TAO Person Re-Identification
Train and deploy TAO ReID models for cross-camera person matching.
The skill limits itself to Read/Bash, requires Docker and NVIDIA Container Toolkit, warns against using TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD with untrusted checkpoints, and defines supported action boundaries. Deductions apply because it does not require user confirmation, document least-privilege data access, address person-image privacy, disclose external data flows, provide rollback, or define recovery after failure; W&B is enabled by default and may create an external data flow.
SKILL.md, skill_info.yaml, templates, and schemas document actions, inputs, outputs, constraints, and several failure patterns, including deterministic settings and diagnostic guidance. Deductions apply because the source contains no executable reproduction, CI, or key-path test coverage; template defaults may conflict with the documented AutoML metric policy, and file-versus-directory input semantics may cause runtime failures.
The name, trigger phrases, supported and unsupported actions, dataset requirements, GPU requirements, and single-node limitation are relatively clear for TAO, Docker, and NVIDIA-GPU ReID workflows. Deductions apply because non-ReID boundaries, Chinese-language interaction, local-network adaptation, dataset structure assumptions, real data acquisition, and confirmation conditions are not fully specified.
The skill provides front matter, version 0.1.0, author information, an Apache-2.0 declaration, schemas, templates, error patterns, benchmark material, and parameter mappings; repository material describes an update mechanism. Deductions apply because recommended Instructions and Examples sections are missing, the author format is incomplete, there is no changelog, explicit maintainer, dependency-version policy, installation troubleshooting, or FAQ, and the supplied license metadata is NOASSERTION.
The documentation covers train, evaluate, inference, and export, with data-source, checkpoint, output-file, and parameter examples that could directly guide a TAO ReID workflow. Deductions apply because there is only one positive evaluation task, while the benchmark reports 20% Codex correctness, 0% Codex discoverability, and 44% Codex effectiveness; no representative artifacts or independent result verification are supplied, so stable end-to-end benefit is unconfirmed.
The committed evidence includes SKILL.md, schemas, templates, configuration mappings, an evaluation report, and one evaluation case, providing some auditability. Deductions apply because there is no executable test suite or CI covering key actions, the evaluation has one task and no negative cases, cross-source corroboration is absent, and this conclusion is based entirely on static review.
- The workflow handles person re-identification data but does not specify privacy controls, retention, access restrictions, or cross-border data handling.
- W&B is enabled by default; confirm before execution whether metrics, code, or metadata may be uploaded and whether this complies with network and organizational policy.
- AutoML routing, checkpoint selection, the Docker image, and data sources lack end-to-end verification in this static review; validate first with isolated data and recoverable output directories.
- The benchmark is extremely small and reports 0% Codex discoverability, so verify explicitly that the skill is selected for relevant requests.
What does this skill do, and when should you use it?
This skill guides NVIDIA TAO Toolkit workflows for person re-identification, using metric learning to produce discriminative embeddings for cross-camera matching. It supports training, evaluation, inference, and ONNX export, with AutoML enabled by default for training. Evaluation requires both query and gallery/test data for retrieval metrics such as CMC and mAP. Docker, NVIDIA Container Toolkit, and at least one GPU are required.
Reads and applies TAO ReID workflow specifications; maps train, test, and query archives to the required action fields; launches direct training or the default AutoML route; resumes training from checkpoints; evaluates retrieval performance using CMC and mAP; exports checkpoints to ONNX; writes inference JSON results, CMC plots, and sampled-match plots; and resolves model files from parent job result directories.
- A computer-vision engineer wants to train a cross-camera ReID model on custom identity data.
- An evaluation engineer needs to measure CMC and mAP using query and gallery datasets.
- A deployment engineer needs to export a trained ReID checkpoint as an ONNX file.
- A researcher needs to run retrieval inference and save matching results as JSON.
- A training operator wants to resume from a checkpoint or optimize cmc_rank_1 with AutoML.
What are this skill's strengths and limitations?
- Covers training, evaluation, inference, and ONNX export in one workflow.
- Clearly specifies the required train, query, and gallery/test data for each action.
- Supports AutoML with cmc_rank_1 as the default optimization metric.
- Documents multi-GPU behavior, checkpoint handoff, and common failure patterns.
- Requires Docker, NVIDIA Container Toolkit, and at least one GPU.
- Uses a single-node-oriented workflow with no explicit num_nodes configuration.
- Evaluation and inference require both query and test/gallery datasets with a shared identity namespace.
- The supplied sources do not provide a specific TAO image, complete launch command, or platform test matrix.
How do you install this skill?
Install the skill with the NVIDIA skills CLI:
npx skills add nvidia/skills --skill tao-train-reid --yes
The README does not specify the exact local destination; the CLI prompts for or applies its installation location.
How do you use this skill?
After installation, trigger the skill with prompts such as “train ReID,” “person re-identification,” or “cross-camera person matching.” Training, evaluation, inference, and export require the relevant dataset or checkpoint paths. Training uses AutoML by default; set automl_policy: off for a run that should use direct model training. Docker, NVIDIA Container Toolkit, and a GPU are required. The supplied sources do not document a complete TAO container command or image name.