What does this skill do, and when should you use it?
This skill orchestrates a three-phase workflow: AutoML baseline tuning, DEFT-based iterative data improvement, and AutoML refinement on the augmented dataset. It primarily targets NVIDIA TAO Toolkit PCB AOI ChangeNet workflows, while supporting the same pattern for other DEFT applications. The skill handles handoffs of specifications, checkpoints, and augmented training data; it does not reimplement AutoML or DEFT. It requires Docker and nvidia-container-toolkit and is designed to run continuously after one consolidated user approval.
Reads training and validation datasets, workspace specifications, and downstream preflight requirements; invokes tao-run-automl for the baseline hyperparameter sweep; selects the winning hyperparameters and checkpoint, merges the specification, and pre-seeds DEFT state so its duplicate baseline training is skipped; invokes tao-run-deft-aoi or a matching DEFT skill for RCA, routing, SDG, mining, dataset assembly, and plain-training iterations; reads the best checkpoint and final iteration label from deft_state.json; reruns AutoML on the augmented training CSV using the DEFT winner as a warm start; and registers the final checkpoint in deft_state.json for the inference handoff.
- A PCB AOI team wants to improve a visual ChangeNet model end to end with AutoML and DEFT.
- A vision-training user wants to tune hyperparameters before running a DEFT data-improvement loop.
- A team needs to retune a model on DEFT-augmented data while fine-tuning from the loop's best checkpoint.
- A user is working with a non-AOI DEFT application that follows the same AutoML → DEFT → AutoML pattern.
How do you install this skill?
- Before execution, verify all downstream skills, scripts, images, GPUs, datasets, and credential requirements; this review did not execute those checks.
- After approval the workflow is intended to run autonomously and performs file writes plus container login/pull operations; establish backups and a rollback plan first.
- Do not select AOI models from val_loss alone; inspect class prediction counts and validate with FAR at 100% recall or an equivalent deployment KPI.
- Verify NGC/Hugging Face reachability in the target network and confirm compliant handling of datasets and credentials.
- Shell / CLI
- Network access
- Local filesystem
Dockernvidia-container-toolkit
Install the skill from the NVIDIA/skills collection with the skills CLI: npx skills add nvidia/skills --skill tao-run-automl-deft-pipeline --yes. The README also supports targeting a client, for example by adding --agent codex. The skill becomes available when the agent next loads skills and encounters a relevant task.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/tao-run-automl-deft-pipeline" ~/.claude/skills/
rm -rf "$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:
- run the AOI workflow,
- AutoML + DEFT,
- improve my AOI ChangeNet model.
Ask the agent to “run the AOI workflow,” “AOI end-to-end,” “AutoML + DEFT,” or “improve my AOI ChangeNet model.” Before side effects, it presents one consolidated summary covering the host, credentials, images, datasets, three-phase configuration, and compute estimate, then requests approval. After approval it runs the phases sequentially. Confirm the workspace, dataset paths, specifications, and approximate time per training job; exact downstream parameters come from tao-run-automl and the selected DEFT skill.
What are this skill's strengths and limitations?
- Clearly composes baseline AutoML, DEFT data improvement, and post-DEFT AutoML refinement.
- Reuses the Phase 1 winning checkpoint, avoiding a separate DEFT baseline retrain.
- Warm-starts Phase 3 from the DEFT winner while retaining the same validation set for comparison.
- Includes guidance for preflight, leakage prevention, class imbalance, and run-to-run noise.
- The workflow is sequential; both AutoML phases run multiple trainings and DEFT iterations usually dominate cost.
- Requires Docker, nvidia-container-toolkit, tao-run-automl, and a DEFT downstream skill.
- A non-recoverable DEFT hard stop causes Phase 3 to be skipped.
- The SKILL.md does not enumerate all downstream workflow requirements or provide an independent runtime estimate or platform test matrix for this skill.
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 AutoML + DEFT Three-Phase Pipeline this page ✓ NVIDIA · Official | 51 · Use with care | ★ 3.8k | today | Apache-2.0 |
| TAO Visual ChangeNet Inspection ✓ NVIDIA · Official | 48 · Use with care | ★ 3.8k | today | Apache-2.0 |
| TAO AutoML Tuning ✓ NVIDIA · Official | 54 · 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 |
| TAO Person Re-Identification ✓ NVIDIA · Official | 49 · Use with care | ★ 3.8k | today | Apache-2.0 |
Compared with running tao-run-deft-aoi alone, this skill adds AutoML before and after the DEFT loop. Compared with running tao-run-automl alone, it adds the iterative DEFT data-improvement stage. Requests for DEFT only or AutoML only should use the corresponding downstream skill.
How did FollowSkills review this skill?
The document requires one explicit confirmation before docker pulls, login, job launches, or result-directory writes, forbids printing credential values, and calls for leakage checks and KPI-test isolation. These support a moderate trust score. Deductions apply because Bash/Write, docker login/pull, and autonomous cross-phase execution remain broadly privileged; dependency security, backups, rollback, and least-privilege boundaries are not defined, and baseline_spec.yaml is overwritten directly.
The three phases, handoff artifacts, stop conditions, preflight checks, and hard-stop behavior are described in useful detail, including skipping Phase 3 after an unrecoverable DEFT failure. Deductions apply because execution depends on downstream skills, scripts, and environment details not included here; no committed tests or CI cover the key paths, and older downstream versions may introduce an extra approval gate. Static calibration limits this score to 10 or below.
Trigger phrases, intended scenarios, AOI versus non-AOI routing, and non-fit cases are clearly stated. Deductions apply because non-AOI support is presented as a generic pattern without concrete interface guarantees, and there is no evidence of Chinese-language support or mainland-China network reachability. Core operation depends on Docker, NVIDIA containers, NGC/HF credentials, and downstream skills.
The metadata gives the name, Apache-2.0 license, author, version, and dependencies. The documentation is layered around preflight, phases, handoffs, pitfalls, quick start, and references, and includes a benchmark and update note. Deductions apply because the evaluation found missing recommended Instructions and Examples sections, the author format is nonconforming, and changelog, maintenance ownership, and formal update procedures are incomplete. The skill card also does not perfectly align with the frontmatter license wording.
The stated deliverable is a final AutoML checkpoint, and the skill specifies phase inputs, artifact handoffs, metric safeguards, warm-start behavior, and regression protection for an end-to-end training workflow. The benchmark report provides limited task-level correctness and effectiveness evidence. Deductions apply because no directly reviewable training artifacts or complete outcome logs are included, the workflow is costly, and multiple external skills are required. Static calibration limits this score to 7 or below.
The files provide detailed command patterns, state fields, path conventions, metrics, and an evaluation report, giving the workflow some auditability. Deductions apply because the benchmark covers only one positive task, has no negative-trigger coverage, and includes no committed key-path tests or independently reproducible execution logs. The reported results therefore provide limited rather than corroborated evidence. Static calibration limits this score to 5 or below.
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 →