code/methods/base.py

"""Method base contract for the MacroLens unified API.

Every method class implements this contract:

    class ConcreteMethod(Method):
        name = "..."
        family = "..."
        tasks = frozenset({...})

        def __init__(self, *, task: str, config: MethodConfig | None = None, **kwargs): ...
        def fit(self, X, y, *, seed: int = 42) -> "Method": ...
        def predict(self, X) -> np.ndarray | pd.DataFrame: ...
        def save(self, path: pathlib.Path) -> None: ...

        @classmethod
        def load(cls, path: pathlib.Path) -> "Method": ...

        @classmethod
        def default_config(cls) -> MethodConfig: ...

        def hyperparams(self) -> dict[str, Any]: ...   # actual config (post-init)
        def lib_versions(self) -> dict[str, str]: ...

Hard rules (enforced by ``tests/test_layer_isolation.py`` and
``tests/test_method_contract.py``):

* Methods do NOT call ``pd.read_parquet`` or any file IO.
* Methods do NOT import the eval module.
* Methods do NOT subsample, filter, or join with canonical indices.
* Methods do NOT consume ``meta`` -- they take only ``X`` (and at fit
  time, ``y``).
* ``predict`` return shape is task-determined per the
  ``method-task coverage matrix`` in the plan; the runner asserts shape.

Per-family serialisation is handled by mixins
(``_JoblibSaveMixin``, ``_TorchSaveMixin``, ``_HFSaveMixin``) defined in
this module. Concrete methods inherit ONE of them.
"""

from __future__ import annotations

import abc
import importlib.metadata
import json
import pathlib
from typing import TYPE_CHECKING, Any, ClassVar

import numpy as np
import pandas as pd

if TYPE_CHECKING:
    from ._config import MethodConfig


# ── Public Method ABC ──────────────────────────────────────────────────────


class Method(abc.ABC):
    """Abstract base class for every MacroLens method.

    Subclasses MUST set the four class-level attributes below and implement
    :meth:`fit`, :meth:`predict`, :meth:`save`, :meth:`load`, and
    :meth:`default_config`. The :meth:`hyperparams` and :meth:`lib_versions`
    methods have sensible defaults inherited from the appropriate save-mixin.
    """

    name: ClassVar[str]
    family: ClassVar[str]
    tasks: ClassVar[frozenset[str]]
    schema_version: ClassVar[int] = 1

    # populated by __init__ in concrete subclasses
    task: str
    config: "MethodConfig"

    @abc.abstractmethod
    def fit(self, X: Any, y: Any, *, seed: int = 42) -> "Method":
        """Fit the method on (X, y). Returns self for chaining."""

    @abc.abstractmethod
    def predict(self, X: Any) -> np.ndarray | pd.DataFrame:
        """Emit predictions for X. Shape determined by self.task."""

    @abc.abstractmethod
    def save(self, path: pathlib.Path) -> None:
        """Persist fitted state to ``path`` (a directory). Writes
        ``manifest.json`` plus family-specific artefacts.
        """

    @classmethod
    @abc.abstractmethod
    def load(cls, path: pathlib.Path) -> "Method":
        """Reconstruct a fitted method from ``path``. Raises if the
        manifest's ``schema_version`` is incompatible.
        """

    @classmethod
    @abc.abstractmethod
    def default_config(cls) -> "MethodConfig":
        """Return the default ``MethodConfig`` for this class."""

    # ── default helpers (override only if necessary) ──

    def hyperparams(self) -> dict[str, Any]:
        """Return the actual config used (post-init)."""
        return self.config.model_dump()

    def lib_versions(self) -> dict[str, str]:
        """Return a dict of {package: version} for libraries this method
        uses. Default returns the versions of numpy / pandas / scikit-learn /
        torch / transformers (whichever are installed). Subclasses may
        override to add or restrict.
        """
        pkgs = ["numpy", "pandas", "scikit-learn", "torch", "transformers",
                "lightgbm", "vllm", "peft", "chronos-forecasting", "uni2ts",
                "timesfm"]
        out: dict[str, str] = {}
        for p in pkgs:
            try:
                out[p] = importlib.metadata.version(p)
            except importlib.metadata.PackageNotFoundError:
                pass
        return out

    # ── manifest helpers used by save/load ──

    def _manifest(self) -> dict[str, Any]:
        """Return the manifest dict written by ``save``."""
        return {
            "name": self.name,
            "family": self.family,
            "tasks": sorted(self.tasks),
            "schema_version": self.schema_version,
            "task": self.task,
            "hyperparams": self.hyperparams(),
            "lib_versions": self.lib_versions(),
        }

