Jerry Jiarui XU
Jerry Jiarui XU
@pzhren Maybe some of the link is invalid. But it should be fine.
Hi @slyviacassell I installed `rename` by running ` sudo apt-get install rename` ``` (base) ➜ ~ rename -h Usage: rename [ -h|-m|-V ] [ -v ] [ -0 ] [...
Hi @pzhren You need to follow the instruction [here](https://gitlab.com/jfolz/yfcc100m/-/tree/master) and set up your own AWS credential.
It seems that your credential is not correctly setup
Hi @Davidleeeeee The onehot is to make sure each group feature is exclusive from each other. Otherwise, the classification accuracy will drop.
Thx! Will do!
Hi @dingjiansw101 (1) `pre_assign_attn` is used to aggregate information from image tokens to group tokens. It is omitted in the paper for simplicity. (2) Yes.
Hi @dingjiansw101 Yes. You are correct. The only difference is the cross attention and self attention. `pre_assign_attn` is there for some legacy reasons. We didn't ablate whether it is import...
Hi @XiaominLi1997 The model is trained with embedding size 256 only.
Hi @scm-later The model trained on single GPU may not have good accuracy since the contrastive learning needs large batch size.