Skip to content

Overview

Overview

Scikit-Learn (SkLearn) supports many data science styles. On one end, there are free-form scripts that use SkLearn utility functions sparingly. On the other end, there are pipelines, which pack SkLearn transformers and the final SkLearn model into an integral, easily commandable (ie. fit, predict) and shareable entity.

PMML can support both styles. However, from the conversion perspective, the pipeline style is preferable to functional styles, because it enables easier automation.

Workflow

A typical PMML-oriented SkLearn pipeline has the following stages:

  • Feature mapping (E). Select and order raw features.
  • Feature transformation (T). Transform features to improve their statistical properties, derive new features based on existing features.
  • Feature selection (L). Filter transformed features.
  • Model. Estimate the mathematical relationship between the selected features and the target.

For example: 

from sklearn.compose import ColumnTransformer
from sklearn.feature_selection import SelectKBest
from sklearn.linear_model import LogisticRegression
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import OneHotEncoder, StandardScaler

# Combined feature mapper and transformer
transformer = ColumnTransformer([
    ("continuous", StandardScaler(), continuous_cols),
    ("categorical", OneHotEncoder(), categorical_cols)
], remainder = "drop")

# Feature selector
selector = SelectKBest(k = ...)

# Model
classifier = LogisticRegression()

pipeline = Pipeline([
    ("transformer", transformer),
    ("selector", selector),
    ("classifier", classifier)
])
pipeline.fit(X, y)

Export to PMML

Convert SkLearn artifacts to PMML using the sklearn2pmml package.

This package provides two main interaction points.

The sklearn2pmml.sklearn2pmml(obj, pmml_path) utility function works on live objects: 

from sklearn2pmml import sklearn2pmml

sklearn2pmml(pipeline, "Pipeline.pmml")

The sklearn2pmml command-line application works on previously serialized objects: 

sklearn2pmml -i Pipeline.pkl -o Pipeline.pmml

Resources