diff --git a/paddle/framework/CMakeLists.txt b/paddle/framework/CMakeLists.txt
index 5df14ae78d432084ae1dc68a44c9b42c50da7d11..c9cf45e9d7d02d85f52cc16ed25cb5a651cc1a2d 100644
--- a/paddle/framework/CMakeLists.txt
+++ b/paddle/framework/CMakeLists.txt
@@ -56,5 +56,6 @@ cc_library(paddle_pybind SHARED
     uniform_random_op
     gaussian_random_op
     fill_zeros_like_op
-    scale_op)
+    scale_op
+    minus_op)
 endif(WITH_PYTHON)
diff --git a/paddle/framework/pybind.cc b/paddle/framework/pybind.cc
index 3aaf0de1506c03cd3f85745e1a918481b2b9509f..b4b7921d3363a66e132c56d2bb5f4a8281b9dcbc 100644
--- a/paddle/framework/pybind.cc
+++ b/paddle/framework/pybind.cc
@@ -44,6 +44,7 @@ USE_OP(gaussian_random);
 USE_OP(uniform_random);
 USE_OP(scale);
 USE_OP_ITSELF(identity);
+USE_OP(minus);
 
 namespace paddle {
 namespace framework {
diff --git a/paddle/operators/CMakeLists.txt b/paddle/operators/CMakeLists.txt
index 0ba598823b9d090108d9ada719393c4ed417dd62..61f7a4070f28d86e740f51cf202062dc22d71091 100644
--- a/paddle/operators/CMakeLists.txt
+++ b/paddle/operators/CMakeLists.txt
@@ -69,3 +69,4 @@ op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc
 op_library(uniform_random_op
         SRCS uniform_random_op.cc uniform_random_op.cu)
 op_library(scale_op SRCS scale_op.cc scale_op.cu DEPS net_op)
+op_library(minus_op SRCS minus_op.cc minus_op.cu DEPS scale_op)
diff --git a/paddle/operators/minus_op.cc b/paddle/operators/minus_op.cc
new file mode 100644
index 0000000000000000000000000000000000000000..c660ab5d3297fb93d8c1d4e417c73dbc0d4fbef6
--- /dev/null
+++ b/paddle/operators/minus_op.cc
@@ -0,0 +1,84 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/operators/minus_op.h"
+#include "paddle/operators/net_op.h"
+
+namespace paddle {
+namespace operators {
+
+class MinusOp : public framework::OperatorWithKernel {
+ public:
+  MinusOp(const std::string &type, const VarNameMap &inputs,
+          const VarNameMap &outputs, const framework::AttributeMap &attrs)
+      : OperatorWithKernel(type, inputs, outputs, attrs) {}
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    auto *left_tensor = ctx.Input<framework::Tensor>("X");
+    auto *right_tensor = ctx.Input<framework::Tensor>("Y");
+
+    PADDLE_ENFORCE_EQ(
+        framework::product(left_tensor->dims()),
+        framework::product(right_tensor->dims()),
+        "Minus operator must take two tensor with same num of elements");
+    ctx.Output<framework::Tensor>("Out")->Resize(left_tensor->dims());
+  }
+};
+
+class MinusOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  MinusOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "The left tensor of minus operator.").NotInGradient();
+    AddInput("Y", "The right tensor of minus operator.").NotInGradient();
+    AddOutput("Out", "The output tensor of minus operator.").NotInGradient();
+
+    AddComment(R"DOC(Minus Operator
+
+Equation: Out = X - Y
+)DOC");
+  }
+};
+template <typename AttrType>
+class MinusGradOp : public NetOp {
+ public:
+  MinusGradOp(const std::string &type, const VarNameMap &inputs,
+              const VarNameMap &outputs, const framework::AttributeMap &attrs)
+      : NetOp(type, inputs, outputs, attrs) {
+    auto out_grad = Input(framework::GradVarName("Out"));
+    auto x_grad = Output(framework::GradVarName("X"));
+    auto y_grad = Output(framework::GradVarName("Y"));
+
+    // x_grad = out_grad
+    AddOp(framework::OpRegistry::CreateOp("identity", {{"X", {out_grad}}},
+                                          {{"Out", {x_grad}}}, {}));
+
+    framework::AttributeMap scale_attr;
+    scale_attr["scale"] = static_cast<AttrType>(-1);
+    AddOp(framework::OpRegistry::CreateOp("scale", {{"X", {out_grad}}},
+                                          {{"Out", {y_grad}}}, scale_attr));
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+USE_OP(scale);
+USE_OP_ITSELF(identity);
+namespace ops = paddle::operators;
+REGISTER_OP(minus, ops::MinusOp, ops::MinusOpMaker, minus_grad,
+            ops::MinusGradOp<float>);
+REGISTER_OP_CPU_KERNEL(minus,
+                       ops::MinusKernel<paddle::platform::CPUPlace, float>);
diff --git a/paddle/operators/minus_op.cu b/paddle/operators/minus_op.cu
new file mode 100644
index 0000000000000000000000000000000000000000..a8375cc6301b2c1a917299c3933b03226bb72907
--- /dev/null
+++ b/paddle/operators/minus_op.cu
@@ -0,0 +1,18 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/operators/minus_op.h"
+
+REGISTER_OP_GPU_KERNEL(
+    minus, paddle::operators::MinusKernel<paddle::platform::GPUPlace, float>);
diff --git a/paddle/operators/minus_op.h b/paddle/operators/minus_op.h
new file mode 100644
index 0000000000000000000000000000000000000000..6310a4fd5141516cff4fc7acbe1d17913a1b5506
--- /dev/null
+++ b/paddle/operators/minus_op.h
@@ -0,0 +1,39 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#pragma once
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename Place, typename T>
+class MinusKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* left_tensor = context.Input<framework::Tensor>("X");
+    auto* right_tensor = context.Input<framework::Tensor>("Y");
+    auto* out_tensor = context.Output<framework::Tensor>("Out");
+
+    out_tensor->mutable_data<T>(context.GetPlace());
+    auto& dev = context.GetEigenDevice<Place>();
+    framework::EigenVector<T>::Flatten(*out_tensor).device(dev) =
+        framework::EigenVector<T>::Flatten(*left_tensor) -
+        framework::EigenVector<T>::Flatten(*right_tensor);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle