Extending Skyulf-Core

Skyulf-core is intentionally simple: calculators learn parameters, appliers apply them.

Add a new preprocessing node

1. Create a new module in skyulf.preprocessing.
2. Implement a Calculator and an Applier.
3. Register the node type string in the FeatureEngineer dispatcher.

Example skeleton

from typing import Any, Dict, Tuple, Union

import pandas as pd

from skyulf.preprocessing.base import BaseApplier, BaseCalculator
from skyulf.utils import pack_pipeline_output, unpack_pipeline_input


class MyNodeCalculator(BaseCalculator):
    def fit(
        self,
        df: Union[pd.DataFrame, Tuple[pd.DataFrame, pd.Series]],
        config: Dict[str, Any],
    ) -> Dict[str, Any]:
        X, _, _ = unpack_pipeline_input(df)
        # Learn something from X...
        return {"type": "my_node", "columns": config.get("columns", [])}


class MyNodeApplier(BaseApplier):
    def apply(
        self,
        df: Union[pd.DataFrame, Tuple[pd.DataFrame, pd.Series]],
        params: Dict[str, Any],
    ) -> Union[pd.DataFrame, Tuple[pd.DataFrame, pd.Series]]:
        X, y, is_tuple = unpack_pipeline_input(df)
        # Apply transformation...
        return pack_pipeline_output(X, y, is_tuple)

Add a new modeling estimator

1. Implement a new BaseModelCalculator and BaseModelApplier (or subclass SklearnCalculator/SklearnApplier).
2. Add a mapping entry in SkyulfPipeline._init_model_estimator().

Testing guidance

Prefer integration tests that run:

Calculator.fit → Applier.apply on a small synthetic dataframe.