Hello, Ronghang Hu, thanks for releasing the code.
When I see the code in the task of VQA in the GQA dataset, I find that the val_balance and testdev_balance have some answers that don't exist in train_balance.
So, whether this is a problem?
I can't download the CLEVR-Ref+ dataset from the link you provided, while I got the data from this link. However, when I download data from the link, there is no test images, test_refexps.json and test_scenes.json in the CLEVR-Ref+-v1.0 Dataset. Where can I download the data related to the test set?
In addition, what is the difference between CLEVR-Ref+-CoGenT-v1.0 Dataset and CLEVR-Ref+-v1.0 Dataset? And which one do you used? Is the valB in CLEVR-Ref+-CoGenT-v1.0 Dataset means the test set?
Can you briefly explain why hadamard product is used to perform attention calculation in Eqn 2 and for computing conditioning in Eqn 4 and 5 in the paper?