diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py
index 4692f20c1bbd410ea4461223cb8cba0b03c833eb..195245a12f9d6af78986ee2984e0b5caa2927e99 100644
--- a/python/paddle/fluid/framework.py
+++ b/python/paddle/fluid/framework.py
@@ -442,11 +442,16 @@ class Variable(object):
 
     @property
     def _stop_gradient(self):
-        return self._ivar.stop_gradient
+        if _in_imperative_mode():
+            return self._ivar.stop_gradient
+        else:
+            return self.stop_gradient
 
     @_stop_gradient.setter
     def _stop_gradient(self, s):
-        self._ivar.stop_gradient = s
+        if _in_imperative_mode():
+            self._ivar.stop_gradient = s
+        self.stop_gradient = s
 
     @property
     def persistable(self):
diff --git a/python/paddle/fluid/imperative/nn.py b/python/paddle/fluid/imperative/nn.py
index 543f57389003fd969b0504fd5175a1b44f9db06d..dc90603c3716bef347a5212a3c6fb66522e49c14 100644
--- a/python/paddle/fluid/imperative/nn.py
+++ b/python/paddle/fluid/imperative/nn.py
@@ -333,7 +333,6 @@ class BatchNorm(layers.Layer):
             dtype=self._dtype,
             default_initializer=Constant(1.0))
         if use_global_stats and self._helper.param_attr.learning_rate == 0.:
-            self._scale.stop_gradient = True
             self._scale._stop_gradient = True
 
         self._bias = self._helper.create_parameter(
@@ -342,7 +341,6 @@ class BatchNorm(layers.Layer):
             dtype=self._dtype,
             is_bias=True)
         if use_global_stats and self._helper.bias_attr.learning_rate == 0.:
-            self._bias.stop_gradient = True
             self._bias._stop_gradient = True
 
         self._mean = self._helper.create_parameter(
@@ -353,7 +351,6 @@ class BatchNorm(layers.Layer):
                 do_model_average=do_model_average_for_mean_and_var),
             shape=param_shape,
             dtype=self._dtype)
-        self._mean.stop_gradient = True
         self._mean._stop_gradient = True
 
         self._variance = self._helper.create_parameter(
@@ -364,7 +361,6 @@ class BatchNorm(layers.Layer):
                 do_model_average=do_model_average_for_mean_and_var),
             shape=param_shape,
             dtype=self._dtype)
-        self._variance.stop_gradient = True
         self._variance._stop_gradient = True
 
         self._in_place = in_place