diff --git a/demo/image_classification/prediction.py b/demo/image_classification/prediction.py
index 9a86aafcb2..49c0ff600c 100755
--- a/demo/image_classification/prediction.py
+++ b/demo/image_classification/prediction.py
@@ -126,7 +126,7 @@ class ImageClassifier():
         # For oversampling, average predictions across crops.
         # If not, the shape of output[name]: (1, class_number),
         # the mean is also applicable.
-        return output[output_layer].mean(0)
+        return output[output_layer]['value'].mean(0)
 
     def predict(self, image=None, output_layer=None):
         assert isinstance(image, basestring)
diff --git a/demo/model_zoo/resnet/classify.py b/demo/model_zoo/resnet/classify.py
index 4631816c43..6074cc1d3a 100755
--- a/demo/model_zoo/resnet/classify.py
+++ b/demo/model_zoo/resnet/classify.py
@@ -156,7 +156,7 @@ class ImageClassifier():
             # For oversampling, average predictions across crops.
             # If not, the shape of output[name]: (1, class_number),
             # the mean is also applicable.
-            res[name] = output[name].mean(0)
+            res[name] = output[name]['value'].mean(0)
 
         return res
 
diff --git a/paddle/api/Arguments.cpp b/paddle/api/Arguments.cpp
index 41beed38a8..a3f4bfffc9 100644
--- a/paddle/api/Arguments.cpp
+++ b/paddle/api/Arguments.cpp
@@ -38,6 +38,13 @@ Arguments* Arguments::createByPaddleArgumentVector(void* ptr) {
   return args;
 }
 
+Arguments* Arguments::createByPaddleArgument(const void* ptr) {
+  auto p = (paddle::Argument*)(ptr);
+  auto args = new Arguments();
+  args->m->outputs.push_back(*p);
+  return args;
+}
+
 Matrix* Arguments::getSlotValue(size_t idx) const throw(RangeError) {
   auto& a = m->getArg(idx);
   return Matrix::createByPaddleMatrixPtr(&a.value);
diff --git a/paddle/api/GradientMachine.cpp b/paddle/api/GradientMachine.cpp
index 66115f8293..538ca2999f 100644
--- a/paddle/api/GradientMachine.cpp
+++ b/paddle/api/GradientMachine.cpp
@@ -144,12 +144,12 @@ Parameter* GradientMachine::getParameter(size_t i) throw(RangeError) {
 
 void GradientMachine::randParameters() { m->machine->randParameters(); }
 
-Matrix* GradientMachine::getLayerOutput(const std::string& layerName) const
+Arguments* GradientMachine::getLayerOutput(const std::string& layerName) const
     throw(UnsupportError) {
-  auto nn = std::dynamic_pointer_cast<paddle::NeuralNetwork>(m->machine);
+  auto nn = m->machine;
   if (nn) {
-    auto mat = nn->getLayerOutput(layerName);
-    return Matrix::createByPaddleMatrixPtr(&mat);
+    auto arg = nn->getLayerOutput(layerName);
+    return Arguments::createByPaddleArgument(&arg);
   } else {
     throw UnsupportError();
   }
diff --git a/paddle/api/PaddleAPI.h b/paddle/api/PaddleAPI.h
index 80c50cdb08..1831b8e170 100644
--- a/paddle/api/PaddleAPI.h
+++ b/paddle/api/PaddleAPI.h
@@ -454,6 +454,7 @@ public:
 
 private:
   static Arguments* createByPaddleArgumentVector(void* ptr);
+  static Arguments* createByPaddleArgument(const void* ptr);
   void* getInternalArgumentsPtr() const;
 
 private:
@@ -769,7 +770,7 @@ public:
 
   void randParameters();
 
-  Matrix* getLayerOutput(const std::string& layerName) const
+  Arguments* getLayerOutput(const std::string& layerName) const
       throw(UnsupportError);
 
   /**
@@ -956,7 +957,7 @@ public:
 
   Arguments* getForwardOutput();
 
-  Matrix* getLayerOutput(const std::string& layerName);
+  Arguments* getLayerOutput(const std::string& layerName) const;
 };
 
 /// the N-Best results generated from one input sequence.
diff --git a/paddle/api/Trainer.cpp b/paddle/api/Trainer.cpp
index d83dc380be..84e4ca054a 100644
--- a/paddle/api/Trainer.cpp
+++ b/paddle/api/Trainer.cpp
@@ -131,12 +131,11 @@ void Trainer::testOneDataBatch(size_t batchSize, const Arguments& args) {
 void TrainerPrivate::finishTestPeriod() { tester_->finishTestPeriod(); }
 void Trainer::finishTestPeriod() { m->finishTestPeriod(); }
 
