pyproject.toml

[project]
name = "varunanet"
version = "0.1.0"
description = "SAR-based flood detection and situational awareness system"
requires-python = ">=3.10"
# Real dependencies get added phase by phase, as the code that actually
# needs them gets written -- not all at once up front. numpy is here first
# because the data contract module operates on numpy arrays; rasterio reads
# the GeoTIFF chips the Sen1Floods11 dataset ships as.
dependencies = [
    "numpy",
    "rasterio",
    "pysheds",
    "affine",
    "matplotlib",
    "scikit-image",
    "scikit-learn",
    # Phase 3 (primary CNN): torch/torchvision are the training backbone,
    # and segmentation_models_pytorch supplies ready-made encoder/decoder
    # architectures (U-Net, U-Net++, DeepLabV3+) with ImageNet-pretrained
    # weights, so the project spends effort on the experimental design
    # instead of re-deriving U-Net from scratch.
    "torch",
    "torchvision",
    "segmentation-models-pytorch",
    # Phase 4 (architecture study, spec section 4.3): SegFormer is the
    # transformer comparison against the CNN family above -- HuggingFace
    # ships it pretrained, segmentation-models-pytorch doesn't.
    "transformers",
    # Config-driven training loop (spec section 8). omegaconf comes in as
    # hydra-core's own dependency, not listed separately.
    "hydra-core",
    # Experiment tracking, offline mode by default -- SLURM compute nodes
    # are often air-gapped (spec section 8).
    "wandb",
    # Gradient-boosted tree baselines (benchmarks/ml_ensemble.py), alongside
    # the Random Forest already covered by scikit-learn above. On macOS these
    # need the OpenMP runtime, which isn't a Python dependency: `brew install
    # libomp` (see infra/Dockerfile.train for the Linux equivalent).
    "xgboost",
    "lightgbm",
    # Phase 5 (serving pipeline, spec section 5): the trained model is
    # exported to ONNX and served through ONNX Runtime so the inference
    # service never needs torch. That is the whole point of the split
    # between infra/Dockerfile.train (CUDA, heavy) and Dockerfile.serve
    # (slim) -- onnxruntime is a ~50MB wheel against torch's ~2GB tree.
    # onnxscript is not optional: torch's current exporter imports it, and
    # without it torch.onnx.export fails with ModuleNotFoundError rather
    # than anything that points at the real cause.
    "onnx",
    "onnxruntime",
    "onnxscript",
    # Vectorizing the flood mask to polygons for PostGIS
    # (inference/vectorize.py). shapely does the geometry; pyproj computes
    # true ellipsoidal areas, which is not optional -- Sen1Floods11 chips
    # are EPSG:4326, so a planar area on them is in square degrees and a
    # fully flooded scene reports zero hectares. pyproj arrives as a
    # rasterio dependency but is listed explicitly because this project
    # imports it directly.
    "shapely",
    "pyproj",
    # The inference service itself (inference/service.py). uvicorn is the
    # ASGI server the container runs; httpx backs fastapi's TestClient, so
    # the HTTP surface can be tested without binding a port.
    "fastapi",
    "uvicorn[standard]",
    "httpx",
    # Real Assam Sentinel-1 ingestion (spec section 15.4, Track B Step 3):
    # data/fetch_assam_scene.py and data/build_assam_demo.py pull real
    # scenes via Google Earth Engine and download them over plain HTTP.
    "earthengine-api",
    "requests",
]

[project.optional-dependencies]
# Tools for local development only; not needed to run the project itself.
dev = [
    "ruff",
    "black",
    "pytest",
    "jupyter",
    "nbconvert",
    "ipykernel",
]

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

[tool.setuptools]
# Only these folders contain importable Python code. Without this list,
# setuptools would also try to package non-Python folders like frontend/.
packages = ["data", "models", "training", "benchmarks", "inference"]

[tool.black]
line-length = 100
target-version = ["py310"]

[tool.ruff]
line-length = 100
# training/kaggle_train.ipynb runs on Kaggle, not in this repo's own test
# suite (see VarunaNet_Spec.md's environment notes: black already skips
# .ipynb files the same way, since black[jupyter] isn't installed) --
# lines like a real "!pip install ..." cell or an f-string printing several
# metrics at once read better long and un-wrapped than split just to satisfy
# a line-length rule meant for this project's actual source files.
extend-exclude = ["training/kaggle_train.ipynb"]

[tool.ruff.lint]
# Default rule set plus import-sorting, so `ruff check` also catches
# unsorted/unused imports instead of needing a separate isort run.
select = ["E", "F", "I"]

[tool.pytest.ini_options]
testpaths = ["tests"]