[BUG] `Core.get_params()` returns fitted attributes and internal state, violating sklearn estimator API

[dkstlzk]

Core.get_params() currently returns self.__dict__ when deep=True.
This exposes fitted attributes and internal state (e.g. coef_, statistics_) in the parameter dictionary.

However, according to the scikit-learn estimator API, get_params() should return only constructor parameters, not attributes created during fit(). Returning fitted attributes can break compatibility with utilities such as sklearn.clone and GridSearchCV.

Minimal Example

from pygam import LinearGAM
import numpy as np

X = np.random.rand(100,1)
y = np.sin(X[:,0])

gam = LinearGAM().fit(X,y)
params = gam.get_params(deep=True)

print("coef_" in params)        # True
print("statistics_" in params)  # True

[dkstlzk]

I took a closer look at the implementation and traced the behavior to pygam/core.py.

In Core.get_params, when deep=True, the method returns self.__dict__. Since self.__dict__ includes fitted attributes (e.g. coef_, statistics_, etc.), these end up being exposed in the parameter dictionary even though they are created during fit().

Because classes such as LinearGAM inherit from Core, this behavior propagates across all GAM estimators.

Additionally, returning self.__dict__ directly means the returned dictionary is a reference to the internal attribute storage. This could allow external mutation of the estimator state if the returned dictionary is modified.

This appears to be inconsistent with the scikit-learn estimator API, where get_params() should return only constructor parameters.

If maintainers agree, I’d be happy to explore a fix that constructs the parameter dictionary from the estimator's __init__ signature (similar to how scikit-learn implements get_params).