"""Leak-free preprocessing pipeline (Phase 3). The saved pipeline contains three steps: 1. ``FeatureEngineer`` - stateless row-local feature engineering: calendar features from ``datetime`` and sine/cosine encoding of ``wind_direction``. No statistic is learned, so it cannot leak. 2. ``ColumnTransformer`` - median imputation + standardisation for numeric columns; mode imputation + one-hot encoding for categorical columns. 3. The estimator (added by :func:`build_pipeline`). Every *learned* transformation is fitted on the training split only and is re-fitted inside each cross-validation fold when the whole pipeline is passed to ``GridSearchCV`` / ``RandomizedSearchCV``. Never pass pre-transformed arrays to a CV search - pass the raw frame (``INPUT_COLUMNS``) and this pipeline. Lag/rolling columns are deterministic, causal history features: they only look backwards, use no fitted parameter, and are built identically at training and inference time by ``src.features.build_features`` / ``src.features.inference``. """ from __future__ import annotations import sys from pathlib import Path import numpy as np import pandas as pd from sklearn.base import BaseEstimator, TransformerMixin from sklearn.compose import ColumnTransformer from sklearn.impute import SimpleImputer from sklearn.pipeline import Pipeline from sklearn.preprocessing import OneHotEncoder, StandardScaler if __package__ in (None, ""): sys.path.insert(0, str(Path(__file__).resolve().parents[2])) from src.utils.config import ( # noqa: E402 CATEGORICAL_FEATURES, INPUT_COLUMNS, NUMERIC_FEATURES, TARGET, TIME_COL, ) class FeatureEngineer(BaseEstimator, TransformerMixin): """Stateless, row-local feature engineering. Input : DataFrame with ``datetime`` and ``wind_direction``. Output: the same frame plus ``hour``, ``day_of_week``, ``month``, ``is_weekend``, ``wind_dir_sin``, ``wind_dir_cos``. """ def fit(self, X, y=None): # noqa: N803 - sklearn signature return self def transform(self, X): # noqa: N803 - sklearn signature if not isinstance(X, pd.DataFrame): raise TypeError("FeatureEngineer expects a pandas DataFrame") missing = [c for c in (TIME_COL, "wind_direction") if c not in X.columns] if missing: raise ValueError(f"FeatureEngineer missing columns: {missing}") df = X.copy() dt = pd.to_datetime(df[TIME_COL]) df["hour"] = dt.dt.hour.astype("int8") df["day_of_week"] = dt.dt.dayofweek.astype("int8") df["month"] = dt.dt.month.astype("int8") df["is_weekend"] = (dt.dt.dayofweek >= 5).astype("int8") radians = np.deg2rad(df["wind_direction"].astype(float)) df["wind_dir_sin"] = np.sin(radians) df["wind_dir_cos"] = np.cos(radians) return df def get_feature_names_out(self, input_features=None): return np.asarray(INPUT_COLUMNS, dtype=object) def build_preprocessor( numeric_features: list[str] | None = None, categorical_features: list[str] | None = None, ) -> ColumnTransformer: """Numeric: median impute -> standardise; categorical: mode -> one-hot.""" numeric_features = list(numeric_features or NUMERIC_FEATURES) categorical_features = list(categorical_features or CATEGORICAL_FEATURES) numeric_pipe = Pipeline( steps=[ ("imputer", SimpleImputer(strategy="median")), ("scaler", StandardScaler()), ] ) categorical_pipe = Pipeline( steps=[ ("imputer", SimpleImputer(strategy="most_frequent")), ( "onehot", OneHotEncoder(handle_unknown="ignore", sparse_output=False), ), ] ) return ColumnTransformer( transformers=[ ("num", numeric_pipe, numeric_features), ("cat", categorical_pipe, categorical_features), ], remainder="drop", verbose_feature_names_out=False, ) def build_feature_pipeline( numeric_features: list[str] | None = None, categorical_features: list[str] | None = None, ) -> Pipeline: """Feature engineering + imputation/scaling/encoding (no estimator).""" return Pipeline( steps=[ ("features", FeatureEngineer()), ( "preprocess", build_preprocessor(numeric_features, categorical_features), ), ] ) def build_pipeline( estimator, numeric_features: list[str] | None = None, categorical_features: list[str] | None = None, ) -> Pipeline: """Full deployable pipeline: raw inputs -> features -> transforms -> model.""" return Pipeline( steps=[ ("prep", build_feature_pipeline(numeric_features, categorical_features)), ("model", estimator), ] ) def get_feature_names(pipeline) -> list[str]: """Readable output names after the preprocessing steps.""" if isinstance(pipeline, Pipeline) and "preprocess" in pipeline.named_steps: pre = pipeline.named_steps["preprocess"] elif isinstance(pipeline, Pipeline) and "prep" in pipeline.named_steps: pre = pipeline.named_steps["prep"].named_steps["preprocess"] else: pre = pipeline return list(pre.get_feature_names_out()) def get_xy(df: pd.DataFrame) -> tuple[pd.DataFrame, np.ndarray]: """Raw pipeline input and label for a split (use this for CV searches).""" return df[INPUT_COLUMNS], df[TARGET].to_numpy() def fit_on_train( train: pd.DataFrame, val: pd.DataFrame, test: pd.DataFrame, target: str = TARGET, ): """Fit the preprocessing pipeline on train and transform all three splits. Returns (fitted preprocessing Pipeline, X_train, y_train, X_val, y_val, X_test, y_test). For a quick inspection only; model training/CV should use :func:`build_pipeline` with the raw frames. """ prep = build_feature_pipeline() X_train = prep.fit_transform(train[INPUT_COLUMNS]) X_val = prep.transform(val[INPUT_COLUMNS]) X_test = prep.transform(test[INPUT_COLUMNS]) return ( prep, X_train, train[target].to_numpy(), X_val, val[target].to_numpy(), X_test, test[target].to_numpy(), ) def smoke_test() -> None: """Quick self-check: scaler statistics must match the training split.""" from src.utils.config import SPLITS_DIR if not (SPLITS_DIR / "train.parquet").exists(): raise FileNotFoundError("run `python -m src.preprocessing.split` first") train = pd.read_parquet(SPLITS_DIR / "train.parquet") val = pd.read_parquet(SPLITS_DIR / "val.parquet") test = pd.read_parquet(SPLITS_DIR / "test.parquet") prep, Xtr, ytr, Xva, yva, Xte, yte = fit_on_train(train, val, test) print("X_train", Xtr.shape, "| X_val", Xva.shape, "| X_test", Xte.shape) print("steps:", list(prep.named_steps)) column_pre = prep.named_steps["preprocess"] scaler = column_pre.named_transformers_["num"].named_steps["scaler"] imputer = column_pre.named_transformers_["num"].named_steps["imputer"] engineered_train = prep.named_steps["features"].transform(train[INPUT_COLUMNS]) expected = imputer.transform(engineered_train[NUMERIC_FEATURES]).mean(axis=0) if not np.allclose(scaler.mean_, expected, atol=1e-8): raise AssertionError("scaler was not fitted on the training split only") engineered_all = prep.named_steps["features"].transform( pd.concat([train, val, test])[INPUT_COLUMNS] ) full_mean = np.nanmean(engineered_all[NUMERIC_FEATURES].to_numpy(), axis=0) drift = float(np.max(np.abs(scaler.mean_ - full_mean))) print(f"scaler mean == imputed train mean: OK | max drift vs full-data mean: {drift:.4f}") calendar = prep.named_steps["features"].transform(train[INPUT_COLUMNS].head(3)) print("FeatureEngineer added:", [c for c in ("hour", "is_weekend", "wind_dir_sin") if c in calendar]) print("feature names:", len(get_feature_names(prep))) if __name__ == "__main__": smoke_test()