FIX builtin PrettyPrinter usage on estimators (#12938)

NicolasHug · jnothman · commit 07b0a420a9df · 2019-01-14T17:29:57.000+11:00
diff --git a/sklearn/utils/_pprint.py b/sklearn/utils/_pprint.py
@@ -321,7 +321,10 @@ def _pprint_key_val_tuple(self, object, stream, indent, allowance, context,
         self._format(v, stream, indent + len(rep) + len(middle), allowance,
                      context, level)
 
-    _dispatch = pprint.PrettyPrinter._dispatch
+    # Note: need to copy _dispatch to prevent instances of the builtin
+    # PrettyPrinter class to call methods of _EstimatorPrettyPrinter (see issue
+    # 12906)
+    _dispatch = pprint.PrettyPrinter._dispatch.copy()
     _dispatch[BaseEstimator.__repr__] = _pprint_estimator
     _dispatch[KeyValTuple.__repr__] = _pprint_key_val_tuple
 
diff --git a/sklearn/utils/tests/test_pprint.py b/sklearn/utils/tests/test_pprint.py
@@ -1,4 +1,5 @@
 import re
+from pprint import PrettyPrinter
 
 from sklearn.utils._pprint import _EstimatorPrettyPrinter
 from sklearn.pipeline import make_pipeline, Pipeline
@@ -311,3 +312,11 @@ def test_length_constraint():
     vectorizer = CountVectorizer(vocabulary=vocabulary)
     repr_ = vectorizer.__repr__()
     assert '...' in repr_
+
+
+def test_builtin_prettyprinter():
+    # non regression test than ensures we can still use the builtin
+    # PrettyPrinter class for estimators (as done e.g. by joblib).
+    # Used to be a bug
+
+    PrettyPrinter().pprint(LogisticRegression())
