TAO Metric-Learning Recognition
Train embedding models for fine-grained visual retrieval and matching.
The skill limits declared tools to Read and Bash and documents Docker/NVIDIA Container Toolkit, data-source construction, checkpoint inference, upload exclusions, and a warning against using the checkpoint-loading environment variable for untrusted checkpoints. It does not clearly define user confirmation, sensitive-data handling, least-privilege Bash, external network/cloud effects, rollback, or the complete data flow, so points are deducted.
Actions, inputs/outputs, templates, parameters, deterministic training settings, and several error patterns are documented. However, there is little actionable recovery guidance or key-path reproduction evidence, dependencies on external TAO/AutoML components remain, and the templates/action descriptions show possible inconsistencies; the static ceiling also applies.
Trigger phrases, train/evaluate/inference/export/deployment scenarios, GPU requirements, and dataset structures are reasonably clear. The skill is nevertheless tightly coupled to TAO, Docker, and NVIDIA GPUs, is single-node oriented, and does not state non-fit boundaries, Chinese-language support, or mainland-China reachability, so points are deducted.
The files provide a name, description, Apache-2.0 metadata, version, author, references, schema/template organization, and publisher maintenance signals. They lack dedicated Instructions, Examples, FAQs, changelog, and a clear maintenance/update path; the included benchmark also reports hierarchy, author-format, and missing recommended-section findings, so points are deducted.
The documentation covers core training through export and TensorRT deployment configuration and supplies dataset and parameter examples, which can reduce setup work. Representative outputs and end-to-end static verification are limited, while AutoML, SDK, container, and dataset dependencies add cost, so the score remains below the static maximum.
Structured skill_info, templates, an evaluation task, and one external benchmark report provide limited audit evidence. The benchmark has only one task and one attempt, with no skill-specific CI/test coverage or independent multi-source corroboration; static evidence is insufficient for more than 5 points.
- Before execution, confirm Bash, Docker, GPU, dataset paths, and cloud-storage permissions, especially the example S3 source.
- Do not set TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1 for untrusted checkpoints; it may increase deserialization risk.
- Reconcile the evaluate/export/inference templates with skill_info, the TAO version, and referenced files before use; the static material does not prove end-to-end operation.
- Mainland-China reachability and Chinese operating guidance are not documented; external TAO/container images and data sources may be inaccessible.
What does this skill do, and when should you use it?
This skill supports NVIDIA TAO Toolkit workflows for metric-learning recognition, including training, evaluation, export, inference, and TensorRT deployment. It learns embeddings for fine-grained matching use cases such as retail product recognition with triplet or contrastive losses. Training is AutoML-enabled by default, with a per-run option to disable AutoML. It requires Docker, nvidia-container-toolkit, and at least one NVIDIA GPU.
Reads the skill metadata, training schemas, and deployment references; constructs required dataset paths and spec_overrides for each action; routes eligible training requests through TAO AutoML or direct model training; configures reference/query datasets for retrieval evaluation; guides checkpoint inference, ONNX export, inference, and TensorRT operations; and documents the trusted-checkpoint setting needed for certain PyTorch 2.6 checkpoint loads.
- A computer-vision team training retrieval embeddings for retail products or other fine-grained categories.
- A TAO user evaluating retrieval performance with reference and query datasets for unknown classes.
- A deployment engineer exporting a trained model to ONNX or generating a TensorRT engine.
- An application developer running retrieval inference on test images from an unknown-class dataset.
What are this skill's strengths and limitations?
- Covers training, evaluation, export, inference, and TensorRT deployment flows.
- Documents model options, dataset requirements, checkpoint inference, and important hyperparameters.
- Provides an AutoML-by-default training policy with a per-run opt-out.
- Clearly defines reference/query requirements for retrieval evaluation.
- Requires Docker, nvidia-container-toolkit, and NVIDIA GPU hardware; the minimum is one GPU, with two GPUs and 16GB or more VRAM per GPU recommended.
- The documented multi-GPU setup is Lightning-managed and single-node oriented, with no explicit multi-node configuration.
- The supplied source does not include a complete container setup, dataset acquisition workflow, or end-to-end runnable example.
- The PyTorch 2.6 checkpoint workaround is intended only for checkpoints from the same trusted TAO workflow and should not be used for untrusted checkpoints.
How do you install this skill?
Install the individual skill with the command documented in the repository README:
npx skills add nvidia/skills --skill tao-train-metric-learning-recognition --yes
For Codex-specific installation:
npx skills add nvidia/skills --skill tao-train-metric-learning-recognition --agent codex --yes
How do you use this skill?
After installation, activate it with a task prompt such as “train metric learning for retail product recognition” or “run ml-recog inference with retrieval embeddings.” Provide Docker, nvidia-container-toolkit, an NVIDIA GPU, and datasets matching the documented training or evaluation requirements. Training requests must include the required data-source overrides; evaluation requires reference and query datasets. The source does not document a complete image setup, dataset download procedure, or end-to-end command sequence.