scikit-learn · jnboehm · Mar 17, 2020 · Mar 17, 2020 · Mar 17, 2020 · Mar 17, 2020
diff --git a/doc/whats_new/v0.23.rst b/doc/whats_new/v0.23.rst
@@ -661,6 +661,14 @@ Changelog
 - |FIX| :func:`utils.all_estimators` now only returns public estimators.
   :pr:`15380` by `Thomas Fan`_.
 
+:mod:`sklearn.pipeline`
+.......................
+
+- |Fix| :class:`pipeline.Pipeline` can now use estimators that only
+  expose `fit_transform` but not a separate `fit` and `transform`,
+  such as `manifold.TSNE`.
+  :pr:`16714` by :user:`Niklas Böhm <jnboehm>`.
+
 Miscellaneous
 .............
 

diff --git a/sklearn/pipeline.py b/sklearn/pipeline.py
@@ -154,19 +154,24 @@ def _validate_steps(self):
         for t in transformers:
             if t is None or t == 'passthrough':
                 continue
-            if (not (hasattr(t, "fit") or hasattr(t, "fit_transform")) or not
-                    hasattr(t, "transform")):
+            if (not hasattr(t, "fit_transform") and
+                    (not (hasattr(t, "fit") and hasattr(t, "transform")))):
                 raise TypeError("All intermediate steps should be "
-                                "transformers and implement fit and transform "
-                                "or be the string 'passthrough' "
-                                "'%s' (type %s) doesn't" % (t, type(t)))
+                                "transformers and implement fit and "
+                                "transform, fit_transform or be the string "
+                                "'passthrough'. '%s' (type %s) doesn't"
+                                % (t, type(t)))
+            elif not hasattr(t, "transform"):
+                warnings.warn("Intermediate step '%s' (type %s) does not have "
+                              "transform, pipeline is not reusable on "
+                              "test data." % (t, type(t)))
 
         # We allow last estimator to be None as an identity transformation
         if (estimator is not None and estimator != 'passthrough'
                 and not hasattr(estimator, "fit")):
             raise TypeError(
-                "Last step of Pipeline should implement fit "
-                "or be the string 'passthrough'. "
+                "Last step of Pipeline should implement fit, "
+                "fit_transform or be the string 'passthrough'. "
                 "'%s' (type %s) doesn't" % (estimator, type(estimator)))
 
     def _iter(self, with_final=True, filter_passthrough=True):

diff --git a/sklearn/tests/test_pipeline.py b/sklearn/tests/test_pipeline.py
@@ -31,6 +31,7 @@
 from sklearn.feature_selection import SelectKBest, f_classif
 from sklearn.dummy import DummyRegressor
 from sklearn.decomposition import PCA, TruncatedSVD
+from sklearn.manifold import TSNE
 from sklearn.datasets import load_iris
 from sklearn.preprocessing import StandardScaler
 from sklearn.feature_extraction.text import CountVectorizer
@@ -71,6 +72,11 @@ def set_params(self, **params):
         return self
 
 
+class OnlyFitTrans(NoFit):
+    def fit_transform(self, X, y=None):
+        return X
+
+
 class NoInvTransf(NoTrans):
     def transform(self, X):
         return X
@@ -164,8 +170,8 @@ def test_pipeline_init():
     # Check that we can't instantiate pipelines with objects without fit
     # method
     assert_raises_regex(TypeError,
-                        'Last step of Pipeline should implement fit '
-                        'or be the string \'passthrough\''
+                        'Last step of Pipeline should implement fit, '
+                        'fit_transform or be the string \'passthrough\''
                         '.*NoFit.*',
                         Pipeline, [('clf', NoFit())])
     # Smoke test with only an estimator
@@ -393,6 +399,58 @@ def test_pipeline_methods_preprocessing_svm():
         pipe.score(X, y)
 
 
+def test_pipeline_methods_pca_tsne():
+    # test that only fit_transform needs to be present in order to
+    # run a pipeline with fit_transform.
+    # Don't require transform to be present, explicitly.
+    pca = PCA(n_components=2, random_state=0)
+    tsne = TSNE(random_state=0)
+    separate_emb = tsne.fit_transform(pca.fit_transform(iris.data))
+
+    pca_for_pipeline = PCA(n_components=2, random_state=0)
+    tsne_for_pipeline = TSNE(random_state=0)
+    msg = ("Intermediate step '%s' (type %s) does not have "
+           "transform, pipeline is not reusable on test data."
+           % (tsne_for_pipeline, type(tsne_for_pipeline)))
+
+    pipe = make_pipeline(pca_for_pipeline, tsne_for_pipeline,
+                         'passthrough')
+
+    with pytest.warns(UserWarning, match=re.escape(msg)):
+        pipeline_emb = pipe.fit_transform(iris.data)
+
+    assert_array_almost_equal(pipeline_emb, separate_emb)
+
+    error_estimator = NoTrans()
+    msg = ("All intermediate steps should be "
+           "transformers and implement fit and "
+           "transform, fit_transform or be the string "
+           "'passthrough'. '%s' (type %s) doesn't"
+           % (error_estimator, type(error_estimator)))
+    with pytest.raises(TypeError, match=re.escape(msg)):
+        make_pipeline(error_estimator, 'passthrough')
+
+
+def test_fit_predict_on_nonreusable_pipeline():
+    oft = OnlyFitTrans()
+
+    km = KMeans(random_state=0)
+    km_for_pipeline = KMeans(random_state=0)
+
+    separate_pred = km.fit_predict(iris.data)
+
+    pipe = make_pipeline(oft, km_for_pipeline)
+
+    msg = ("Intermediate step '%s' (type %s) does not have "
+           "transform, pipeline is not reusable on test data."
+           % (oft, type(oft)))
+
+    with pytest.warns(UserWarning, match=re.escape(msg)):
+        pipeline_pred = pipe.fit_predict(iris.data)
+
+    assert_array_almost_equal(pipeline_pred, separate_pred)
+
+
 def test_fit_predict_on_pipeline():
     # test that the fit_predict method is implemented on a pipeline
     # test that the fit_predict on pipeline yields same results as applying
@@ -492,7 +550,7 @@ def test_feature_union():
     # test error if some elements do not support transform
     assert_raises_regex(TypeError,
                         'All estimators should implement fit and '
-                        'transform.*\\bNoTrans\\b',
+                        'transform.*\\bNoTrans\\b.*',
                         FeatureUnion,
                         [("transform", Transf()), ("no_transform", NoTrans())])