diff --git a/paddle/operators/cos_sim_op.cc b/paddle/operators/cos_sim_op.cc
index 3760d0b161cae82fe02bdfca5304fc8b3bea6881..c033af3b741ae26ad9d37b2164f87aa6e8651c6e 100644
--- a/paddle/operators/cos_sim_op.cc
+++ b/paddle/operators/cos_sim_op.cc
@@ -90,8 +90,8 @@ class CosSimOpGrad : public framework::OperatorWithKernel {
 
     auto *x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
     auto *y_grad = ctx.Output<Tensor>(framework::GradVarName("Y"));
-    x_grad->Resize(x_dims);
-    y_grad->Resize(y_dims);
+    if (x_grad) x_grad->Resize(x_dims);
+    if (y_grad) y_grad->Resize(y_dims);
   }
 };
 
diff --git a/paddle/operators/cos_sim_op.h b/paddle/operators/cos_sim_op.h
index 69d35d8bc2b383e40e366a96e39521a26fc78a82..9e3ff26815644e11d8f9c9a3c3a3840159401c17 100644
--- a/paddle/operators/cos_sim_op.h
+++ b/paddle/operators/cos_sim_op.h
@@ -28,30 +28,30 @@ template <typename Place, typename T>
 class CosSimKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    auto* x = context.Input<Tensor>("X");
-    auto* y = context.Input<Tensor>("Y");
-    auto* z = context.Output<Tensor>("Out");
-    auto* x_norm = context.Output<Tensor>("XNorm");
-    auto* y_norm = context.Output<Tensor>("YNorm");
+    auto* input_x = context.Input<Tensor>("X");
+    auto* input_y = context.Input<Tensor>("Y");
+    auto* output_z = context.Output<Tensor>("Out");
+    auto* output_x_norm = context.Output<Tensor>("XNorm");
+    auto* output_y_norm = context.Output<Tensor>("YNorm");
 
-    z->mutable_data<T>(context.GetPlace());
-    x_norm->mutable_data<T>(context.GetPlace());
-    y_norm->mutable_data<T>(context.GetPlace());
+    output_z->mutable_data<T>(context.GetPlace());
+    output_x_norm->mutable_data<T>(context.GetPlace());
+    output_y_norm->mutable_data<T>(context.GetPlace());
 
-    auto dims = x->dims();
+    auto dims = input_x->dims();
     int size = static_cast<int>(framework::product(dims));
     auto new_dims = framework::make_ddim({dims[0], size / dims[0]});
-    auto X = EigenMatrix<T>::From(*x, new_dims);
-    auto Y = EigenMatrix<T>::From(*y, new_dims);
-    auto Z = EigenMatrix<T>::From(*z);
-    auto XNorm = EigenMatrix<T>::From(*x_norm);
-    auto YNorm = EigenMatrix<T>::From(*y_norm);
+    auto x = EigenMatrix<T>::From(*input_x, new_dims);
+    auto y = EigenMatrix<T>::From(*input_y, new_dims);
+    auto z = EigenMatrix<T>::From(*output_z);
+    auto x_norm = EigenMatrix<T>::From(*output_x_norm);
+    auto y_norm = EigenMatrix<T>::From(*output_y_norm);
 
     auto place = context.GetEigenDevice<Place>();
-    auto XY = (X * Y).sum(Eigen::array<int, 1>({1}));
-    XNorm.device(place) = (X * X).sum(Eigen::array<int, 1>({1})).sqrt();
-    YNorm.device(place) = (Y * Y).sum(Eigen::array<int, 1>({1})).sqrt();
-    Z.device(place) = XY / XNorm / YNorm;
+    auto xy = (x * y).sum(Eigen::array<int, 1>({1}));
+    x_norm.device(place) = x.square().sum(Eigen::array<int, 1>({1})).sqrt();
+    y_norm.device(place) = y.square().sum(Eigen::array<int, 1>({1})).sqrt();
+    z.device(place) = xy / x_norm / y_norm;
   }
 };
 
@@ -59,41 +59,44 @@ template <typename Place, typename T>
 class CosSimGradKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    auto* x = context.Input<Tensor>("X");
-    auto* y = context.Input<Tensor>("Y");
-    auto* z = context.Input<Tensor>("Out");
-    auto* x_norm = context.Input<Tensor>("XNorm");
-    auto* y_norm = context.Input<Tensor>("YNorm");
-    auto* grad_x = context.Output<Tensor>(framework::GradVarName("X"));
-    auto* grad_y = context.Output<Tensor>(framework::GradVarName("Y"));
-    auto* grad_z = context.Input<Tensor>(framework::GradVarName("Out"));
+    auto* input_x = context.Input<Tensor>("X");
+    auto* input_y = context.Input<Tensor>("Y");
+    auto* input_z = context.Input<Tensor>("Out");
+    auto* input_x_norm = context.Input<Tensor>("XNorm");
+    auto* input_y_norm = context.Input<Tensor>("YNorm");
+    auto* output_grad_x = context.Output<Tensor>(framework::GradVarName("X"));
+    auto* output_grad_y = context.Output<Tensor>(framework::GradVarName("Y"));
+    auto* input_grad_z = context.Input<Tensor>(framework::GradVarName("Out"));
 
