diff --git a/python/paddle/v2/framework/backward.py b/python/paddle/v2/framework/backward.py
index 6827792cb351243f926aeca5f37324dc987d6a79..678efd5d20585355a684bb2df16fdb57a69e0eeb 100644
--- a/python/paddle/v2/framework/backward.py
+++ b/python/paddle/v2/framework/backward.py
@@ -19,8 +19,20 @@ def append_backward_ops(loss, parameter_list=None, no_grad_set=None):
     :rtype: list[Variable]
     """
     assert isinstance(loss, framework.Variable)
-    param_grad_map = loss.block.program.append_backward(loss, no_grad_set or
-                                                        set())
+
+    if no_grad_set is None:
+        program = loss.block.program
+        assert isinstance(program, framework.Program)
+        no_grad_set = list()
+        for block in program.blocks:
+            assert isinstance(block, framework.Block)
+            for var in block.vars.itervalues():
+                assert isinstance(var, framework.Variable)
+                if var.stop_gradient:
+                    no_grad_set.append(var.name)
+        no_grad_set = set(no_grad_set)
+
+    param_grad_map = loss.block.program.append_backward(loss, no_grad_set)
     if parameter_list is not None:
         parameters = parameter_list
     else:
diff --git a/python/paddle/v2/framework/framework.py b/python/paddle/v2/framework/framework.py
index a26d8b517db0f23c3189604c65a74f602b87e542..dd23c47961b6b27ffca32dc7fe496d715411c858 100644
--- a/python/paddle/v2/framework/framework.py
+++ b/python/paddle/v2/framework/framework.py
@@ -21,6 +21,7 @@ class Variable(object):
                  dtype=None,
                  lod_level=None,
                  persistable=None,
+                 stop_gradient=False,
                  **kwargs):
         self.block = block
 
@@ -89,6 +90,7 @@ class Variable(object):
 
         self.block.vars[name] = self
         self.op = None
+        self.stop_gradient = stop_gradient
 
     def __str__(self):
         protostr = self.desc.serialize_to_string()
diff --git a/python/paddle/v2/framework/layers.py b/python/paddle/v2/framework/layers.py
index 967a85f1a5dccc10b265fd96050dd70bd9653b62..0739b2d2e2964f7f3565baf068a80c89938d193a 100644
--- a/python/paddle/v2/framework/layers.py
+++ b/python/paddle/v2/framework/layers.py
@@ -99,7 +99,7 @@ def data(name,
         shape = [-1] + shape  # append batch size as -1
 
     return helper.create_global_variable(
-        name=name, shape=shape, dtype=data_type, type=type)
+        name=name, shape=shape, dtype=data_type, type=type, stop_gradient=True)
 
 
 def _convert_(name):
diff --git a/python/paddle/v2/framework/tests/test_recurrent_op.py b/python/paddle/v2/framework/tests/test_recurrent_op.py
index d2c43168aa309a6e1bb8c1ae93d159620e5a270f..001de349d1f62e661335e63f7519c59697a1030c 100644
--- a/python/paddle/v2/framework/tests/test_recurrent_op.py
+++ b/python/paddle/v2/framework/tests/test_recurrent_op.py
@@ -125,11 +125,13 @@ class RecurrentOpTest1(unittest.TestCase):
             name='x',
             append_batch_size=False,
             **self.p_info)
+        x.stop_gradient = False
         h_boot = data(
             shape=[self.input_dim],
             data_type='float32',
             name='h_boot',
             **self.p_info)
+        h_boot.stop_gradient = False
 
         rnn = StaticRNN(main_program=self.main_program)
         with rnn.step():
@@ -256,11 +258,13 @@ class RecurrentOpTest2(RecurrentOpTest1):
             name='x',
             append_batch_size=False,
             **self.p_info)
+        x.stop_gradient = False
         h_boot = data(
             shape=[self.input_dim],
             data_type='float32',
             name='h_boot',
             **self.p_info)
+        h_boot.stop_gradient = False
 
         rnn = StaticRNN(main_program=self.main_program)
         with rnn.step():
@@ -353,18 +357,21 @@ class RecurrentOpTest3(RecurrentOpTest1):
             name='x',
             append_batch_size=False,
             **self.p_info)
+        x.stop_gradient = False
         h_boot1 = data(
             shape=[self.batch_size, self.input_dim],
             data_type='float32',
             name='h_boot1',
             append_batch_size=False,
             **self.p_info)
+        h_boot1.stop_gradient = False
         h_boot2 = data(
             shape=[self.batch_size, self.input_dim],
             data_type='float32',
             name='h_boot2',
             append_batch_size=False,
             **self.p_info)
+        h_boot2.stop_gradient = False
 
         rnn = StaticRNN(main_program=self.main_program)
         with rnn.step():