nn.softmax的参数使用和数据使用

nn.Softmax ，dim=0

对于二维数组的dim=0时，将每列的每个元素和这一列的自然指数之和进行比，

import torch.nn as nn
import numpy as np
import math

arr = np.linspace(1,16,16).reshape(4,4)
arr = torch.tensor(arr)
sofmax = nn.Softmax(dim=0)
out = sofmax(arr)
print(arr)
print(out)
son = math.exp(1)
pat = 0
for i in range(1,14,4):
    pat+=math.exp(i)

print(son/pat)
'''
tensor([[ 1.,  2.,  3.,  4.],
        [ 5.,  6.,  7.,  8.],
        [ 9., 10., 11., 12.],
        [13., 14., 15., 16.]], dtype=torch.float64)
tensor([[6.0317e-06, 6.0317e-06, 6.0317e-06, 6.0317e-06],
        [3.2932e-04, 3.2932e-04, 3.2932e-04, 3.2932e-04],
        [1.7980e-02, 1.7980e-02, 1.7980e-02, 1.7980e-02],
        [9.8168e-01, 9.8168e-01, 9.8168e-01, 9.8168e-01]], dtype=torch.float64)
6.031677857384872e-06
'''

dim=0，修改第一个值为20后

只是第一列发生变化，

arr = np.linspace(1,16,16).reshape(4,4)

arr = np.array([[ 20.,  2.,  3.,  4.],
        [ 5.,  6.,  7.,  8.],
        [ 9., 10., 11., 12.],
        [13., 14., 15., 16.]])
arr = torch.tensor(arr)

修改dim=1

对红框中的一行进行计算，

import torch.nn as nn
import numpy as np
import math

arr = np.linspace(1,16,16).reshape(4,4)

arr = np.array([[ 1.,  2.,  3.,  4.],
        [ 5.,  6.,  7.,  8.],
        [ 9., 10., 11., 12.],
        [13., 14., 15., 16.]])
arr = torch.tensor(arr)
sofmax = nn.Softmax(dim=1)
out = sofmax(arr)
print(arr)
print(out)
son = math.exp(1)
pat = 0
for i in range(1,5):
    pat+=math.exp(i)

print(son/pat)