(5)tensorflow张量的变换

张量的变换

功能	代码
改变视图	tf.reshape(x,shape)
增加维度	tf.expand_dims(x, axis)
删除维度	tf.squeeze(x, axis)
交换维度	tf.transpose(x, perm)
复制数据	tf.tile(x, multiples)

改变视图

• tf.reshape(x,shape)形成新形状
• 视图变换只需要满足新视图的元素总量与存储区域大小相等即可
• x.ndim可以查询维度，x.shape查询形状

import tensorflow as tf
x = tf.range(96)
y = tf.reshape(x,[2,4,4,3])
print(x,'\n',y)

out:

tf.Tensor(
[ 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47
 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71
 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95], shape=(96,), dtype=int32) 
 tf.Tensor(
[[[[ 0  1  2]
   [ 3  4  5]
   [ 6  7  8]
   [ 9 10 11]]

  [[12 13 14]
   [15 16 17]
   [18 19 20]
   [21 22 23]]

  [[24 25 26]
   [27 28 29]
   [30 31 32]
   [33 34 35]]

  [[36 37 38]
   [39 40 41]
   [42 43 44]
   [45 46 47]]]


 [[[48 49 50]
   [51 52 53]
   [54 55 56]
   [57 58 59]]

  [[60 61 62]
   [63 64 65]
   [66 67 68]
   [69 70 71]]

  [[72 73 74]
   [75 76 77]
   [78 79 80]
   [81 82 83]]

  [[84 85 86]
   [87 88 89]
   [90 91 92]
   [93 94 95]]]], shape=(2, 4, 4, 3), dtype=int32)

import tensorflow as tf
x = tf.range(96)
y = tf.reshape(x,[2,4,4,-1])
print(x,'\n',y)

out:
tf.Tensor(
[ 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47
 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71
 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95], shape=(96,), dtype=int32) 
 tf.Tensor(
[[[[ 0  1  2]
   [ 3  4  5]
   [ 6  7  8]
   [ 9 10 11]]

  [[12 13 14]
   [15 16 17]
   [18 19 20]
   [21 22 23]]

  [[24 25 26]
   [27 28 29]
   [30 31 32]
   [33 34 35]]

  [[36 37 38]
   [39 40 41]
   [42 43 44]
   [45 46 47]]]


 [[[48 49 50]
   [51 52 53]
   [54 55 56]
   [57 58 59]]

  [[60 61 62]
   [63 64 65]
   [66 67 68]
   [69 70 71]]

  [[72 73 74]
   [75 76 77]
   [78 79 80]
   [81 82 83]]

  [[84 85 86]
   [87 88 89]
   [90 91 92]
   [93 94 95]]]], shape=(2, 4, 4, 3), dtype=int32)

import tensorflow as tf
x = tf.range(96)
y = tf.reshape(x,[2,4,4,-1])

print(y.ndim)
print(y.shape)

out:
4
(2, 4, 4, 3)

增、 删维度

• tf.expand_dims(x, axis)可以插入一个指定的新维度（类似指针）
• axis 可为正负
• tf.squeeze(x, axis)函数， axis 参数为待删除的维度的索引号，缺省时删除所有为1的维度

import tensorflow as tf
x = tf.range(96)
y = tf.reshape(x,[2,4,4,3])
r = tf.expand_dims(y,axis = 4)
r1 = tf.expand_dims(y,axis=-4)
print(r.shape)
print(r1.shape)

out:
(2, 4, 4, 3, 1)
(2, 1, 4, 4, 3)

import tensorflow as tf
x = tf.range(96)
y = tf.reshape(x,[2,4,4,3])
r = tf.expand_dims(y,axis = 4)
r1 = tf.expand_dims(y,axis=-4)
print(r.shape)
print(r1.shape)
r2 = tf.squeeze(r,axis = 4)
r3 = tf.squeeze(r1)
print(r2.shape)
print(r3.shape)

out:
(2, 4, 4, 3, 1)
(2, 1, 4, 4, 3)
(2, 4, 4, 3)
(2, 4, 4, 3)

交换维度

• tf.transpose(x, perm) ,perm为列表形式的新维数
• 以[b, h,w, c]转换到[b, c, h,w]为例：tf.transpose(x, perm)，perm=[0,3,1,2]

import tensorflow as tf
x = tf.range(27)
y = tf.reshape(x,[3,3,3])
y1 = tf.transpose(y,[0,2,1])
print(y,'\n',y1)

out:
tf.Tensor(
[[[ 0  1  2]
  [ 3  4  5]
  [ 6  7  8]]

 [[ 9 10 11]
  [12 13 14]
  [15 16 17]]

 [[18 19 20]
  [21 22 23]
  [24 25 26]]], shape=(3, 3, 3), dtype=int32) 
 tf.Tensor(
[[[ 0  3  6]
  [ 1  4  7]
  [ 2  5  8]]

 [[ 9 12 15]
  [10 13 16]
  [11 14 17]]

 [[18 21 24]
  [19 22 25]
  [20 23 26]]], shape=(3, 3, 3), dtype=int32)

复制数据

• tf.tile(x, multiples)
• 以x作为单位基准方块，multiples指定单位基准方块的拓展，multiples = [2,2]时如下
  

import tensorflow as tf
a = tf.range(9)
x = tf.reshape(a,[3,3])
b = tf.tile(x,multiples=[2,2])
b1 = tf.tile(x, multiples=[2,1])
print(x,'\n',b,'\n',b1)

out:
tf.Tensor(
[[0 1 2]
 [3 4 5]
 [6 7 8]], shape=(3, 3), dtype=int32) 
 tf.Tensor(
[[0 1 2 0 1 2]
 [3 4 5 3 4 5]
 [6 7 8 6 7 8]
 [0 1 2 0 1 2]
 [3 4 5 3 4 5]
 [6 7 8 6 7 8]], shape=(6, 6), dtype=int32) 
 tf.Tensor(
[[0 1 2]
 [3 4 5]
 [6 7 8]
 [0 1 2]
 [3 4 5]
 [6 7 8]], shape=(6, 3), dtype=int32)

Broadcasting

• tf.broadcast_to(x, new_shape)
• Broadcasting和 tf.tile 复制的最终效果是一样的，操作对用户透明，但是 Broadcasting 机制节省了大量计算资源，建议在运算过程中尽可能地利用 Broadcasting 机制提高计算效率