diff --git a/python/paddle/fluid/backward.py b/python/paddle/fluid/backward.py
index 10742af4ff432e0ba2ed2d2bae7f75ceb9ae3168..5d4fb2d1586428728287e6d92d461ed4be69f16f 100644
--- a/python/paddle/fluid/backward.py
+++ b/python/paddle/fluid/backward.py
@@ -1243,7 +1243,8 @@ def append_backward(loss,
             p_g_list6 = fluid.backward.append_backward(loss=avg_loss, parameter_list=all_weights, no_grad_set=set(all_weights))
 
     """
-    assert isinstance(loss, framework.Variable)
+    check_type(loss, 'loss', framework.Variable,
+               'fluid.backward.append_backward')
 
     if loss.op is None:
         # the loss is from a cloned program. Find loss op manually.
@@ -1254,7 +1255,8 @@ def append_backward(loss,
                       int(core.op_proto_and_checker_maker.OpRole.Loss))
 
     if callbacks is not None:
-        isinstance(callbacks, list)
+        check_type(callbacks, 'callbacks', list,
+                   'fluid.backward.append_backward')
 
     program = loss.block.program
     root_block = program.block(0)
@@ -1370,20 +1372,17 @@ def append_backward(loss,
     program._sync_with_cpp()
 
     if parameter_list is not None:
-        if not isinstance(parameter_list, (list, tuple, set)):
-            raise TypeError(
-                "The type of parameter_list argument must be list or tuple or set, but received %s."
-                % (type(parameter_list)))
+        check_type(parameter_list, 'parameter_list', (list, tuple, set),
+                   'fluid.backward.append_backward')
         parameters = []
         for i, param in enumerate(parameter_list):
+            check_type(param, 'parameter_list[%s]' % i, (framework.Variable,
+                                                         six.string_types),
+                       'fluid.backward.append_backward')
             if isinstance(param, framework.Variable):
                 parameters.append(param.name)
             elif isinstance(param, six.string_types):
                 parameters.append(param)
-            else:
-                raise TypeError(
-                    "The type of parameter_list's member must be paddle.fluid.Variable or str, but received %s."
-                    % (type(param)))
     else:
         params = program.global_block().all_parameters()
         parameters = [param.name for param in params if param.trainable]
@@ -1716,8 +1715,6 @@ def gradients(targets, inputs, target_gradients=None, no_grad_set=None):
                'fluid.backward.gradients')
     check_type(target_gradients, 'target_gradients', (
         framework.Variable, list, type(None)), 'fluid.backward.gradients')
-    check_type(no_grad_set, 'no_grad_set', (set, type(None)),
-               'fluid.backward.gradients')
 
     outs = calc_gradient(targets, inputs, target_gradients, no_grad_set)
     return _as_list(outs)
diff --git a/python/paddle/fluid/tests/unittests/test_backward.py b/python/paddle/fluid/tests/unittests/test_backward.py
index 3b064c2a7a98eef801f87e7d775cff259730edc0..2a4d024aa432b77052ed27cc8eb73d0f0b55fa71 100644
--- a/python/paddle/fluid/tests/unittests/test_backward.py
+++ b/python/paddle/fluid/tests/unittests/test_backward.py
@@ -287,6 +287,46 @@ class TestSimpleNetWithErrorNoGradSet(TestBackward):
             self._check_error_no_grad_set(self.net, [test, "test", 3])
 
 
+class TestAppendBackwardWithError(unittest.TestCase):
+    def build_net(self):
+        x = fluid.data(name='x', shape=[None, 13], dtype='int64')
+        y = fluid.data(name='y', shape=[None, 1], dtype='float32')
+        x_emb = fluid.embedding(x, size=[100, 256])
+        y_predict = fluid.layers.fc(input=x_emb, size=1, name='my_fc')
+        loss = fluid.layers.square_error_cost(input=y_predict, label=y)
+        avg_loss = fluid.layers.mean(loss)
+        param_names = [
+            param.name
+            for param in fluid.default_main_program().block(0).all_parameters()
+        ]
+
+        return avg_loss, param_names
+
+    def setUp(self):
+        main_program = fluid.Program()
+        with fluid.program_guard(main_program):
+            self.avg_loss, self.param_names = self.build_net()
+
+    def test_loss_type_error(self):
+        with self.assertRaises(TypeError):
+            fluid.backward.append_backward(loss=self.avg_loss.name)
+
+    def test_parameter_list_type_error(self):
+        with self.assertRaises(TypeError):
+            self.param_names[0] = np.random.random([10])
+            fluid.backward.append_backward(
+                loss=self.avg_loss, parameter_list=self.param_names)
+
+    def test_callback_type_error(self):
+        with self.assertRaises(TypeError):
+
+            def callback(block, context):
+                return
+
+            fluid.backward.append_backward(
+                loss=self.avg_loss, callbacks=callback)
+
+
 # TODO(Aurelius84): add conditional network test
 class ConditionalNet(BackwardNet):
     def __init__(self):