CrossEntropyLoss计算过程

import torch
import torch.nn as nn
import math
print('不使用softmax计算CrossEntropyLoss：')
criterion = nn.CrossEntropyLoss()
output = torch.randn(3, 5, requires_grad=True)
label = torch.empty(3, dtype=torch.long).random_(5)
loss = criterion(output, label)

print("网络输出为3个5类:")
print(output)
print("要计算loss的类别:")
print(label)
print("计算loss的结果:")
print(loss)

first = [0, 0, 0]
for i in range(3):
    first[i] = -output[i][label[i]]
second = [0, 0, 0]
for i in range(3):
    for j in range(5):
        second[i] += math.exp(output[i][j])
res = 0
for i in range(3):
    res += (first[i] + math.log(second[i]))
print("自己的计算结果：")
print(res / 3)
print('~~' * 50)
print('CrossEntropyLoss包含sotfmax，下面使用softmax计算CrossEntropyLoss：')
print('softmax后输出：')
softmax = nn.Softmax(dim=1)
softmax_output = softmax(output)
print(softmax_output)
sum_soft = 0
for i in range(3):
    sum_soft += -math.log(softmax_output[i][label[i]])
print('softmax后计算结果：')
print(sum_soft/3)
print("计算loss的结果:")
print(loss)