TAO OCRNet Scene Text Recognition
Train and deploy NVIDIA TAO OCRNet models for cropped text recognition.
The skill limits declared tools to Read and Bash and states Docker plus NVIDIA Container Toolkit requirements. However, Bash, GPU use, output paths, and default Weights & Biases enablement may create external effects; data flow, sensitive-data handling, user confirmation, isolation, and rollback are not specified, so points are deducted. No malware or other clear red-line behavior is shown.
The documentation covers actions, inputs, checkpoint selection, dataset requirements, error patterns, and deterministic settings, supporting a fairly complete static happy path. No tests or CI cover key paths, dependency versions are unspecified, and referenced deployment materials and generated resources are not fully verifiable in the supplied files, so points are deducted.
The audience, OCRNet scenarios, trigger phrases, input formats, and supported actions are clear. Non-fit boundaries, Chinese-language support, and mainland-China environment fit are not sufficiently described, while Docker/GPU prerequisites narrow environmental coverage, so points are deducted.
The skill includes front matter, version, author, license, layered documentation, action tables, parameter examples, and error patterns. It lacks explicit maintenance ownership and update path, changelog, and FAQ; the benchmark also reports missing recommended Instructions/Examples sections, an author-format issue, and folder-structure findings. License metadata is also NOASSERTION, so points are deducted.
The documentation covers conversion, training, evaluation, inference, export, pruning, quantization, retraining, and deployment, with directly adaptable configuration examples, so the core value is clear. Static materials do not verify usable outputs, and the single-task benchmark reports limited Codex correctness and effectiveness, so points are deducted.
The supplied material provides revision context, action configurations, an evaluation task, and a limited benchmark report, giving some auditability. Coverage is only one task, with no key-path test suite or CI evidence enabling independent reproduction, so points are deducted.
- Before execution, confirm whether external logging or network access such as W&B is permitted, and ensure input data and encryption keys cannot enter logs or remote services.
- Verify Docker, NVIDIA Container Toolkit, TAO/PyTorch/TensorRT versions and the referenced files that are not included; static evidence cannot ensure commands will run.
- Mainland-China network reachability and end-to-end Chinese character support are not established; separately validate character lists, GT files, and BOM handling.
- Training, export, pruning, and quantization write model artifacts and output paths; confirm paths, backups, and recovery procedures before execution.
What does this skill do, and when should you use it?
This skill supports NVIDIA TAO OCRNet for recognizing text in cropped text-region images. It covers dataset conversion, training, evaluation, inference, export, pruning, quantization, pruned-model retraining, and TensorRT deployment, with CTC and attention-based decoders. AutoML is enabled at the model layer and is the default route for training when the packaged schema and template are available, but it can be disabled for an individual run. Docker, nvidia-container-toolkit, and an NVIDIA GPU with at least 8GB of VRAM are required.
Reads packaged schemas, specification templates, deployment references, and model metadata; guides preparation of cropped text images, ground-truth files, character lists, and LMDB datasets; runs TAO OCRNet actions including dataset_convert, train, evaluate, inference, export, prune, quantize, and retrain; resolves checkpoints and exported artifacts from parent jobs; reads tao-deploy-ocrnet.md for TensorRT engine generation and TensorRT-backed evaluation or inference; and produces LMDB datasets, checkpoints, ONNX files, pruned PTH files, calibration caches, and TensorRT engines.
- A vision engineer training OCRNet on custom cropped text-region images.
- A model developer comparing CTC and attention-based decoders for scene text recognition.
- A deployment team exporting a trained OCRNet model to ONNX and generating a TensorRT engine.
- A model-optimization engineer pruning, quantizing, or retraining OCRNet from a pruned artifact.
- A user training or evaluating against a defined character set with a corresponding character-list file.
What are this skill's strengths and limitations?
- Covers the OCRNet lifecycle from dataset conversion through deployment.
- Supports both CTC and attention-based decoders.
- Documents checkpoint selection, parent-job artifact inference, and per-action data requirements.
- Supports AutoML while allowing a per-run opt-out.
- Requires Docker, nvidia-container-toolkit, and an NVIDIA GPU with at least 8GB of VRAM.
- Raw image folders and ground-truth files must be converted to LMDB before training, quantization, or retraining.
- The training setup is single-node oriented; multi-GPU execution is Lightning-managed and there is no explicit num_nodes setting.
- The supplied source does not provide independent test results or concrete performance benchmarks for this individual skill.
How do you install this skill?
Install the NVIDIA skills collection with the skills CLI: npx skills add nvidia/skills. Install only this skill with: npx skills add nvidia/skills --skill tao-train-ocrnet --yes. To target Codex, use: npx skills add nvidia/skills --skill tao-train-ocrnet --agent codex --yes. Use it in an environment with Docker, nvidia-container-toolkit, and an NVIDIA GPU.
How do you use this skill?
After installation, give the agent a concrete request such as “train OCRNet” or “run scene text recognition inference on cropped text images.” Prepare cropped images, gt_new.txt files, and a character list; when using raw folders and ground-truth files, run dataset_convert separately for training and validation before supplying the resulting LMDB paths. For TensorRT gen_trt_engine, TensorRT evaluate, or TensorRT inference, read references/tao-deploy-ocrnet.md first.