ENH add support for 'passthrough' in FeatureUnion

doc/whats_new/v1.1.rst

@@ -53,6 +53,12 @@ Changelog
   :pr:`20880` by :user:`Guillaume Lemaitre <glemaitre>`
   and :user:`András Simon <simonandras>`.
 
+:mod:`sklearn.pipeline`
+.......................
+
+- |Enhancement| Added support for "passthrough" in :class:`FeatureUnion`.
+  Setting a transformer to "passthrough" will pass the features unchanged.
+  :pr:`20860` by :user:`Shubhraneel Pal <shubhraneel>`.
 
 Code and Documentation Contributors
 -----------------------------------

sklearn/pipeline.py

@@ -17,6 +17,7 @@
 from joblib import Parallel
 
 from .base import clone, TransformerMixin
+from .preprocessing import FunctionTransformer
 from .utils._estimator_html_repr import _VisualBlock
 from .utils.metaestimators import available_if
 from .utils import (
@@ -853,21 +854,24 @@ class FeatureUnion(TransformerMixin, _BaseComposition):
 
     Parameters of the transformers may be set using its name and the parameter
     name separated by a '__'. A transformer may be replaced entirely by
-    setting the parameter with its name to another transformer,
-    or removed by setting to 'drop'.
+    setting the parameter with its name to another transformer, removed by
+    setting to 'drop' or disabled by setting to 'passthrough' (features are
+    passed without transformation).
 
     Read more in the :ref:`User Guide <feature_union>`.
 
     .. versionadded:: 0.13
 
     Parameters
     ----------
-    transformer_list : list of tuple
-        List of tuple containing `(str, transformer)`. The first element
-        of the tuple is name affected to the transformer while the
-        second element is a scikit-learn transformer instance.
-        The transformer instance can also be `"drop"` for it to be
-        ignored.
+    transformer_list : list of (str, transformer) tuples
+        List of transformer objects to be applied to the data. The first
+        half of each tuple is the name of the transformer. The transformer can
+        be 'drop' for it to be ignored or can be 'passthrough' for features to
+        be passed unchanged.
+
+        .. versionadded:: 1.1
+           Added the option `"passthrough"`.
 
         .. versionchanged:: 0.22
            Deprecated `None` as a transformer in favor of 'drop'.
@@ -977,7 +981,7 @@ def _validate_transformers(self):
 
         # validate estimators
         for t in transformers:
-            if t == "drop":
+            if t in ("drop", "passthrough"):
                 continue
             if not (hasattr(t, "fit") or hasattr(t, "fit_transform")) or not hasattr(
                 t, "transform"
@@ -1004,12 +1008,15 @@ def _iter(self):
         Generate (name, trans, weight) tuples excluding None and
         'drop' transformers.
         """
+
         get_weight = (self.transformer_weights or {}).get
-        return (
-            (name, trans, get_weight(name))
-            for name, trans in self.transformer_list
-            if trans != "drop"
-        )
+
+        for name, trans in self.transformer_list:
+            if trans == "drop":
+                continue
+            if trans == "passthrough":
+                trans = FunctionTransformer()
+            yield (name, trans, get_weight(name))
 
     @deprecated(
         "get_feature_names is deprecated in 1.0 and will be removed "

sklearn/tests/test_pipeline.py

@@ -1004,6 +1004,60 @@ def test_set_feature_union_step_drop(get_names):
     assert not record
 
 
+def test_set_feature_union_passthrough():
+    """Check the behaviour of setting a transformer to `"passthrough"`."""
+    mult2 = Mult(2)
+    mult3 = Mult(3)
+    X = np.asarray([[1]])
+
+    ft = FeatureUnion([("m2", mult2), ("m3", mult3)])
+    assert_array_equal([[2, 3]], ft.fit(X).transform(X))
+    assert_array_equal([[2, 3]], ft.fit_transform(X))
+
+    ft.set_params(m2="passthrough")
+    assert_array_equal([[1, 3]], ft.fit(X).transform(X))
+    assert_array_equal([[1, 3]], ft.fit_transform(X))
+
+    ft.set_params(m3="passthrough")
+    assert_array_equal([[1, 1]], ft.fit(X).transform(X))
+    assert_array_equal([[1, 1]], ft.fit_transform(X))
+
+    # check we can change back
+    ft.set_params(m3=mult3)
+    assert_array_equal([[1, 3]], ft.fit(X).transform(X))
+    assert_array_equal([[1, 3]], ft.fit_transform(X))
+
+    # Check 'passthrough' step at construction time
+    ft = FeatureUnion([("m2", "passthrough"), ("m3", mult3)])
+    assert_array_equal([[1, 3]], ft.fit(X).transform(X))
+    assert_array_equal([[1, 3]], ft.fit_transform(X))
+
+    X = iris.data
+    columns = X.shape[1]
+    pca = PCA(n_components=2, svd_solver="randomized", random_state=0)
+
+    ft = FeatureUnion([("passthrough", "passthrough"), ("pca", pca)])
+    assert_array_equal(X, ft.fit(X).transform(X)[:, :columns])
+    assert_array_equal(X, ft.fit_transform(X)[:, :columns])
+
+    ft.set_params(pca="passthrough")
+    X_ft = ft.fit(X).transform(X)
+    assert_array_equal(X_ft, np.hstack([X, X]))
+    X_ft = ft.fit_transform(X)
+    assert_array_equal(X_ft, np.hstack([X, X]))
+
+    ft.set_params(passthrough=pca)
+    assert_array_equal(X, ft.fit(X).transform(X)[:, -columns:])
+    assert_array_equal(X, ft.fit_transform(X)[:, -columns:])
+
+    ft = FeatureUnion(
+        [("passthrough", "passthrough"), ("pca", pca)],
+        transformer_weights={"passthrough": 2},
+    )
+    assert_array_equal(X * 2, ft.fit(X).transform(X)[:, :columns])
+    assert_array_equal(X * 2, ft.fit_transform(X)[:, :columns])
+
+
 def test_step_name_validation():
     error_message_1 = r"Estimator names must not contain __: got \['a__q'\]"
     error_message_2 = r"Names provided are not unique: \['a', 'a'\]"