diff --git a/paddle/operators/pad_op.cc b/paddle/operators/pad_op.cc
index 5dee8d0f5eabfe85c0fabaa5bde71aed5ede6b9b..dac1c56bdd43a303576e2ee87f9d71b60242fe2c 100644
--- a/paddle/operators/pad_op.cc
+++ b/paddle/operators/pad_op.cc
@@ -27,6 +27,9 @@ class PadOp : public framework::OperatorWithKernel {
   void InferShape(const framework::InferShapeContext &ctx) const override {
     auto dim0 = ctx.Input<Tensor>("X")->dims();
     auto paddings = GetAttr<std::vector<std::pair<int, int>>>("paddings");
+    PADDLE_ENFORCE_EQ(
+        dim0.size(), paddings.size(),
+        "Paddings size should be equal to dimension size of input tensor.");
     std::vector<int> dim1(dim0.size());
     for (int i = 0; i < dim0.size(); ++i) {
       dim1[i] = dim0[i] + paddings[i].first + paddings[i].second;
diff --git a/paddle/operators/pad_op.h b/paddle/operators/pad_op.h
index 9a0a064d7520d8c21c96f7bd5a587b208b1a0534..234019394c86be9c525351154396e3cb280fc8e4 100644
--- a/paddle/operators/pad_op.h
+++ b/paddle/operators/pad_op.h
@@ -14,8 +14,6 @@
 
 #pragma once
 
-#include "paddle/operators/math/math_function.h"
-
 #include "paddle/framework/eigen.h"
 #include "paddle/framework/op_registry.h"
 
@@ -30,12 +28,13 @@ using EigenTensor = framework::EigenTensor<T, D, MajorType, IndexType>;
 
 template <typename Place, typename T, size_t D>
 void PadFunction(const framework::ExecutionContext& context) {
-  auto pads = context.op_.GetAttr<std::vector<std::pair<int, int>>>("paddings");
+  auto pads =
+      context.op().GetAttr<std::vector<std::pair<int, int>>>("paddings");
   Eigen::array<std::pair<int, int>, D> paddings;
   for (int i = 0; i < pads.size(); ++i) {
     paddings[i] = pads[i];
   }
-  T pad_value = context.op_.GetAttr<T>("pad_value");
+  T pad_value = context.op().GetAttr<T>("pad_value");
 
   auto* X = context.Input<Tensor>("X");
   auto* Out = context.Output<Tensor>("Out");
@@ -80,7 +79,8 @@ class PadKernel : public framework::OpKernel {
 
 template <typename Place, typename T, size_t D>
 void PadGradFunction(const framework::ExecutionContext& context) {
-  auto pads = context.op_.GetAttr<std::vector<std::pair<int, int>>>("paddings");
+  auto pads =
+      context.op().GetAttr<std::vector<std::pair<int, int>>>("paddings");
   Eigen::array<std::pair<int, int>, D> paddings;
   for (int i = 0; i < pads.size(); ++i) {
     paddings[0].first = -paddings[0].first;
diff --git a/python/paddle/v2/framework/op.py b/python/paddle/v2/framework/op.py
index 0349407a851ebb48f69d7daef7a318cf348aad5d..359ccec814dc0494fb86509c3a081c74a98c1aee 100644
--- a/python/paddle/v2/framework/op.py
+++ b/python/paddle/v2/framework/op.py
@@ -96,7 +96,7 @@ class OpDescCreationMethod(object):
                     new_attr.strings.extend(user_defined_attr)
                 elif attr.type == framework_pb2.INT_PAIRS:
                     for p in user_defined_attr:
-                        pair = new_attr.pairs.add()
+                        pair = new_attr.int_pairs.add()
                         pair.first = p[0]
                         pair.second = p[1]
                 else:
diff --git a/python/paddle/v2/framework/tests/test_pad_op.py b/python/paddle/v2/framework/tests/test_pad_op.py
index b862033d8ca9bb5f478352a3d0cca71fe6930b45..10aeaa752fe73c9e9abdce3374d466e35a1bbc92 100644
--- a/python/paddle/v2/framework/tests/test_pad_op.py
+++ b/python/paddle/v2/framework/tests/test_pad_op.py
@@ -22,17 +22,22 @@ class TestPadOp(unittest.TestCase):
         }
 
 
-class PadGradOpTest(GradientChecker):
-    def test_pad(self):
-        op = Operator(
+class TestPadGradOp(GradientChecker):
+    def setUp(self):
+        self.op = Operator(
             type="pad",
             X="X",
             Out="Out",
             paddings=[(0, 1), (2, 3)],
             pad_value=0)
-        inputs = {'X': np.random.random((16, 16)).astype("float32"), }
+        self.inputs = {'X': np.random.random((16, 16)).astype("float32"), }
+
+    def test_normal(self):
+        self.check_grad(
+            self.op, self.inputs, set(["X"]), "Out", max_relative_error=0.5)
 
-        self.check_grad(op, inputs, set(["X"]), "Out", max_relative_error=0.5)
+    def test_cpu_gpu_compare(self):
+        self.compare_grad(self.op, self.inputs)
 
 
 if __name__ == '__main__':