You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
------------------ 原始邮件 ------------------
发件人: "Guo ***@***.***>;
发送时间: 2023年6月29日(星期四) 下午2:45
收件人: ***@***.***>;
抄送: ***@***.***>; ***@***.***>;
主题: Re: [OpenLMLab/LOMO] Some confusion about the method of the paper (Issue #27)
没有重复计算,链式法则只会用到上一步的梯度结果,而不是之前所有步的梯度。上一步的还保留,但历史的梯度被动态清空了。
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you authored the thread.Message ID: ***@***.***>
大佬您好,传统梯度反向链式传播会用到上一步的梯度计算结果,但文中的方法在更新后不存储梯度,是否意味着后续梯度计算中多了重复的计算,类似时间换空间的做法。这么理解正确吗?
The text was updated successfully, but these errors were encountered: