What does this skill do, and when should you use it?
This skill provides NVIDIA TAO Toolkit guidance for Visual ChangeNet in AOI and PCB inspection. It supports PASS/NO_PASS binary classification from image pairs and pixel-level change masks from before/after pairs. The instructions cover training, evaluation, inference, export, quantization, and conditional TensorRT deployment workflows, along with data formats, schemas, troubleshooting, and parent-model resolution. Docker, nvidia-container-toolkit, and an NVIDIA GPU with at least 16GB of VRAM are required.
Reads packaged TAO schemas, templates, and skill metadata; constructs dataset paths and spec overrides for classification or segmentation; runs TAO Toolkit or local Docker commands for training, evaluation, inference, and export; consumes image archives plus a four-column CSV for classification and A/, B/, list/, label/ paired directories for segmentation; stages the Hugging Face C-RADIO checkpoint locally and mounts it into the container; and produces classification outputs, change masks, checkpoints, exported models, or deployment artifacts when the relevant actions are packaged. It also routes eligible training runs through AutoML by default and explains when direct training is required.
- An AOI engineer trains a PASS/NO_PASS classifier from paired PCB inspection images.
- A machine-vision team trains a pixel-level change-segmentation model from before/after image pairs.
- A TAO user evaluates, runs inference with, or exports an existing Visual ChangeNet checkpoint.
- A deployment engineer generates a TensorRT engine when a compatible ONNX parent model already exists.
- A platform engineer diagnoses pretrained-weight paths, dataset-format errors, out-of-memory failures, or segmentation CUDA-extension issues.
How do you install this skill?
- The invocation examples expose NGC_API_KEY to the container and mount datasets, results, and model files; verify least-privilege credentials, data boundaries, GPU scope, and cleanup/rollback before use.
- The first segmentation run compiles custom CUDA operators, and backbone staging depends on Hugging Face; validate network access, image availability, storage, and caching in advance.
- skill_info.yaml and the main document differ on some segment data sources, actions, and templates; they should be reconciled and covered by reproducible tests before publication.
- Shell / CLI
- Network access
- Local filesystem
Dockernvidia-container-toolkitPythonhuggingface_hubNVIDIA TAO Toolkit containerNVIDIA GPU with 16GB+ VRAM
Install the single skill with the skills CLI:
npx skills add nvidia/skills --skill tao-train-visual-changenet --yes
Or install the NVIDIA collection:
npx skills add nvidia/skills
The README does not document a separate manual-copy procedure. After installation, the skill is available when the agent loads skills and encounters a relevant request.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/tao-train-visual-changenet" ~/.claude/skills/
rm -rf "$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 Visual ChangeNet for PCB defect detection
After installation, use a trigger such as “train Visual ChangeNet for PCB defect detection”, “ChangeNet classify”, or “ChangeNet segment”. Specify the task mode, dataset locations, spec file, output directory, GPU/platform settings, and parent checkpoint. Classification requires an image directory and a four-column CSV; segmentation requires A/, B/, list/, and label/ directories. AutoML is the default for eligible training runs; request “disable AutoML” or “plain training” for a direct-training run. For local Docker, resolve the TAO image from versions.yaml, use --shm-size=8g, and mount the local C-RADIO safetensors checkpoint.
What are this skill's strengths and limitations?
- Supports both paired-image classification and change segmentation for AOI workflows.
- Documents dataset layouts, CSV semantics, spec overrides, and important training parameters.
- Explicitly addresses the need to stage and mount C-RADIO weights before launch.
- Includes AutoML routing, parent-model inference, troubleshooting, and local-Docker guidance.
- Requires Docker, nvidia-container-toolkit, a TAO runtime, and an NVIDIA GPU.
- The C-RADIO checkpoint must be downloaded from Hugging Face before training and is approximately 393 MB.
- Segmentation compiles the MultiScaleDeformableAttention CUDA extension on first run, taking about five minutes.
- Whether quantization, TensorRT generation, and deployment actions are runnable depends on packaged metadata, schemas, and available ONNX artifacts; the source does not provide complete test results.
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 Visual ChangeNet Inspection this page ✓ NVIDIA · Official | 48 · Use with care | ★ 3.8k | today | Apache-2.0 |
| TAO AutoML + DEFT Three-Phase Pipeline ✓ NVIDIA · Official | 51 · Use with care | ★ 3.8k | today | Apache-2.0 |
| TAO MAE Vision Pretraining ✓ NVIDIA · Official | 48 · Use with care | ★ 3.8k | today | Apache-2.0 |
| TAO Optical Inspection Training ✓ NVIDIA · Official | 48 · Use with care | ★ 3.8k | today | Apache-2.0 |
| TAO Depth Anything v2 ✓ NVIDIA · Official | 50 · Use with care | ★ 3.8k | today | Apache-2.0 |
How did FollowSkills review this skill?
The documentation transparently describes Docker, GPU, S3, Hugging Face, NGC_API_KEY, and host-volume data flows, and limits declared tools to Read/Bash. However, --gpus all, secret injection through environment variables, and external model downloads lack least-privilege guidance, confirmation requirements, sensitive-data handling, complete isolation, and rollback procedures, so points are deducted.
The skill documents routing, data formats, model-weight staging, error patterns, and troubleshooting, and says to report blocking conditions when schemas or ONNX artifacts are missing. Static evidence provides no executable reproduction, comprehensive test suite, or precise failure-output contract, and some action/data-source descriptions are inconsistent, so the score is reduced and kept below the static ceiling.
The audience, AOI/PCB scenarios, classification and segmentation inputs, and trigger phrases are reasonably clear. Non-fit boundaries, Chinese-language support, and broader hardware/storage compatibility are not specified; core staging depends on Hugging Face and execution depends on NVIDIA Docker/NGC, creating possible mainland-China reachability concerns, so points are deducted.
The skill includes a name, description, Apache-2.0 license, author, version, references, parameter guidance, and troubleshooting. It lacks dedicated Instructions and Examples sections, a changelog, explicit maintenance responsibility, and an update path; the benchmark also reports folder/file convention findings, so points are deducted.
The documentation gives fairly complete configuration and data requirements for training, evaluation, inference, export, quantization, and segmentation, which should be directly useful. However, no representative output is statically verifiable here, and the single benchmark task with limited coverage does not establish result correctness, completeness, or comparative benefit, so the score is capped and reduced.
A fixed revision, committed evals.json, and benchmark summary provide some traceability. There are no raw logs, broad test coverage, independent reproduction artifacts, or corroborating sources, so only limited credit is justified.
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 →