'binary_crossentropy' & 'categorical_crossentropy' in keras

最新推荐文章于 2022-08-26 17:06:23 发布

yusisc

最新推荐文章于 2022-08-26 17:06:23 发布

阅读量3.6k

点赞数

分类专栏： tensorflow

本文链接：https://blog.youkuaiyun.com/yusisc/article/details/81179353

版权

本文探讨了Keras中binary_crossentropy和categorical_crossentropy两种损失函数的区别。binary_crossentropy适用于二元多标签分类问题，而categorical_crossentropy用于多类别互斥分类问题。深入理解这两个损失函数的底层原理有助于解决分类任务中的混淆问题。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

In model.compile(*) of keras, I met binary_crossentropy & categorical_crossentropy. These two kinds of loss somehow made me confused.
Checking their underlying will reveal the mechanism of these two kinds of loss.

loss	refer to
binary_crossentropy	K.mean(K.binary_crossentropy(y_true, y_pred), axis=-1)
categorical_crossentropy	tf.nn.softmax_cross_entropy_with_logits(labels=target, logits=output)

The problem is what is binary_crossentropy and softmax_cross_entropy_with_logits in TensorFlow.

binary_crossentropy (and tf.nn.sigmoid_cross_entropy_with_logits under the hood) is for binary multi-label classification (labels are independent).
categorical_crossentropy (and tf.nn.softmax_cross