TAO PyT Image Classification
Train, evaluate, optimize, and deploy PyTorch image classifiers with NVIDIA TAO.
The skill limits declared tools to Read/Bash and documents data paths, container use, upload exclusions, NVIDIA ownership, and Apache-2.0 metadata. However, Bash, Docker, GPU access, model files, and potentially enabled Weights & Biases create external effects; stepwise confirmation, sensitive-data handling, network-flow disclosure, rollback, and recovery are not specified, so points are deducted.
The skill provides action definitions, generated schemas, templates, parameter mappings, deterministic training settings, and common error patterns, while the benchmark report claims a pass. Static evidence lacks executable tests or real CI coverage for key paths; deployment depends on a referenced file whose contents are unavailable here, and failure handling remains limited, so points are deducted.
Trigger phrases, target scenarios, supported workflows, GPU requirements, and data formats are relatively clear. Non-fit boundaries, Chinese-language support, mainland-China environment fit, and reachability of Docker images or external services are not addressed, so points are deducted.
The documentation includes front matter, version, author, license, schemas, templates, parameter tables, error patterns, and a benchmark report; repository context also gives publishing, synchronization, and security-reporting paths. It lacks the recommended Instructions/Examples sections, a detailed FAQ, a skill-level changelog, pinned dependency versions, and clearly assigned maintenance ownership, so points are deducted.
The documentation covers the core workflow from constructing data paths through training, checkpoint handoff, and deployment, and the benchmark reports one positive task passing. However, there is only one task and no representative artifact or statically verifiable output; direct usability and comparative benefit over alternatives are insufficiently evidenced, so points are deducted within the static ceiling.
Schemas, templates, skill_info, an evaluation task, and an external evaluation report provide some auditable evidence. There is no committed test suite or CI evidence covering this skill's key paths, and the benchmark has only one task, so verifiability is limited.
- Before execution, confirm Docker image provenance, GPU permissions, model and data paths, and whether Weights & Biases will transmit data over the network.
- Add executable CI/tests for key actions, include the referenced deployment documentation, and document failure recovery and version-pinned dependencies.
- Mainland-China users may face reachability issues for container images, NVIDIA services, or external experiment-tracking services; no alternatives are documented.
What does this skill do, and when should you use it?
This skill supports NVIDIA TAO Toolkit workflows for PyTorch image classification, including training, evaluation, inference, distillation, quantization, export, and TensorRT deployment. It covers backbones such as FAN, EfficientNet, and ResNet, and includes AutoML-aware training policy guidance. Docker, nvidia-container-toolkit, and a GPU are required. The instructions use packaged schemas, specification templates, deployment references, and checkpoint-resolution rules to construct and run model jobs.
Reads skill metadata, JSON schemas, YAML specification templates, and the TensorRT deployment reference; constructs or adjusts TAO specifications for training, evaluation, inference, distillation, quantization, export, and TensorRT engine generation; requires action-specific paths for training, validation, test, class-label, and calibration data; runs TAO PyT workflows; resolves checkpoints from parent jobs; and produces or consumes model checkpoints, ONNX files, and TensorRT engines for evaluation or inference.
- A computer-vision engineer wants to train a classifier on custom image data with a FAN, EfficientNet, or ResNet backbone.
- A team needs validation, test-set evaluation, or inference for a TAO image-classification checkpoint.
- A deployment engineer needs to export a classifier to ONNX and generate a TensorRT engine.
- A model-optimization specialist needs knowledge distillation or quantization with a calibration dataset.
- A researcher wants AutoML enabled by default or wants to disable it for one training run.
What are this skill's strengths and limitations?
- Covers the classification workflow from training through TensorRT deployment.
- Supports multiple backbones, distillation, quantization, and AutoML policy handling.
- Documents action-specific data requirements, parameter inference, checkpoint handoff, and common error patterns.
- States the required Docker and NVIDIA container runtime environment clearly.
- Requires a GPU, Docker, and nvidia-container-toolkit; the documented hardware guidance recommends at least 16GB of VRAM per GPU.
- Archives must be downloaded and extracted before local Docker workflows can use the image directories.
- The supplied material provides no standalone test suite, benchmark results, or verified platform matrix for this individual skill.
- FSDP is unsupported; multi-node use also depends on orchestrator-provided environment variables.
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-image-classification --yes
The README says the CLI prompts for the installation destination but does not specify a fixed directory.
How do you use this skill?
After installation, give the agent a concrete request such as "train image classifier with TAO classification-pyt", "evaluate this TAO classification model", or "export the classifier and generate a TensorRT engine". Provide training, validation, and test data plus classes.txt, and specify the backbone, GPU settings, output directory, and AutoML policy as needed. For local Docker, point image-directory fields to extracted class-root folders rather than passing the tar.gz archives directly.