generate/generate.py

from __future__ import annotations

import argparse
import hashlib
import json
import math
import random
import shutil
import struct
from dataclasses import dataclass
from datetime import datetime, timedelta, timezone
from pathlib import Path
from typing import Annotated, Any

import pyarrow as pa
import rumi
import taco
from pydantic import BaseModel, Field

ROOT = Path(__file__).resolve().parents[1]
DATA_ROOT = ROOT / "data"
MANIFEST_PATH = ROOT / "manifest.json"
CHECKSUMS_PATH = ROOT / "checksums.sha256"
LOCK_PATH = ROOT / "source-lock.json"
GENERATOR_SEED = 20260909
SPLITS = ("train", "train", "train", "validation", "test", "test")
TimestampUTC = Annotated[datetime, pa.timestamp("us", tz="UTC")]


class FixtureSample(BaseModel):
    case_id: str = Field(description="Logical fixture case identifier")
    sample_key: str = Field(description="Stable sample key inside the fixture case")
    synthetic: bool = Field(description="Whether semantic metadata is synthetic")
    seed: int = Field(description="Deterministic seed used for synthetic metadata")


class RumiAsset(BaseModel):
    header: bytes = Field(description="External Rumi header required for stateless reads")
    source_fixture: str = Field(description="Name of the pinned upstream Rumi fixture")
    data_sha256: str = Field(description="SHA-256 checksum of the Rumi container")
    header_sha256: str = Field(description="SHA-256 checksum of the external Rumi header")
    array_sha256: str = Field(description="SHA-256 checksum of the decoded logical array")
    shape: list[int] = Field(description="Logical Rumi array shape")
    dtype: str = Field(description="Logical Rumi array data type")
    frame_layout: str = Field(description="Rumi frame layout")


class GeoPoint(BaseModel):
    longitude: float = Field(ge=-180, le=180, description="Synthetic longitude in degrees")
    latitude: float = Field(ge=-90, le=90, description="Synthetic latitude in degrees")
    synthetic: bool = Field(description="Whether the point is synthetic")


class TemporalInfo(BaseModel):
    time_start: TimestampUTC = Field(description="Synthetic interval start")
    time_end: TimestampUTC | None = Field(default=None, description="Synthetic interval end")
    synthetic: bool = Field(description="Whether the temporal metadata is synthetic")


@dataclass(frozen=True)
class SourceFixture:
    name: str
    data: Path
    header: Path
    manifest: dict[str, Any]

    def metadata(self) -> RumiAsset:
        return RumiAsset(
            header=self.header.read_bytes(),
            source_fixture=self.name,
            data_sha256=self.manifest["data_sha256"],
            header_sha256=self.manifest["header_sha256"],
            array_sha256=self.manifest["array_sha256"],
            shape=self.manifest["shape"],
            dtype=self.manifest["dtype"],
            frame_layout=self.manifest["frame_layout"],
        )


@dataclass(frozen=True)
class AssetRecipe:
    path: str | None
    source: str


@dataclass(frozen=True)
class CaseDefinition:
    number: int
    slug: str
    description: str
    structure: tuple[str, ...] | None
    coordinate_profile: str
    task: str
    recipes: tuple[tuple[AssetRecipe, ...], ...]

    @property
    def case_id(self) -> str:
        return f"{self.number:02d}-{self.slug}"


def _asset(path: str | None, source: str) -> AssetRecipe:
    return AssetRecipe(path, source)


def _cases() -> tuple[CaseDefinition, ...]:
    s2 = ("s2-00-tile", "s2-01-tile", "s2-02-tile")
    single = ("s2-00-tile", "s2-01-tile", "s2-02-tile", "s1-00", "worldcover-00", "emit-00")
    secondary = ("worldcover-00", "s1-00", "alphaearth-00", "worldcover-00", "s1-00", "emit-00")
    sequence_sources = ("s2-00-tile", "s2-00-planar", "s2-00-chunky", "s2-01-tile", "s2-02-tile")
    sequence_lengths = (2, 3, 4, 5, 2, 4)
    optional_counts = (0, 1, 2, 0, 2, 1)

