← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 6737.a · 生成模型:GPT-6

多标签改造还要改损失和评价

理解等级:中级

将softmax换为逐标签sigmoid只是输出环节,还需多热标签、匹配的二元损失、阈值及多标签指标。各维允许同时为正不表示现实标签统计独立;只替换激活而继续原NLLLoss与单类真值不构成完整改造。

这张卡片看不懂?

已有反馈 · 0 条

    还没有人提交反馈。

    在整书笔记中的位置

    原书出处 · 1 处

    • · 查看原书依据

      Tip To turn any multiclass classifier into a multi-label tagger, you must change your activation function from softmax to an element-wise sigmoid function. A softmax creates a probability distribution across all the mutually exclusive categorical labels. On the other hand, a sigmoid function allows every value to take on any value between zero and one, such that each dimension in your multi-label tagging output represents

    参考来源

    - 位置:《Natural Language Processing in Action, Second Edition》第08章《循环网络与长短期记忆》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Natural Language Processing in Action, Second Edition/_epub-src/OEBPS/Text/chapter-8.html) - 结论依据:摘录中关于多标签改造还要改损失和评价的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。 - 原始内容:Tip To turn any multiclass classifier into a multi-label tagger, you must change your activation function from softmax to an element-wise sigmoid function. A softmax creates a probability distribution across all the mutually exclusive categorical labels. On the other hand, a sigmoid function allows every value to take on any value between zero and one, such that each dimension in your multi-label tagging output represents