diff --git a/paddle/fluid/operators/dropout_op.cc b/paddle/fluid/operators/dropout_op.cc
index b5023f391c88ebe77e7d9dd753f3cebdf5a41533..3c28ef30922e6d6ba09b96282619eef15867631e 100644
--- a/paddle/fluid/operators/dropout_op.cc
+++ b/paddle/fluid/operators/dropout_op.cc
@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/dropout_op.h"
+#include <string>
 
 namespace paddle {
 namespace operators {
@@ -57,15 +58,29 @@ class DropoutOpMaker : public framework::OpProtoAndCheckerMaker {
                   "will be dropped.")
         .SetDefault(false);
     AddAttr<int>("seed", "Dropout random seed.").SetDefault(0);
-    AddAttr<bool>("dropout_implementation",
-                  "When it's True, In the training, after set some value"
-                  "to 0 (probability is dropout_prob),"
-                  "all the value will divide (1-dropout_prob)"
-                  "By using this way, will do nothing in the inference program"
-                  "The dropout op can be removed in the inference program."
-                  "The inference program will be more efficient"
-                  "When it's False, same as original")
-        .SetDefault(false);
+    AddAttr<std::string>(
+        "dropout_implementation",
+        "[\"downgrade_in_infer\"|\"upscale_in_train\"]"
+        "There are two kinds of ways to implement dropout"
+        "(the mask below is a tensor have the same shape with input"
+        "the value of mask is 0 or 1, the ratio of 0 is dropout_prob)"
+        "1. downgrade_in_infer(default), downgrade the outcome at inference "
+        "time"
+        "   train: out = input * mask"
+        "   inference: out = input * dropout_prob"
+        "2. upscale_in_train, upscale the outcome at training time, do nothing "
+        "in inference"
+        "   train: out = input * mask / ( 1.0 - dropout_prob )"
+        "   inference: out = input"
+        "   dropout op can be removed from the program. the program will be "
+        "efficient")
+        .SetDefault("downgrade_in_infer")
+        .AddCustomChecker([](const std::string& type) {
+          PADDLE_ENFORCE(
+              type == "downgrade_in_infer" || type == "upscale_in_train",
+              "dropout_implementation can only be downgrade_in_infer or "
+              "upscale_in_train");
+        });
 
     AddComment(R"DOC(
 Dropout Operator.
diff --git a/paddle/fluid/operators/dropout_op.cu b/paddle/fluid/operators/dropout_op.cu
index a3d264ac13624384a20e21b895750a6032369e47..e011f47e086183a4ef3a3373c17acd6c21b6cf7e 100644
--- a/paddle/fluid/operators/dropout_op.cu
+++ b/paddle/fluid/operators/dropout_op.cu
@@ -17,6 +17,7 @@ limitations under the License. */
 #include <thrust/iterator/counting_iterator.h>
 #include <thrust/random.h>
 #include <thrust/transform.h>
+#include <string>
 #include "paddle/fluid/operators/dropout_op.h"
 #include "paddle/fluid/platform/float16.h"
 
@@ -27,7 +28,7 @@ template <typename T>
 __global__ void RandomGenerator(const size_t n, const int seed,
                                 const float dropout_prob, const T* src,
                                 T* mask_data, T* dst,
-                                bool dropout_implementation) {
+                                bool is_upscale_in_train) {
   thrust::minstd_rand rng;
   rng.seed(seed);
   thrust::uniform_real_distribution<float> dist(0, 1);
@@ -48,7 +49,7 @@ __global__ void RandomGenerator(const size_t n, const int seed,
     if (dist(rng) < dropout_prob) {
       mask = static_cast<T>(0);
     } else {
-      if (dropout_implementation) {
+      if (is_upscale_in_train) {
         mask = static_cast<T>(1.0f / (1.0f - dropout_prob));
       } else {
         mask = static_cast<T>(1);
@@ -72,7 +73,8 @@ class GPUDropoutKernel : public framework::OpKernel<T> {
     y->mutable_data<T>(context.GetPlace());
     float dropout_prob = context.Attr<float>("dropout_prob");
 
-    auto dropout_implementation = context.Attr<bool>("dropout_implementation");
+    auto dropout_implementation =
+        context.Attr<std::string>("dropout_implementation");
     auto& place = *context.template device_context<Place>().eigen_device();
     if (!context.Attr<bool>("is_test")) {
       auto* mask = context.Output<Tensor>("Mask");
@@ -90,11 +92,11 @@ class GPUDropoutKernel : public framework::OpKernel<T> {
       RandomGenerator<
           T><<<grid, threads, 0, context.cuda_device_context().stream()>>>(
           size, seed, dropout_prob, x_data, mask_data, y_data,
-          dropout_implementation);
+          (dropout_implementation == "upscale_in_train"));
     } else {
       auto X = EigenMatrix<T>::Reshape(*x, 1);
       auto Y = EigenMatrix<T>::Reshape(*y, 1);
-      if (dropout_implementation) {
+      if (dropout_implementation == "upscale_in_train") {
         Y.device(place) = X;
       } else {
         Y.device(place) = X * static_cast<T>(1.0f - dropout_prob);
diff --git a/paddle/fluid/operators/dropout_op.h b/paddle/fluid/operators/dropout_op.h
index bc86aeb7f0944e56184929989def61e821a33b05..6c629b7b6d255828023ed25680675ca104a33e12 100644
--- a/paddle/fluid/operators/dropout_op.h
+++ b/paddle/fluid/operators/dropout_op.h
@@ -14,6 +14,7 @@ limitations under the License. */
 #pragma once
 