    return (
        CaseDefinition(
            1,
            "single-rumi",
            "Null-structure samples with no coordinate metadata.",
            None,
            "none",
            "other",
            tuple((_asset(None, source),) for source in single),
        ),
        CaseDefinition(
            2,
            "paired-rumi",
            "Fixed Rumi pairs with synthetic point coordinates.",
            ("primary.rumi", "secondary.rumi"),
            "point",
            "other",
            tuple(
                (_asset("primary.rumi", s2[index % len(s2)]), _asset("secondary.rumi", secondary[index]))
                for index in range(6)
            ),
        ),
        CaseDefinition(
            3,
            "multisensor",
            "Optical and radar-shaped assets with regular synthetic STAC grids.",
            ("optical.rumi", "radar.rumi"),
            "stac",
            "classification",
            tuple((_asset("optical.rumi", s2[index % len(s2)]), _asset("radar.rumi", "s1-00")) for index in range(6)),
        ),
        CaseDefinition(
            4,
            "change-detection",
            "Nested before/after/label-shaped assets with synthetic irregular footprints.",
            ("before/image.rumi", "after/image.rumi", "change.rumi"),
            "istac",
            "change-detection",
            tuple(
                (
                    _asset("before/image.rumi", s2[index % len(s2)]),
                    _asset("after/image.rumi", s2[(index + 1) % len(s2)]),
                    _asset("change.rumi", "worldcover-00"),
                )
                for index in range(6)
            ),
        ),
        CaseDefinition(
            5,
            "fixed-time-series",
            "Three fixed time steps with regular synthetic STAC intervals.",
            ("t0.rumi", "t1.rumi", "t2.rumi"),
            "stac-interval",
            "regression",
            tuple(
                (
                    _asset("t0.rumi", "s2-00-tile"),
                    _asset("t1.rumi", "s2-01-tile"),
                    _asset("t2.rumi", "s2-02-tile"),
                )
                for _ in range(6)
            ),
        ),
        CaseDefinition(
            6,
            "variable-sequence",
            "Variable-length Rumi sequences without coordinate metadata.",
            ("frame*[2,5].rumi",),
            "none",
            "classification",
            tuple(
                tuple(
                    _asset(f"frame{position}.rumi", sequence_sources[(sample + position) % len(sequence_sources)])
                    for position in range(length)
                )
                for sample, length in enumerate(sequence_lengths)
            ),
        ),
        CaseDefinition(
            7,
            "optional-assets",
            "A required image plus zero to two optional assets and nullable synthetic points.",
            ("image.rumi", "aux*[0,2].rumi"),
            "optional-point",
            "other",
            tuple(
                (
                    _asset("image.rumi", s2[index % len(s2)]),
                    *(
                        _asset(f"aux{position}.rumi", ("s1-00", "worldcover-00")[position])
                        for position in range(optional_counts[index])
                    ),
                )
                for index in range(6)
            ),
        ),
        CaseDefinition(
            8,
            "nested-fusion",
            "Nested input and target folders carrying synthetic folder-level STAC metadata.",
            ("inputs/optical.rumi", "inputs/radar.rumi", "targets/landcover.rumi"),
            "folder-stac",
            "segmentation",
            tuple(
                (
                    _asset("inputs/optical.rumi", s2[index % len(s2)]),
                    _asset("inputs/radar.rumi", "s1-00"),
                    _asset("targets/landcover.rumi", "worldcover-00"),
                )
                for index in range(6)
            ),
        ),
        CaseDefinition(
            9,
            "high-dimensional",
            "Single named cubes with synthetic temporal-only metadata.",
            ("cube.rumi",),
            "temporal-only",
            "other",
            tuple(
                (_asset("cube.rumi", source),)
                for source in ("era5-t2m-00-time", "alphaearth-00", "emit-00", "worldcover-00", "s1-00", "s2-00-tile")
            ),
        ),
        CaseDefinition(
            10,
            "layout-equivalence",
            "Equivalent Rumi tile, planar, and chunky encodings sharing one synthetic STAC grid.",
            ("scene.rumi",),
            "shared-stac",
            "other",
            tuple(
                (_asset("scene.rumi", source),)
                for source in (
                    "s2-00-tile",
                    "s2-00-planar",
                    "s2-00-chunky",
                    "s2-00-tile",
                    "s2-00-planar",
                    "s2-00-chunky",
                )
            ),
        ),
    )


