Data & Analysis ✓ NVIDIA · Official masked-autoencoderself-supervised-learningcomputer-visionimage-classificationpytorchnvidia-taoautoml

TAO MAE Vision Pretraining

Guides NVIDIA TAO through masked-autoencoder pretraining and fine-tuning for visual representations.

FollowSkills review · FSRS-2.0
Not recommended
48/ 100 5-point scale 2.4 / 5
Trust13 / 25 · 2.6/5

The skill limits declared tools to Read and Bash and states Docker/NVIDIA Container Toolkit requirements; however, it handles datasets, checkpoints, S3 paths, and encryption_key values without documenting sensitive-data protection, least-privilege confirmation, network flows, rollback, or recovery, so points are deducted. No malware, credential theft, covert exfiltration, destructive default, or other red-line behavior is evident.

Reliability7 / 20 · 1.8/5

SKILL.md, skill_info.yaml, and the templates consistently describe commands, inputs, outputs, and several error patterns, and deterministic training settings are stated. However, static review did not execute key paths; dependency versions, environment reproduction, abnormal-input feedback, and complete test coverage are thin. The static cap keeps this at or below 10, so points are deducted.

Adaptability10 / 15 · 3.3/5

Trigger phrases, MAE pretrain/finetune scenarios, action scope, data format, hardware requirements, and the unsupported pretrain-checkpoint inference case are relatively clear. Boundaries and output contracts are incomplete, Chinese-language guidance is absent, the Docker/NVIDIA GPU barrier is substantial, and mainland-China reachability of images or data sources is not established, so points are deducted.

Convention8 / 15 · 2.7/5

The skill provides front matter, version 0.1.0, author, an Apache-2.0 declaration, parameter tables, templates, and error patterns. It lacks explicit Instructions and Examples sections; maintenance ownership, changelog, dependency pinning, and update path are incomplete. The benchmark also flags the author format, while supplied license metadata is NOASSERTION, so a high score is not justified.

Effectiveness6 / 15 · 2.0/5

The documentation covers data-source mappings, AutoML routing, stage selection, checkpoint propagation, export, and a deployment entry point, making the core TAO MAE workflow plausible. However, no representative final artifacts are statically verified, deployment reference content is incomplete, and high GPU/memory costs plus runtime assumptions may require manual correction. The static cap keeps this at or below 7, so points are deducted.

Verifiability4 / 10 · 2.0/5

A dated one-task NVSkills-Eval report provides summary metrics and static findings. It has only one positive task, no negative-trigger coverage, no raw logs, and no committed tests covering this skill's key paths; independent reproduction and cross-source corroboration are therefore limited. The static cap keeps this at or below 5.

Evidence confidence:Low Reviewed Jul 20, 2026 Reviewed revision 55f18499943e
The upstream repository has new commits since this review. The score still applies to the reviewed revision shown and may not cover the latest changes.
Before you use it
  • Before execution, confirm Docker, NVIDIA Container Toolkit, GPU memory, the TAO container image, and data-source availability and reachability. Do not place real encryption keys, credentials, or sensitive data in prompts, logs, S3 examples, or shared configuration; the skill does not define their protection or cleanup. Verify that train.stage, model.arch, num_classes, input size, and checkpoint agree, and that archives are extracted as required; otherwise zero-sample loaders or shape mismatches may occur. AutoML is enabled by default and may incur additional compute; although an on/off override is described, complete user confirmation, cost limits, and rollback procedures are not specified.
See the full review method →

What does this skill do, and when should you use it?

This skill guides an AI agent through NVIDIA TAO workflows for self-supervised MAE pretraining and image-classification fine-tuning. It covers train, evaluate, inference, and export actions, with TensorRT engines handled through the TAO Deploy workflow. Training supports pretrain and finetune stages and defaults to an AutoML route when the packaged schema and template are available. Docker and nvidia-container-toolkit are required; pretraining requires at least two GPUs and is recommended for eight GPUs with 24GB or more of VRAM per GPU.

Reads the MAE skill instructions, deployment guide, dataset requirements, generated schemas, and specification templates; constructs mandatory data-source paths for train, evaluate, and inference actions; guides execution of the parent PyTorch mae CLI for train, evaluate, inference, and export; routes eligible training runs through tao-run-automl; applies checkpoint, output-directory, encryption-key, and ONNX-path inference rules; and uses the separate TAO Deploy workflow to build TensorRT engines.

  1. A computer-vision engineer wants to pretrain an MAE on image archives to learn visual representations without labels.
  2. A model developer has pretrained MAE weights and wants to fine-tune them for image classification.
  3. An MLOps engineer needs to evaluate a fine-tuned checkpoint or run classification inference.
  4. A deployment engineer wants to export the model to ONNX and build a TensorRT engine through TAO Deploy.
  5. A TAO user wants AutoML enabled by default but controllable per training run.

What are this skill's strengths and limitations?

Pros
  • Covers pretraining, fine-tuning, evaluation, inference, export, and TensorRT deployment workflows.
  • Defines action-specific archive formats, mandatory data-source fields, and important parameter constraints.
  • Supports ConvNeXtV2, ViT, and Hiera architecture families and provides an AutoML training policy.
  • Documents checkpoint and parent-job result-directory inference rules.
Limitations
  • MAE pretraining requires at least two GPUs, is recommended for eight GPUs, and needs 24GB or more of VRAM per GPU.
  • Requires Docker and nvidia-container-toolkit, so it targets NVIDIA GPU container environments.
  • Pretraining checkpoints cannot be used directly for inference or classification-style evaluation; a fine-tuned checkpoint is required.
  • A local tar archive path passed directly to the MAE CLI can produce a zero-sample dataloader; archives must be extracted for direct local Docker runs.

