BUG Using `as_sklearn` can lead to a broken scikit-learn model

[betatim]

When using the as_sklearn and from_sklearn functionality to convert a cuml estimator to the equivalent scikit-learn estimator the handling of default values is inconsistent.

The following works:

import cuml
from sklearn import linear_model
from sklearn.datasets import make_classification

X, y = make_classification(n_features=32, random_state=42)

lr = cuml.LogisticRegression()
lr.fit(X, y)

lr_from_cuml = lr.as_sklearn()
print(lr_from_cuml.solver)  # set to 'lbfgs', scikit-learn default

However, explicitly setting solver leads to a broken scikti-learn model

lr = cuml.LogisticRegression(solver="qn")
lr.fit(X, y)

lr_from_cuml = lr.as_sklearn()
print(lr_from_cuml.solver)  # set to 'qn', not supported in scikit-learn

There are two problems here, one that doing the same thing (not passing a value and passing the default value) leads to different results and two that the user can end up with a broken scikit-learn model.

Part of the cause for this is a subtle inconsistency between how as_sklearn and from_sklearn work. The former uses only the explicitly passed constructor arguments when creating the scikit-learn instance. The latter uses get_params to get all the parameters. We should use get_params/all constructor arguments irrespective of whether the user passed them or not in both cases.

Doing this means we need to use our translation machinery "in reverse" or create a new set of translations. The machinery we have right now for cuml.accel works for scikit-learn -> cuml only.

[betatim]

Other places where this causes problems include

• https://scikit-learn.org/1.3/modules/generated/sklearn.cluster.KMeans.html#sklearn.cluster.KMeans (n_init parameter going through a deprecation cycle leads to the from_sklearn estimator to use "warn" as value). The default in scikit-learn v1.3 is 10 which means passing nothing leads to a different algorithm compared to passing "auto".
• https://scikit-learn.org/stable/modules/generated/sklearn.decomposition.TruncatedSVD.html#sklearn.decomposition.TruncatedSVD uses n_components=2 as default but in cuml the default is n_components=1. Without passing n_components the round trip doesn't work and the scikit-learn estimator is not the same as the cuml one

There are probably more example like these. The solution is to pass all hyper-parameters when constructing the class and applying a (version dependent) translation to the values.

We need a distinct translation mechanism for cuml -> scikit-learn as we can't invert the _hyperparam_interop_translator dictionary used to perform the scikit-learn -> cuml translation (several scikit-learn values map to the same cuml value).

[csadorf]

@jcrist Is this something that would be resolved by your revisions to the cuml.accel proxy layer?

[jcrist]

No, this is orthogonal to that. The only relevant thing I think is that by splitting out how we handle cuml.accel the as_sklearn/from_sklearn functionality can focus on handling only what's relevant for this API, rather than having to think about how that might affect cuml.accel.