diff --git a/paddle/fluid/operators/math/softmax.cc b/paddle/fluid/operators/math/softmax.cc
index 78c65af24a8c5fa57e33415acc3018790bf70790..6300836e500cfee844b3a7193522c06653f5adb0 100644
--- a/paddle/fluid/operators/math/softmax.cc
+++ b/paddle/fluid/operators/math/softmax.cc
@@ -19,8 +19,10 @@ namespace paddle {
 namespace operators {
 namespace math {
 
-template class SoftmaxFunctor<platform::CPUDeviceContext, float>;
-template class SoftmaxFunctor<platform::CPUDeviceContext, double>;
+template class SoftmaxFunctor<platform::CPUDeviceContext, float,true>;
+template class SoftmaxFunctor<platform::CPUDeviceContext, float,false>;
+template class SoftmaxFunctor<platform::CPUDeviceContext, double,true>;
+template class SoftmaxFunctor<platform::CPUDeviceContext, double,false>;
 template class SoftmaxGradFunctor<platform::CPUDeviceContext, float>;
 template class SoftmaxGradFunctor<platform::CPUDeviceContext, double>;
 
diff --git a/paddle/fluid/operators/math/softmax.h b/paddle/fluid/operators/math/softmax.h
index da1f0b672d3a5fb5da8f4d72892be21964bdbc0d..bf698dc2f753f0002557af07ad7ea976c85edada 100644
--- a/paddle/fluid/operators/math/softmax.h
+++ b/paddle/fluid/operators/math/softmax.h
@@ -19,7 +19,7 @@ namespace paddle {
 namespace operators {
 namespace math {
 
-template <typename DeviceContext, typename T>
+template <typename DeviceContext, typename T, bool is_test>
 class SoftmaxFunctor {
  public:
   void operator()(const DeviceContext& context, const framework::Tensor* X,
diff --git a/paddle/fluid/operators/math/softmax_impl.h b/paddle/fluid/operators/math/softmax_impl.h
index dd9971ba091cc3ece86654f65c335b98087f45ed..6a0a6c2e46d6d0338eb880a9e8a07f8c3cc08493 100644
--- a/paddle/fluid/operators/math/softmax_impl.h
+++ b/paddle/fluid/operators/math/softmax_impl.h
@@ -32,8 +32,8 @@ struct ValueClip {
   }
 };
 
-template <typename DeviceContext, typename T>
-void SoftmaxFunctor<DeviceContext, T>::operator()(const DeviceContext& context,
+template <typename DeviceContext, typename T, bool is_test>
+void SoftmaxFunctor<DeviceContext, T, is_test>::operator()(const DeviceContext& context,
                                                   const framework::Tensor* X,
                                                   framework::Tensor* Y) {
   auto logits = EigenMatrix<T>::From(*X);
@@ -65,6 +65,42 @@ void SoftmaxFunctor<DeviceContext, T>::operator()(const DeviceContext& context,
                                                  .broadcast(one_by_class));
 }
 
+template <typename DeviceContext, typename T>
+class SoftmaxFunctor<DeviceContext, T, true> {
+void operator()(const DeviceContext& context,
+                const framework::Tensor* X,
+                framework::Tensor* Y) {
+  auto logits = EigenMatrix<T>::From(*X);
+  auto softmax = EigenMatrix<T>::From(*Y);
+
+  const int kBatchDim = 0;
+  const int kClassDim = 1;
+
+  const int batch_size = logits.dimension(kBatchDim);
+  const int num_classes = logits.dimension(kClassDim);
+
+  Eigen::DSizes<int, 1> along_class(kClassDim);
+  Eigen::DSizes<int, 2> batch_by_one(batch_size, 1);
+  Eigen::DSizes<int, 2> one_by_class(1, num_classes);
+
+  auto shifted_logits = (logits -
+                         logits.maximum(along_class)
+                             .eval()
+                             .reshape(batch_by_one)
+                             .broadcast(one_by_class));
+
+  softmax.device(*context.eigen_device()) = shifted_logits.exp();
+  softmax.device(*context.eigen_device()) = (softmax *
+                                             softmax.sum(along_class)
+                                                 .inverse()
+                                                 .eval()
+                                                 .reshape(batch_by_one)
+                                                 .broadcast(one_by_class));
+}
+};
+
+
+
 template <typename DeviceContext, typename T>
 void SoftmaxGradFunctor<DeviceContext, T>::operator()(
     const DeviceContext& context, const framework::Tensor* y,
diff --git a/paddle/fluid/operators/softmax_op.h b/paddle/fluid/operators/softmax_op.h
index cf1eeb017d666f605a431aa54637d8cbc99c7c46..5bc72aac485acceffd90964582d2a8d748f84adc 100644
--- a/paddle/fluid/operators/softmax_op.h
+++ b/paddle/fluid/operators/softmax_op.h
@@ -35,8 +35,14 @@ class SoftmaxKernel : public framework::OpKernel<T> {
     Tensor X_2d = framework::ReshapeToMatrix(*X, rank - 1);
     Tensor Out_2d = framework::ReshapeToMatrix(*Out, rank - 1);
 
-    math::SoftmaxFunctor<DeviceContext, T>()(
-        context.template device_context<DeviceContext>(), &X_2d, &Out_2d);
+    const bool is_test = context.Attr<bool>("is_test");
+    if( is_test == true) {
+      math::SoftmaxFunctor<DeviceContext, T,true>()(
+          context.template device_context<DeviceContext>(), &X_2d, &Out_2d);
+    } else {
+      math::SoftmaxFunctor<DeviceContext, T,false>()(
+          context.template device_context<DeviceContext>(), &X_2d, &Out_2d);
+    }
   }
 };
 
diff --git a/paddle/fluid/operators/softmax_with_cross_entropy_op.h b/paddle/fluid/operators/softmax_with_cross_entropy_op.h
index e9aba3b37b8cc01d4fe5de5200579d4e93f67e56..2eec8541c84c00b1c951a9d393d3f54aa97d0971 100644
--- a/paddle/fluid/operators/softmax_with_cross_entropy_op.h
+++ b/paddle/fluid/operators/softmax_with_cross_entropy_op.h
@@ -42,7 +42,7 @@ class SoftmaxWithCrossEntropyKernel : public framework::OpKernel<T> {
 
     auto& dev_ctx =
         context.template device_context<platform::CPUDeviceContext>();
-    math::SoftmaxFunctor<platform::CPUDeviceContext, T>()(dev_ctx, logits,
+    math::SoftmaxFunctor<platform::CPUDeviceContext, T, false>()(dev_ctx, logits,
                                                           softmax);
     math::CrossEntropyFunctor<platform::CPUDeviceContext, T>()(
         dev_ctx, loss, softmax, labels, context.Attr<bool>("soft_label"),