[build-system]

requires = ["hatchling"]

build-backend = "hatchling.build"

[project]

name = "raitap"

version = "0.14.0"

authors = [

{ name = "Stanislas Laurent", email = "lauresta@students.zhaw.ch" },

{ name = "Jonas Vonderhagen", email = "vondejon@students.zhaw.ch" },

{ name = "Philipp Denzel", email = "denp@zhaw.ch" },

{ name = "Oliver Forster", email = "foro@zhaw.ch" },

]

maintainers = [

{ name = "Stanislas Laurent", email = "lauresta@students.zhaw.ch" },

{ name = "Jonas Vonderhagen", email = "vondejon@students.zhaw.ch" },

{ name = "Philipp Denzel", email = "denp@zhaw.ch" },

{ name = "Oliver Forster", email = "foro@zhaw.ch" },

]

description = "Python library to assess the responsibility level of AI models for integration into MLOps workflows."

readme = "README.md"

license = { file = "LICENSE" }

# Hydra 1.3.2 is part of the published runtime and currently breaks on Python 3.14.

requires-python = ">=3.11,<3.14"

keywords = [

"artificial intelligence",

"deep learning",

"neural networks",

"technical assessment",

"responsible AI",

]

classifiers = [

"Programming Language :: Python :: 3",

"Programming Language :: Python :: 3.11",

"Programming Language :: Python :: 3.12",

"Programming Language :: Python :: 3.13",

"Typing :: Typed",

"Operating System :: OS Independent",

"Topic :: Scientific/Engineering :: Artificial Intelligence",

"License :: OSI Approved :: GNU General Public License v3 (GPLv3)",

]

dependencies = [

# Core runtime.

# Upper bound on hydra-core / omegaconf: hydra-zen 0.16 reads the removed

# `omegaconf._utils.get_yaml_loader` symbol, which is gone in omegaconf 2.4

# (shipped as 2.4.0.devN). hydra-core 1.4.0.devN bundles the new omegaconf

# too. Cap both at <next-major until hydra-zen catches up. Defense-in-depth:

# the cap stops the broken pre-releases even if a downstream project relaxes

# pre-release handling. Consumers should scope pre-releases to triton-xpu

# (`triton-xpu>=3.0.0rc0`), not a global `prerelease = "allow"` — see

# docs/using-raitap/installing.

"hydra-core>=1.3.2,<1.4",

"omegaconf>=2.3,<2.4",

"hydra-zen>=0.13",

"numpy>=2.4.2",

"packaging>=23.0",

# Data and visualisation.

"matplotlib==3.10.*", # pinned: bumping invalidates committed pytest-mpl baselines (FreeType/Agg rendering)

"pandas>=2.0.0",

"pandas-stubs~=3.0.0",

"pillow>=12.2.0",

# Terminal output (cross-platform ANSI: handles Windows console via colorama-style shim).

"rich>=13.7.0",

]

[project.scripts]

raitap = "raitap.cli:main"

docs-preview = "raitap.docs_preview:main"

# `raitap-deps` removed: dep inference is now integrated into `raitap` itself.

# Use `raitap --dry-run` to preview the inferred install, `raitap --custom-deps`

# to bypass inference entirely.

[project.urls]

Documentation = "https://caiivs.github.io/raitap/"

Homepage = "https://caiivs.github.io/raitap/"

Repository = "https://github.com/CAIIVS/raitap/"

Issues = "https://github.com/CAIIVS/raitap/issues"

Source = "https://github.com/CAIIVS/raitap/"

[project.optional-dependencies]

# Backend module

torch-cpu = ["torch>=2.8.0,<2.9.0", "torchvision>=0.20.0,<0.24.0"]

torch-cuda = ["torch>=2.8.0,<2.9.0", "torchvision>=0.20.0,<0.24.0"]

torch-intel = [

# Intel XPU stays on torch>=2.10.0 (unchanged from main): the only torch in the

# 2.8 window on the xpu index is 2.8.0+xpu, whose pinned pytorch-triton-xpu==3.4.0

# is not published there, so the 2.8 floor is unresolvable for XPU. auto-LiRPA

# needs torch<2.9 and has no XPU support anyway, so it is declared mutually

# exclusive with the intel extras in [tool.uv] conflicts below.

"torch>=2.10.0; python_full_version < '3.14' and sys_platform != 'darwin'",

"torchvision>=0.20.0; python_full_version < '3.14' and sys_platform != 'darwin'",

