What does this skill do, and when should you use it?
This skill supports NVIDIA TAO Toolkit workflows for SegFormer semantic segmentation. It covers training, evaluation, export, inference, quantization, and the dedicated TensorRT deployment path. It documents dataset layouts, important parameters, distributed training, checkpoint handoff, AutoML policy, and common failure patterns. It fits users with segmentation data and NVIDIA GPU infrastructure who want to operate TAO through Docker.
Reads the skill's TAO configuration guidance, generated schemas, specification templates, and deployment reference; instructs the agent to construct dataset paths and spec_overrides for training, evaluation, export, inference, and quantization; uses the TAO PyT CLI for standard model actions and the SegFormer deploy workflow for TensorRT engine generation, TensorRT evaluation, and TensorRT inference; selects AutoML or direct training according to the run policy; resolves appropriate model checkpoints from training result directories for downstream actions; and diagnoses issues involving val_miou, dataset layout, class counts, image size, distributed training, and GPU memory.
- A computer-vision engineer trains a binary or multiclass SegFormer model on a custom UNet-format segmentation dataset.
- A model developer evaluates a trained model, exports ONNX, runs inference, or performs quantization.
- A deployment engineer generates a TensorRT engine and runs TensorRT evaluation or inference through the SegFormer deployment workflow.
- An AutoML user runs TAO SegFormer hyperparameter search optimized for maximum val_miou.
- A team trains across multiple GPUs or nodes using Lightning-managed distributed execution.
How do you install this skill?
- Before training, W&B synchronization, or deployment, obtain explicit confirmation for data transfer, network access, output-directory effects, and GPU consumption; disable W&B when unnecessary.
- Verify the TAO container, Docker/NVIDIA Container Toolkit, TAO version, and referenced files in the target environment; mainland-China reachability is not established.
- Maintainers should clarify how the skill license metadata aligns with the repository's dual-license statement and which content each license covers.
- Shell / CLI
- Local filesystem
Dockernvidia-container-toolkitNVIDIA GPU
Install the skill from the NVIDIA/skills repository with the skills CLI:
npx skills add nvidia/skills --skill tao-train-segformer --yes
The README does not specify a manual folder-copy procedure; the CLI places the skill in the selected agent's skill location.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/tao-train-segformer" ~/.claude/skills/
rm -rf "$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 a SegFormer semantic segmentation model
- export and quantize a TAO SegFormer model.
After installation, ask a compatible agent for a concrete task such as “train a SegFormer semantic segmentation model” or “export and quantize a TAO SegFormer model.” Provide an extracted dataset root containing images/<split> and masks/<split>, plus the required checkpoint, class count, image size, and output directory. Training defaults to AutoML; say “disable AutoML” or “plain training” to turn it off for that run. Use the deployment workflow described in references/tao-deploy-segformer.md for TensorRT engine generation, TensorRT evaluation, and TensorRT inference. The supplied material does not document exact container launch commands or a complete end-to-end example.
What are this skill's strengths and limitations?
- Covers SegFormer training, evaluation, export, inference, and quantization.
- Defines per-action dataset paths, important spec fields, and checkpoint handoff behavior.
- Documents the default AutoML policy, val_miou objective, multi-GPU settings, and common troubleshooting patterns.
- Clarifies that gen_trt_engine is not exposed by the parent PyT CLI and identifies the dedicated deployment route.
- Requires Docker, nvidia-container-toolkit, and at least one NVIDIA GPU.
- Only segmentation data and the UNet format are explicitly supported; other formats are not evidenced.
- The complete deployment reference files and container-version details are not included in the supplied material.
- No test suite, complete GPU compatibility list, or full end-to-end command sequence is provided.
- TensorBoard is unsupported for SegFormer segmentation training and must remain disabled.
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 SegFormer Semantic Segmentation this page ✓ NVIDIA · Official | 52 · Use with care | ★ 3.8k | today | Apache-2.0 |
| TAO OneFormer Universal Segmentation ✓ NVIDIA · Official | 53 · Use with care | ★ 3.8k | today | Apache-2.0 |
| TAO Grounding DINO Open-Vocabulary Detection ✓ NVIDIA · Official | 56 · Use with care | ★ 3.8k | today | Apache-2.0 |
| TAO PyT Image Classification ✓ NVIDIA · Official | 48 · Use with care | ★ 3.8k | today | Apache-2.0 |
| TAO Deformable DETR Object Detection ✓ NVIDIA · Official | 46 · Use with care | ★ 3.8k | today | Apache-2.0 |
The source provides no direct comparison with other segmentation models or tools. It only characterizes SegFormer as relatively lightweight with hierarchical feature extraction and documents TAO and TensorRT workflows.
How did FollowSkills review this skill?
The skill scopes tools to Read and Bash and states Docker/NVIDIA container requirements; no malware, credential theft, or covert exfiltration is evident. However, training, uploads, W&B (default enable: true in templates), and TensorRT flows may create external or persistent effects, while user confirmation, sensitive-data handling, least-privilege guidance, and rollback are absent, so points are deducted.
Actions, dataset paths, parameter mappings, checkpoint selection, and common error patterns are documented, making the happy path plausible. Static materials contain no key-path test suite or reproducible execution logs; dependency versions, image availability, and abnormal-input feedback remain incomplete. The static ceiling also limits this score to 10 or less, so points are deducted.
The target scenarios, supported actions, data format, GPU requirements, and several boundaries are reasonably clear. Semantic non-trigger conditions, input/output contracts, Chinese support, and mainland-China reachability are not documented; the core Docker, NVIDIA Toolkit, and TAO container dependencies have limited environment-fit evidence, so points are deducted.
The skill has layered references, version 0.1.0, author and license fields, templates, parameter tables, limitations, and error patterns. It lacks the recommended Instructions/Examples sections, a changelog, explicit maintenance ownership, and an update path; the skill license metadata is also not fully reconciled with the repository's documentation/source dual-license statement, so points are deducted.
The skill covers training, evaluation, export, inference, quantization, and deployment, with concrete data-path, spec-override, AutoML-routing, and checkpoint-handoff guidance that could complete the core task. Only one positive benchmark task is supplied, and static evidence does not verify directly usable outputs or compare alternatives and cost; the static ceiling limits this to 7, so points are deducted.
Versioned configuration, action definitions, templates, parameter mappings, and a single benchmark report provide limited auditability. There is no independently reproducible key-path test suite or execution evidence, and coverage is only one positive task; therefore the score remains low and below the static ceiling of 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 →