How do you install this skill?

Install the individual skill with the repository's documented skills CLI flow:

npx skills add nvidia/skills --skill tao-train-mask-auto-encoder --yes

To target Codex explicitly:

npx skills add nvidia/skills --skill tao-train-mask-auto-encoder --agent codex --yes

The runtime also requires Docker and nvidia-container-toolkit.

How do you use this skill?

After installation, give the agent a relevant request such as “pretrain MAE,” “self-supervised vision pretraining,” “Masked Autoencoder,” or “MAE fine-tune.” Supply the required training and validation data sources for training, and the required validation or test archive for evaluation and inference. Use a fine-tuned checkpoint for classification evaluation and inference. For direct local Docker runs, extract images_train.tar.gz, images_val.tar.gz, or images_test.tar.gz and point the specification to the extracted folders. Before generating a TensorRT engine, the agent should read references/tao-deploy-mask-auto-encoder.md.

FAQ

Is this suitable for an environment without NVIDIA GPUs?
No. The skill requires Docker and nvidia-container-toolkit, and MAE pretraining requires at least two GPUs, with eight GPUs recommended and at least 24GB of VRAM per GPU.
Can AutoML be disabled?
Yes. Set automl_policy: off for the individual training run, or use phrases such as “turn off AutoML,” “disable AutoML,” or “no HPO.” This does not change the model's AutoML metadata.
Can a pretraining checkpoint be used directly for inference?
No. The MAE prediction dataloader does not support train.stage: pretrain. Use a finetune checkpoint for inference and classification-style evaluation.

More skills from this repository

All from NVIDIA/skills

Data & Analysis ✓ NVIDIA · Official

TAO PyT Image Classification

Train, evaluate, optimize, and deploy PyTorch image classifiers with NVIDIA TAO.

Data & Analysis ✓ NVIDIA · Official

TAO SegFormer Semantic Segmentation

Guidance for training, evaluating, exporting, and deploying TAO SegFormer models.

Data & Analysis ✓ NVIDIA · Official

TAO Person Re-Identification

Train and deploy TAO ReID models for cross-camera person matching.

Data & Analysis ✓ NVIDIA · Official

TAO Visual ChangeNet Inspection

Train paired-image classifiers and change-segmentation models for PCB defect inspection.

Data & Analysis ✓ NVIDIA · Official

TAO OneFormer Universal Segmentation

Train and deploy one task-conditioned model for panoptic, instance, and semantic image segmentation.

Dev & Engineering ✓ NVIDIA · Official

TAO Hugging Face Model Integrator

Connect Hugging Face computer-vision models to NVIDIA TAO training, ONNX export, and TensorRT deployment workflows.

Data & Analysis ✓ NVIDIA · Official

TAO Depth Anything v2

Train, evaluate, export, and deploy monocular depth models from single RGB images with NVIDIA TAO.

Data & Analysis ✓ NVIDIA · Official

TAO Grounding DINO Open-Vocabulary Detection

Train and deploy text-guided object detectors without a fixed class vocabulary.

Data & Analysis ✓ NVIDIA · Official

TAO AutoML Tuning

Run governed hyperparameter optimization for NVIDIA TAO networks across supported SDK platforms.

Data & Analysis ✓ NVIDIA · Official

TAO AutoML + DEFT Three-Phase Pipeline

Chains hyperparameter tuning, iterative data improvement, and final refinement for AOI and other DEFT applications.

Data & Analysis ✓ NVIDIA · Official

Mask Grounding DINO Training & Deployment

Train and deploy text-guided open-vocabulary instance segmentation models.

Data & Analysis ✓ NVIDIA · Official

TAO Deformable DETR Object Detection

Guides end-to-end training, evaluation, export, quantization, and deployment of lightweight Deformable DETR detectors in NVIDIA TAO.

Data & Analysis ✓ NVIDIA · Official

TAO NVDINOv2 Visual Representation Training

Train reusable visual features with label-free teacher-student distillation.

Data & Analysis ✓ NVIDIA · Official

TAO MAL Weakly Supervised Segmentation

Train segmentation models from point or box annotations and produce masks with minimal labeling.

Data & Analysis ✓ NVIDIA · Official

TAO Pose Classification

Classify pose-keypoint sequences into action categories with ST-GCN.

Data & Analysis ✓ NVIDIA · Official

TAO Mask2Former Segmentation

Guides NVIDIA TAO workflows for training and deploying Mask2Former segmentation models.

Data & Analysis ✓ NVIDIA · Official

Hugging Face GPU Fine-Tuning

Build reproducible Hugging Face fine-tuning pipelines on local NVIDIA GPUs.

Data & Analysis ✓ NVIDIA · Official

TAO Video Action Recognition

Train, evaluate, export, and run inference for RGB, optical-flow, and multi-stream video action models with NVIDIA TAO.

Data & Analysis ✓ NVIDIA · Official

TAO VCN Classification Gap Analysis

Find the weakest VCN classification samples and turn them into augmentation targets.

Data & Analysis ✓ NVIDIA · Official

TAO ChangeNet Root Cause Analyzer

Uses image evidence to identify why NVIDIA TAO Visual ChangeNet classification experiments fail.

Related skills