# Pin triton-xpu>=3.0.0 — its only PyPI release (0.0.2) is yanked. The

# 3.x wheels live on the pytorch-intel index (routed via [tool.uv.sources]

# below) and are pulled in transitively by torch==*+xpu. See #110.

"triton-xpu>=3.0.0; python_full_version < '3.14' and sys_platform != 'darwin'",

]

# TODO remove bundled torch deps once ONNX no longer relies on torch tensors internally

onnx-cpu = [

"onnx>=1.21.0",

"onnxruntime>=1.24.4",

"torch>=2.8.0,<2.9.0",

"torchvision>=0.20.0,<0.24.0",

]

onnx-cuda = [

"onnx>=1.21.0",

"onnxruntime-gpu>=1.24.4",

"torch>=2.8.0,<2.9.0",

"torchvision>=0.20.0,<0.24.0",

]

onnx-intel = [

"onnx>=1.21.0; sys_platform != 'darwin'",

"onnxruntime-openvino>=1.24.1; sys_platform != 'darwin'",

"openvino>=2025.1.0; sys_platform == 'win32'",

# Intel XPU stays on torch>=2.10.0 — see the torch-intel comment above.

"torch>=2.10.0; python_full_version < '3.14' and sys_platform != 'darwin'",

"torchvision>=0.20.0; python_full_version < '3.14' and sys_platform != 'darwin'",

"triton-xpu>=3.0.0; python_full_version < '3.14' and sys_platform != 'darwin'",

]

# Transparency module

# opencv-python: SHAP modern-API image masker (inpaint/blur strategies); #267.

shap = ["shap>=0.46.0", "opencv-python>=4.9"]

captum = ["captum>=0.7.0"]

quantus = ["quantus>=0.5.3"]

transparency = ["raitap[shap,captum]"]

# Tree/tabular model backend (XGBoost) — enables shap.TreeExplainer end-to-end.

# Named per-library (like shap/captum); a `tree` umbrella can re-export this and

# future tree libs (LightGBM/sklearn) once they land. Kept out of the

# transparency umbrella: TreeExplainer needs both shap and xgboost.

# scikit-learn: XGBoostBackend loads via the sklearn-API ``XGBClassifier``, which

# imports scikit-learn — so the backend cannot load a model without it.

# torch: raitap's pipeline is torch-based (data tensors + the tree backend's

# numpy<->torch bridge), so a tree config needs torch even though XGBoost does the

# compute. Bundled here and routed to the CPU index in [tool.uv.sources] (like

# onnx-cpu bundles torch); XGBoost runs on CPU here, so the CPU build suffices and

# this extra stays hardware-variant-free.

xgboost = ["xgboost>=2.0", "scikit-learn>=1.3", "torch>=2.8.0,<2.9.0"]

# Text model backend (HuggingFace) — AutoModelForSequenceClassification +

# AutoTokenizer loading for the text input modality. Does not bundle torch:

# relies on whatever torch-* extra is already installed for the pipeline.

text = ["transformers>=4.40"]

# Robustness module

torchattacks = ["torchattacks>=3.5.1"]

foolbox = ["foolbox>=3.3.4", "numba>=0.59"]

# Average-case (statistical-sampling) adapter: the ImageNet-C common-corruption

# suite. Pure numpy/scipy/skimage/opencv — no torch, lazy-imported.

imagecorruptions = ["imagecorruptions>=1.1.2"]

# Marabou formal-verification adapter (Linux/macOS x86-64, Python 3.11 only).

# maraboupy 2.0.0 publishes cp311 wheels only — no cp312/cp313, no Windows.

# Marker below keeps ``uv sync`` working on other interpreters by silently

# skipping the extra. Not bundled into the ``robustness`` umbrella because

# installing Marabou adds a non-trivial native dep empirical-only users don't need.

marabou = [

"maraboupy>=2.0.0,<3.0; python_full_version < '3.12' and sys_platform != 'win32'",

"onnx>=1.15",

]