-Matrix* Trainer::getLayerOutput(const std::string& layerName) {
-  auto nn = std::dynamic_pointer_cast<paddle::NeuralNetwork>(
-      this->m->getGradientMachine());
+Arguments* Trainer::getLayerOutput(const std::string& layerName) const {
+  auto nn = this->m->getGradientMachine();
   CHECK(nn) << "trainerInternal_.getGradientMachine() is not NeuralNetwork";
-  auto m = nn->getLayerOutput(layerName);
-  return Matrix::createByPaddleMatrixPtr(&m);
+  auto arg = nn->getLayerOutput(layerName);
+  return Arguments::createByPaddleArgument(&arg);
 }
 
 void Trainer::forwardOneBatch(size_t batchSize) {
diff --git a/paddle/gserver/gradientmachines/GradientMachine.h b/paddle/gserver/gradientmachines/GradientMachine.h
index 0829968d87..bc2f2f8563 100644
--- a/paddle/gserver/gradientmachines/GradientMachine.h
+++ b/paddle/gserver/gradientmachines/GradientMachine.h
@@ -134,6 +134,10 @@ public:
     backward(callback);
   }
 
+  virtual Argument getLayerOutput(const std::string& layerName) {
+    return *((Argument*)nullptr);
+  }
+
   // see comment in Layer.h for the function with the same name
   virtual void resetState() {}
 
diff --git a/paddle/gserver/gradientmachines/MultiGradientMachine.cpp b/paddle/gserver/gradientmachines/MultiGradientMachine.cpp
index 80f223824d..123273f916 100644
--- a/paddle/gserver/gradientmachines/MultiGradientMachine.cpp
+++ b/paddle/gserver/gradientmachines/MultiGradientMachine.cpp
@@ -282,6 +282,18 @@ void MultiGradientMachine::forwardBackward(const std::vector<Argument>& inArgs,
   backwardImp(callback);
 }
 
+Argument MultiGradientMachine::getLayerOutput(const std::string& layerName) {
+  std::vector<Argument> args;
+  args.reserve(threads_.size());
+
+  for (auto& thread : threads_) {
+    args.push_back(thread->getGradientMachine()->getLayerOutput(layerName));
+  }
+  outLayerArgs_.concat(args, false /* use_gpu */, outArgStream_, passType_);
+
+  return outLayerArgs_;
+}
+
 void MultiGradientMachine::backwardImp(const UpdateCallback& callback) {
   for (size_t i = 0; i < parameters_.size(); i++) {
     if (!parameters_[i]->useGpu() || parameters_[i]->isStatic()) continue;
diff --git a/paddle/gserver/gradientmachines/MultiGradientMachine.h b/paddle/gserver/gradientmachines/MultiGradientMachine.h
index 9be15ef4bc..838a52b515 100644
--- a/paddle/gserver/gradientmachines/MultiGradientMachine.h
+++ b/paddle/gserver/gradientmachines/MultiGradientMachine.h
@@ -189,6 +189,8 @@ public:
                        PassType passType,
                        const UpdateCallback& callback);
 
+  virtual Argument getLayerOutput(const std::string& layerName);
+
   virtual void onPassEnd();
 
   virtual void finish();
@@ -314,6 +316,8 @@ protected:
   std::vector<Argument> outArgs_;
   hl_stream_t outArgStream_;
 
+  Argument outLayerArgs_;
+
   /// ParameterType which needs to be merged from each GPU
   std::vector<ParameterType> mergeTypes_;
   int numDevices_;         /* number of gpu devices */
diff --git a/paddle/gserver/gradientmachines/NeuralNetwork.cpp b/paddle/gserver/gradientmachines/NeuralNetwork.cpp
index 273a9111c3..4512aacc81 100644
--- a/paddle/gserver/gradientmachines/NeuralNetwork.cpp
+++ b/paddle/gserver/gradientmachines/NeuralNetwork.cpp
@@ -293,11 +293,10 @@ void NeuralNetwork::backward(const UpdateCallback& callback) {
   }
 }
 
-MatrixPtr NeuralNetwork::getLayerOutput(const std::string& layerName) {
-  auto it = layerMap_.find(layerName);
-  CHECK(it != layerMap_.end()) << "Cannot find layer: " << layerName;
-  return it->second->getOutputValue();
+Argument NeuralNetwork::getLayerOutput(const std::string& layerName) {
+  return getLayer(layerName)->getOutput();
 }
+
 void NeuralNetwork::onPassEnd() {
   for (auto& layer : layers_) {
     layer->onPassEnd();
diff --git a/paddle/gserver/gradientmachines/NeuralNetwork.h b/paddle/gserver/gradientmachines/NeuralNetwork.h
index 25af4abcf8..e7b6c43840 100644
--- a/paddle/gserver/gradientmachines/NeuralNetwork.h
+++ b/paddle/gserver/gradientmachines/NeuralNetwork.h
@@ -87,7 +87,8 @@ public:
 
   virtual void backward(const UpdateCallback& callback = nullptr);
 
