知识卡片
一致层形状便于组合而非无限容量
常见Transformer块让序列长度与隐藏宽度在块间保持一致,方便残差与堆叠。增加层仍有参数、计算、优化及数据需求;RNN和适当填充的CNN也可堆叠,一致形状不是Transformer独有或可无限扩大的证明。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
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
9.1.2 A LEGO set for language As with CNNs and RNNs (LSTMs and GRUs), each layer of a transformer gives you a deeper and deeper representation of the meaning or thought of the input text. But unlike CNNs and RNNs, each layer of the encoder half of a transformer outputs a tensor that is the exact same size and shape as the previous layers, providing
参考来源
- 位置:《Natural Language Processing in Action, Second Edition》第09章《可堆叠的Transformer》(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-9.html)
- 结论依据:摘录中关于一致层形状便于组合而非无限容量的定义、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:9.1.2 A LEGO set for language As with CNNs and RNNs (LSTMs and GRUs), each layer of a transformer gives you a deeper and deeper representation of the meaning or thought of the input text. But unlike CNNs and RNNs, each layer of the encoder half of a transformer outputs a tensor that is the exact same size and shape as the previous layers, providing