Description

The Pipeline documentation states that:

All estimators in a pipeline, except the last one, must be transformers (i.e. must have a transform method). The last estimator may be any type (transformer, classifier, etc.)

However, trying to use GridSearchCV with a Pipeline that only includes transformers will fail (in the hypothetical scenario that you are trying to select hyper-parameters for a transformer only; this can be useful if the transformer in question is a post-processing step that takes some hyper-parameters).

Steps/Code to Reproduce

For instance, the following (dumb example) will raise an error:

from sklearn.preprocessing import MinMaxScaler


Y = np.random.randn(1000,2)
Y = np.clip(X, 0., 1.)
X = 3. * Y

pipe = Pipeline(steps=[
    ('sc', MinMaxScaler())
])

param_grid = {'sc__feature_range': [(0., 1.), (0., 5.), (3., 10.)]}

gs = GridSearchCV(
    estimator=pipe,
    param_grid=param_grid,
    scoring='r2',
    cv=3
)
gs.fit(X, Y)
print(gs.best_params_)

Output:

AttributeError: 'MinMaxScaler' object has no attribute 'predict'

However, if you include a dumb estimator that simply passes its input to the output at the end of the pipeline, it will run with no issues:

from sklearn.base import BaseEstimator
from sklearn.preprocessing import MinMaxScaler


class PassThroughEstimator(BaseEstimator):
    def fit(self, X, y=None):
        return self

    def predict(self, X, y=None):
        return X


Y = np.random.randn(1000,2)
Y = np.clip(X, 0., 1.)
X = 3. * Y

pipe = Pipeline(steps=[
    ('sc', MinMaxScaler()),
    ('pse', PassThroughEstimator())
])

param_grid = {'sc__feature_range': [(0., 1.), (0., 5.), (3., 10.)]}

gs = GridSearchCV(
    estimator=pipe,
    param_grid=param_grid,
    scoring='r2',
    cv=3
)
gs.fit(X, Y)
print(gs.best_params_)

Output:

{'sc__feature_range': (0.0, 1.0)}

I realise that this is not a common case (needing to only fit a transformer with hyper-parameters), but is this intended behaviour?

On a related note, is it possible to include a post-processing transformer (e.g. smoothing for time-series regression or even a scaler in case the target has been pre-processed) at the end of a Pipeline and still be able to use GridSearchCV? According to the documentation it shouldn't be, since in that case not all layers preceding the last one are transformers. See also #4143.

Versions

System:
python: 3.6.8 |Anaconda, Inc.| (default, Feb 21 2019, 18:30:04) [MSC v.1916 64 bit (AMD64)]
executable: C:\Users\nak142\Miniconda3\envs\myo\python.exe
machine: Windows-10-10.0.18362-SP0

BLAS:
macros:
lib_dirs:
cblas_libs: cblas

Python deps:
pip: 19.1.1
setuptools: 41.0.1
sklearn: 0.21.2
numpy: 1.16.4
scipy: 1.2.1
Cython: 0.29.12
pandas: 0.24.2