-
Notifications
You must be signed in to change notification settings - Fork 725
Open
Labels
bugSomething isn't workingSomething isn't working
Description
想咨询一下序列推荐问题。
recbole是做了数据增强,就是给定序列,会形成多个子序列。比如sasrec模型,原始代码是seq2seq形式计算损失:即给定序列计算多个损失。recbole似乎是给定子序列只计算1个损失。这两者是否是等价的?但gru4rec应该是给定序列计算一个损失,recbole做了增强后就会导致算更多的损失吧?
其次,模型显存问题,recbole的sasrec我发现使用celoss,显存占用很低,复现sasrec原始模型显存占用大(如果使用celoss显存会更大)。recbole是用到了什么显存优化吗?
Reactions are currently unavailable
Metadata
Metadata
Assignees
Labels
bugSomething isn't workingSomething isn't working