PyTorch 交叉熵损失函数内部原理简单实现

import numpy as np
import torch


# 分类标签[2, 0, 1] 映射成独热编码
def labels_to_one_hot(label, dim):
    # label 标签，dim 特征数
    hot_encode = np.zeros((len(label), dim))
    hot_encode[np.arange(len(label)), label] = 1
    return hot_encode


Y = np.array([2, 0, 1])
Y_pred = np.array([[0.8, 0.2, 0.3],
                   [0.2, 0.3, 0.5],
                   [0.2, 0.2, 0.5]])
one_hot = labels_to_one_hot(Y, Y_pred.shape[1])
print(one_hot)
loss = 0
# 分别对每个样本求loss
for y, y_pred in zip(one_hot, Y_pred):
    # soft_max
    soft_y_pred = np.exp(y_pred) / np.exp(y_pred).sum()
    # 累加loss
    loss += (-y * np.log(soft_y_pred)).sum()
print(loss / 3)  # 求均值