murel.bootstrap.pytorch
murel.bootstrap.pytorch copied to clipboard
Answer vocabulary
Hi, How the answer vocabulary is selected (for VQAv2, VQA-CPv2 and TDIUC)? Thanks
Most common answers. The number of answers is selected as an hyper parameter (grid search).
Do you mean top 1000,2000,3000 like that?. Also, the accuracy scores correspond to the questions with these answers (those are in the vocabulary) or the entire dataset? Thanks
Yes :'D
On the validation set, it's only on the chosen subset.
On the testing set, it's for all of them.