Input embedding for the ablation experiments

deanpl...@gmail.com

unread,

Mar 20, 2019, 12:57:04 AM3/20/19

to Visual Commonsense Reasoning

Hi Rowan,

Thanks for your great work! Can I know what's the input of the models in the ablation experiments? e.g. Did you use the same bert embedding in the "no query" model in your ablation experiment as the bert embeddings included in the repo? Thanks!

Rowan Zellers

unread,

Mar 20, 2019, 12:40:32 PM3/20/19

to Visual Commonsense Reasoning

thanks!! So the input of the model is technically the same (ie, BERT still), however, we used a different BERT embedding. The reason is that the bert embeddings included in the repo are conditioned on the query, so in order to actually do a 'no query' experiment, we had to compute BERT embeddings separately for just the responses in isolation.

let me know if that helps!!

deanpl...@gmail.com

unread,

Mar 20, 2019, 12:56:22 PM3/20/19

to Visual Commonsense Reasoning

Thanks for your response! Can I know if it's possible to also release the bert embeddings for the no-query mode? I think it might be quite expensive to compute the bert embeddings of the dataset ...

deanpl...@gmail.com

unread,

Mar 21, 2019, 5:29:06 AM3/21/19

to Visual Commonsense Reasoning

Never mind... It seems it's not that expensive to calculate the BERT embeddings and I have successfully generated them using your script. Thanks anyway!