hengyuan-hu / bottom-up-attention-vqa

An efficient PyTorch implementation of the winning entry of the 2017 VQA Challenge.
GNU General Public License v3.0
754 stars 181 forks source link

Code for calculating accuracy metric #27

Closed erobic closed 6 years ago

erobic commented 6 years ago

The README file says that accuracy was calculated with the VQA metric. Is the code for that calculation in the repo? I am unsure how to use the upper bound score alongside the actual score to get to this metric.

jnhwkim commented 6 years ago

https://github.com/GT-Vision-Lab/VQA/issues/1 https://github.com/hengyuan-hu/bottom-up-attention-vqa/pull/18

erobic commented 6 years ago

I see now that the hard-coded scores make sense. Should the final accuracy then be: score/upper_bound? (I am running it in a different dataset and the upper bound is much lower than 100)

ZhuFengdaaa commented 6 years ago

@erobic the answer label is so disperse that many questions whose answers with the highest frequent are <= 2 ? You can try some text preprocess tricks written in compute_softscore.py

erobic commented 6 years ago

Right, it was some rough idea I wanted to try, but am discarding it for now. Thanks for the help though!