diff --git a/python/pyspark/mllib/regression.py b/python/pyspark/mllib/regression.py
index 015a7860116c9..414a0ada80787 100644
--- a/python/pyspark/mllib/regression.py
+++ b/python/pyspark/mllib/regression.py
@@ -163,7 +163,8 @@ def _regression_train_wrapper(train_func, modelClass, data, initial_weights):
     first = data.first()
     if not isinstance(first, LabeledPoint):
         raise ValueError("data should be an RDD of LabeledPoint, but got %s" % first)
-    initial_weights = initial_weights or [0.0] * len(data.first().features)
+    if initial_weights is None:
+        initial_weights = [0.0] * len(data.first().features)
     weights, intercept = train_func(data, _convert_to_vector(initial_weights))
     return modelClass(weights, intercept)
 
diff --git a/python/pyspark/mllib/tests.py b/python/pyspark/mllib/tests.py
index 5328d99b69684..155019638f806 100644
--- a/python/pyspark/mllib/tests.py
+++ b/python/pyspark/mllib/tests.py
@@ -323,6 +323,13 @@ def test_regression(self):
         self.assertTrue(gbt_model.predict(features[2]) <= 0)
         self.assertTrue(gbt_model.predict(features[3]) > 0)
 
+        try:
+            LinearRegressionWithSGD.train(rdd, initialWeights=array([1.0, 1.0]))
+            LassoWithSGD.train(rdd, initialWeights=array([1.0, 1.0]))
+            RidgeRegressionWithSGD.train(rdd, initialWeights=array([1.0, 1.0]))
+        except ValueError:
+            self.fail()
+
 
 class StatTests(PySparkTestCase):
     # SPARK-4023