diff --git a/paddle/fluid/operators/similarity_focus_op.cc b/paddle/fluid/operators/similarity_focus_op.cc
index 25a023aed2bbc9cd459fca2f724148f5e29e246b..768b6903b741c58d565f5b98da3349996f9ae76a 100644
--- a/paddle/fluid/operators/similarity_focus_op.cc
+++ b/paddle/fluid/operators/similarity_focus_op.cc
@@ -35,14 +35,17 @@ class SimilarityFocusOpMaker : public framework::OpProtoAndCheckerMaker {
 SimilarityFocus Operator.
 
 Generate a similarity focus mask with the same shape of input using the following method:
-1. Extract the 4-D matrix(here the first dimension is BatchSize) corresponding
-   to the axis according to the indexes. For example, if axis=1 and indexes=[a],
-   it will get the matrix T=X[:, a, :, :]. In this case, if the shape of input X
-   is (BatchSize, A, B, C), the shape of matrix T is (BatchSize, B, C).
-2. For each index, find the largest numbers in the matrix T, so that the same
-   row and same column has at most one number(obviously there will be min(B, C)
-   numbers), and mark the corresponding position of the 3-D similarity focus mask
-   as 1, otherwise as 0. Do elementwise-or for each index.
+1. Extract the 3-D tensor(here the first dimension is BatchSize) corresponding 
+   to the axis according to the indexes. For example, if axis=1 and indexes=[a], 
+   it will get the matrix T=X[:, a, :, :]. In this case, if the shape of input X 
+   is (BatchSize, A, B, C), the shape of tensor T is (BatchSize, B, C).
+2. For each index, find the largest numbers in the tensor T, so that the same 
+   row and same column has at most one number(what it means is that if the 
+   largest number has been found in the i-th row and the j-th column, then 
+   the numbers in the i-th or j-th column will be skipped. Obviously there 
+   will be min(B, C) numbers), and mark the corresponding position of the 
+   3-D similarity focus mask as 1, otherwise as 0. Do elementwise-or for 
+   each index.
 3. Broadcast the 3-D similarity focus mask to the same shape of input X.
 
 Refer to `Similarity Focus Layer <http://www.aclweb.org/anthology/N16-1108>`_
diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py
index a1ef1ca00974c1338679e8ed6eefe1263d59f1e0..be0e75161bbff10dd351c24f7a56e022f20ee036 100644
--- a/python/paddle/fluid/layers/nn.py
+++ b/python/paddle/fluid/layers/nn.py
@@ -7560,14 +7560,17 @@ def similarity_focus(input, axis, indexes, name=None):
     SimilarityFocus Operator
 
     Generate a similarity focus mask with the same shape of input using the following method:
-    1. Extract the 4-D matrix(here the first dimension is BatchSize) corresponding 
+    1. Extract the 3-D tensor(here the first dimension is BatchSize) corresponding 
        to the axis according to the indexes. For example, if axis=1 and indexes=[a], 
        it will get the matrix T=X[:, a, :, :]. In this case, if the shape of input X 
-       is (BatchSize, A, B, C), the shape of matrix T is (BatchSize, B, C).
-    2. For each index, find the largest numbers in the matrix T, so that the same 
-       row and same column has at most one number(obviously there will be min(B, C) 
-       numbers), and mark the corresponding position of the 3-D similarity focus mask 
-       as 1, otherwise as 0. Do elementwise-or for each index.
+       is (BatchSize, A, B, C), the shape of tensor T is (BatchSize, B, C).
+    2. For each index, find the largest numbers in the tensor T, so that the same 
+       row and same column has at most one number(what it means is that if the 
+       largest number has been found in the i-th row and the j-th column, then 
+       the numbers in the i-th or j-th column will be skipped. Obviously there 
+       will be min(B, C) numbers), and mark the corresponding position of the 
+       3-D similarity focus mask as 1, otherwise as 0. Do elementwise-or for 
+       each index.
     3. Broadcast the 3-D similarity focus mask to the same shape of input X.
 
     Refer to `Similarity Focus Layer <http://www.aclweb.org/anthology/N16-1108>`_
@@ -7624,9 +7627,9 @@ def similarity_focus(input, axis, indexes, name=None):
     Args:
         input(Variable): The input tensor variable(default float). It should 
             be a 4-D tensor with shape [BatchSize, A, B, C].
-        axis(int): Indicating the dimension to be select. It can only be
+        axis(int): Indicating the dimension to be selected. It can only be
             1, 2 or 3.
-        indexes(list): indicating the indexes of the selected dimension.
+        indexes(list): Indicating the indexes of the selected dimension.
 
     Returns:
         Variable: A tensor variable with the same shape and same type 
@@ -7649,7 +7652,11 @@ def similarity_focus(input, axis, indexes, name=None):
     if len(indexes) == 0:
         raise ValueError("indexes can not be empty.")
 
-    out = helper.create_tmp_variable(dtype=helper.input_dtype())
+    if name is None:
+        out = helper.create_variable_for_type_inference(dtype=input.dtype)
+    else:
+        out = helper.create_variable(
+            name=name, dtype=input.dtype, persistable=False)
     helper.append_op(
         type='similarity_focus',
         inputs={'X': input},