diff --git a/paddle/operators/mean_op.h b/paddle/operators/mean_op.h
index fcb703e63bd5a82f9ffac2bf64e06fd0218dbdaa..9848af280b62729bef9243052ceae0b7d8f4c6f5 100644
--- a/paddle/operators/mean_op.h
+++ b/paddle/operators/mean_op.h
@@ -55,9 +55,10 @@ class MeanGradKernel : public framework::OpKernel {
     IG->mutable_data<T>(context.GetPlace());
 
     T ig_size = (T)framework::product(IG->dims());
+    Eigen::DSizes<int, 1> bcast(ig_size);
 
     EigenVector<T>::Flatten(*IG).device(context.GetEigenDevice<Place>()) =
-        EigenScalar<T>::From(*OG) / ig_size;
+        (EigenVector<T>::From(*OG) / ig_size).broadcast(bcast);
   }
 };
 
diff --git a/python/paddle/v2/framework/tests/test_mean_op.py b/python/paddle/v2/framework/tests/test_mean_op.py
index b5d52b90567bcd0c9f376147145d8638049f7bab..f32b3160d651a290823223c46c45bb3b6950a505 100644
--- a/python/paddle/v2/framework/tests/test_mean_op.py
+++ b/python/paddle/v2/framework/tests/test_mean_op.py
@@ -1,5 +1,6 @@
 import unittest
 from op_test_util import OpTestMeta
+from gradient_checker import GradientChecker, create_op
 import numpy as np
 
 
@@ -12,5 +13,12 @@ class TestMeanOp(unittest.TestCase):
         self.outputs = {'Out': np.mean(self.inputs['X'])}
 
 
+class MeanGradOpTest(GradientChecker):
+    def test_normal(self):
+        op = create_op("mean")
+        inputs = {"X": np.random.random((10, 10)).astype("float32")}
+        self.check_grad(op, inputs, set("X"), "Out")
+
+
 if __name__ == '__main__':
     unittest.main()