transE icon indicating copy to clipboard operation
transE copied to clipboard

Published 20 hours ago • verified publisher icon wuxiyu

update函数中梯度更新问题

Open wywgreat opened this issue 4 years ago • 1 comments

根据原文中loss function的推导，我认为应该为如下形式

headEntityVector = headEntityVector + 2 * self.learingRate * (relationVectorBeforeBatch- tailEntityVectorBeforeBatch) tailEntityVector = tailEntityVector + 2 * self.learingRate * (headEntityVectorBeforeBatch - relationVectorBeforeBatch)

WithCorrupted的形式类似，求解答。

wywgreat avatar Apr 27 '20 09:04 wywgreat

我倒是认为headEntityVector是可以放大loss，加快权重更新

DC-Lin avatar Oct 30 '20 09:10 DC-Lin