[人工智能]理解 pytorch 的 CrossEntropyLoss

简介
为了更好地理解 pytorch 的 CrossEntropyLoss，于是打算进行简单的实现。

官方文档：
https://pytorch.org/docs/stable/nn.html?highlight=crossentropyloss#torch.nn.CrossEntropyLoss

官网 loss 的公式：

?

x 的维度是 (batch_size, C)
class 的维度是 (batch_size)
(这里的 C 是分类的个数）
?

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

class CrossEntropyLoss():

    def __init__(self, weight=None, size_average=True):
        """
        初始化参数，因为要实现 torch.nn.CrossEntropyLoss 的两个比较重要的参数

        :param weight: 给予每个类别不同的权重
        :param size_average: 是否要对 loss 求平均
        """

        self.weight = weight
        self.size_average = size_average


    def __call__(self, input, target):
        """
        计算损失
        这个方法让类的实例表现的像函数一样，像函数一样可以调用

        :param input: (batch_size, C)，C是类别的总数
        :param target: (batch_size, 1)
        :return: 损失
        """

        batch_loss = 0.
        for i in range(input.shape[0]):
            # print('***',input[i, target[i]],i,target[i],np.exp(input[i, :]))
            numerator = np.exp(input[i, target[i]])     # 分子
            denominator = np.sum(np.exp(input[i, :]))   # 分母

            # 计算单个损失
            loss = -np.log(numerator / denominator)
            if self.weight:
                loss = self.weight[target[i]] * loss
            print("单个损失： ",loss)

            # 损失累加
            batch_loss += loss

        # 整个 batch 的总损失是否要求平均
        if self.size_average == True:
            batch_loss /= input.shape[0]

        return batch_loss
if __name__ == "__main__":
    input = np.array([[-1.5616, -0.7906,  1.4143, -0.0957,  0.1657],
        [-1.4285,  0.3045,  1.5844, -2.1508,  1.8181],
        [ 1.0205, -1.3493, -1.2965,  0.1715, -1.2118]])
    target = np.array([2, 0, 3])

    criterion = CrossEntropyLoss()

    # 类中实现了 __call__，所以类实例可以像函数一样可以调用
    loss = criterion(input, target)

    print()
    print("输出： ", loss)

    #torch.nn中库函数
    test_loss = nn.CrossEntropyLoss()
    test_input = torch.from_numpy(input)
    test_target = torch.from_numpy(target).long()
    test_out = test_loss(test_input,test_target)
    print('test loss:',test_out)

?参考：https://blog.csdn.net/qq_41805511/article/details/99438838