TAO Mask2Former Segmentation
Guides NVIDIA TAO workflows for training and deploying Mask2Former segmentation models.
The skill explicitly requires Docker and nvidia-container-toolkit and limits declared tools to Read/Bash. It documents data sources, parent-job artifacts, and encryption-key mappings, with no evidence of malware, credential theft, or destructive defaults. Deductions apply because Bash is broad, there is no user-confirmation or rollback procedure, and data may flow through S3, pretrained-weight downloads, and enabled W&B without sufficient sensitive-data or external-effect disclosure.
The documentation covers training, evaluation, inference, export, quantization, multi-GPU operation, and several failure patterns, with deterministic training defaults. Deductions apply because packaged templates default to values inconsistent with the documented COCO panoptic workflow, such as ADE dataset defaults; AutoML, deployment references, and older PyTorch behavior depend on external components, and no complete reproducible test suite is supplied. Static calibration limits this score to 10.
Triggers, scope, data formats, hardware requirements, and major usage constraints are reasonably clear, covering panoptic, instance, and semantic tasks. Deductions apply because Chinese-language guidance and semantic non-trigger boundaries are absent, while core operation depends on Docker, NVIDIA GPUs, TAO containers, S3, and GitHub; overseas dependencies may be difficult to reach from mainland China.
The skill has useful metadata, versioning, author attribution, parameter guidance, templates, troubleshooting, deployment notes, and a stated Apache-2.0 license; the repository README describes synchronization and installation entry points. Deductions apply for missing Instructions, Examples, FAQ, changelog, and explicit maintainer/update ownership, incomplete author formatting, and schema-folder packaging findings recorded in the benchmark.
The skill directly addresses the Mask2Former workflow from training through deployment and supplies data-path mappings, configuration templates, parent-model inference, and troubleshooting guidance, which can reduce manual configuration. Deductions apply because representative outputs were not independently verified here, users must still provide data and model artifacts, and runtime/template version differences may require rework; static calibration caps this dimension at 7.
The supplied files include a pinned revision context, schema metadata, action commands, templates, and documented failure cases; the benchmark records one single-task evaluation. Deductions apply because the benchmark has only one positive task and no negative coverage, provides no independently verifiable CI or comprehensive test suite, and its results are repository claims rather than corroborated evidence. Static calibration caps this dimension at 5.
- Before use, verify the TAO container version, packaged schemas/templates, dataset format, and COCO category-ID settings; template defaults are not always aligned with the prose requirements.
- Training and inference may access S3, GitHub pretrained weights, and W&B; avoid unauthorized sensitive data or credentials and confirm network, residency, and upload policies.
- Use an isolated environment for an initial small-scale validation and retain checkpoints and output directories for manual recovery.
What does this skill do, and when should you use it?
This skill supports NVIDIA TAO Mask2Former workflows for universal image segmentation: panoptic, instance, and semantic segmentation. It covers training, evaluation, inference, ONNX export, quantization, and TensorRT deployment. The instructions define dataset files, specification overrides, AutoML routing, multi-GPU settings, export constraints, and troubleshooting patterns. It requires Docker, nvidia-container-toolkit, and an NVIDIA GPU, with substantial memory recommended.
Reads the skill’s deployment reference, generated TAO schemas, specification templates, and AutoML configuration; constructs required COCO or COCO panoptic data-source paths for each action; and guides train, evaluate, inference, export, quantize, and TensorRT deployment workflows. It documents configuration for pretrained Swin weights, class counts, segmentation modes, GPU allocation, checkpoints, ONNX files, TensorRT engines, and inferred parent-job or output-directory parameters.
- A computer-vision engineer training panoptic segmentation on COCO panoptic data can use it to configure datasets, classes, epochs, and GPUs.
- An evaluation engineer validating a trained Mask2Former checkpoint can use it to prepare the required evaluation sources and checkpoint settings.
- A deployment engineer exporting Mask2Former to ONNX and TensorRT can use it to follow the deployment reference and compatible export dimensions.
- A team running semantic, instance, or panoptic inference can use it to select model.mode and provide the required annotation files.
- An engineer facing CUDA out-of-memory errors can use it to reduce image resolution or select a smaller Swin variant.
What are this skill's strengths and limitations?
- Covers Mask2Former training, evaluation, inference, export, quantization, and TensorRT deployment workflows.
- Provides explicit COCO panoptic file requirements, specification keys, and per-action data rules.
- Documents AutoML policy handling, checkpoint inference, multi-GPU configuration, and recurring failure modes.
- States that TensorRT supports FP32 and FP16 and provides compatible default export dimensions.
- Requires Docker, nvidia-container-toolkit, and at least one NVIDIA GPU; four GPUs and 24GB or more of VRAM per GPU are recommended.
- Mask2Former is memory-intensive, with per-GPU batch size 1 as the default.
- Every action requires constructed data-source paths, so users must prepare the required COCO or COCO panoptic files.
- TensorRT does not support INT8, and very small export dimensions may trigger an ONNX shape-inference error.
How do you install this skill?
Install the named skill with the CLI command documented in the repository README:
npx skills add nvidia/skills --skill tao-train-mask2former --yes
The CLI selects or prompts for the installation target; cloning the repository or manually copying the folder is not required. The source does not document additional installation-directory details.
How do you use this skill?
After installation, ask a compatible client for a concrete task such as "train a Mask2Former panoptic segmentation model," "evaluate this Mask2Former checkpoint," or "export Mask2Former to ONNX and deploy it with TensorRT." Before generating a TensorRT engine or running TensorRT evaluation or inference, read references/tao-deploy-mask2former.md. Every training, evaluation, inference, and quantization action requires the data-source paths specified by the skill.