np.tile 和np.newaxis

1. output
2.  
   array([[ 0.24747071, -0.43886742],
3.  
   [-0.03916734, -0.70580089],
4.  
   [ 0.00462337, -0.51431584],
5.  
   ...,
6.  
   [ 0.15071507, -0.57029653],
7.  
   [ 0.06246116, -0.33766761],
8.  
   [ 0.08218585, -0.59906501]], dtype=float32)
9.  
    
10.  
    ipdb> np.shape(output)
11.  
    (64, 2)
12.  
     
13.  
    ipdb> np.max(output, axis=1)[:,np.newaxis]
14.  
    array([[ 0.24747071],
15.  
    [-0.03916734],
16.  
    [ 0.00462337],
17.  
    ...,
18.  
    [ 0.15071507],
19.  
    [ 0.06246116],
20.  
    [ 0.08218585]], dtype=float32)
21.  
     
22.  
    ipdb> np.tile(np.max(output, axis=1)[:,np.newaxis], [1,2]))
23.  
    *** SyntaxError: invalid syntax (<stdin>, line 1)
24.  
     
25.  
    ipdb> np.tile(np.max(output, axis=1)[:,np.newaxis], [1,2])
26.  
    array([[ 0.24747071, 0.24747071],
27.  
    [-0.03916734, -0.03916734],
28.  
    [ 0.00462337, 0.00462337],
29.  
    ...,
30.  
    [ 0.15071507, 0.15071507],
31.  
    [ 0.06246116, 0.06246116],
32.  
    [ 0.08218585, 0.08218585]], dtype=float32)
33.  
     
34.  
    ipdb> output
35.  
    array([[ 0.24747071, -0.43886742],
36.  
    [-0.03916734, -0.70580089],
37.  
    [ 0.00462337, -0.51431584],
38.  
    ...,
39.  
    [ 0.15071507, -0.57029653],
40.  
    [ 0.06246116, -0.33766761],
41.  
    [ 0.08218585, -0.59906501]], dtype=float32)
42.  
     
43.  
    ipdb> np.max(output, axis=1)
44.  
    array([ 0.24747071, -0.03916734, 0.00462337, ..., 0.15071507,
45.  
    0.06246116, 0.08218585], dtype=float32)
46.  
    ipdb> np.exp(output - np.tile(np.max(output, axis=1)[:,np.newaxis], [1,2]))
47.  
    array([[ 1.        ,  0.50341612],
48.  
           [ 1.        ,  0.51343411],
49.  
           [ 1.        ,  0.59515154],
50.  
           ...,
51.  
           [ 1.        ,  0.48626012],
52.  
           [ 1.        ,  0.67023373],
53.  
           [ 1.        ,  0.50598365]], dtype=float32)

1- np.newaxis

np.newaxis的功能是插入新维度，看下面的例子：

a=np.array([1,2,3,4,5])
print a.shape

print a

输出结果

(5,)
[1 2 3 4 5]

可以看出a是一个一维数组,

x_data=np.linspace(-1,1,300)[:,np.newaxis]
a=np.array([1,2,3,4,5])
b=a[np.newaxis,:]
print a.shape,b.shape
print a

print b

输出结果：

(5,) (1, 5)
[1 2 3 4 5]
[[1 2 3 4 5]]

x_data=np.linspace(-1,1,300)[:,np.newaxis]
a=np.array([1,2,3,4,5])
b=a[:,np.newaxis]
print a.shape,b.shape
print a
print b

输出结果

(5,) (5, 1)
[1 2 3 4 5]
[[1]
 [2]
 [3]
 [4]
 [5]]

可以看出np.newaxis分别是在行或列上增加维度，原来是（6，）的数组，在行上增加维度变成（1,6）的二维数组，在列上增加维度变为(6,1)的二维数组

2. np.tile

函数原型：numpy.tile(A,reps) #简单理解是此函数将A进行重复输出

其中A和reps都是array_like的参数，A可以是：array,list,tuple,dict,matrix以及基本数据类型int,string,float以及bool类型,reps的类型可以是tuple,list,dict,array,int,bool,但不可以是float,string,matrix类型。

计较常用的形式有两种，是将A简单进行一维重复输出，和将A进行二维重复后输出。
一维重复：

1.  
   import numpy as np
2.  
   a = [[1,2,3],[4,5,5]]
3.  
   b = np.tile(a,3)
4.  
   print(b)
5.  
    