 #include <random>
+#include <string>
 
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/op_registry.h"
@@ -36,7 +37,8 @@ class CPUDropoutKernel : public framework::OpKernel<T> {
     auto* y_data = y->mutable_data<T>(context.GetPlace());
     float dropout_prob = context.Attr<float>("dropout_prob");
 
-    auto dropout_implementation = context.Attr<bool>("dropout_implementation");
+    auto dropout_implementation =
+        context.Attr<std::string>("dropout_implementation");
     if (!context.Attr<bool>("is_test")) {
       auto* mask = context.Output<Tensor>("Mask");
       auto* mask_data = mask->mutable_data<T>(context.GetPlace());
@@ -57,7 +59,7 @@ class CPUDropoutKernel : public framework::OpKernel<T> {
           mask_data[i] = 0;
           y_data[i] = 0;
         } else {
-          if (dropout_implementation) {
+          if (dropout_implementation == "upscale_in_train") {
             mask_data[i] = 1.0f / static_cast<T>(1.0f - dropout_prob);
             y_data[i] = x_data[i] / static_cast<T>(1.0f - dropout_prob);
           } else {
@@ -71,7 +73,7 @@ class CPUDropoutKernel : public framework::OpKernel<T> {
       auto Y = EigenMatrix<T>::Reshape(*y, 1);
       auto& place =
           *context.template device_context<DeviceContext>().eigen_device();
-      if (dropout_implementation) {
+      if (dropout_implementation == "upscale_in_train") {
         Y.device(place) = X;
       } else {
         Y.device(place) = X * static_cast<T>(1.0f - dropout_prob);
diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py
index 83446e4bd16dd2a324785b29e40ddf50ba1d9d77..98f4539feb65624e7cc1566c733b750d1324dee4 100644
--- a/python/paddle/fluid/layers/nn.py
+++ b/python/paddle/fluid/layers/nn.py
@@ -985,7 +985,7 @@ def dropout(x,
             is_test=False,
             seed=None,
             name=None,
-            dropout_implementation=False):
+            dropout_implementation="downgrade_in_infer"):
     """
     Computes dropout.
 
@@ -1005,13 +1005,20 @@ def dropout(x,
                     units will be dropped. DO NOT use a fixed seed in training.
         name (str|None): A name for this layer(optional). If set None, the layer
                          will be named automatically.
-        dropout_implementation(bool): A Flag indicating whether divide (1-dropout_prob). 
-                                      When it's True, all the units will divide (1-dropout_prob)
-                                      after set some units to zero in the train program.
-                                      And do nothing in the inference program.
-                                      The dropout op can be removed in the inference program.
-                                      The inference program will be more efficient
-                                      When it's False, same as original
+        dropout_implementation(string): ['downgrade_in_infer'(defauld)|'upscale_in_train']
+                                        1. downgrade_in_infer(default), downgrade the outcome at inference
+                                           train: out = input * mask
+                                           inference: out = input * dropout_prob
+                                           (make is a tensor same shape with input, value is 0 or 1
+                                            ratio of 0 is dropout_prob)
+                                        2. upscale_in_train, upscale the outcome at training time
+                                           train: out = input * mask / ( 1.0 - dropout_prob )
+                                           inference: out = input
+                                           (make is a tensor same shape with input, value is 0 or 1
+                                            ratio of 0 is dropout_prob)
+                                           dropout op can be removed from the program. 
+                                           the program will be efficient
+                                        
 
 
     Returns:
diff --git a/python/paddle/fluid/tests/unittests/test_dropout_op.py b/python/paddle/fluid/tests/unittests/test_dropout_op.py
index ecfacb3277b9d75bb8f198cc3f7ad2cf935a1efb..be3c5f3b9558ec522803ed9a5acedea75cda6ccc 100644
--- a/python/paddle/fluid/tests/unittests/test_dropout_op.py
+++ b/python/paddle/fluid/tests/unittests/test_dropout_op.py
@@ -93,7 +93,7 @@ class TestDropoutOp6(TestDropoutOp):
             'dropout_prob': 1.0,
             'fix_seed': True,
             'is_test': False,
-            'div_prob_in_train': True
+            'dropout_implementation': 'upscale_in_train'
         }
         self.outputs = {
             'Out': np.zeros((32, 64)).astype('float32'),
@@ -109,7 +109,7 @@ class TestDropoutOp7(TestDropoutOp):
             'dropout_prob': 0.0,
             'fix_seed': True,
             'is_test': False,
-            'div_prob_in_train': True
+            'dropout_implementation': 'upscale_in_train'
         }
         self.outputs = {
             'Out': self.inputs['X'],
@@ -125,7 +125,7 @@ class TestDropoutOp8(OpTest):
             'dropout_prob': 0.35,
             'fix_seed': True,
             'is_test': True,
-            'div_prob_in_train': True
+            'dropout_implementation': 'upscale_in_train'
         }
         self.outputs = {'Out': self.inputs['X']}
 
@@ -140,7 +140,7 @@ class TestDropoutOp9(OpTest):
         self.attrs = {
             'dropout_prob': 0.75,
             'is_test': True,
-            'div_prob_in_train': True
+            'dropout_implementation': 'upscale_in_train'
         }
         self.outputs = {'Out': self.inputs['X']}