Wetsuit

A Scikit-Learn wrapper for H2O Estimators.

Why Wetsuit

While H2O Estimators have the .fit() and .predict() methods of the Scikit-Learn API, they don't always function as expected, especially with Pipeline objects. This package contains two estimators to remedy.

For example. the H2OEstimator.fit() method expects two H2OFrame objects, vice pandas DataFrame or numpy NDArray objects. Wetsuit provides two classes classes that wrap H2OEstimator objects and handle type conversion automatically, within the .fit() and .predict() methods:

• WetsuitRegressor
• WetsuitClassifier

Install

Create a virtual environment with Python >= 3.7 and install from PyPI:

pip install wetsuit

Use

Basic Pipeline

Here's an example that shows Scikit-Learn Pipeline compatibility. To align with the H2O API, we must instantiate the WetsuitClassifier with a list of feature names and the name of the response variable (these can also be indices). From there, you can plug in to a basic Pipeline object.

import h2o
from h2o.estimators import H2OXGBoostEstimator
from sklearn.datasets import load_iris
from sklearn.preprocessing import StandardScaler
from sklearn.pipeline import Pipeline

import wetsuit

h2o.init()

data = load_iris()

cls = wetsuit.WetsuitClassifier(H2OXGBoostEstimator(), data['feature_names'], 'target')
pl = Pipeline([
    ('scaler', StandardScaler()),
    ('cls', cls)
])
pl.fit(data['data'], data['target'])
fitted = pl.predict(data['data'])

h2o.cluster().shutdown()

Note: If you're doing feature selection within the pipeline, it's best instantiate the WetsuitClassifier from within the pipeline, so that you can dynamically pass a list of selected features using a selector's .get_feature_names_out() method.

Documentation

Documentation hosted on Github Pages: https://chris-santiago.github.io/wetsuit/