Describe the workflow you want to enable

Currently, partial dependence plots (PDP) and ICE profiles can be conveniently visualized by calling PartialDependenceDisplay.from_estimator. This works as long as the model is a valid scikit-learn estimator.

Since PDP/ICE actually only require the availability of a prediction function, I'd love to be able to pass only a prediction function taking an array-like input X and providing numeric output values.

Describe your proposed solution

Add method PartialDependenceDisplay.from_predict_fun that would enable above solution.

Trivial (but not working) example:

import numpy as np
from sklearn.inspection import PartialDependenceDisplay

X = np.array([range(10)]).reshape(5, -1)

def predict_fun(X):
    return np.repeat(3, X.shape[0])

PartialDependenceDisplay.from_predict_fun(
    predict_fun,
    X
)

Describe alternatives you've considered, if relevant

Define custom estimator.

Additional context

No response