About

The scikit_ext package contains various scikit-learn extensions, built entirely on top of sklearn base classes. The package is separated into two modules: estimators and scorers. Full documentation can be found here.

Installation

Package Index on PyPI To install:

pip install scikit-ext

Estimators

• MultiGridSearchCV: Extension to native sklearn GridSearchCV for multiple estimators and param_grids. Accepts a list of estimators and param_grids, iterating through each fitting a GridSearchCV model for each estimator/param_grid. Chooses the best fitted GridSearchCV model. Inherits sklearn’s BaseSearchCV class, so attributes and methods are all similar to GridSearchCV.

• PrunedPipeline: Extension to native sklearn Pipeline intended for text learning pipelines with a vectorization step and a feature selection step. Instead of remembering all vectorizer vocabulary elements and selecting appropriate features at prediction time, the extension prunes the vocabulary after fitting to only include elements who will ultimately survive the feature selection filter applied later in the pipeline. This reduces memory and improves prediction latency. Predictions will be identical to those made with a trained Pipeline model. Inherits sklearn’s Pipeline class, so attributes and methods are all similar to Pipeline.

• ZoomGridSearchCV: Extension to native sklearn GridSearchCV. Fits multiple GridSearchCV models, updating the param_grid after each iteration. The update looks at successful parameter values for each grid key. A new list of values is created which expands the resolution of the search values centered around the best performing value of the previous fit. This allows the standard grid search process to start with a small number of distant values for each parameter, and zoom in as the better performing corner of the hyperparameter search space becomes clear.

• IterRandomEstimator: Meta-Estimator intended primarily for unsupervised estimators whose fitted model can be heavily dependent on an arbitrary random initialization state. It is best used for problems where a fit_predict method is intended, so the only data used for prediction will be the same data on which the model was fitted.

• OptimizedEnsemble: An optimized ensemble class. Will find the optimal n_estimators parameter for the given ensemble estimator, according to the specified input parameters.

• OneVsRestAdjClassifier: One-Vs-Rest multiclass strategy. The adjusted version is a custom extension which overwrites the inherited predict_proba method with a more flexible method allowing custom normalization for the predicted probabilities. Any norm argument that can be passed directly to sklearn.preprocessing.normalize is allowed. Additionally, norm=None will skip the normalization step alltogeter. To mimick the inherited OneVsRestClassfier behavior, set norm=’l2’. All other methods are inherited from OneVsRestClassifier.

Scorers

• TimeScorer: Score using estimated prediction latency of estimator.

• MemoryScorer: Score using estimated memory of pickled estimator object.

• CombinedScorer: Score combining multiple scorers by averaging their scores.

• cluster_distribution_score: Scoring function which scores the resulting cluster distribution accross classes. A more even distribution indicates a higher score.

Authors

Evan Harris

License

This project is licensed under the MIT License - see the LICENSE file for details