thunlp / Fast-TransX

An Efficient implementation of TransE and its extended models for Knowledge Representation Learning
MIT License
402 stars 108 forks source link

L2 or L1? 看论文中score function是算欧式距离,为什么代码中是求L1? #26

Closed HongxuChenUQ closed 6 years ago

THUCSTHanxu13 commented 6 years ago

其实这些方法用L1和L2都可以,个人经验L1效果好一些,实现的时候为了统一就都设定为L1了,L2的话修改下剃度就可以了。