# auto-LiRPA certified-robustness adapter (sound, incomplete bound propagation).

# Resolved from GitHub master via [tool.uv.sources] below — auto-LiRPA has no

# PyPI release supporting torch 2.x (PyPI tops out at 0.3, torch<1.13). The

# requirement is listed by bare name (no `@ git+...`) so it stays PyPI-legal:

# a direct-URL Requires-Dist would make raitap's own wheel unpublishable. Not

# folded into the ``robustness`` umbrella for the same reason ``marabou`` isn't —

# a git-only dep makes ``pip install raitap[robustness]`` unresolvable for

# non-uv users, and bound-propagation verification is a niche opt-in.

auto-lirpa = ["auto-LiRPA"]

robustness = ["raitap[torchattacks,foolbox,marabou,imagecorruptions]"]

# Job launcher / batch scheduler support

launcher = ["hydra-submitit-launcher>=1.2.0"]

# Tracking module

mlflow = ["mlflow>=3.9.0"]

tracking = ["raitap[mlflow]"]

# Metrics module

metrics = ["torchmetrics>=1.8.2", "faster-coco-eval>=1.6.3"]

# Reporting module

pdf = ["borb==3.0.7"]

html = ["jinja2>=3.1"]

reporting = ["raitap[pdf,html]"]

[dependency-groups]

lint = ["ruff>=0.8.0", "pyright>=1.1.390"]

test = ["pytest>=9.0.3", "pytest-cov>=7.1.0", "pytest-mpl>=0.19.0"]

tooling = ["pre-commit>=4.0.0", "commitizen>=4.13.10"]

docs = [

"sphinx>=9.1.0; python_full_version >= '3.13'",

"myst-parser>=5.0.0; python_full_version >= '3.13'",

"furo>=2025.12.19; python_full_version >= '3.13'",

"sphinx-copybutton>=0.5.2; python_full_version >= '3.13'",

"sphinx-design>=0.6.1; python_full_version >= '3.13'",

"sphinx-autobuild>=2025.8.25; python_full_version >= '3.13'",

"sphinxcontrib-mermaid>=1.0.0; python_full_version >= '3.13'",

"sphinx-llms-txt>=0.2.0; python_full_version >= '3.13'",

"sphinx-sitemap>=2.6.0; python_full_version >= '3.13'",

]

dev = [

{ include-group = "lint" },

{ include-group = "test" },

{ include-group = "tooling" },

{ include-group = "docs" },

]

# uv dependency resolution rules.

[tool.uv]

# PyPI wheel for borb 3.0.x is invalid (duplicate ZIP entries); build from sdist instead.

no-binary-package = ["borb"]

override-dependencies = [

"numpy>=2.4",

"Pillow>=12.0",

"scikit-image>=0.26",

"blis>=1.0.2",

"thinc>=8.3.6,<9",

"spacy>=3.8.0",

# torchattacks 3.5.1 pins requests<2.26 which conflicts with sphinx 9.1+; relax.

"requests>=2.30.0",

# auto-LiRPA leaks test deps into install_requires and hard-pins

# ``pytest==8.1.1``, conflicting with raitap's ``pytest>=9.0.3`` dev group.

# Override the exact pin to raitap's floor (its other leaked test deps —

# pylint / pytest-order / pytest-mock — are open ranges and resolve cleanly).

"pytest>=9.0.3",

]

conflicts = [

[

{ extra = "torch-cpu" },

{ extra = "torch-cuda" },

{ extra = "torch-intel" },

],

[

{ extra = "onnx-cpu" },

{ extra = "onnx-cuda" },

{ extra = "onnx-intel" },

],

[

{ extra = "onnx-cpu" },

{ extra = "torch-cuda" },

],

[

{ extra = "onnx-cpu" },

{ extra = "torch-intel" },

],

[

{ extra = "onnx-cuda" },

{ extra = "torch-cpu" },

],

[

{ extra = "onnx-cuda" },

{ extra = "torch-intel" },

],

[

{ extra = "onnx-intel" },

{ extra = "torch-cpu" },

],

[

{ extra = "onnx-intel" },

{ extra = "torch-cuda" },

],

# xgboost bundles the CPU torch build (for the tensor pipeline), so it is

# mutually exclusive with any extra that pulls a cuda/xpu torch build.

[

{ extra = "xgboost" },

{ extra = "torch-cuda" },

],

[

{ extra = "xgboost" },

{ extra = "torch-intel" },

],

[

{ extra = "xgboost" },

{ extra = "onnx-cuda" },

],

[

{ extra = "xgboost" },

{ extra = "onnx-intel" },

],

# auto-LiRPA needs torch<2.9 but the Intel XPU extras require torch>=2.10.0

# (the 2.8 xpu wheel's triton dep is unpublished) — and auto-LiRPA has no XPU

# support regardless. Declare them mutually exclusive so uv never resolves the

# unsolvable "auto-lirpa + intel" fork.

[

{ extra = "auto-lirpa" },

{ extra = "torch-intel" },

],

[

{ extra = "auto-lirpa" },

{ extra = "onnx-intel" },

],

]

