知识卡片
瓶颈适配器先降维再升维
瓶颈适配器把原宽度投影到较小空间,经非线性再映回原宽度,常结合残差并训练新增模块。插入位置、宽度和初始化影响资源与行为,不能把某张图的位置当所有适配器唯一结构。
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
- · 进阶微调技术
原书出处 · 1 处
- · 查看原书依据
to the LLM architecture. Adapters can be integrated into the LLM architecture in a variety of ways, but in Transformers, the common way is to insert them at each layer of the Transformer. To reduce the number of parameters, the width of the adapter module should be much less than the width of the underlying Transformer model. This constitutes a down-projection, also called a bottleneck.
参考来源
- 位置:《Designing Large Language Model Applications》第07章《进阶微调技术》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch07.html)
- 结论依据:摘录中关于瓶颈适配器先降维再升维的定义、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:to the LLM architecture. Adapters can be integrated into the LLM architecture in a variety of ways, but in Transformers, the common way is to insert them at each layer of the Transformer. To reduce the number of parameters, the width of the adapter module should be much less than the width of the underlying Transformer model. This constitutes a down-projection, also called a bottleneck.