-    grad_x->mutable_data<T>(context.GetPlace());
-    grad_y->mutable_data<T>(context.GetPlace());
-
-    auto dims = x->dims();
+    auto dims = input_x->dims();
     int size = static_cast<int>(framework::product(dims));
     auto new_dims = framework::make_ddim({dims[0], size / dims[0]});
-    auto X = EigenMatrix<T>::From(*x, new_dims);
-    auto Y = EigenMatrix<T>::From(*y, new_dims);
-    auto Z = EigenMatrix<T>::From(*z);
-    auto X_norm = EigenMatrix<T>::From(*x_norm);
-    auto Y_norm = EigenMatrix<T>::From(*y_norm);
-    auto dX = EigenMatrix<T>::From(*grad_x, new_dims);
-    auto dY = EigenMatrix<T>::From(*grad_y, new_dims);
-    auto dZ = EigenMatrix<T>::From(*grad_z);
+    auto x = EigenMatrix<T>::From(*input_x, new_dims);
+    auto y = EigenMatrix<T>::From(*input_y, new_dims);
+    auto z = EigenMatrix<T>::From(*input_z);
+    auto x_norm = EigenMatrix<T>::From(*input_x_norm);
+    auto y_norm = EigenMatrix<T>::From(*input_y_norm);
+    auto dz = EigenMatrix<T>::From(*input_grad_z);
 
     Eigen::DSizes<int, 2> bcast(1, new_dims[1]);
-    auto Z_bcast = Z.broadcast(bcast);
-    auto dZ_bcast = dZ.broadcast(bcast);
+    auto z_bcast = z.broadcast(bcast);
+    auto dz_bcast = dz.broadcast(bcast);
     auto place = context.GetEigenDevice<Place>();
-    auto X_snorm_bcast = X_norm.square().eval().broadcast(bcast);
-    auto Y_snorm_bcast = Y_norm.square().eval().broadcast(bcast);
-    auto norm_prod_bcast = (X_norm * Y_norm).eval().broadcast(bcast);
-    dX.device(place) =
-        dZ_bcast * (Y / norm_prod_bcast - Z_bcast * X / X_snorm_bcast);
-    dY.device(place) =
-        dZ_bcast * (X / norm_prod_bcast - Z_bcast * Y / Y_snorm_bcast);
+    auto x_snorm_bcast = x_norm.square().eval().broadcast(bcast);
+    auto y_snorm_bcast = y_norm.square().eval().broadcast(bcast);
+    auto norm_prod_bcast = (x_norm * y_norm).eval().broadcast(bcast);
+    if (output_grad_x) {
+      output_grad_x->mutable_data<T>(context.GetPlace());
+      auto dx = EigenMatrix<T>::From(*output_grad_x, new_dims);
+      dx.device(place) =
+          dz_bcast * (y / norm_prod_bcast - z_bcast * x / x_snorm_bcast);
+    }
+    if (output_grad_y) {
+      output_grad_y->mutable_data<T>(context.GetPlace());
+      auto dy = EigenMatrix<T>::From(*output_grad_y, new_dims);
+      dy.device(place) =
+          dz_bcast * (x / norm_prod_bcast - z_bcast * y / y_snorm_bcast);
+    }
   }
 };
 
diff --git a/python/paddle/v2/framework/tests/test_cos_sim_op.py b/python/paddle/v2/framework/tests/test_cos_sim_op.py
index a19be47f764d707ba9db9d3c4a3ad989f13b9f7a..32013a7999a4be42e5974b9ac751d5d911730994 100644
--- a/python/paddle/v2/framework/tests/test_cos_sim_op.py
+++ b/python/paddle/v2/framework/tests/test_cos_sim_op.py
@@ -24,26 +24,36 @@ class TestCosSimOp(unittest.TestCase):
         }
 
 
-class CosSimGradOpTest(GradientChecker):
-    def test_cos_sim_2d(self):
-        op = create_op("cos_sim")
-        inputs = {
+class TestCosSimGradOp(GradientChecker):
+    def setUp(self):
+        self.op = create_op("cos_sim")
+        self.inputs = {
             'X': np.random.random((10, 5)).astype("float32"),
             'Y': np.random.random((10, 5)).astype("float32")
         }
-        self.compare_grad(op, inputs)
+
+    def test_cpu_gpu_compare(self):
+        self.compare_grad(self.op, self.inputs)
+
+    def test_normal(self):
         self.check_grad(
-            op, inputs, set(["X", "Y"]), "Out", max_relative_error=0.05)
+            self.op, self.inputs, ["X", "Y"], "Out", max_relative_error=0.05)
 
-    def test_cos_sim_3d(self):
-        op = create_op("cos_sim")
-        inputs = {
-            'X': np.random.random((10, 5, 2)).astype("float32"),
-            'Y': np.random.random((10, 5, 2)).astype("float32")
-        }
-        self.compare_grad(op, inputs)
+    def test_ignore_x(self):
+        self.check_grad(
+            self.op,
+            self.inputs, ["Y"],
+            "Out",
+            max_relative_error=0.05,
+            no_grad_set={"X"})
+
+    def test_ignore_y(self):
         self.check_grad(
-            op, inputs, set(["X", "Y"]), "Out", max_relative_error=0.05)
+            self.op,
+            self.inputs, ["X"],
+            "Out",
+            max_relative_error=0.05,
+            no_grad_set={"Y"})
 
 
 if __name__ == '__main__':