Gong haojie

Results 2 issues of Gong haojie

Since some model have BN,dropout layers, why dont use model.eval() in the val function? am really confused, hope someone can help me. thx in advance.

在提取留言语义上我看见您通过average over all tokens 来获得一个768维的vector。 我想问的问题是:经过bert,为什么您不使用[CLS]token来代表语义然后再聚类,而是使用取平均呢?