flickr 30k features download about ban-vqa HOT 4 CLOSED

I extracted the flickr bottom-up attention features separately from @jnhwkim.
And as he mentioned, I guess the environment or hyperparameter settings were different from @jnhwkim's work.

Sorry that I think I blew the previous google drive link.
I re-uploaded bottom-up attention feature files corresponding 29906.

https://drive.google.com/file/d/1Rdu8UDEi4Z0lnMaqwQYJKBeNTnSY-dsw/view?usp=sharing

from ban-vqa.

@jaesuny thanks for the reply and the link!
I keep the commit 54f044c since it is used for the camera-ready (I believe). Please refer to @jaesuny 's data to double-check if you consider it.

from ban-vqa.

Please refer to #43 for the context.

In short, I believe 30722 is from July 2018 (or earlier and the camera-ready version I believe), and 29906 is from @jaesuny 's reproduction in Nov 2018. The discrepancy is from the adaptive number (10-100) of detected boxes by the bottom-up attention networks tweaking by some hyperparameters I guess. I updated those numbers to match with the released data in 54f044c.

from ban-vqa.

Thank you for your help @jnhwkim @jaesuny !

from ban-vqa.