# Map runtime extras to the correct package indexes.

[tool.uv.sources]

# auto-LiRPA's torch-2.x support lives only on GitHub master (no PyPI release

# past 0.3 / torch<1.13). uv sources are NOT written into wheel metadata, so

# resolving from git here keeps the published raitap wheel PyPI-legal while the

# `auto-lirpa` extra installs the working version for dev/CI.

auto-LiRPA = { git = "https://github.com/Verified-Intelligence/auto_LiRPA" }

torch = [

{ index = "pytorch-cpu", extra = "torch-cpu" },

{ index = "pytorch-cpu", extra = "onnx-cpu" },

{ index = "pytorch-cpu", extra = "xgboost" },

{ index = "pytorch-cuda", extra = "torch-cuda" },

{ index = "pytorch-cuda", extra = "onnx-cuda" },

{ index = "pytorch-intel", extra = "torch-intel" },

{ index = "pytorch-intel", extra = "onnx-intel" },

]

torchvision = [

{ index = "pytorch-cpu", extra = "torch-cpu" },

{ index = "pytorch-cpu", extra = "onnx-cpu" },

{ index = "pytorch-cuda", extra = "torch-cuda" },

{ index = "pytorch-cuda", extra = "onnx-cuda" },

{ index = "pytorch-intel", extra = "torch-intel" },

{ index = "pytorch-intel", extra = "onnx-intel" },

]

# Route triton-xpu at the pytorch-intel index where the 3.x wheels live;

# PyPI only hosts the yanked 0.0.2. See #110.

triton-xpu = [

{ index = "pytorch-intel", extra = "torch-intel" },

{ index = "pytorch-intel", extra = "onnx-intel" },

]

# Explicit indexes used by runtime extras.

[[tool.uv.index]]

name = "pytorch-cpu"

url = "https://download.pytorch.org/whl/cpu"

explicit = true

[[tool.uv.index]]

name = "pytorch-cuda"

url = "https://download.pytorch.org/whl/cu126"

explicit = true

[[tool.uv.index]]

name = "pytorch-intel"

url = "https://download.pytorch.org/whl/xpu"

explicit = true

[tool.hatch.build.targets.wheel]

packages = ["src/raitap", "src/hydra_plugins"]

include = [

"src/raitap/py.typed",

"src/raitap/configs/**/*.yaml",

"src/raitap/configs/**/*.yml",

"src/raitap/reporting/templates/**/*",

"src/hydra_plugins/**/*.py",

]

exclude = ["src/raitap/**/test_*.py", "src/raitap/**/conftest.py", "src/raitap/tests/_fake_*plugin/**", "src/raitap/configs/zhaw/*.yaml", "src/raitap/configs/zhaw/*.yml"]

[tool.hatch.build.targets.wheel.force-include]

# Ship pyproject.toml inside the wheel so the auto-deps bootstrap can read

# raitap's own extras / conflicts / python pin from any install layout

# (dev checkout *and* PyPI wheel). Without this, the bootstrap-from-zero

# path in ``raitap.deps.bootstrap`` resolves ``_PYPROJECT`` to ``Lib/``

# under a wheel install and dies with a "No such file or directory" frame.

"pyproject.toml" = "raitap/_pyproject.toml"

[tool.hatch.build.targets.sdist]

# Keep PyPI sdist focused on installable library code; everything else

# (contributor-only assets, standalone example, docs) stays in the repo.

include = [

"src/raitap/**",

"src/hydra_plugins/**",

"pyproject.toml",

"README.md",

"LICENSE",

"CHANGELOG.md",

]

exclude = [

"src/raitap/**/test_*.py",

"src/raitap/**/conftest.py",

"src/raitap/tests/_fake_*plugin/**",

"src/raitap/configs/zhaw/*.yaml",

"src/raitap/configs/zhaw/*.yml",

"contributor-configs/**",

"example/**",

"docs/**",

"scripts/**",

"mlartifacts/**",

"mlflow*/**",

"outputs/**",

"tmp/**",

"conductor/**",

]

[tool.ruff]

line-length = 100

target-version = "py311"

extend-exclude = ["assets", "usecases", "outputs", "mlartifacts"]

[tool.ruff.lint]

select = [

"E",

"W",

"F",

"I",

"N",

"UP",

"B",

"C4", # flake8-comprehensions

"SIM",

"RUF",

"ANN001",

"ANN201",

"ANN202",

"TCH", # flake8-type-checking

"PYI",

"PGH003",

]

ignore = []

[tool.ruff.lint.per-file-ignores]

"__init__.py" = ["F401"] # Allow unused imports in __init__.py

[tool.ruff.format]

quote-style = "double"

indent-style = "space"

[tool.commitizen]

name = "cz_customize"

tag_format = "v$version"

version_scheme = "semver"

version_provider = "uv"

update_changelog_on_bump = false

major_version_zero = true

[tool.commitizen.customize]

schema = "<type>(<scope>): <subject>"

# If a scope is present, it must be one of the allowed values below.

schema_pattern = "^(feat|fix|docs|style|refactor|perf|test|build|ci|chore|revert)(\\((transparency|tracking|metrics|docs|infra|deps|misc|config|model|data|reporting|robustness)\\))?(!)?:\\s.+$"

[tool.pyright]

include = ["src"]

exclude = [

"**/.*",

"**/.venv",

"**/__pycache__",

"**/node_modules",

"**/build",

"**/dist",

"assets",

"usecases",

"outputs",

"mlartifacts",

]

pythonVersion = "3.11"

typeCheckingMode = "standard"

reportMissingImports = true

reportMissingTypeStubs = false

reportUnusedImport = true

reportUnusedVariable = true

[tool.pytest.ini_options]

norecursedirs = ["assets", "usecases", "outputs", "mlartifacts", "src/raitap/tests/_fake_plugin", "src/raitap/tests/_fake_broken_plugin"]

testpaths = ["src"]

pythonpath = ["src"]

python_files = ["test_*.py"]

python_classes = ["Test*"]

python_functions = ["test_*"]

markers = [

"e2e: end-to-end tests that are resource-intensive; run once per PR, not on every commit",

"mpl: matplotlib baseline-regression tests that require committed reference images",

"runtime: runtime resolution and backend hardware-label tests for platform-specific CI selection",

"cuda: tests that require real CUDA hardware and run only on the GPU CI lane",

"slow: integration tests that download model weights; kept in Quality Gate but opt-out via -m 'not slow'",

"parity: assert raitap output equals a direct third-party-library call",

"visual: pytest-mpl pixel baseline regression of rendered figures",

"robustness: parity/e2e tests for the robustness (adversarial/formal) assessors",

"negative: tests asserting failure paths (missing baseline, invalid config, ...)",

]

addopts = ["-v", "--strict-markers", "--strict-config", "--tb=short"]

filterwarnings = [

# captum's "Input Tensor.*required_grads" is also silenced at runtime by the

# adapter (raitap.transparency.explainers.captum_explainer) — duplicated here

# because pytest resets the warnings filter between tests.

"ignore:Input Tensor.*required_grads:UserWarning",

# captum uses deprecated matplotlib cm.get_cmap; will be fixed upstream

"ignore:The get_cmap function was deprecated:matplotlib.MatplotlibDeprecationWarning",

# shap uses np.random.seed internally; will be fixed upstream

"ignore:The NumPy global RNG was seeded:FutureWarning",

# tight_layout incompatibility in ShapImageVisualiser; cosmetic only

"ignore:This figure includes Axes that are not compatible with tight_layout:UserWarning",

]

[tool.coverage.run]

source = ["src"]

omit = ["*/tests/*", "*/__pycache__/*", "*/.venv/*"]

[tool.coverage.report]

precision = 2

show_missing = true

skip_covered = false

exclude_lines = [

"pragma: no cover",

"def __repr__",

"raise AssertionError",

"raise NotImplementedError",

"if __name__ == .__main__.:",

"if TYPE_CHECKING:",

"@abstractmethod",

]