FIX changed_only=True with kwargs parameters

sklearn/utils/_pprint.py

@@ -66,6 +66,7 @@
 from inspect import signature
 import pprint
 from collections import OrderedDict
+from collections import defaultdict
 
 from ..base import BaseEstimator
 from .._config import get_config
@@ -94,6 +95,11 @@ def _changed_params(estimator):
                         estimator.__init__)
     init_params = signature(init_func).parameters
     init_params = {name: param.default for name, param in init_params.items()}
+    # We make init_params a defaultdict to be nice with third party libraries
+    # that have **kwargs in __init__. The parameters in kwargs would cause a
+    # KeyError otherwise.
+    init_params = defaultdict(lambda: None, init_params)
+
     for k, v in params.items():
         if (repr(v) != repr(init_params[k]) and
                 not (is_scalar_nan(init_params[k]) and is_scalar_nan(v))):

sklearn/utils/tests/test_pprint.py

@@ -2,6 +2,7 @@
 from pprint import PrettyPrinter
 
 import numpy as np
+import pytest
 
 from sklearn.utils._pprint import _EstimatorPrettyPrinter
 from sklearn.linear_model import LogisticRegressionCV
@@ -538,3 +539,18 @@ def test_builtin_prettyprinter():
     # Used to be a bug
 
     PrettyPrinter().pprint(LogisticRegression())
+
+def test_kwargs_in_init():
+    # Make sure the changed_only=True mode is OK when an argument is passed as
+    # kwargs.
+    # Non-regression test for
+    # https://github.com/scikit-learn/scikit-learn/issues/17206
+
+    pytest.importorskip("lightgbm")
+    from lightgbm import LGBMClassifier  # noqa
+
+    # metric is part of **kwargs
+    est = LGBMClassifier(metric='auc', max_depth=10)
+
+    expected = "LGBMClassifier(max_depth=10, metric='auc')"
+    assert expected == est.__repr__()