Reference Issues/PRs

Fixes #28780

What does this implement/fix? Explain your changes.

When FunctionTransformer:

• has feature_names_out set to None
• has set_output called with "pandas" or "polars"
• is fitted

Then get_feature_names_out() should return the output dataframe's columns. However, it raises an AttributeError instead.

A specific example, copied from the issue's discussions:

from sklearn.preprocessing import FunctionTransformer
import pandas as pd

my_transformer = FunctionTransformer(
    lambda X : pd.DataFrame(
        {
            f"{str(col)}^{power}" : X[col]**power
            for col in X
            for power in range(2,4)
        }
    )
   # no features_names_out
)
X = pd.DataFrame({
    "feature 1" : [1,2,3,4,5],
    "feature 2" : [3,4,5,6,7]
})
my_transformer.set_output(transform="pandas")
my_transformer.fit_transform(X)
# raises: AttributeError: This 'FunctionTransformer' has no attribute 'get_feature_names_out'
my_transformer.get_feature_names_out()

Any other comments?

To get the output column names, I saw three options:

1. Retrieve them when fit is called by running func on the fitted data
2. Retrieve them when transform is called by recording the output columns.
3. Construct a dummy dataframe when get_feature_names_out is called with feature names, and apply the function to it.

(1) means calling the function twice during a fit_transform, whereas (2) means making transform a stateful operation. (3) means that the specific example above, when called without an argument, would still fail.

I went with (1): during fit, call func with a smaller slice of the input in order to get output dataframe column names. If that's not desirable let me know and I can change the approach.

The fix works both for pandas and polars dataframes.