TAO DINO Object Detection
Run NVIDIA TAO workflows for training and deploying DINO 2D object detectors.
The skill limits declared tools to Read and Bash and explicitly documents Docker, NVIDIA Container Toolkit, dataset URIs, checkpoints, result directories, data-source handling, parent-model inference, and common failures. Points are deducted because Bash plus Docker/S3 can cause external training, export, and deployment effects, while explicit user confirmation, isolation, sensitive-data handling, rollback, and recovery procedures are not specified.
SKILL.md and its references give fairly consistent guidance for training, evaluation, export, inference, quantization, distillation, AutoML, data sources, and diagnosable failures such as missing validation data, invalid class counts, OOM, and checkpoint-path errors. Points are deducted because the supplied material does not provide complete executable reproduction, automated tests, or a real suite covering abnormal inputs; static calibration caps this at 10.
The name, trigger phrases, supported actions, COCO formats, required inputs, non-production smoke profile, TensorRT routing, and AutoML override are clearly stated. Points are deducted because non-fit scenarios, Chinese-language support, and mainland-China network reachability are not documented, while core workflows depend on Docker, NVIDIA GPUs, S3, and TAO SDK infrastructure.
The skill provides a reference map, progressive reference files, parameter guidance, dataset contracts, error documentation, version 0.1.0, Apache-2.0 licensing, and NVIDIA authorship; the benchmark also records schema findings. Points are deducted for missing dedicated Instructions and Examples sections, incomplete author formatting, no clear changelog, maintenance owner, or update path, and documented folder/schema convention findings.
The skill covers the main DINO workflows and supplies data-source overrides, checkpoint inference, AutoML metric extraction, TensorRT routing, and troubleshooting guidance; BENCHMARK.md reports correctness and effectiveness results for one task. Points are deducted because the sample contains only one task and does not statically establish directly usable outputs from real training, export, or deployment; environment, data, and result validation remain necessary. Static calibration caps this at 7.
The pinned repository revision, skill_info.yaml, evaluation task, and BENCHMARK.md provide traceability for some claims and expected behavior. Points are deducted because there is only one evaluation task, no complete test suite or CI coverage, and no independent multi-source corroboration; static calibration caps this at 5.
- Before execution, explicitly confirm external effects involving Docker, GPUs, S3/TAO jobs, and output directories, and constrain credential and data access.
- The AutoML smoke profile uses fixed S3 datasets and is documented as unsuitable for production benchmarking; it is not evidence of user-data or production performance.
- The skill depends on NVIDIA GPUs, TAO SDK/containers, and remote S3 paths, so core workflows may be unreachable in mainland-China or GPU-less environments.
- Before publication, add broader multi-task tests, recovery guidance, a maintenance/update path, dedicated Examples/Instructions sections, and complete author metadata.
What does this skill do, and when should you use it?
This skill guides NVIDIA TAO Toolkit workflows for DINO, a Transformer-based 2D object detector with denoising training and multi-scale features. It covers training, evaluation, export, distillation, quantization, TensorRT deployment, and inference. Training requires a train dataset URI, a validation dataset URI, and a valid class count, with COCO or coco_raw data supported. Docker and NVIDIA Container Toolkit are required.
Reads the skill’s dataset contracts, action parameters, error catalog, AutoML guidance, and deployment references; builds data-source and spec overrides for each action; handles COCO and coco_raw datasets; guides use of pretrained backbone or full-model checkpoints; runs or orchestrates training, evaluation, export, distillation, quantization, TensorRT engine generation, and inference; uses mAP50 for quick operational monitoring and val_mAP for COCO or paper-style comparisons.
- A computer-vision developer training a TAO DINO 2D detector on COCO-format data.
- An engineer evaluating or running inference with a trained DINO checkpoint.
- A deployment team exporting DINO or generating a TensorRT engine.
- A user running AutoML/HPO while maintaining DINO’s dataset and multi-GPU constraints.
- A TAO user applying teacher-student distillation, quantization, or diagnosing common CUDA and data-configuration failures.
What are this skill's strengths and limitations?
- Covers the DINO workflow from training through deployment and inference.
- Explicitly documents the always-required validation dataset, class-count constraint, data-source overrides, and multi-GPU consistency rule.
- Includes AutoML/HPO, checkpoint inference, distillation, and common-error guidance.
- Supports COCO and coco_raw datasets and documents the standard archive layout.
- Requires Docker and NVIDIA Container Toolkit.
- Training requires dataset URIs and a correct class count; validation data cannot be omitted.
- Data-source overrides are mandatory for every action, so configuration mistakes can cause startup failures or CUDA device-side asserts.
- The supplied source does not provide a standalone test suite or specific benchmark results for this skill.
How do you install this skill?
Install the skill through the CLI command documented by the repository README: npx skills add nvidia/skills --skill tao-train-dino --yes. You can list available skills with npx skills add nvidia/skills --list. The source does not document a separate package or additional version requirement for this skill.
How do you use this skill?
After installation, prompt an Agent that has loaded the skill with phrases such as “train DINO”, “DETR object detection”, “TAO 2D detection”, or “DINO with distillation”. For training, provide a train dataset URI, validation dataset URI, and num_classes; the standard dataset artifacts are typically images.tar.gz and annotations.json. If running AutoML without inputs, the skill supplies a local smoke-run profile. Before TensorRT engine generation, TensorRT evaluation, or TensorRT inference, read references/tao-deploy-dino.md. The source does not provide a complete local Docker command.