What does this skill do, and when should you use it?
This skill guides an AI agent through NVIDIA TAO Toolkit workflows for OCDNet scene text detection. OCDNet detects arbitrary-oriented text regions in natural images with a differentiable binarization approach. It covers training, evaluation, inference, export, pruning, quantization, retraining from pruned outputs, and TensorRT deployment workflows. The runtime requires Docker, nvidia-container-toolkit, and an NVIDIA GPU with at least 8GB of VRAM.
Reads deployment guidance, generated TAO Core schemas, and YAML templates; constructs extracted dataset paths for each action; runs OCDNet train, evaluate, export, inference, prune, quantize, and default_specs workflows; routes eligible training runs through AutoML; manages checkpoint handoff between training, evaluation, export, pruning, and quantization; reads the deployment reference before TensorRT engine generation or TensorRT evaluation and inference; and produces checkpoints, exported ONNX files, pruned models, or quantized models.
- TAO users training an OCDNet scene text detector on datasets containing img/ and gt/ directories.
- Computer-vision engineers evaluating a trained model or detecting arbitrarily oriented text in natural images.
- Deployment engineers exporting ONNX or running TensorRT engine generation and inference workflows.
- Model-optimization practitioners pruning, retraining from pruned graphs, or quantizing OCDNet checkpoints.
- Experimenters using AutoML with hmean or train_loss as the optimization metric.
How do you install this skill?
- Execution uses Bash, Docker, GPU resources, and TAO runtime components, creating external computation and file outputs; confirm authorization, data locations, output paths, and resource budgets before running.
- The documentation requires dataset extraction and checkpoint handling but does not define cleanup, rollback, or disk-quota behavior after failures.
- Version compatibility and mainland-China reachability for AutoML, TAO containers, TensorRT deployment, and related dependencies are not established by the selected skill.
- The benchmark contains only one positive task and no execution logs, so it does not establish runtime correctness or production-grade results.
- Shell / CLI
- Local filesystem
Dockernvidia-container-toolkitNVIDIA GPU with 8GB+ VRAM
Install the skill with the NVIDIA skills CLI: npx skills add nvidia/skills --skill tao-train-ocdnet --yes. Prepare Docker and nvidia-container-toolkit in the runtime environment, along with an NVIDIA GPU providing at least 8GB of VRAM. The source does not document additional manual copying steps.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/tao-train-ocdnet" ~/.claude/skills/
rm -rf "$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 scene text detection inference with this OCDNet checkpoint.
After installation, give the agent a concrete request such as “train OCDNet” or “run scene text detection inference with this OCDNet checkpoint.” Provide extracted training and validation dataset paths; each split should contain img/ and gt/, or use a UTF-8 datalist file. Pass data sources through spec_overrides; do not pass tar.gz archives directly. Before TensorRT engine generation, TensorRT evaluation, or TensorRT inference, read references/tao-deploy-ocdnet.md. For a one-epoch training run, set train.lr_scheduler.args.warmup_epoch to 0.
What are this skill's strengths and limitations?
- Covers the OCDNet workflow from training through deployment.
- Documents per-action dataset paths, checkpoint handoff rules, and common failure patterns.
- Supports AutoML, single-GPU training, multi-GPU or multi-node strategies, pruning, and quantization.
- Requires Docker, nvidia-container-toolkit, and an NVIDIA GPU.
- Dataset archives are not unpacked at runtime, so users must prepare the correct extracted directories.
- The source provides no skill-specific benchmark values, runtime estimates, or broad platform coverage evidence.
- The PyT CLI does not expose retrain or gen_trt_engine subcommands; those operations belong to defined workflows.
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 |
|---|---|---|---|---|
| OCDNet Scene Text Detection this page ✓ NVIDIA · Official | 53 · 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 DINO Object Detection ✓ NVIDIA · Official | 58 · Recommended | ★ 3.8k | today | Apache-2.0 |
| TAO PointPillars 3D Detection ✓ NVIDIA · Official | 52 · 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 |
How did FollowSkills review this skill?
The skill limits declared tools to Read and Bash and clearly specifies data paths, checkpoint handoff, and output directories, providing some permission and data-flow scoping. It lacks pre-execution confirmation, sensitive-data guidance, network/container boundary details, rollback, and recovery procedures, so 10 points are deducted.
It documents training, evaluation, inference, export, pruning, quantization, retraining, and several failure patterns, including AutoML and checkpoint constraints. However, it depends on Docker, NVIDIA Container Toolkit, and TAO runtime components without committed key-path reproduction tests; the static ceiling therefore applies and further uncertainty reduces the score.
The name, trigger phrases, input formats, action scope, and multi-GPU settings are reasonably clear for NVIDIA GPU-based TAO OCDNet workflows. Non-fit cases, version/platform boundaries, Chinese-language support, and mainland-China reachability are not established, and container availability is unspecified, so 5 points are deducted.
The skill has front matter, Apache-2.0 licensing, version and author metadata, tags, parameter tables, template references, error patterns, and checkpoint guidance. It lacks dedicated Instructions, Examples, FAQs, changelog, explicit maintenance ownership, and an update path; the supplied benchmark also records hierarchy, recommended-section, and author-format findings, so 5 points are deducted.
The documentation provides actionable parameter mappings, dataset paths, and checkpoint rules for the OCDNet actions, so the core task is plausibly supported. No statically verifiable representative outputs or quality results are supplied, and full use depends on external TAO/Docker infrastructure; the score is limited to 6.
Schema/template mappings, skill_info, an evaluation task, and a benchmark report provide some audit trails. Coverage is thin—one task, no verifiable execution logs, and no committed key-path test suite—and the benchmark report is not independent reproduction evidence, so 1 point is deducted from the static ceiling.
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 →