context is not passed with `set_config`

[gunsodo]

From SLEP018 on, scikit-learn has released a global context setter with a simple set_config API. One of the use case is to propagate the transformed values through the pipeline as a Pandas DataFrame (with set_config(transform_output="pandas")).

I got this issue trying to run a HPO using TuneGridSearchCV which does not preserve the context set previously. A simple replication can be done using:

---

model.py

import sklearn
from sklearn.pipeline import Pipeline
from sklearn.linear_model import Ridge
from sklearn.preprocessing import StandardScaler
from sklearn.feature_selection import VarianceThreshold

# This sets the global context.
sklearn.set_config(transform_output="pandas")

class CustomVarianceThreshold(VarianceThreshold):
    def fit(self, X, y = None):
        assert X.columns is not None
        return super().fit(X, y)

    def transform(self, X):
        assert X.columns is not None
        return super().transform(X)

def MODEL():
    return Pipeline([
        ("scaler", StandardScaler()),
        ("selector", CustomVarianceThreshold()),
        ("regressor", Ridge()),
    ])

---

main.py

import numpy as np
import sys
from sklearn.datasets import load_iris
from sklearn.model_selection import GridSearchCV
from tune_sklearn import TuneGridSearchCV

from model import MODEL


PARAMS = {
    "regressor__alpha": np.linspace(0, 1, 100)
}


# This works fine.
def run_single():
    model = MODEL()

    data = load_iris(as_frame=True)
    model.fit(data.data, data.target)
    model.predict(data.data)


# This also works fine. HPO is enabled, but Ray is not the backend.
def run_hpo():
    model = MODEL()

    cv = GridSearchCV(model, PARAMS, n_jobs=-1)
    data = load_iris(as_frame=True)
    cv.fit(data.data, data.target)

    best_model = cv.best_estimator_
    best_model.predict(data.data)


# This breaks because `columns` is not an attribute of X.
def run_hpo_with_ray():
    model = MODEL()

    cv = TuneGridSearchCV(model, PARAMS, n_jobs=-1)
    data = load_iris(as_frame=True)
    cv.fit(data.data, data.target)

    best_model = cv.best_estimator_
    best_model.predict(data.data)

Although there is a workaround which is to define model with extra method (replacing StandardScaler() with StandardScaler().set_output(transform="pandas")), I think it would be nice if a global context setting via set_config integrates well with tune_sklearn.

I have checked the issues but did not find any pre-existing issue/documentation. Please let me know if this is a duplicate and I apologize if that is the case.