tf.keras.losses.CategoricalCrossentropy 多分类 交叉熵 损失函数示例

import numpy as np 
import tensorflow as tf

y_true = np.array([0, 1, 0])  # 独热编码
y_pred = np.array([0.5, 0.2, 0.3])  # 预测概率
delta = 1e-7

计算公式

$loss=\frac{1}{n}{\sum }_i^{}tru{e}_i\times \left(-lo{g}_e^{pre{d}_i+l{e}^{-7}}\right)$

n：样本数

true : y_true

pred : y_pred

计算过程

$-[(0\times lo{g}_e^{0.5+l{e}^{-7}})+(1\times lo{g}_e^{0.2+l{e}^{-7}})+(0\times lo{g}_e^{0.3+l{e}^{-7}})]=1.6094374124342252$

tf.keras.losses.CategoricalCrossentropy 实现

cce = tf.keras.losses.CategoricalCrossentropy()
cce(y_true, y_pred).numpy()

1.6094377040863037

numpy 实现

y_pred = y_pred+delta  # 添加一个微小值可以防止负无限大(np.log(0))的发生。
print(y_pred)

[0.5000001 0.2000001 0.3000001]

y_pred_log = np.log(y_pred)  # log表示以e为底数的自然对数
print(y_pred_log)

[-0.69314698 -1.60943741 -1.20397247]

print(y_true*y_pred_log)

[-0.         -1.60943741 -0.        ]

loss = -np.sum(y_true*y_pred_log)
print(loss)

1.6094374124342252