def _digest(path: Path) -> str:
    value = hashlib.sha256()
    with path.open("rb") as stream:
        for chunk in iter(lambda: stream.read(1024 * 1024), b""):
            value.update(chunk)
    return value.hexdigest()


def _seed(case_id: str, sample_index: int, salt: str = "sample") -> int:
    value = f"{GENERATOR_SEED}:{case_id}:{sample_index}:{salt}".encode()
    return int.from_bytes(hashlib.sha256(value).digest()[:8], "big") % (2**63 - 1)


def _point(longitude: float, latitude: float) -> bytes:
    return struct.pack("<BIdd", 1, 1, longitude, latitude)


def _polygon(west: float, south: float, east: float, north: float) -> bytes:
    ring = ((west, south), (east, south), (east, north), (west, north), (west, south))
    return struct.pack("<BIII", 1, 3, 1, len(ring)) + b"".join(struct.pack("<dd", x, y) for x, y in ring)


def _mercator(longitude: float, latitude: float) -> tuple[float, float]:
    radius = 6_378_137.0
    x = radius * math.radians(longitude)
    y = radius * math.log(math.tan(math.pi / 4 + math.radians(latitude) / 2))
    return x, y


def _spatial_values(case_id: str, sample_index: int, salt: str = "sample") -> dict[str, Any]:
    rng = random.Random(_seed(case_id, sample_index, salt))
    longitude = rng.uniform(-165.0, 165.0)
    latitude = rng.uniform(-70.0, 70.0)
    crs = "EPSG:4326" if rng.randrange(2) == 0 else "EPSG:3857"
    if crs == "EPSG:4326":
        center_x, center_y = longitude, latitude
        pixel_size = rng.uniform(0.0001, 0.005)
    else:
        center_x, center_y = _mercator(longitude, latitude)
        pixel_size = rng.uniform(5.0, 100.0)
    start = datetime(2020, 1, 1, tzinfo=timezone.utc) + timedelta(
        days=rng.randrange(0, 365 * 6), hours=rng.randrange(0, 24)
    )
    return {
        "longitude": longitude,
        "latitude": latitude,
        "crs": crs,
        "center_x": center_x,
        "center_y": center_y,
        "pixel_size": pixel_size,
        "time_start": start,
    }


def _stac(
    case_id: str,
    sample_index: int,
    shape: list[int],
    *,
    interval: bool = False,
    shared: bool = False,
    folder: bool = False,
    salt: str = "sample",
) -> taco.metadata.sample.STAC:
    spatial_index = 0 if shared else sample_index
    values = _spatial_values(case_id, spatial_index, salt)
    height, width = shape[-2:]
    west = values["center_x"] - width * values["pixel_size"] / 2
    north = values["center_y"] + height * values["pixel_size"] / 2
    model = taco.metadata.folder.STAC if folder else taco.metadata.sample.STAC
    kwargs: dict[str, Any] = {}
    if interval:
        kwargs["time_end"] = values["time_start"] + timedelta(hours=6 + sample_index)
    return model(
        crs=values["crs"],
        tensor_shape=shape,
        geotransform=(west, values["pixel_size"], 0.0, north, 0.0, -values["pixel_size"]),
        time_start=values["time_start"],
        centroid=_point(values["longitude"], values["latitude"]),
        **kwargs,
    )


def _istac(case_id: str, sample_index: int) -> taco.metadata.sample.ISTAC:
    values = _spatial_values(case_id, sample_index)
    radius = 0.1 if values["crs"] == "EPSG:4326" else 10_000.0
    return taco.metadata.sample.ISTAC(
        crs=values["crs"],
        geometry=_polygon(
            values["center_x"] - radius,
            values["center_y"] - radius,
            values["center_x"] + radius,
            values["center_y"] + radius,
        ),
        centroid=_point(values["longitude"], values["latitude"]),
        time_start=values["time_start"],
        time_end=values["time_start"] + timedelta(days=1 + sample_index),
    )


def _temporal(case_id: str, sample_index: int) -> TemporalInfo:
    values = _spatial_values(case_id, sample_index)
    return TemporalInfo(
        time_start=values["time_start"],
        time_end=values["time_start"] + timedelta(hours=sample_index + 1),
        synthetic=True,
    )


def _sample_groups(case: CaseDefinition) -> dict[str, Any]:
    groups: dict[str, Any] = {
        "fixture": FixtureSample,
        "ml": taco.metadata.sample.Split,
    }
    if case.structure is None:
        groups["rumi"] = RumiAsset
    if case.coordinate_profile == "point":
        groups["geo"] = GeoPoint
    elif case.coordinate_profile in {"stac", "stac-interval", "shared-stac"}:
        groups["stac"] = taco.metadata.sample.STAC
    elif case.coordinate_profile == "istac":
        groups["istac"] = taco.metadata.sample.ISTAC
    elif case.coordinate_profile == "optional-point":
        groups["geo"] = GeoPoint | None
    elif case.coordinate_profile == "temporal-only":
        groups["temporal"] = TemporalInfo
    return groups


def _contract(case: CaseDefinition) -> taco.Contract:
    levels = [taco.Level("sample", **_sample_groups(case))]
    if case.structure is not None:
        if case.number == 4:
            levels.extend(
                (
                    taco.Level("children", rumi=RumiAsset | None),
                    taco.Level("children/before", rumi=RumiAsset),
                    taco.Level("children/after", rumi=RumiAsset),
                )
            )
        elif case.number == 8:
            levels.extend(
                (
                    taco.Level("children", stac=taco.metadata.folder.STAC),
                    taco.Level("children/inputs", rumi=RumiAsset),
                    taco.Level("children/targets", rumi=RumiAsset),
                )
            )
        else:
            levels.append(taco.Level("children", rumi=RumiAsset))
    return taco.Contract(structure=case.structure, metadata=taco.MetadataSchema(*levels))


def _collection(case: CaseDefinition) -> taco.Collection:
    return taco.Collection(
        contract=_contract(case),
        id=case.case_id,
        dataset_version="1.0.0",
        title=f"TACO fixture: {case.slug}",
        description=case.description + " Payload contents and semantic metadata are test-only.",
        licenses=["other"],
        providers=["Asterisk Labs"],
        tasks=[case.task],
    )


def _metadata(
    case: CaseDefinition,
    sample_index: int,
    sources: dict[str, SourceFixture],
) -> taco.Metadata:
    recipe = case.recipes[sample_index]
    seed = _seed(case.case_id, sample_index)
    groups: dict[str, BaseModel] = {
        "fixture": FixtureSample(
            case_id=case.case_id,
            sample_key=f"sample-{sample_index}",
            synthetic=True,
            seed=seed,
        ),
        "ml": taco.metadata.sample.Split(split=SPLITS[sample_index]),
    }
    first = sources[recipe[0].source]
    if case.structure is None:
        groups["rumi"] = first.metadata()
    if case.coordinate_profile == "point":
        values = _spatial_values(case.case_id, sample_index)
        groups["geo"] = GeoPoint(longitude=values["longitude"], latitude=values["latitude"], synthetic=True)
    elif case.coordinate_profile == "stac":
        groups["stac"] = _stac(case.case_id, sample_index, first.manifest["shape"])
    elif case.coordinate_profile == "stac-interval":
        groups["stac"] = _stac(case.case_id, sample_index, first.manifest["shape"], interval=True)
    elif case.coordinate_profile == "istac":
        groups["istac"] = _istac(case.case_id, sample_index)
    elif case.coordinate_profile == "optional-point" and sample_index % 3 != 2:
        values = _spatial_values(case.case_id, sample_index)
        groups["geo"] = GeoPoint(longitude=values["longitude"], latitude=values["latitude"], synthetic=True)
    elif case.coordinate_profile == "temporal-only":
        groups["temporal"] = _temporal(case.case_id, sample_index)
    elif case.coordinate_profile == "shared-stac":
        groups["stac"] = _stac(case.case_id, sample_index, first.manifest["shape"], shared=True)
    return taco.Metadata(**groups)


def _samples(case: CaseDefinition, sources: dict[str, SourceFixture]) -> tuple[taco.Sample, ...]:
    samples = []
    for sample_index, recipes in enumerate(case.recipes):
        assets = [
            taco.Asset(
                sources[recipe.source].data,
                path=recipe.path,
                metadata=None if case.structure is None else taco.Metadata(rumi=sources[recipe.source].metadata()),
            )
            for recipe in recipes
        ]
        folders: list[taco.Folder] = []
        if case.number == 8:
            first_shape = sources[recipes[0].source].manifest["shape"]
            folders = [
                taco.Folder(
                    path,
                    metadata=taco.Metadata(
                        stac=_stac(
                            case.case_id,
                            sample_index,
                            first_shape,
                            folder=True,
                            salt=f"folder:{path}",
                        )
                    ),
                )
                for path in ("inputs", "targets")
            ]
        samples.append(
            taco.Sample(
                assets=assets,
                metadata=_metadata(case, sample_index, sources),
                folders=folders,
            )
        )
    return tuple(samples)


def _load_sources(source_root: Path) -> tuple[dict[str, SourceFixture], dict[str, Any]]:
    lock = json.loads(LOCK_PATH.read_text(encoding="utf-8"))
    manifest_path = source_root / "manifest.json"
    if not manifest_path.is_file():
        raise FileNotFoundError(f"Rumi fixture manifest not found: {manifest_path}")
    actual_manifest_hash = _digest(manifest_path)
    if actual_manifest_hash != lock["manifest_sha256"]:
        raise RuntimeError(
            f"Rumi manifest checksum mismatch: {actual_manifest_hash}; expected {lock['manifest_sha256']}"
        )
    manifest = json.loads(manifest_path.read_text(encoding="utf-8"))
    sources = {}
    for item in manifest["fixtures"]:
        data = source_root / item["data"]
        header = source_root / item["header"]
        if _digest(data) != item["data_sha256"]:
            raise RuntimeError(f"Rumi data checksum mismatch: {data}")
        if _digest(header) != item["header_sha256"]:
            raise RuntimeError(f"Rumi header checksum mismatch: {header}")
        sources[item["name"]] = SourceFixture(item["name"], data, header, item)
    return sources, lock


def _write(
    collection: taco.Collection,
    output: Path,
    samples: tuple[taco.Sample, ...],
    **options: Any,
) -> taco.writer.BuildResult:
    with taco.open_writer(collection, output, **options) as writer:
        writer.extend(samples)
        return writer.run()


def _relative(path: Path) -> str:
    return path.resolve().relative_to(ROOT).as_posix()


def _dataset_record(
    case: CaseDefinition,
    topology: str,
    entrypoint: Path,
    parts: tuple[Path, ...] = (),
) -> dict[str, Any]:
    return {
        "id": f"{case.case_id}/{topology}",
        "case": case.case_id,
        "topology": topology,
        "container": "folder" if topology == "folder" else ("zip" if topology == "single-zip" else "tacocat"),
        "path": _relative(entrypoint),
        "parts": [_relative(path) for path in parts],
        "samples": len(case.recipes),
        "assets": sum(len(recipe) for recipe in case.recipes),
    }


def _case_record(case: CaseDefinition) -> dict[str, Any]:
    return {
        "id": case.case_id,
        "description": case.description,
        "structure": None if case.structure is None else list(case.structure),
        "coordinate_profile": case.coordinate_profile,
        "task": case.task,
        "samples": [
            {
                "sample_key": f"sample-{index}",
                "split": SPLITS[index],
                "seed": _seed(case.case_id, index),
                "assets": [{"path": asset.path, "source_fixture": asset.source} for asset in assets],
            }
            for index, assets in enumerate(case.recipes)
        ],
    }