6.  
   #输出为
7.  
   #[[1 2 3 1 2 3 1 2 3]
8.  
   # [4 5 5 4 5 5 4 5 5]]

二维重复：#上面的一维重复相当于 b = np.tile(a,[1,3])

1.  
   import numpy as np
2.  
   a = [[1,2,3],[4,5,5]]
3.  
   b = np.tile(a,[2,3])
4.  
   print(b)
5.  
    
6.  
   #输出为：
7.  
   #[[1 2 3 1 2 3 1 2 3]
8.  
   # [4 5 5 4 5 5 4 5 5]
9.  
   # [1 2 3 1 2 3 1 2 3]
10.  
    # [4 5 5 4 5 5 4 5 5]]

2.1 np.tile

numpy.tile()是个什么函数呢，说白了，就是把数组沿各个方向复制

比如 a = np.array([0,1,2]),    np.tile(a,(2,1))就是把a先沿x轴（就这样称呼吧）复制1倍，即没有复制，仍然是 [0,1,2]。 再把结果沿y方向复制2倍，即最终得到

array([[0,1,2],

[0,1,2]])

同理：

>>> b = np.array([[1, 2], [3, 4]])
>>> np.tile(b, 2) #沿X轴复制2倍
array([[1, 2, 1, 2],
       [3, 4, 3, 4]])
>>> np.tile(b, (2, 1))#沿X轴复制1倍（相当于没有复制），再沿Y轴复制2倍
array([[1, 2],
       [3, 4],
       [1, 2],
       [3, 4]])

numpy.tile()具体细节，如下：

numpy.tile(A, reps)

Construct an array by repeating A the number of times given by reps.

If reps has length d, the result will have dimension of max(d, A.ndim).

If A.ndim < d, A is promoted to be d-dimensional by prepending new axes. So a shape (3,) array is promoted to (1, 3) for 2-D replication, or shape (1, 1, 3) for 3-D replication. If this is not the desired behavior, promote A to d-dimensions manually before calling this function.

If A.ndim > d, reps is promoted to A.ndim by pre-pending 1’s to it. Thus for an A of shape (2, 3, 4, 5), a repsof (2, 2) is treated as  (1, 1, 2, 2).

Note : Although tile may be used for broadcasting, it is strongly recommended to use numpy’s broadcasting operations and functions.

Parameters:	A : array_like The input array. reps : array_like The number of repetitions of A along each axis.
Returns:	c : ndarray The tiled output array.

See also

repeat

Repeat elements of an array.

broadcast_to

Broadcast an array to a new shape

Examples

>>>

>>> a = np.array([0, 1, 2])
>>> np.tile(a, 2)
array([0, 1, 2, 0, 1, 2])
>>> np.tile(a, (2, 2))
array([[0, 1, 2, 0, 1, 2],
       [0, 1, 2, 0, 1, 2]])
>>> np.tile(a, (2, 1, 2))
array([[[0, 1, 2, 0, 1, 2]],
       [[0, 1, 2, 0, 1, 2]]])

>>>

>>> b = np.array([[1, 2], [3, 4]])
>>> np.tile(b, 2)
array([[1, 2, 1, 2],
       [3, 4, 3, 4]])
>>> np.tile(b, (2, 1))
array([[1, 2],
       [3, 4],
       [1, 2],
       [3, 4]])

>>>

>>> c = np.array([1,2,3,4])
>>> np.tile(c,(4,1))
array([[1, 2, 3, 4],
       [1, 2, 3, 4],
       [1, 2, 3, 4],
       [1, 2, 3, 4]])

原文：https://blog.csdn.net/wuguangbin1230/article/details/77678785