This rule raises an issue when trying to access a Scikit-Learn transformer used in a pipeline with caching directly.

Why is this an issue?

When using a pipeline with a cache and passing the transformer objects as an instance from a variable, it is possible to access the transformer objects directly.

This is an issue since all the transformers are cloned when the Pipeline is fitted, and therefore, the objects outside the Pipeline are not updated and will yield unexpected results.

How to fix it

Replace the direct access to the transformer with an access to the named_steps attribute of the pipeline.

Code examples

Noncompliant code example

from sklearn.datasets import load_diabetes
from sklearn.preprocessing import RobustScaler
from sklearn.neighbors import KNeighborsRegressor
from sklearn.pipeline import Pipeline

diabetes = load_diabetes()
scaler = RobustScaler()
knn = KNeighborsRegressor(n_neighbors=5)

pipeline = Pipeline([
    ('scaler', scaler),
    ('knn', knn)
], memory="cache")

pipeline.fit(diabetes.data, diabetes.target)
print(scaler.center_) # Noncompliant : raises an AttributeError

Compliant solution

from sklearn.datasets import load_diabetes
from sklearn.preprocessing import RobustScaler
from sklearn.neighbors import KNeighborsRegressor
from sklearn.pipeline import Pipeline

diabetes = load_diabetes()
scaler = RobustScaler()
knn = KNeighborsRegressor(n_neighbors=5)

pipeline = Pipeline([
    ('scaler', scaler),
    ('knn', knn)
], memory="cache")

pipeline.fit(diabetes.data, diabetes.target)
print(pipeline.named_steps['scaler'].center_) # Compliant

Resources

Documentation