如何理解深度学习中的tensor（涉及numpy的二维，三维以及四维数据的理解）numpy-axis和keepdims

https://www.cnblogs.com/douzujun/p/10289560.html

https://blog.csdn.net/fangjian1204/article/details/53055219 

---

转载自：

https://blog.csdn.net/holmes_mx/article/details/82813865

 

如何理解深度学习中的tensor（涉及numpy的二维，三维以及四维数据的理解）

作者：

holmes_MX

---

自己加的一些理解：

 

import numpy as np
c = [ [[[0.1, 0.2, 0.3], [0.4, 0.5, 0.6], [0.7, 0.8, 0.9], [1.0, 1.1, 1.2]],
[[0.11, 0.21, 0.31], [0.41, 0.51, 0.61], [0.71, 0.81, 0.91], [1.01, 1.11, 1.21]] ],
      [[[0.12, 0.22, 0.32], [0.42, 0.52, 0.62], [0.72, 0.82, 0.92], [1.02, 1.12, 1.22]],
       [[0.112, 0.212, 0.312], [0.412, 0.512, 0.612], [0.712, 0.812, 0.912], [1.012, 1.112, 1.212]]]
      ]
c_array = np.array( c )

print(  c_array )
print( c_array.shape )

c_array_1 = c_array.transpose(0,2,3,1)
print(  c_array_1 )
print( c_array_1.shape )

N,H,W,C = c_array_1.shape 
c_array_2 = c_array_1.reshape(N*H*W,C)
print(  c_array_2 )
print( c_array_2.shape )

 

[[[[0.1   0.2   0.3  ]
   [0.4   0.5   0.6  ]
   [0.7   0.8   0.9  ]
   [1.    1.1   1.2  ]]

  [[0.11  0.21  0.31 ]
   [0.41  0.51  0.61 ]
   [0.71  0.81  0.91 ]
   [1.01  1.11  1.21 ]]]


 [[[0.12  0.22  0.32 ]
   [0.42  0.52  0.62 ]
   [0.72  0.82  0.92 ]
   [1.02  1.12  1.22 ]]

  [[0.112 0.212 0.312]
   [0.412 0.512 0.612]
   [0.712 0.812 0.912]
   [1.012 1.112 1.212]]]]
(2, 2, 4, 3)
[[[[0.1   0.11 ]
   [0.2   0.21 ]
   [0.3   0.31 ]]

  [[0.4   0.41 ]
   [0.5   0.51 ]
   [0.6   0.61 ]]

  [[0.7   0.71 ]
   [0.8   0.81 ]
   [0.9   0.91 ]]

  [[1.    1.01 ]
   [1.1   1.11 ]
   [1.2   1.21 ]]]


 [[[0.12  0.112]
   [0.22  0.212]
   [0.32  0.312]]

  [[0.42  0.412]
   [0.52  0.512]
   [0.62  0.612]]

  [[0.72  0.712]
   [0.82  0.812]
   [0.92  0.912]]

  [[1.02  1.012]
   [1.12  1.112]
   [1.22  1.212]]]]
(2, 4, 3, 2)
[[0.1   0.11 ]
 [0.2   0.21 ]
 [0.3   0.31 ]
 [0.4   0.41 ]
 [0.5   0.51 ]
 [0.6   0.61 ]
 [0.7   0.71 ]
 [0.8   0.81 ]
 [0.9   0.91 ]
 [1.    1.01 ]
 [1.1   1.11 ]
 [1.2   1.21 ]
 [0.12  0.112]
 [0.22  0.212]
 [0.32  0.312]
 [0.42  0.412]
 [0.52  0.512]
 [0.62  0.612]
 [0.72  0.712]
 [0.82  0.812]
 [0.92  0.912]
 [1.02  1.012]
 [1.12  1.112]
 [1.22  1.212]]
(24, 2)

---

1. 二维数据的理解

 理解为：只有(Height, Width)，而且理解的是对于每一【 】从后向前看。即第一个[0.1， 0.2， 0.3]是width维度的。

b = [[0.1, 0.2, 0.3], [0.4, 0.5, 0.6]]
b_array = np.array( b )
 
print(b_array)
print(b_array.shape)
print( b_array[1,2] )
 
### result
 
[[0.1 0.2 0.3]
 [0.4 0.5 0.6]]
(2, 3)
0.6

2. 对三维数据的理解

 理解为：只有(Height, Width, Channel)，而且理解的是对于每一【 】从后向前看。即第一个[1， 2， 3, 3.5]是Channel维度的。

a = [ [[1, 2, 3, 3.5], [4, 5, 6, 6.5]], 
[[7, 8, 9, 9.5], [10, 11, 12, 12.5]], 
[[13, 14, 15, 15.5], [16, 17, 18, 18.5]]]
 
a_array = np.array( a )
print(a_array)
print( a_array.shape )
print( a_array[0, 1, 3] )
 
 
### result
[[[ 1.   2.   3.   3.5]
  [ 4.   5.   6.   6.5]]
 
 [[ 7.   8.   9.   9.5]
  [10.  11.  12.  12.5]]
 
 [[13.  14.  15.  15.5]
  [16.  17.  18.  18.5]]]
(3, 2, 4)
6.5

可视化上述的三维数据(字迹丑大家见谅，理解原理优先)：

3. 对四维数据的理解

 理解为：(Batch_size, Height, Width, Channel)，而且理解的是对于每一【 】从后向前看。即第一个[0.1， 0.2， 0.3]是Channel维度的。

c = [ [[[0.1, 0.2, 0.3], [0.4, 0.5, 0.6], [0.7, 0.8, 0.9], [1.0, 1.1, 1.2]],
[[0.11, 0.21, 0.31], [0.41, 0.51, 0.61], [0.71, 0.81, 0.91], [1.01, 1.11, 1.21]] ],
      [[[0.12, 0.22, 0.32], [0.42, 0.52, 0.62], [0.72, 0.82, 0.92], [1.02, 1.12, 1.22]],
       [[0.112, 0.212, 0.312], [0.412, 0.512, 0.612], [0.712, 0.812, 0.912], [1.012, 1.112, 1.212]]]
      ]
c_array = np.array( c )
print(  c_array )
print( c_array.shape )
print( c_array[1, 0, 3, 1] )
 
 
## result
[[[[0.1   0.2   0.3  ]
   [0.4   0.5   0.6  ]
   [0.7   0.8   0.9  ]
   [1.    1.1   1.2  ]]
 
  [[0.11  0.21  0.31 ]
   [0.41  0.51  0.61 ]
   [0.71  0.81  0.91 ]
   [1.01  1.11  1.21 ]]]
 
 
 [[[0.12  0.22  0.32 ]
   [0.42  0.52  0.62 ]
   [0.72  0.82  0.92 ]
   [1.02  1.12  1.22 ]]
 
  [[0.112 0.212 0.312]
   [0.412 0.512 0.612]
   [0.712 0.812 0.912]
   [1.012 1.112 1.212]]]]
(2, 2, 4, 3)
1.12

4. tensor的简单操作
4.1 numpy 中对tensor的操作

以三维tensor为例：

numpy.sum( tensor, axis ) numpy.mean( tensor, axis)

其中axis = -1是对最里面的一个维度操作。如numpy.sum( tensor, axis = -1 )， 即对第2维度进行操作，即对channel进行相加。

实例为：

import numpy as np
a = [[[1, 2, 3,], [4, 5, 6]], [[1.1, 2.1, 3.1], [4.1, 5.1, 6.1]]]
print( np.sum(a, axis=-1) )
print( np.sum(a, axis=0) )
print( np.sum(a, axis=1) )
print( np.sum(a, axis=2) )
 
 
## output
#axis = -1  ## channel
[[ 6.  15. ]
 [ 6.3 15.3]]
 
# axis = 0  ## height
[[ 2.1  4.1  6.1]
 [ 8.1 10.1 12.1]]
 
# axis = 1   ## width
[[5.  7.  9. ]
 [5.2 7.2 9.2]]
 
# axis = 2  ## channel
[[ 6.  15. ]
 [ 6.3 15.3]]