TAO Pose Classification
Classify pose-keypoint sequences into action categories with ST-GCN.
The skill restricts declared tools to Read/Bash and documents Docker, GPU, data paths, checkpoint handling, and encryption-key mappings; no malware, credential theft, or destructive default is evident. Points are deducted for missing user confirmation, sensitive-data handling, pinned dependency versions, external data-transfer boundaries, rollback, and cleanup guidance.
Actions, templates, parameters, deterministic settings, and error patterns are fairly detailed, including missing schemas, label shapes, graph layouts, and checkpoint issues. Points are deducted because key paths were not executed; deploy capability is described inconsistently, inference output declaration does not fully align with skill_info, and abnormal-input feedback is incomplete.
Trigger phrases, target scenarios, input formats, action scope, and excluded deploy/prune/quantize/standalone-retrain capabilities are stated. Points are deducted because Docker, NVIDIA container tooling, GPU hardware, and many examples depend on S3; Chinese-language interaction and mainland-China network reachability are not addressed.
The package includes front matter, license, version, author, tags, per-action templates, schema guidance, parameter tables, error patterns, and benchmark documentation. Points are deducted for missing recommended Instructions, Examples, and FAQ sections, invalid validator-style author formatting, no skill-level changelog or explicit maintenance/update path, and repository license metadata marked NOASSERTION.
Training, evaluation, export, inference, and optional conversion workflows include concrete parameters, dataset files, checkpoint handoff, and troubleshooting; the benchmark reports favorable results on a limited task. Points are deducted because the benchmark covers documentation planning rather than real model execution, output files, or result quality, leaving substantial environment and data preparation.
Revision-pinned source files, schema/template configuration, an eval definition, and a benchmark report provide auditable documentation and limited behavioral evidence. Points are deducted for the single-task evaluation, absence of committed key-path tests or CI reproduction, and the fact that no commands were executed in this static review.
- This review did not execute Docker, the TAO CLI, training, evaluation, export, or inference; the benchmark report is not proof of real model-result validation.
- Before use, verify the container image, TAO version, S3 access, dataset format, checkpoint selection, and encryption-key handling.
- The documentation contains a deploy-capability inconsistency, and the inference output file is not fully declared in skill_info; verify the actual runner and schemas first.
- For human pose data, add privacy, access-control, retention, and cross-border-transfer review outside this skill.
What does this skill do, and when should you use it?
This skill supports NVIDIA TAO Toolkit workflows for classifying skeleton sequences from pose-keypoint data with ST-GCN. It covers dataset conversion, training, evaluation, export, and inference, with optional AutoML routing for training. Preconverted `.npy` and `.pkl` datasets can be used directly without conversion. The documented environment requires Docker, NVIDIA Container Toolkit, and at least one NVIDIA GPU with about 8GB of VRAM.
Reads TAO pose-classification configuration, dataset paths, label maps, and model checkpoints; runs pose_classification dataset_convert only for raw DeepStream BodyPose JSON; trains or resumes models with pose_classification train; evaluates checkpoints, exports ONNX files, and writes inference results; routes training through tao-skill-bank:tao-run-automl when AutoML is enabled for the run and the packaged training schema and template are available. The documented monitoring metric is val_loss, minimized for AutoML selection.
- A computer-vision practitioner has converted `train_data.npy`, `train_label.pkl`, `val_data.npy`, and `val_label.pkl` and wants to train a six-class action model.
- A DeepStream user has raw BodyPose JSON and needs to convert it before TAO pose-classification training.
- A model developer needs to evaluate a trained pose-classification checkpoint on validation data.
- An engineer needs to export a selected `.pth` checkpoint as an ONNX file.
- A researcher needs to run classification inference on `test_data.npy` and `test_label.pkl` and save the output file.
What are this skill's strengths and limitations?
- Covers conversion, training, resume training, evaluation, ONNX export, and inference.
- Clearly distinguishes raw DeepStream JSON from already converted `.npy` / `.pkl` datasets.
- Documents required data paths, label maps, graph layout, checkpoints, and output files.
- Allows AutoML to be enabled or disabled per training run.
- The workflow is single-node; multi-GPU launch is managed by Lightning.
- It does not expose deploy, prune, quantize, or standalone retrain actions.
- It requires Docker, NVIDIA Container Toolkit, and an NVIDIA GPU; CPU-only execution is not documented.
- AutoML requires packaged, parseable training schemas and templates; otherwise direct training is required.
How do you install this skill?
Install the individual skill with the repository README’s skills CLI command: npx skills add nvidia/skills --skill tao-train-pose-classification --yes. The CLI handles the installation destination; the source does not specify a separate directory. Before running it, prepare Docker, NVIDIA Container Toolkit, and a suitable NVIDIA GPU.
How do you use this skill?
After installation, give the Agent a request such as train pose classification, skeleton action recognition, ST-GCN, or keypoint sequence classifier. Training defaults to automl_policy: on, while a new launch request can specify automl_policy: off. Use existing train_data.npy, train_label.pkl, val_data.npy, and val_label.pkl directly; run dataset conversion only for raw DeepStream BodyPose JSON. For downstream evaluation, export, or inference, provide the exact intended .pth checkpoint and the action-specific dataset paths.