v1.0.3#

JATIC Maturity I

Release v1.0.3 has been assessed against and meets the requirements associated with JATIC Maturity Level I.

This patch release contains fixes to container security scanning, missing documentation cross-references, container compliance pipeline components, and notebook install cells. Key areas include adding the missing medium-severity vulnerability scan to the container pipeline, completing the cross-references between how-to guides and tutorials, refreshing container compliance components, and resolving a torch / torchvision ABI drift that broke GPU saliency notebook execution.

Fixes#

  • Fixed missing medium-severity vulnerability scanning in the container CI pipeline by adding Trivy scans.

  • Fixed a medium-severity vulnerability in Dockerfile by updating pip to 25.3.

  • Fixed missing cross-references between how-to guides and tutorials so each how-to links to the tutorials that exercise its perturbers and each tutorial points at the underlying perturber API how-to. Fixed inconsistent “Related How-To Guides” section headings across tutorial notebooks.

  • Updated container compliance pipeline components in .gitlab-ci/pipelines/.gitlab-compliance.yml.

  • Fixed an outdated six-month roadmap (ROADMAP.md and docs/figures/nrtk-road-map.png).

  • Fixed several repository-level documentation and CI items, including CONTRIBUTING.md, README.md, ROADMAP.md, .markdownlint.yaml, .gitlab-ci/.gitlab-docs.yml, the nrtk_xaitk_workflow README, and license-header / docstring updates across a few notebook utility modules and perturber implementations.

  • Fixed torch / torchvision ABI drift in image_classification_perturbation_saliency.ipynb and object_detection_perturbation_saliency.ipynb install cells by pulling torchvision from the same --extra-index-url as torch. Without this, ultralytics would transitively install a mismatched torchvision wheel and the notebooks would fail at import time on Linux/CUDA hosts.

  • Fixed stale notebook outputs in both XAITK saliency notebooks by re-executing them on a CUDA runner.

  • Fixed pyright flagging public torch API (torch.tensor, torch.Generator, torch.zeros, torch.float32, etc.) as not exported from torch. Pyright 1.1.409 introduced a regression in reportPrivateImportUsage handling (microsoft/pyright#11404). Capped the linting dependency to pyright<1.1.409 (rather than excluding only the known-bad version) in pyproject.toml, locking the supported install range to 1.1.397–1.1.408 so this release’s dev environment keeps installing cleanly across its support lifetime regardless of future pyright releases.

On this page