-  MatrixPtr getLayerOutput(const std::string& layerName);
+  virtual Argument getLayerOutput(const std::string& layerName);
+
   const LayerPtr& getLayer(const std::string& layerName) const {
     auto it = layerMap_.find(layerName);
     CHECK(it != layerMap_.end()) << "Unknown layer " << layerName;
diff --git a/paddle/gserver/layers/CosSimLayer.cpp b/paddle/gserver/layers/CosSimLayer.cpp
index a6c0300acf..57ba124e40 100644
--- a/paddle/gserver/layers/CosSimLayer.cpp
+++ b/paddle/gserver/layers/CosSimLayer.cpp
@@ -42,7 +42,7 @@ void CosSimLayer::forward(PassType passType) {
   /* malloc memory for the output_ if necessary */
   int batchSize = getInputValue(0)->getHeight();
   int size = getSize();
-  CHECK_EQ(forward_.size(), 1) << "Only one forward function needed";
+  CHECK_EQ(forward_.size(), 1UL) << "Only one forward function needed";
 
   {
     REGISTER_TIMER_INFO("CosFwResetTimer", getName().c_str());
@@ -68,7 +68,7 @@ void CosSimLayer::forward(PassType passType) {
 void CosSimLayer::backward(const UpdateCallback& callback) {
   /* activation */ {
     REGISTER_TIMER_INFO("CosBpAtvTimer", getName().c_str());
-    CHECK_EQ(backward_.size(), 1) << "Only one backward function needed";
+    CHECK_EQ(backward_.size(), 1UL) << "Only one backward function needed";
 
     const auto outG = this->getOutputGrad();
     const auto outV = this->getOutputValue();
diff --git a/paddle/gserver/layers/CosSimVecMatLayer.cpp b/paddle/gserver/layers/CosSimVecMatLayer.cpp
index aabafd473a..0f887d8adf 100644
--- a/paddle/gserver/layers/CosSimVecMatLayer.cpp
+++ b/paddle/gserver/layers/CosSimVecMatLayer.cpp
@@ -112,7 +112,7 @@ bool CosSimVecMatLayer::init(const LayerMap& layerMap,
 
 void CosSimVecMatLayer::forward(PassType passType) {
   Layer::forward(passType);
-  CHECK_EQ(forward_.size(), 1) << "Only one forward function needed";
+  CHECK_EQ(forward_.size(), 1UL) << "Only one forward function needed";
 
   MatrixPtr inV0 = getInputValue(0);
   MatrixPtr inV1 = getInputValue(1);
@@ -145,7 +145,7 @@ void CosSimVecMatLayer::forward(PassType passType) {
 }
 
 void CosSimVecMatLayer::backward(const UpdateCallback& callback) {
-  CHECK_EQ(backward_.size(), 1) << "Only one forward function needed";
+  CHECK_EQ(backward_.size(), 1UL) << "Only one forward function needed";
 
   MatrixPtr inV0 = getInputValue(0);
   MatrixPtr inV1 = getInputValue(1);
diff --git a/paddle/math/tests/test_RowBuffer.cpp b/paddle/math/tests/test_RowBuffer.cpp
index 5f66f22ef7..8cc4c69a1a 100644
--- a/paddle/math/tests/test_RowBuffer.cpp
+++ b/paddle/math/tests/test_RowBuffer.cpp
@@ -17,10 +17,10 @@ limitations under the License. */
 
 TEST(RowBuffer, testAutoGrow) {
   paddle::RowBuffer buf(128);
-  ASSERT_EQ(128, buf.getWidth());
+  ASSERT_EQ(128UL, buf.getWidth());
   ASSERT_TRUE(buf.isAutoGrowth());
   buf.resize(2);
-  ASSERT_EQ(2, buf.getRowCount());
+  ASSERT_EQ(2UL, buf.getRowCount());
   for (size_t i = 0; i < buf.getWidth() * 2; ++i) {
     buf.data()[i] = i;
   }
@@ -35,7 +35,7 @@ TEST(RowBuffer, testAutoGrow) {
     data[i] = i;
   }
 
-  ASSERT_EQ(3, buf.getRowCount());
+  ASSERT_EQ(3UL, buf.getRowCount());
   for (size_t i = 0; i < buf.getRowCount() - 1; ++i) {
     for (size_t j = 0; j < buf.getWidth(); ++j) {
       ASSERT_NEAR(i * buf.getWidth() + j, buf.get(i)[j], 1e-5);
@@ -51,7 +51,7 @@ TEST(RowBuffer, testWithMemBuf) {
       std::make_shared<paddle::CpuMemoryHandle>(128 * 2 * sizeof(real));
   paddle::RowBuffer buf(mem, 128);
   ASSERT_TRUE(!buf.isAutoGrowth());
-  ASSERT_EQ(2, buf.getRowCount());
+  ASSERT_EQ(2UL, buf.getRowCount());
   for (size_t i = 0; i < buf.getWidth() * 2; ++i) {
     buf.data()[i] = i;
   }
diff --git a/paddle/py_paddle/util.py b/paddle/py_paddle/util.py
index ce105d249a..a708def1d2 100644
--- a/paddle/py_paddle/util.py
+++ b/paddle/py_paddle/util.py
@@ -208,7 +208,7 @@ def __monkeypatch_gradient_machine__():
 
         output = dict()
         for name in layerNames:
-            output[name] = __matrix_to_numpy__(self.getLayerOutput(name))
+            output[name] = __arguments_to_numpy__(0, self.getLayerOutput(name))
         return output
 
     swig_paddle.GradientMachine.getLayerOutputs = getLayerOutputs