def _build_case(case: CaseDefinition, sources: dict[str, SourceFixture]) -> list[dict[str, Any]]:
    collection = _collection(case)
    samples = _samples(case, sources)
    root = DATA_ROOT / case.case_id
    records = []

    folder = _write(collection, root / "folder", samples)
    records.append(_dataset_record(case, "folder", folder.path))

    single = _write(collection, root / "single-zip" / "dataset.zip", samples)
    records.append(_dataset_record(case, "single-zip", single.path))

    by_size = _write(collection, root / "by-size" / "dataset.zip", samples, partition_size=1)
    records.append(_dataset_record(case, "by-size", by_size.path, by_size.parts))

    by_split = _write(collection, root / "by-split" / "dataset.zip", samples, partition_by="ml:split")
    records.append(_dataset_record(case, "by-split", by_split.path, by_split.parts))

    manual_root = root / "manual-catalog"
    manual_parts = []
    for label, subset in zip(("one", "two", "three"), (samples[:1], samples[1:3], samples[3:]), strict=True):
        result = _write(collection, manual_root / f"part-{label}.zip", subset)
        manual_parts.append(result.path)
    catalog = taco.consolidate(manual_parts, output=manual_root)
    records.append(_dataset_record(case, "manual-catalog", catalog, tuple(manual_parts)))
    return records


def _write_checksums() -> None:
    files = sorted(path for path in DATA_ROOT.rglob("*") if path.is_file())
    files.append(MANIFEST_PATH)
    lines = [f"{_digest(path)}  {_relative(path)}" for path in files]
    CHECKSUMS_PATH.write_text("\n".join(lines) + "\n", encoding="utf-8")


def build(source_root: Path, *, clean: bool) -> dict[str, Any]:
    source_root = source_root.expanduser().resolve()
    if clean and DATA_ROOT.exists():
        if DATA_ROOT.parent != ROOT or DATA_ROOT.name != "data" or not LOCK_PATH.is_file():
            raise RuntimeError(f"refusing to clean unsafe output path: {DATA_ROOT}")
        shutil.rmtree(DATA_ROOT)
    if DATA_ROOT.exists() and any(DATA_ROOT.iterdir()):
        raise FileExistsError(f"output is not empty; rerun with --clean: {DATA_ROOT}")
    DATA_ROOT.mkdir(parents=True, exist_ok=True)

    sources, lock = _load_sources(source_root)
    cases = _cases()
    selected = {asset.source for case in cases for sample in case.recipes for asset in sample}
    missing = sorted(selected - set(sources))
    if missing:
        raise RuntimeError(f"Rumi source fixtures are missing: {missing}")

    datasets = []
    for case in cases:
        print(f"building {case.case_id}", flush=True)
        datasets.extend(_build_case(case, sources))

    manifest = {
        "schema_version": 1,
        "generator_seed": GENERATOR_SEED,
        "taco_version": taco.__version__,
        "rumi_version": rumi.__version__,
        "source": lock,
        "logical_cases": [_case_record(case) for case in cases],
        "datasets": datasets,
        "source_fixtures": {
            name: {
                key: fixture.manifest[key]
                for key in ("data_sha256", "header_sha256", "array_sha256", "shape", "dtype", "frame_layout")
            }
            for name, fixture in sorted(sources.items())
            if name in selected
        },
    }
    if len(datasets) != 50:
        raise AssertionError(f"expected 50 datasets, built {len(datasets)}")
    MANIFEST_PATH.write_text(json.dumps(manifest, indent=2, ensure_ascii=False) + "\n", encoding="utf-8")
    _write_checksums()
    return manifest


def main() -> None:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument(
        "--source",
        type=Path,
        default=ROOT.parent / "rumi-api-fixtures",
        help="checkout of the pinned rumi-api-fixtures dataset",
    )
    parser.add_argument("--clean", action="store_true", help="replace the generated data directory")
    args = parser.parse_args()
    manifest = build(args.source, clean=args.clean)
    print(f"built {len(manifest['datasets'])} dataset entrypoints in {DATA_ROOT}")


if __name__ == "__main__":
    main()