Merge pull request #947 from reyoung/feature/clean_bn_code

Clean BatchNorm Code.

paddle/gserver/layers/BatchNormalizationLayer.cpp

@@ -59,24 +59,14 @@ void BatchNormalizationLayer::calMeanAndStd(const MatrixPtr& mat) {
 
 void BatchNormalizationLayer::calMovingMeanAndVar() {
   // calculating and saving moving mean and variance
-  MatrixPtr movingMean = movingMean_->getW();
-  MatrixPtr movingVar = movingVar_->getW();
-
-  if (!useGpu_ && FLAGS_trainer_count > 1) {
-    auto mvMean = std::dynamic_pointer_cast<SharedCpuMatrix>(movingMean);
-    auto mvVar = std::dynamic_pointer_cast<SharedCpuMatrix>(movingVar);
-    CHECK(mvMean && mvVar);
-
-    mvMean->add(*savedMean_, movingAvgFraction_, 1.0 - movingAvgFraction_);
-    mvVar->add(*savedInvVar_, movingAvgFraction_, 1.0 - movingAvgFraction_);
-  } else {
+  auto& movingMean = movingMean_->getW();
+  auto& movingVar = movingVar_->getW();
   // movingMean =  movingMean * movingAvgFraction_
   //            + savedMean_ * (1 - movingAvgFraction_)
   movingMean->add(*savedMean_, movingAvgFraction_, 1.0 - movingAvgFraction_);
   // movingVar =  movingVar * movingAvgFraction_
   //           + savedInvVar_ * (1 - movingAvgFraction_)
   movingVar->add(*savedInvVar_, movingAvgFraction_, 1.0 - movingAvgFraction_);
-  }
 }
 
 void BatchNormalizationLayer::setMeanAndStd() {

paddle/math/Matrix.h

@@ -1973,8 +1973,8 @@ public:
 
 public:
   virtual void mul(CpuSparseMatrix* a, CpuMatrix* b, real scaleAB, real scaleT);
-  void add(Matrix& b, real p1, real p2);
-  void add(real p1, real p2);
+  virtual void add(Matrix& b, real p1, real p2);
+  virtual void add(real p1, real p2);
 
 private:
   using Matrix::mul;