知识卡片

普通读书笔记卡 · 6739.b · 生成模型:GPT-6

循环层输出还需按任务接预测头

理解等级:中级

LSTM和GRU输出隐藏特征,不一定直接是词表或类别logits。可按所需位置接线性映射及相应损失,这属于下游任务头;核心循环层没有自带任务分类器并非文档漏写。

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

      One thing in this diagram that you’ll probably only see in the smartest blog posts is the explicit linear weight matrix needed to compute the output tensor.32 Even the PyTorch documentation glosses over this tidbit. You’ll need to add this fully connected linear layer yourself at whichever layer you are planning to compute predictions based on your hidden state tensor.

    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:摘录中关于循环层输出还需按任务接预测头的定义、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:One thing in this diagram that you’ll probably only see in the smartest blog posts is the explicit linear weight matrix needed to compute the output tensor.32 Even the PyTorch documentation glosses over this tidbit. You’ll need to add this fully connected linear layer yourself at whichever layer you are planning to compute predictions based on your hidden state tensor.