知识卡片
冻结参数与固定行为有别
冻结某模块可保留其参数并减少更新成本,但系统输出仍会因上游模块或输入改变而变化。冻结也不自动覆盖所有带状态的行为;实施时应明确运行模式和统计量策略。冻结的特征是否适用于新域,需要目标任务评估,而不是由“通用”名称保证。
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
- · 迁移学习与适配
原书出处 · 1 处
- · 查看原书依据
Figure 37.2 shows that you don't have to finetune all layers in your network but can selectively finetune some and initialize others from scratch. In fact this strategy can be applied quite generally, where we initialize some parts of a computation graph from pretrained models and initialize others from scratch. There are many reasons why you might want to do this, or not. Sometimes you
参考来源
- 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第37章《迁移学习与适配》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)/_epub-src/OPS/xhtml/c037.xhtml)
- 结论依据:摘录中关于冻结参数与固定行为有别的定义、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:Figure 37.2 shows that you don't have to finetune all layers in your network but can selectively finetune some and initialize others from scratch. In fact this strategy can be applied quite generally, where we initialize some parts of a computation graph from pretrained models and initialize others from scratch. There are many reasons why you might want to do this, or not. Sometimes you