Update python getLayerOutputs

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)

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
 

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);

paddle/api/GradientMachine.cpp

@@ -144,12 +144,11 @@ 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 = m->machine;
   if (nn) {
-    auto mat = nn->getLayerOutput(layerName);
-    return Matrix::createByPaddleMatrixPtr(&mat);
+    return Arguments::createByPaddleArgument(&nn->getLayerOutput(layerName));
   } else {
     throw UnsupportError();
   }

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);
 
   /**
@@ -952,7 +953,7 @@ public:
 
   Arguments* getForwardOutput();
 
-  Matrix* getLayerOutput(const std::string& layerName);
+  Arguments* getLayerOutput(const std::string& layerName);
 };
 
 /// the N-Best results generated from one input sequence.

paddle/api/Trainer.cpp

@@ -131,12 +131,10 @@ 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) {
+  auto nn = this->m->getGradientMachine();
   CHECK(nn) << "trainerInternal_.getGradientMachine() is not NeuralNetwork";
-  auto m = nn->getLayerOutput(layerName);
-  return Matrix::createByPaddleMatrixPtr(&m);
+  return Arguments::createByPaddleArgument(&nn->getLayerOutput(layerName));
 }
 
 void Trainer::forwardOneBatch(size_t batchSize) {

paddle/gserver/gradientmachines/GradientMachine.h

@@ -134,8 +134,8 @@ public:
     backward(callback);
   }
 
-  virtual MatrixPtr getLayerOutput(const std::string& layerName) const {
-    return nullptr;
+  virtual const Argument& getLayerOutput(const std::string& layerName) {
+    return *((Argument*)nullptr);
   }
 
   // see comment in Layer.h for the function with the same name

paddle/gserver/gradientmachines/MultiGradientMachine.cpp

@@ -282,33 +282,17 @@ void MultiGradientMachine::forwardBackward(const std::vector<Argument>& inArgs,
   backwardImp(callback);
 }
 
-MatrixPtr MultiGradientMachine::getLayerOutput(
-    const std::string& layerName) const {
-  // each thread has the same neural network
-  auto nn = threads_[0]->getGradientMachine();
-  size_t height = 0;
-  size_t width = nn->getLayerOutput(layerName)->getWidth();
-  std::vector<MatrixPtr> mats;
-  mats.reserve(threads_.size());
-  for (auto& thread : threads_) {
-    MatrixPtr out = thread->getGradientMachine()->getLayerOutput(layerName);
-    mats.push_back(out);
-    height += out->getHeight();
-    CHECK_EQ(width, out->getWidth());
-  }
+const Argument& MultiGradientMachine::getLayerOutput(
+    const std::string& layerName) {
+  std::vector<Argument> args;
+  args.reserve(threads_.size());
 
-  MatrixPtr layerOutput;
-  Matrix::resizeOrCreate(layerOutput, height, width, false, false);
-
-  // copy one layer output from one trainer thread at each time
-  size_t startRow = 0;
-  for (auto& mat : mats) {
-    auto tmpMatrix = layerOutput->subMatrix(startRow, mat->getHeight());
-    tmpMatrix->copyFrom(*mat);
-    startRow += mat->getHeight();
+  for (auto& thread : threads_) {
+    args.push_back(thread->getGradientMachine()->getLayerOutput(layerName));
   }
+  outLayerArgs_.concat(args, false /* use_gpu */, outArgStream_, passType_);
 
-  return layerOutput;
+  return outLayerArgs_;
 }
 
 void MultiGradientMachine::backwardImp(const UpdateCallback& callback) {

paddle/gserver/gradientmachines/MultiGradientMachine.h

@@ -189,7 +189,7 @@ public:
                        PassType passType,
                        const UpdateCallback& callback);
 
-  virtual MatrixPtr getLayerOutput(const std::string& layerName) const;
+  virtual const Argument& getLayerOutput(const std::string& layerName);
 
   virtual void onPassEnd();
 
@@ -316,6 +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 */

paddle/gserver/gradientmachines/NeuralNetwork.cpp

@@ -293,10 +293,8 @@ void NeuralNetwork::backward(const UpdateCallback& callback) {
   }
 }
 
-MatrixPtr NeuralNetwork::getLayerOutput(const std::string& layerName) const {
-  auto it = layerMap_.find(layerName);
-  CHECK(it != layerMap_.end()) << "Cannot find layer: " << layerName;
-  return it->second->getOutputValue();
+const Argument& NeuralNetwork::getLayerOutput(const std::string& layerName) {
+  return getLayer(layerName)->getOutput();
 }
 
 void NeuralNetwork::onPassEnd() {

paddle/gserver/gradientmachines/NeuralNetwork.h

@@ -87,7 +87,7 @@ public:
 
   virtual void backward(const UpdateCallback& callback = nullptr);
 
-  virtual MatrixPtr getLayerOutput(const std::string& layerName) const;
+  virtual const Argument& getLayerOutput(const std::string& layerName);
 
   const LayerPtr& getLayer(const std::string& layerName) const {
     auto it = layerMap_.find(layerName);

paddle/gserver/layers/CosSimLayer.cpp

@@ -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();

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