# =============================================================================
# pyproject.toml — single source of truth for the `captioning` Python package
# -----------------------------------------------------------------------------
# This file follows PEP 621 (project metadata) and PEP 517/518 (build system).
# It replaces a scattered mix of setup.py + requirements.txt + setup.cfg with
# one canonical config. `pip install -e .` installs the package from `src/`.
#
# Why src/ layout?  It prevents accidental imports of the package from the
# repo root during testing — every test exercises the *installed* package,
# the way users will actually import it. This is the layout used by the
# Python Packaging Authority's example projects and recommended by pytest.
# =============================================================================

[build-system]
requires = ["setuptools>=68", "wheel"]
build-backend = "setuptools.build_meta"

# -----------------------------------------------------------------------------
# Project metadata — what `pip show captioning` will display.
# -----------------------------------------------------------------------------
[project]
name = "captioning"
version = "0.1.0"
description = "IEEE-published CNN+Transformer image captioning, restructured into a production-grade multimodal AI platform."
readme = "README.md"
requires-python = ">=3.10,<3.13"
license = { text = "MIT" }
authors = [
    { name = "Apoorv Raj" },
]
keywords = [
    "image-captioning",
    "multimodal",
    "transformer",
    "computer-vision",
    "tensorflow",
    "fastapi",
]
classifiers = [
    "Development Status :: 3 - Alpha",
    "Intended Audience :: Developers",
    "Intended Audience :: Science/Research",
    "License :: OSI Approved :: MIT License",
    "Programming Language :: Python :: 3",
    "Programming Language :: Python :: 3.10",
    "Programming Language :: Python :: 3.11",
    "Programming Language :: Python :: 3.12",
    "Topic :: Scientific/Engineering :: Artificial Intelligence",
]

# -----------------------------------------------------------------------------
# Runtime dependencies for the core ML library + FastAPI backend.
# Pinned hard. Floating versions across TF + transformers + torch is the most
# common source of silent BLEU drift between training runs and deployments.
#
# - tensorflow-cpu (NOT tensorflow): saves ~600 MB and removes the CUDA driver
#   dependency. We are deploying to CPU-only HuggingFace Spaces. If you ever
#   train on a GPU box, install `tensorflow==2.15.0` in that environment only.
# - 2.15.0 specifically: TF 2.16 swapped to Keras 3 by default and broke the
#   `tf.keras.layers.TextVectorization` saving behaviour the IEEE notebook
#   relies on. Stay on 2.15 for v1; upgrade is a deliberate Phase-5+ task.
# - pydantic 2.x: required by FastAPI >= 0.100. Faster and stricter than v1.
# -----------------------------------------------------------------------------
dependencies = [
    "tensorflow-cpu==2.15.0",
    "numpy>=1.26,<2.0",          # NumPy 2.0 broke TF 2.15 binary compat
    "pandas>=2.1,<3.0",
    "pillow>=10.0,<11.0",
    "pyyaml>=6.0,<7.0",
    "pydantic>=2.7,<3.0",
    "pydantic-settings>=2.3,<3.0",
    "fastapi>=0.111,<1.0",
    "uvicorn[standard]>=0.30,<1.0",
    "python-multipart>=0.0.9",   # FastAPI multipart form data (image upload)
    "huggingface-hub>=0.23,<1.0",  # Pulls weights from HF Hub at startup
    "structlog>=24.1,<25.0",     # Structured JSON logs in prod, pretty in dev
    "anyio>=4.3,<5.0",           # Thread-pool offload for sync TF inference
    "tqdm>=4.66,<5.0",
    "click>=8.1,<9.0",           # CLI for scripts/
]

# -----------------------------------------------------------------------------
# Optional dependency groups — installed via `pip install ".[dev,eval]"`.
# Splitting these keeps the production Docker image small (Phase 1 backend
# image is ~1.1 GB; adding `hf` extras takes it to ~2.3 GB which is the
# Phase 3 comparison image).
# -----------------------------------------------------------------------------
[project.optional-dependencies]

# Tier-1 multimodal upgrade: BLIP, ViT-GPT2, GIT models from HuggingFace.
# torch CPU is large (~700 MB); only install when serving the comparison demo.
hf = [
    "transformers==4.41.2",
    "torch==2.3.0",
    "sentencepiece>=0.2.0",
    "accelerate>=0.30,<1.0",
]

# Evaluation metrics. Pulled separately because pycocoevalcap drags Java
# dependencies (METEOR), which we don't want in the serving image.
eval = [
    "sacrebleu>=2.4,<3.0",
    "nltk>=3.8,<4.0",
    "rouge-score>=0.1.2",
    "pycocoevalcap>=1.2",
    "matplotlib>=3.8,<4.0",
]

# Experiment tracking. Local SQLite by default; points at DagsHub in prod.
mlflow = [
    "mlflow>=2.13,<3.0",
]

# Developer tooling: lint, type-check, test. Never deployed.
dev = [
    "ruff>=0.5,<1.0",
    "mypy>=1.10,<2.0",
    "pytest>=8.2,<9.0",
    "pytest-cov>=5.0,<6.0",
    "pytest-asyncio>=0.23,<1.0",
    "httpx>=0.27,<1.0",          # FastAPI TestClient backend
    "pre-commit>=3.7,<4.0",
    "nbstripout>=0.7,<1.0",
    "types-PyYAML",
    "types-requests",
    "pandas-stubs>=2.2,<3.0",
]

# -----------------------------------------------------------------------------
# Where pip should install the package from (the `src/` layout).
# -----------------------------------------------------------------------------
[tool.setuptools.packages.find]
where = ["src"]
include = ["captioning*"]

[tool.setuptools.package-data]
"captioning" = ["py.typed"]      # PEP 561: ship type hints with the package

# =============================================================================
# Tooling configuration — co-located so a single file owns project policy.
# =============================================================================

# ---- Ruff: linter + formatter (replaces black + isort + flake8) -------------
# We prefer Ruff because it runs ~100x faster and is the de-facto modern
# default in the Python ecosystem. One tool, one config, one cache.
[tool.ruff]
line-length = 100
target-version = "py310"
src = ["src", "backend", "scripts", "tests"]
extend-exclude = [
    "notebooks",                 # Notebooks have their own conventions
    "outputs",
    "mlruns",
    "frontend",
]

[tool.ruff.lint]
# Curated rule set — pragmatic defaults, not the full strict catalogue.
select = [
    "E",   # pycodestyle errors
    "W",   # pycodestyle warnings
    "F",   # pyflakes
    "I",   # isort import sorting
    "B",   # flake8-bugbear (likely bugs)
    "UP",  # pyupgrade (modern syntax)
    "SIM", # flake8-simplify
    "RET", # flake8-return
    "PTH", # flake8-use-pathlib (prefer pathlib over os.path)
    "RUF", # Ruff's own rules
]
ignore = [
    "E501",  # line length — formatter handles it; lint warnings are noise
    "B008",  # function call in default arg (FastAPI's Depends() pattern)
]

[tool.ruff.lint.per-file-ignores]
"tests/**" = ["B011"]            # asserts in tests are fine
"scripts/**" = ["T201"]          # print() in CLI scripts is fine

[tool.ruff.format]
quote-style = "double"
indent-style = "space"
docstring-code-format = true

# ---- Mypy: static type checker -----------------------------------------------
# We only enforce types on our own code; third-party untyped libs are tolerated.
[tool.mypy]
python_version = "3.10"
strict = false                   # Start lenient; tighten as types stabilise
warn_unused_configs = true
warn_redundant_casts = true
warn_unused_ignores = true
warn_no_return = true
no_implicit_optional = true
files = ["src/captioning", "backend/app", "scripts"]
mypy_path = ["src", "backend"]
explicit_package_bases = true
namespace_packages = true

[[tool.mypy.overrides]]
module = [
    "tensorflow.*",
    "transformers.*",
    "huggingface_hub.*",
    "PIL.*",
    "nltk.*",
    "sacrebleu.*",
    "rouge_score.*",
    "pycocoevalcap.*",
]
ignore_missing_imports = true

# ---- Pytest -------------------------------------------------------------------
[tool.pytest.ini_options]
minversion = "8.0"
testpaths = ["tests", "backend/app/tests"]
pythonpath = ["backend"]         # Lets `from app.* import ...` resolve in tests
addopts = [
    "-ra",                       # Show short summary for non-passing tests
    "--strict-markers",
    "--strict-config",
    "--showlocals",
]
markers = [
    "slow: tests that take >10 seconds (run with -m slow)",
    "gpu: tests requiring a GPU (skipped in CI by default)",
]
filterwarnings = [
    "ignore::DeprecationWarning:tensorflow.*",
    "ignore::FutureWarning:tensorflow.*",
]

# ---- Coverage -----------------------------------------------------------------
[tool.coverage.run]
branch = true
source = ["src/captioning", "backend/app"]
omit = ["*/tests/*", "*/__init__.py"]

[tool.coverage.report]
exclude_lines = [
    "pragma: no cover",
    "raise NotImplementedError",
    "if TYPE_CHECKING:",
    "if __name__ == .__main__.:",
]
show_missing = true
skip_covered = false