diff --git a/python/paddle/fluid/parallel_executor.py b/python/paddle/fluid/parallel_executor.py
index bb7b7d82f0539e1adc445e343290a5cec9802f8f..6baf648198585022f992709c519038688af293e1 100644
--- a/python/paddle/fluid/parallel_executor.py
+++ b/python/paddle/fluid/parallel_executor.py
@@ -160,7 +160,7 @@ class ParallelExecutor(object):
             build_strategy, num_trainers, trainer_id)
         self.scope = scope
 
-    def run(self, fetch_list, feed=None, feed_dict=None, return_numpy=False):
+    def run(self, fetch_list, feed=None, feed_dict=None, return_numpy=True):
         """
         Run a parallel executor with fetch_list.
 
@@ -197,7 +197,7 @@ class ParallelExecutor(object):
             feed_dict: Alias for feed parameter, for backward compatibility.
                 This parameter has been deprecated. Default None.
             return_numpy(bool): Whether converts the fetched tensor to numpy.
-                Default: False.
+                Default: True.
 
         Returns:
             List: The fetched result list.
diff --git a/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py b/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py
index 829c5a1a5fd099543e9e98b9587d4f316a91b587..21f2037ad408b0a92718c0ea2bae5e8bf563c665 100644
--- a/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py
+++ b/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py
@@ -81,7 +81,6 @@ class TestParallelExecutorBase(unittest.TestCase):
             begin = time.time()
             first_loss, = run_executor(
                 exe=exe, feed=feed_dict, fetch_list=[loss.name])
-            first_loss = np.array(first_loss)
 
             for i in xrange(iter):
                 run_executor(exe=exe, feed=feed_dict, fetch_list=[])
@@ -94,8 +93,6 @@ class TestParallelExecutorBase(unittest.TestCase):
                 print "%.4f Instance per second" % (
                     (batch_size * iter + 2) / (end - begin))
 
-            last_loss = np.array(last_loss)
-
             print first_loss, last_loss
             # self.assertGreater(first_loss[0], last_loss[0])
             return first_loss, last_loss
diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py
index 1ea7a6a5682318fb5f4ef8b3a08911df3cd44acf..63fb58c6927fa387b3b19147b9dc9d24bb8e5132 100644
--- a/python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py
+++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py
@@ -169,9 +169,8 @@ class TestCRFModel(unittest.TestCase):
             data = train_data()
             for i in xrange(10):
                 cur_batch = next(data)
-                print map(np.array,
-                          pe.run(feed=feeder.feed(cur_batch),
-                                 fetch_list=[avg_cost.name]))[0]
+                print pe.run(feed=feeder.feed(cur_batch),
+                             fetch_list=[avg_cost.name])[0]
 
     @unittest.skip(reason="CI hangs")
     def test_update_sparse_parameter_all_reduce(self):
diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_fetch_feed.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_fetch_feed.py
index 3b18072c7b04f118e4c001b51df51f6c048806c6..1f5d2f16773efb7537de85abec88344f8e0daa9f 100644
--- a/python/paddle/fluid/tests/unittests/test_parallel_executor_fetch_feed.py
+++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_fetch_feed.py
@@ -130,7 +130,7 @@ class TestFeedParallel(unittest.TestCase):
             use_cuda=use_cuda, loss_name=loss.name, main_program=main)
 
         for batch_id, data in enumerate(reader()):
-            loss_np = np.array(pe.run(feed=data, fetch_list=[loss.name])[0])
+            loss_np = pe.run(feed=data, fetch_list=[loss.name])[0]
             print batch_id, loss_np
             if batch_id == 2:
                 break
diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_test_while_train.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_test_while_train.py
index 31ba8c1d6096c9c89e0695c8eca8e16a5e303a61..252793944462244539084a288e5259f216359650 100644
--- a/python/paddle/fluid/tests/unittests/test_parallel_executor_test_while_train.py
+++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_test_while_train.py
@@ -70,10 +70,9 @@ class ParallelExecutorTestingDuringTraining(unittest.TestCase):
 
             for i in xrange(5):
                 test_loss, = test_exe.run([loss.name], feed=feed_dict)
-                test_loss = np.array(test_loss)
 
                 train_loss, = train_exe.run([loss.name], feed=feed_dict)
-                train_loss = np.array(train_loss)
+
                 self.assertTrue(
                     np.allclose(
                         train_loss, test_loss, atol=1e-8),