知识卡片
训练位置可并行但层间仍依赖
在给定训练序列及遮罩后,Transformer可同时计算多个词位,不必像普通RNN逐词递推隐藏状态。但后一层仍依赖前一层,自回归生成下一词也依赖已选前缀;不能说所有层和所有生成位置一次并行完成。
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
Because there is no recurrence within the inner guts of the transformer, it doesn’t need to be unrolled. This gives transformers a huge advantage over RNNs. The individual neurons and layers in a transformer can be run in parallel all at once. For an RNN, you had to run the functions for the neurons and layers one at a time in sequence. Unrolling all these
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:摘录中关于训练位置可并行但层间仍依赖的定义、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:Because there is no recurrence within the inner guts of the transformer, it doesn’t need to be unrolled. This gives transformers a huge advantage over RNNs. The individual neurons and layers in a transformer can be run in parallel all at once. For an RNN, you had to run the functions for the neurons and layers one at a time in sequence. Unrolling all these