to run TAO Optical Inspection training, validation, and testing on preconverted data.
- A model developer needs to select a checkpoint from a training or AutoML job and export an ONNX model.
- A deployment engineer needs to generate a TensorRT engine and run TensorRT inference for an optical-inspection model.
- A validation engineer needs to troubleshoot dataset.csv paths, extracted image layouts, or batch-size failures.
How do you install this skill?
- Do not run Bash, Docker, or S3-related actions before confirming data locations, credential scope, and output directories.
- Training and downstream actions depend on preconverted images.tar.gz and dataset.csv files, valid checkpoints, and a working TAO/Docker/GPU environment; installation, version compatibility, and rollback guidance are incomplete.
- Do not treat the benchmark PASS as evidence of successful real training or deployment; it covers only one documentation-planning task.
- For mainland-China networks or Chinese-language workflows, verify reachability and localization support for Docker images, TAO dependencies, and external services before use.
- Shell / CLI
- Local filesystem
Dockernvidia-container-toolkitNVIDIA GPU with at least 8GB VRAM
Install the NVIDIA skills collection with: npx skills add nvidia/skills. Install only this skill with: npx skills add nvidia/skills --skill tao-train-optical-inspection --yes. The README does not document a separate manual-copy procedure for this skill. Before running it, provide Docker, nvidia-container-toolkit, and a suitable NVIDIA GPU.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/tao-train-optical-inspection" ~/.claude/skills/
rm -rf "$tmp"Generated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- train optical inspection,
- AOI defect detection,
- PCB / manufacturing inspection,
After loading the skill in an Agent Skills-compatible client, prompt it with requests such as “train optical inspection,” “AOI defect detection,” or “PCB / manufacturing inspection,” and provide dataset paths, the requested action, output directory, GPU settings, and checkpoint requirements. Use Optical Inspection-format images.tar.gz and dataset.csv splits; training, evaluation, and inference require explicit data-source paths. Before TensorRT gen_trt_engine, TensorRT evaluate, or TensorRT inference, read references/tao-deploy-optical-inspection.md. The source does not document complete standalone CLI commands, dataset-generation commands, or pricing.
What are this skill's strengths and limitations?
- Covers training, evaluation, inference, ONNX export, and TensorRT deployment workflows.
- Supports AutoML with a per-run option to enable or disable it.
- Specifies required data files, checkpoint resolution, GPU settings, and common failure handling.
- Designed for lightweight Siamese networks that typically run on one GPU.
- Requires Docker, nvidia-container-toolkit, and an NVIDIA GPU with at least 8GB of VRAM.
- Targets the Optical Inspection data format and TAO workflow rather than general-purpose defect detection.
- Although the PyT container exposes dataset conversion, this skill does not package a dataset_convert action or template.
- The supplied material provides no standalone test suite, performance benchmark, or operating-cost evidence.
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| TAO Optical Inspection Training this page ✓ NVIDIA · Official | 48 · Use with care | ★ 3.8k | today | Apache-2.0 |
| TAO RT-DETR Object Detection ✓ NVIDIA · Official | 47 · Use with care | ★ 3.8k | today | Apache-2.0 |
| TAO CenterPose Pose Estimation Skill ✓ NVIDIA · Official | 49 · Use with care | ★ 3.8k | today | Apache-2.0 |
| TAO DINO Object Detection ✓ NVIDIA · Official | 58 · Recommended | ★ 3.8k | today | Apache-2.0 |
| TAO Visual ChangeNet Inspection ✓ NVIDIA · Official | 48 · Use with care | ★ 3.8k | today | Apache-2.0 |
How did FollowSkills review this skill?
The skill limits declared tools to Read and Bash and documents Docker, NVIDIA Container Toolkit, data paths, checkpoint resolution, and restrictions on using the force-load environment variable with untrusted checkpoints. It does not define user confirmation, least-privilege Bash procedures, sensitive-data handling for S3 inputs, external-effect isolation, rollback, or recovery. Publisher attribution is traceable, but the supplied license metadata is NOASSERTION, so points are deducted.
SKILL.md gives fairly specific guidance for training, evaluation, inference, export, deployment, dataset formats, checkpoint selection, and common errors, including reporting when schemas are missing. However, the static materials contain no executable tests covering key paths, the action/template branching is complex, and dependency versions plus abnormal-input diagnostics remain limited. The static ceiling prevents a score above 10.
The intended users, AOI/manufacturing inspection scenarios, trigger phrases, input formats, and single-GPU/8GB+ VRAM environment are reasonably clear. It also states single-node limits, dataset-conversion boundaries, and that fake data must not be synthesized. Non-fit cases, semantic false-trigger boundaries, Chinese-language support, and mainland-China network reachability are not adequately addressed, so points are deducted.
The skill includes metadata, version 0.1.0, author, tags, action tables, parameter mappings, templates, and error patterns; repository material supplies Apache/CC licensing, synchronization maintenance context, and a security-reporting path. It lacks the recommended Instructions, Examples, and FAQ sections, the author format fails the benchmark finding, and there is no skill-specific changelog or clearly assigned maintenance owner, so points are deducted.
The skill covers the core workflow from preconverted-data training through evaluation, inference, ONNX export, and TensorRT deployment, with actionable data-source and checkpoint rules. However, no representative final outputs are statically verifiable; the benchmark has only one task, with Codex correctness at 20% and overall effectiveness at 44%. The static ceiling limits the score to 7 or below.
The fixed revision, structured schemas/templates, action configuration, and one benchmark report provide some auditability. Coverage is very thin: one positive task, no negative tasks, and no committed tests covering actual training, evaluation, or deployment paths or independent corroboration. The static ceiling prevents a score above 5.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →