TAO RT-DETR Object Detection
Guides NVIDIA TAO workflows for training and deploying real-time RT-DETR detectors.
The skill restricts tools to Read and Bash and states requirements for Docker, NVIDIA Container Toolkit, GPUs, datasets, and checkpoints. However, it lacks stepwise user confirmation, sensitive-data handling, network/W&B data-flow disclosure, isolation, and rollback guidance. W&B is enabled by default and examples use S3 paths and an encryption key, so least-privilege and transparency are incomplete.
The documentation covers training, distillation, quantization, evaluation, export, inference, deployment, and multi-node configuration, with several error patterns. Static files provide no executable tests or key-path reproduction. External AutoML/TAO Deploy dependencies, ??? placeholders, and possible mismatches between templates and input mappings create reliability uncertainty, so the score remains conservative under the static cap.
The audience, trigger phrases, supported actions, COCO format, GPU/VRAM requirements, and several boundaries are reasonably clear. Input/output contracts, version boundaries, and semantic exclusions are incomplete. Core use depends on Docker, NVIDIA GPUs, TAO containers, and potentially cloud storage, with no mainland-China reachability guidance, so points are deducted.
The skill has useful sections for data requirements, parameters, templates, error patterns, and inference mappings. It lacks the recommended Instructions, Examples, and FAQ sections, has only version 0.1.0, and provides no skill-level changelog, maintenance owner, or update path. The skill declares Apache-2.0 while repository metadata also applies CC-BY-4.0 to documentation, leaving license scope less clear.
The skill directly covers the main TAO RT-DETR workflows and supplies action templates, data-source mappings, checkpoint handoff rules, and troubleshooting advice. The committed benchmark reports one passing task, but the sample size is one, has no negative tasks, and was not executed during this static review. Results still depend on correct environment, external services, and user-supplied paths, so full effectiveness is not justified.
The pinned source set includes templates, skill_info, an evaluation task, and a benchmark report, providing some auditability. It lacks committed tests covering key actions, reproducible CI evidence, and cross-source corroboration; the benchmark has only one task. Static evidence therefore supports only a limited score.
- Before execution, confirm Docker, NVIDIA Container Toolkit, GPU availability, the TAO image, dataset paths, and checkpoint provenance; do not assume sensitive data may be uploaded or that W&B should be enabled.
- Validate the AutoML route, TAO Deploy documentation, template placeholders, input mappings, and RT-DETR version compatibility, and preserve recoverable output directories for failed jobs.
- Mainland-China reachability of S3, container images, and external NVIDIA services is not established by the supplied material.
What does this skill do, and when should you use it?
This skill supports NVIDIA TAO Toolkit workflows for RT-DETR 2D object detection. It covers training, evaluation, inference, export, distillation, and quantization, with rules for datasets, checkpoints, distributed training, and TensorRT deployment. Training is AutoML-enabled by default, with a per-run option to disable it. Docker and NVIDIA Container Toolkit are required.
Reads packaged schema, configuration, and deployment references; runs TAO PyT CLI actions for train, distill, quantize, evaluate, export, and inference; constructs required COCO or COCO raw data-source paths, checkpoints, and output paths; supports multi-GPU and multi-node training through torchrun; exports ONNX with a default 640×640 input and uses the RT-DETR deployment workflow for TensorRT operations; and uses structured training results or logs to identify metrics such as mAP50 and val_mAP.
- A computer-vision engineer needs to train a real-time 2D detector on custom COCO-format data.
- A deployment team needs to export a trained RT-DETR model to ONNX or a TensorRT deployment workflow.
- A model-optimization practitioner needs teacher-model distillation or quantization for RT-DETR.
- A platform engineer needs to run TAO RT-DETR training on one GPU, multiple GPUs, or multiple nodes.
What are this skill's strengths and limitations?
- Covers the main TAO operations from RT-DETR training through deployment optimization.
- Specifies required data sources, checkpoints, and important parameters for each action.
- Supports AutoML, knowledge distillation, quantization, multi-GPU, and multi-node training.
- Includes practical failure patterns and checkpoint handoff guidance.
- Requires Docker, NVIDIA Container Toolkit, and NVIDIA GPU hardware.
- The stated minimum is 1 GPU, with 2 recommended, and 16GB or more VRAM suggested per GPU.
- The supplied skill text does not include an independent test suite, numerical performance benchmarks, or the full deployment reference content.
- TensorRT engine generation is not exposed through the parent PyT CLI gen_trt_engine command and requires the separate deployment workflow.
How do you install this skill?
Install the skill from the NVIDIA collection with: npx skills add nvidia/skills --skill tao-train-rtdetr --yes. Set up Docker and NVIDIA Container Toolkit; the README does not document a separate manual installation procedure for this individual skill.
How do you use this skill?
After installation, ask a compatible Agent to handle a request such as “train RT-DETR” or “train a low-latency object detector with TAO.” Provide dataset locations, class count, GPU settings, and checkpoint information as applicable. AutoML is enabled by default for training; explicitly request “disable AutoML” or set automl_policy: off to use direct training. Supply the action-specific datasets and checkpoints for evaluation, export, inference, distillation, or quantization. Read references/tao-deploy-rtdetr.md before TensorRT actions.
How does this skill compare with similar options?
The skill compares RT-DETR with DINO/GDINO at the configuration level: RT-DETR defaults to 640×640 inputs and three feature levels, making it lighter than DINO's larger inputs and four feature levels; the source provides no standardized benchmark results.