    @staticmethod
    def _check_manifest(path: pathlib.Path, expected_name: str,
                         expected_schema: int) -> dict[str, Any]:
        """Validate ``manifest.json`` at ``path`` and return its contents."""
        m_path = pathlib.Path(path) / "manifest.json"
        if not m_path.exists():
            raise FileNotFoundError(f"manifest.json missing at {m_path}")
        manifest = json.loads(m_path.read_text())
        if manifest.get("name") != expected_name:
            raise ValueError(
                f"checkpoint name mismatch: expected {expected_name}, "
                f"found {manifest.get('name')}"
            )
        if manifest.get("schema_version") != expected_schema:
            raise ValueError(
                f"schema_version mismatch: expected {expected_schema}, "
                f"found {manifest.get('schema_version')} -- run "
                f"tools/migrate_results.py if migrating from v{manifest.get('schema_version')}"
            )
        return manifest


# ── Per-family save/load mixins ───────────────────────────────────────────


class _JoblibSaveMixin:
    """Serialise via ``joblib.dump``/``joblib.load``. For naive +
    classical methods whose state is small (medians, regression coeffs,
    LightGBM Booster).
    """

    def save(self: "Method", path: pathlib.Path) -> None:  # type: ignore[misc]
        import joblib

        path = pathlib.Path(path)
        path.mkdir(parents=True, exist_ok=True)
        (path / "manifest.json").write_text(json.dumps(self._manifest(), indent=2))
        joblib.dump(self.__dict__, path / "state.joblib")

    @classmethod
    def load(cls: type["Method"], path: pathlib.Path) -> "Method":  # type: ignore[misc]
        import joblib

        path = pathlib.Path(path)
        manifest = Method._check_manifest(path, cls.name, cls.schema_version)
        state = joblib.load(path / "state.joblib")
        instance = cls.__new__(cls)
        instance.__dict__.update(state)
        return instance


class _TorchSaveMixin:
    """Serialise PyTorch state_dict + config + scaler. For sequence
    methods (DLinear, ITransformer, ModernTCN).
    """

    def save(self: "Method", path: pathlib.Path) -> None:  # type: ignore[misc]
        import torch

        path = pathlib.Path(path)
        path.mkdir(parents=True, exist_ok=True)
        (path / "manifest.json").write_text(json.dumps(self._manifest(), indent=2))
        # Subclasses set self._model (nn.Module), self._scaler, self._aux
        payload = {
            "state_dict": self._model.state_dict() if hasattr(self, "_model") else None,
            "config": self.config.model_dump(),
            "task": self.task,
            "aux": getattr(self, "_aux", {}),
        }
        torch.save(payload, path / "state.pt")
        if hasattr(self, "_scaler"):
            import joblib
            joblib.dump(self._scaler, path / "scaler.joblib")

    @classmethod
    def load(cls: type["Method"], path: pathlib.Path) -> "Method":  # type: ignore[misc]
        import torch

        path = pathlib.Path(path)
        manifest = Method._check_manifest(path, cls.name, cls.schema_version)
        payload = torch.load(path / "state.pt", weights_only=False)
        instance = cls(task=payload["task"], config=cls.default_config().__class__(**payload["config"]))
        if payload["state_dict"] is not None and hasattr(instance, "_model"):
            instance._model.load_state_dict(payload["state_dict"])
        if hasattr(instance, "_aux"):
            instance._aux = payload.get("aux", {})
        scaler_path = path / "scaler.joblib"
        if scaler_path.exists():
            import joblib
            instance._scaler = joblib.load(scaler_path)
        return instance


class _HFSaveMixin:
    """Serialise via HuggingFace ``save_pretrained``/``from_pretrained``.
    For TSFM/LLM methods. Concrete classes override the ``_hf_save`` and
    ``_hf_load`` hooks for adapter handling (LoRA).
    """

    def save(self: "Method", path: pathlib.Path) -> None:  # type: ignore[misc]
        path = pathlib.Path(path)
        path.mkdir(parents=True, exist_ok=True)
        (path / "manifest.json").write_text(json.dumps(self._manifest(), indent=2))
        self._hf_save(path)  # subclass hook

    @classmethod
    def load(cls: type["Method"], path: pathlib.Path) -> "Method":  # type: ignore[misc]
        path = pathlib.Path(path)
        manifest = Method._check_manifest(path, cls.name, cls.schema_version)
        instance = cls(task=manifest["task"], config=cls.default_config().__class__(**manifest["hyperparams"]))
        instance._hf_load(path)  # subclass hook
        return instance

    def _hf_save(self, path: pathlib.Path) -> None:
        raise NotImplementedError(
            "subclass must implement _hf_save (write model + tokenizer + adapter)"
        )

    def _hf_load(self, path: pathlib.Path) -> None:
        raise NotImplementedError(
            "subclass must implement _hf_load (read model + tokenizer + adapter)"
        )