wrap-glmnet

wrap_glmnet is a small Python wrapper around python-glmnet's LogitNet classifier. It keeps the glmnet elastic-net logistic regression solver, but exposes a more sklearn-like estimator and adds control over glmnet's internal cross-validation behavior.

Why it exists

python-glmnet is useful for regularized logistic regression, but its classifier API has a few rough edges for sklearn workflows. This package provides GlmnetLogitNetWrapper, which is intended to be used in place of glmnet.LogitNet when you need:

• sklearn-compatible cloning and fitted attributes such as n_features_in_ and feature_names_in_
• explicit selection of whether predictions use lambda_1se or the best-performing cross-validation lambda
• sklearn-style class weights
• custom internal cross-validation splitters, including group-aware splitters
• multiclass ROC-AUC and deviance/log-loss scorers for glmnet's internal CV
• held-out CV predicted probabilities and per-fold CV scores
• predictable decision_function and predict_proba output shapes for binary, multiclass, single-lambda, and multi-lambda predictions

How it works

Importing wrap_glmnet patches selected python-glmnet internals so glmnet's lambda-path scoring can use an optional sklearn-style internal_cv splitter. The patched scoring path also forwards groups to scorers that accept it, records _cv_scores_, and, by default, stores held-out predicted probabilities in cv_pred_probs_ with shape (n_samples, n_classes, n_lambdas).

GlmnetLogitNetWrapper delegates the actual model fit to an inner glmnet.LogitNet instance. Constructor keyword arguments not handled by the wrapper are passed through to LogitNet.

By default, predictions, coef_, intercept_, cv_mean_score_final_, and cv_standard_error_final_ use lambda_1se (lambda_best_ in python-glmnet). Set use_lambda_1se=False to use the lambda with the best mean CV score (lambda_max_), or call switch_lambda(...) after fitting to get a copied fitted model with the other lambda choice.

Installation

The package requires Python 3.10 or newer.

pip install wrap_glmnet

For local development:

pip install -r requirements_dev.txt
pip install -e .

Usage

from sklearn.model_selection import StratifiedGroupKFold
from wrap_glmnet import GlmnetLogitNetWrapper

clf = GlmnetLogitNetWrapper(
    alpha=1.0,
    n_lambda=100,
    internal_cv=StratifiedGroupKFold(n_splits=3),
    scoring=GlmnetLogitNetWrapper.rocauc_scorer,
    require_cv_group_labels=True,
)

clf.fit(X_train, y_train, groups=groups)

labels = clf.predict(X_test)
probabilities = clf.predict_proba(X_test)

# Compare the default lambda_1se model to the best-CV-score lambda.
best_cv_lambda_clf = clf.switch_lambda(use_lambda_1se=False)
best_cv_probabilities = best_cv_lambda_clf.predict_proba(X_test)

If internal_cv is not supplied, the wrapper uses glmnet's normal internal CV setup with n_splits (default: 3). If require_cv_group_labels=True, calling fit(...) without groups raises an error.

class_weight accepts sklearn-style values such as a class-to-weight mapping or "balanced". When both class_weight and sample_weight are supplied, the wrapper multiplies them and normalizes the resulting sample weights before fitting.

Important behavior and limitations

• This package is focused on glmnet.LogitNet classification, not every glmnet model type.
• It patches python-glmnet functions at import time, so it relies on python-glmnet internals remaining compatible.
• store_cv_predicted_probabilities=True is the default and can use substantial memory for large datasets or long lambda paths. Disable it if you do not need cv_pred_probs_.
• Multiclass probabilities are computed from the wrapper's decision scores with softmax normalization; binary probabilities use sigmoid-style normalization.

Development

make test
make lint
make docs

Tests cover sklearn cloning, scorer behavior, group-aware CV, lambda switching, stored CV outputs, plotting, and output shapes.

Changelog

0.0.1

• First release on PyPI.