batra-mlp-lab / visdial

[CVPR 2017] Torch code for Visual Dialog

Home Page:https://arxiv.org/abs/1611.08669

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Beam search with length normalized log likelihood?

ruizhaogit opened this issue · comments

Hi guys,

About beam search here, if I understand it correctly, this is the normal beam search without length normalized log likelihood, then it should tend to find a shorter sequence, right? Did you also try the beam search with the length normalized log likelihood?

Thank you!

Best,
Rui

Yup, the current version prefers shorter sequences and we did try both. In terms of retrieval performance as described in the paper, beam search with length normalization performs better on some metrics and worse on others, so there was no clear choice. We picked this and went ahead with it.