diff --git a/paddle/framework/executor.cc b/paddle/framework/executor.cc
index aafef12554fd0877ead14ed67a17db03a8c89eff..51ddb7e58e2505b7df48a9e365854371fcc0a12c 100644
--- a/paddle/framework/executor.cc
+++ b/paddle/framework/executor.cc
@@ -75,15 +75,15 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope) {
     device_context->Wait();
   }
   // // print tensor value
-  for (auto& var : block.vars()) {
-    std::cout << var.name() << std::endl;
-    auto v = scope->FindVar(var.name());
-    const LoDTensor& t = v->Get<LoDTensor>();
-    for (int i = 0; i < t.numel(); ++i) {
-      std::cout << t.data<float>()[i] << " ";
-    }
-    std::cout << std::endl;
-  }
+  // for (auto& var : block.vars()) {
+  //   std::cout << var.name() << std::endl;
+  //   auto v = scope->FindVar(var.name());
+  //   const LoDTensor& t = v->Get<LoDTensor>();
+  //   for (int i = 0; i < t.numel(); ++i) {
+  //     std::cout << t.data<float>()[i] << " ";
+  //   }
+  //   std::cout << std::endl;
+  // }
 }
 
 }  // namespace framework
diff --git a/paddle/framework/executor_test.cc b/paddle/framework/executor_test.cc
index 0856d1f32e9d324444969ea3012311e7f7e7eada..980f5f579c8420423ea3aa1fcee474618a89725a 100644
--- a/paddle/framework/executor_test.cc
+++ b/paddle/framework/executor_test.cc
@@ -25,6 +25,7 @@ limitations under the License. */
 USE_OP(elementwise_add);
 USE_OP(gaussian_random);
 USE_OP(feed);
+USE_OP(fetch);
 
 using std::string;
 using namespace paddle::platform;
@@ -94,6 +95,41 @@ void add_feed_op(string var_name, int index, proto_block* block) {
   Out->add_arguments(var_name);
 }
 
+void add_fetch_op(string var_name, int index, proto_block* block) {
+  std::vector<int> dim{3};
+
+  // insert variable
+  auto a = block->add_vars();
+  a->set_name(var_name);
+  auto a_lt = a->mutable_lod_tensor();
+  a_lt->set_data_type(paddle::framework::DataType::FP32);
+  for (int i : dim) {
+    a_lt->add_dims(i);
+  }
+
+  // insert operation
+  auto op = block->add_ops();
+  op->set_type("fetch");
+
+  // set dims attr
+  auto dims = op->add_attrs();
+  dims->set_name("dims");
+  dims->set_type(paddle::framework::AttrType::INTS);
+  for (int i : dim) {
+    dims->add_ints(i);
+  }
+
+  // set col attr
+  auto col = op->add_attrs();
+  col->set_name("col");
+  col->set_type(paddle::framework::AttrType::INT);
+  col->set_i(index);
+
+  auto Out = op->add_inputs();
+  Out->set_parameter("Input");
+  Out->add_arguments(var_name);
+}
+
 std::once_flag set_variable_flag;
 
 template <typename T>
@@ -119,6 +155,27 @@ void set_feed_variable(const std::vector<std::vector<T>>& inputs) {
   }
 }
 
+template <typename T>
+std::vector<std::vector<T>> get_fetch_variable() {
+  typedef std::vector<paddle::framework::Tensor> FetchOutputs;
+  Variable* g_fetch_value = GetScope()->FindVar("fetch_value");
+  FetchOutputs& fetch_outputs = *(g_fetch_value->GetMutable<FetchOutputs>());
+  auto size = fetch_outputs.size();
+
+  std::vector<std::vector<T>> result;
+  result.reserve(size);
+
+  for (size_t i = 0; i < size; i++) {
+    std::vector<T> tmp;
+    tmp.reserve(fetch_outputs[i].numel());
+    memcpy(tmp.data(), fetch_outputs[i].data<T>(),
+           fetch_outputs[i].numel() * sizeof(T));
+    result.push_back(tmp);
+  }
+
+  return result;
+}
+
 class ExecutorTesterRandom : public ::testing::Test {
  public:
   virtual void SetUp() override {
@@ -181,6 +238,8 @@ class ExecutorTesterFeed : public ::testing::Test {
     Out->set_parameter("Out");
     Out->add_arguments("c");
 
+    add_fetch_op("c", 0, root_block);
+
     std::vector<float> vec1 = {1.0, 2.0, 3.0};
     std::vector<float> vec2 = {4.0, 5.0, 6.0};
     inputs_.push_back(vec1);
@@ -213,8 +272,16 @@ TEST_F(ExecutorTesterFeed, CPU) {
   // 3 mini-batch
   for (int i = 0; i < 3; i++) {
     // need to set feed variable before Executor::Run
+    std::cout << "start mini-batch " << i << std::endl;
     set_feed_variable<float>(inputs_);
     executor->Run(pdesc_, GetScope());
+    std::vector<std::vector<float>> result = get_fetch_variable<float>();
+    for (auto& vec : result) {
+      for (auto& num : vec) {
+        std::cout << num << " ";
+      }
+      std::cout << std::endl;
+    }
   }
 
   delete executor;
diff --git a/paddle/framework/scope.cc b/paddle/framework/scope.cc
index b04120abf2744b0d5af904cf5b5782d8920663b9..2c416570cf62c8a4375293062cfc1e47f1011024 100644
--- a/paddle/framework/scope.cc
+++ b/paddle/framework/scope.cc
@@ -74,7 +74,10 @@ std::unique_ptr<T> make_unique(Args&&... args) {
 framework::Scope* GetScope() {
   static std::unique_ptr<framework::Scope> g_scope =
       make_unique<framework::Scope>();
-  std::call_once(feed_variable_flag, [&]() { g_scope->NewVar("feed_value"); });
+  std::call_once(feed_variable_flag, [&]() {
+    g_scope->NewVar("feed_value");
+    g_scope->NewVar("fetch_value");
+  });
   return g_scope.get();
 }
 
diff --git a/paddle/operators/activation_op.cu b/paddle/operators/activation_op.cu
index 44a6aaf9cb69ee401c90e5695174585f095bd669..93e9f1c694bacba48c4f8c46f90fb5b512bead99 100644
--- a/paddle/operators/activation_op.cu
+++ b/paddle/operators/activation_op.cu
@@ -1,16 +1,16 @@
 /* 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
+   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
+   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. */
+   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. */
 
 #define EIGEN_USE_GPU
 #include "paddle/operators/activation_op.h"
diff --git a/paddle/operators/feed_op.cc b/paddle/operators/feed_op.cc
index 5ae882bc8acfcbcc039bd98e61cafd793ba62edf..a61855cb9946c1dd2b8eef6707c637bd9d91219c 100644
--- a/paddle/operators/feed_op.cc
+++ b/paddle/operators/feed_op.cc
@@ -49,9 +49,9 @@ class FeedOpMaker : public framework::OpProtoAndCheckerMaker {
     AddAttr<int>("data_type", "output data type")
         .SetDefault(framework::DataType::FP32);
     AddAttr<int>("col", "The col in global feed variable").SetDefault(0);
-    AddAttr<std::vector<int>>("dims", "The dimension of random tensor.");
-    AddOutput("Out", "The output of dropout op.");
-    AddComment(R"DOC(Feed data to global feed variable)DOC");
+    AddAttr<std::vector<int>>("dims", "The dimension of feed tensor.");
+    AddOutput("Out", "The output of feed op.");
+    AddComment(R"DOC(Feed data from global feed variable)DOC");
   }
 };
 
diff --git a/paddle/operators/fetch_op.cc b/paddle/operators/fetch_op.cc
new file mode 100644
index 0000000000000000000000000000000000000000..68e8d26dbe11f0a2ca32a28b5968f8f9cfc26b7c
--- /dev/null
+++ b/paddle/operators/fetch_op.cc
@@ -0,0 +1,68 @@
+/* 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/fetch_op.h"
+
+namespace paddle {
+namespace operators {
+
+class FetchOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContextBase* ctx) const override {
+    typedef std::vector<framework::Tensor> FetchOutputs;
+    PADDLE_ENFORCE(ctx->HasInput("Input"), "Input should be not null.");
+    int col = ctx->Attrs().Get<int>("col");
+    framework::Variable* g_fetch_variable =
+        framework::GetScope()->FindVar("fetch_value");
+
+    FetchOutputs* tensors = g_fetch_variable->GetMutable<FetchOutputs>();
+    if (tensors->size() < col) {
+      tensors->resize(col);
+    }
+
+    auto input_dim = ctx->GetInputDim("Input");
+    framework::Tensor tmp;
+    tmp.Resize(input_dim);
+    (*tensors)[col].Resize(input_dim);
+    // need to handle LodTensor later
+  }
+
+  framework::DataType IndicateDataType(
+      const framework::ExecutionContext& ctx) const override {
+    return static_cast<framework::DataType>(Attr<int>("data_type"));
+  }
+};
+
+class FetchOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  FetchOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddAttr<int>("data_type", "output data type")
+        .SetDefault(framework::DataType::FP32);
+    AddAttr<int>("col", "The col in global fetch variable").SetDefault(0);
+    AddAttr<std::vector<int>>("dims", "The dimension of fetch tensor.");
+    AddInput("Input", "The output of fetch op.");
+    AddComment(R"DOC(Fetch data to global fetch variable)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_WITHOUT_GRADIENT(fetch, ops::FetchOp, ops::FetchOpMaker);
+REGISTER_OP_CPU_KERNEL(fetch, ops::FetchKernel<float>);
diff --git a/paddle/operators/fetch_op.cu b/paddle/operators/fetch_op.cu
new file mode 100644
index 0000000000000000000000000000000000000000..2e24d3a8adcdce7e3443f23367a8f06bf6ec86b2
--- /dev/null
+++ b/paddle/operators/fetch_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/feed_op.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(fetch, ops::FetchKernel<float>);
diff --git a/paddle/operators/fetch_op.h b/paddle/operators/fetch_op.h
new file mode 100644
index 0000000000000000000000000000000000000000..95e7986a22bf69bc7ccdc4b2effc61962429cb9b
--- /dev/null
+++ b/paddle/operators/fetch_op.h
@@ -0,0 +1,40 @@
+/* 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 {
+
+using Tensor = framework::Tensor;
+
+template <typename T>
+class FetchKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    typedef std::vector<framework::Tensor> FetchOutputs;
+    Tensor* input = ctx.Output<Tensor>("Input");
+    int col = ctx.template Attr<int>("col");
+    framework::Variable* g_fetch_variable =
+        framework::GetScope()->FindVar("fetch_value");
+    FetchOutputs tensors = g_fetch_variable->Get<FetchOutputs>();
+    tensors[col].mutable_data<T>(platform::CPUPlace());
+    tensors[col].CopyFrom<T>(*input, platform::CPUPlace());
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle