浅谈pytorch中torch.max和F.softmax函数的维度解释

时间：2021-09-17 00:47:24

在利用torch.max函数和F.Ssoftmax函数时，对应该设置什么维度，总是有点懵，遂总结一下：

首先看看二维tensor的函数的例子：

?
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
import torch
import torch.nn.functional as F
 
input = torch.randn(3,4)
print(input)
tensor([[-0.5526, -0.0194, 2.1469, -0.2567],
    [-0.3337, -0.9229, 0.0376, -0.0801],
    [ 1.4721, 0.1181, -2.6214, 1.7721]])
 
b = F.softmax(input,dim=0) # 按列SoftMax,列和为1
print(b)
tensor([[0.1018, 0.3918, 0.8851, 0.1021],
    [0.1268, 0.1587, 0.1074, 0.1218],
    [0.7714, 0.4495, 0.0075, 0.7762]])
 
c = F.softmax(input,dim=1)  # 按行SoftMax,行和为1
print(c)
tensor([[0.0529, 0.0901, 0.7860, 0.0710],
    [0.2329, 0.1292, 0.3377, 0.3002],
    [0.3810, 0.0984, 0.0064, 0.5143]])
 
d = torch.max(input,dim=0)  # 按列取max,
print(d)
torch.return_types.max(
values=tensor([1.4721, 0.1181, 2.1469, 1.7721]),
indices=tensor([2, 2, 0, 2]))
 
e = torch.max(input,dim=1)  # 按行取max，
print(e)
torch.return_types.max(
values=tensor([2.1469, 0.0376, 1.7721]),
indices=tensor([2, 2, 3]))

下面看看三维tensor解释例子：

函数softmax输出的是所给矩阵的概率分布；

b输出的是在dim=0维上的概率分布，b[0][5][6]+b[1][5][6]+b[2][5][6]=1

?
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
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
a=torch.rand(3,16,20)
b=F.softmax(a,dim=0)
c=F.softmax(a,dim=1)
d=F.softmax(a,dim=2)
 
In [1]: import torch as t
In [2]: import torch.nn.functional as F
In [4]: a=t.Tensor(3,4,5)
In [5]: b=F.softmax(a,dim=0)
In [6]: c=F.softmax(a,dim=1)
In [7]: d=F.softmax(a,dim=2)
 
In [8]: a
Out[8]:
tensor([[[-0.1581, 0.0000, 0.0000, 0.0000, -0.0344],
 
     [ 0.0000, -0.0344, 0.0000, -0.0344, 0.0000],
     [-0.0344, 0.0000, -0.0344, 0.0000, -0.0344],
     [ 0.0000, -0.0344, 0.0000, -0.0344, 0.0000]],
 
    [[-0.0344, 0.0000, -0.0344, 0.0000, -0.0344],
     [ 0.0000, -0.0344, 0.0000, -0.0344, 0.0000],
     [-0.0344, 0.0000, -0.0344, 0.0000, -0.0344],
     [ 0.0000, -0.0344, 0.0000, -0.0344, 0.0000]],
 
 
    [[-0.0344, 0.0000, -0.0344, 0.0000, -0.0344],
     [ 0.0000, -0.0344, 0.0000, -0.0344, 0.0000],
     [-0.0344, 0.0000, -0.0344, 0.0000, -0.0344],
     [ 0.0000, -0.0344, 0.0000, -0.0344, 0.0000]]])
 
In [9]: b
Out[9]:
 
tensor([[[0.3064, 0.3333, 0.3410, 0.3333, 0.3333],
     [0.3333, 0.3333, 0.3333, 0.3333, 0.3333],
     [0.3333, 0.3333, 0.3333, 0.3333, 0.3333],
     [0.3333, 0.3333, 0.3333, 0.3333, 0.3333]],
 
    [[0.3468, 0.3333, 0.3295, 0.3333, 0.3333],
     [0.3333, 0.3333, 0.3333, 0.3333, 0.3333],
     [0.3333, 0.3333, 0.3333, 0.3333, 0.3333],
     [0.3333, 0.3333, 0.3333, 0.3333, 0.3333]],
 
    [[0.3468, 0.3333, 0.3295, 0.3333, 0.3333],
     [0.3333, 0.3333, 0.3333, 0.3333, 0.3333],
     [0.3333, 0.3333, 0.3333, 0.3333, 0.3333],
     [0.3333, 0.3333, 0.3333, 0.3333, 0.3333]]])
 
 
In [10]: b.sum()
Out[10]: tensor(20.0000)
 
In [11]: b[0][0][0]+b[1][0][0]+b[2][0][0]
Out[11]: tensor(1.0000)
 
In [12]: c.sum()
Out[12]: tensor(15.)
 
In [13]: c
Out[13]:
tensor([[[0.2235, 0.2543, 0.2521, 0.2543, 0.2457],
 
     [0.2618, 0.2457, 0.2521, 0.2457, 0.2543],
 
     [0.2529, 0.2543, 0.2436, 0.2543, 0.2457],
 
     [0.2618, 0.2457, 0.2521, 0.2457, 0.2543]],
 
 
    [[0.2457, 0.2543, 0.2457, 0.2543, 0.2457],
 
     [0.2543, 0.2457, 0.2543, 0.2457, 0.2543],
 
     [0.2457, 0.2543, 0.2457, 0.2543, 0.2457],
 
     [0.2543, 0.2457, 0.2543, 0.2457, 0.2543]],
 
 
    [[0.2457, 0.2543, 0.2457, 0.2543, 0.2457],
 
     [0.2543, 0.2457, 0.2543, 0.2457, 0.2543],
 
     [0.2457, 0.2543, 0.2457, 0.2543, 0.2457],
 
     [0.2543, 0.2457, 0.2543, 0.2457, 0.2543]]])
 
In [14]: n=t.rand(3,4)
 
In [15]: n
Out[15]:
 
tensor([[0.2769, 0.3475, 0.8914, 0.6845],
    [0.9251, 0.3976, 0.8690, 0.4510],
    [0.8249, 0.1157, 0.3075, 0.3799]])
 
In [16]: m=t.argmax(n,dim=0)
 
In [17]: m
Out[17]: tensor([1, 1, 0, 0])
 
In [18]: p=t.argmax(n,dim=1)
 
In [19]: p
Out[19]: tensor([2, 0, 0])
 
In [20]: d.sum()
Out[20]: tensor(12.0000)
 
In [22]: d
Out[22]:
 
tensor([[[0.1771, 0.2075, 0.2075, 0.2075, 0.2005],
 
     [0.2027, 0.1959, 0.2027, 0.1959, 0.2027],
 
     [0.1972, 0.2041, 0.1972, 0.2041, 0.1972],
 
     [0.2027, 0.1959, 0.2027, 0.1959, 0.2027]],
 
 
    [[0.1972, 0.2041, 0.1972, 0.2041, 0.1972],
 
     [0.2027, 0.1959, 0.2027, 0.1959, 0.2027],
 
     [0.1972, 0.2041, 0.1972, 0.2041, 0.1972],
 
     [0.2027, 0.1959, 0.2027, 0.1959, 0.2027]],
 
 
    [[0.1972, 0.2041, 0.1972, 0.2041, 0.1972],
 
     [0.2027, 0.1959, 0.2027, 0.1959, 0.2027],
 
     [0.1972, 0.2041, 0.1972, 0.2041, 0.1972],
 
     [0.2027, 0.1959, 0.2027, 0.1959, 0.2027]]])
 
In [23]: d[0][0].sum()
Out[23]: tensor(1.)

补充知识：多分类问题torch.nn.Softmax的使用

为什么谈论这个问题呢？是因为我在工作的过程中遇到了语义分割预测输出特征图个数为16，也就是所谓的16分类问题。

因为每个通道的像素的值的大小代表了像素属于该通道的类的大小，为了在一张图上用不同的颜色显示出来，我不得不学习了torch.nn.Softmax的使用。

首先看一个简答的例子，倘若输出为(3, 4, 4)，也就是3张4x4的特征图。

?
1
2
3
import torch
img = torch.rand((3,4,4))
print(img)

输出为：

?
1
2
3
4
5
6
7
8
9
10
11
12
tensor([[[0.0413, 0.8728, 0.8926, 0.0693],
     [0.4072, 0.0302, 0.9248, 0.6676],
     [0.4699, 0.9197, 0.3333, 0.4809],
     [0.3877, 0.7673, 0.6132, 0.5203]],
    [[0.4940, 0.7996, 0.5513, 0.8016],
     [0.1157, 0.8323, 0.9944, 0.2127],
     [0.3055, 0.4343, 0.8123, 0.3184],
     [0.8246, 0.6731, 0.3229, 0.1730]],
    [[0.0661, 0.1905, 0.4490, 0.7484],
     [0.4013, 0.1468, 0.2145, 0.8838],
     [0.0083, 0.5029, 0.0141, 0.8998],
     [0.8673, 0.2308, 0.8808, 0.0532]]])

我们可以看到共三张特征图，每张特征图上对应的值越大，说明属于该特征图对应类的概率越大。

?
1
2
3
4
import torch.nn as nn
sogtmax = nn.Softmax(dim=0)
img = sogtmax(img)
print(img)

输出为：

?
1
2
3
4
5
6
7
8
9
10
11
12
tensor([[[0.2780, 0.4107, 0.4251, 0.1979],
     [0.3648, 0.2297, 0.3901, 0.3477],
     [0.4035, 0.4396, 0.2993, 0.2967],
     [0.2402, 0.4008, 0.3273, 0.4285]],
    [[0.4371, 0.3817, 0.3022, 0.4117],
     [0.2726, 0.5122, 0.4182, 0.2206],
     [0.3423, 0.2706, 0.4832, 0.2522],
     [0.3718, 0.3648, 0.2449, 0.3028]],
    [[0.2849, 0.2076, 0.2728, 0.3904],
     [0.3627, 0.2581, 0.1917, 0.4317],
     [0.2543, 0.2898, 0.2175, 0.4511],
     [0.3880, 0.2344, 0.4278, 0.2686]]])

可以看到，上面的代码对每张特征图对应位置的像素值进行Softmax函数处理， 图中标红位置加和=1，同理，标蓝位置加和=1。

我们看到Softmax函数会对原特征图每个像素的值在对应维度(这里dim=0，也就是第一维)上进行计算，将其处理到0～1之间，并且大小固定不变。

print(torch.max(img,0))

输出为：

?
1
2
3
4
5
6
7
8
9
torch.return_types.max(
values=tensor([[0.4371, 0.4107, 0.4251, 0.4117],
    [0.3648, 0.5122, 0.4182, 0.4317],
    [0.4035, 0.4396, 0.4832, 0.4511],
    [0.3880, 0.4008, 0.4278, 0.4285]]),
indices=tensor([[1, 0, 0, 1],
    [0, 1, 1, 2],
    [0, 0, 1, 2],
    [2, 0, 2, 0]]))

可以看到这里3x4x4变成了1x4x4，而且对应位置上的值为像素对应每个通道上的最大值，并且indices是对应的分类。

清楚理解了上面的流程，那么我们就容易处理了。

看具体案例，这里输出output的大小为：16x416x416.

?
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
output = torch.tensor(output)
 
sm = nn.Softmax(dim=0)
output = sm(output)
 
mask = torch.max(output,0).indices.numpy()
 
# 因为要转化为RGB彩色图，所以增加一维
rgb_img = np.zeros((output.shape[1], output.shape[2], 3))
for i in range(len(mask)):
  for j in range(len(mask[0])):
    if mask[i][j] == 0:
      rgb_img[i][j][0] = 255
      rgb_img[i][j][1] = 255
      rgb_img[i][j][2] = 255
    if mask[i][j] == 1:
      rgb_img[i][j][0] = 255
      rgb_img[i][j][1] = 180
      rgb_img[i][j][2] = 0
    if mask[i][j] == 2:
      rgb_img[i][j][0] = 255
      rgb_img[i][j][1] = 180
      rgb_img[i][j][2] = 180
    if mask[i][j] == 3:
      rgb_img[i][j][0] = 255
      rgb_img[i][j][1] = 180
      rgb_img[i][j][2] = 255
    if mask[i][j] == 4:
      rgb_img[i][j][0] = 255
      rgb_img[i][j][1] = 255
      rgb_img[i][j][2] = 180
    if mask[i][j] == 5:
      rgb_img[i][j][0] = 255
      rgb_img[i][j][1] = 255
      rgb_img[i][j][2] = 0
    if mask[i][j] == 6:
      rgb_img[i][j][0] = 255
      rgb_img[i][j][1] = 0
      rgb_img[i][j][2] = 180
    if mask[i][j] == 7:
      rgb_img[i][j][0] = 255
      rgb_img[i][j][1] = 0
      rgb_img[i][j][2] = 255
    if mask[i][j] == 8:
      rgb_img[i][j][0] = 255
      rgb_img[i][j][1] = 0
      rgb_img[i][j][2] = 0
    if mask[i][j] == 9:
      rgb_img[i][j][0] = 180
      rgb_img[i][j][1] = 0
      rgb_img[i][j][2] = 0
    if mask[i][j] == 10:
      rgb_img[i][j][0] = 180
      rgb_img[i][j][1] = 255
      rgb_img[i][j][2] = 255
    if mask[i][j] == 11:
      rgb_img[i][j][0] = 180
      rgb_img[i][j][1] = 0
      rgb_img[i][j][2] = 180
    if mask[i][j] == 12:
      rgb_img[i][j][0] = 180
      rgb_img[i][j][1] = 0
      rgb_img[i][j][2] = 255
    if mask[i][j] == 13:
      rgb_img[i][j][0] = 180
      rgb_img[i][j][1] = 255
      rgb_img[i][j][2] = 180
    if mask[i][j] == 14:
      rgb_img[i][j][0] = 0
      rgb_img[i][j][1] = 180
      rgb_img[i][j][2] = 255
    if mask[i][j] == 15:
      rgb_img[i][j][0] = 0
      rgb_img[i][j][1] = 0
      rgb_img[i][j][2] = 0
 
cv2.imwrite('output.jpg', rgb_img)

最后保存得到的图为：

浅谈pytorch中torch.max和F.softmax函数的维度解释

以上这篇浅谈pytorch中torch.max和F.softmax函数的维度解释就是小编分享给大家的全部内容了，希望能给大家一个参考，也希望大家多多支持服务器之家。

原文链接：https://blog.csdn.net/Jasminexjf/article/details/90402990

标签：

相关文章