Vilbert Bug Fixes for VQA task

[arjunakula]

Hi,

This is Arjun Akula from UCLA. There are few bugs in the original ViLBERT code for VQA. I fixed the issues and creating a new PULL request. thanks!

[alice-cool]

Dear scholar,
Could you provide the google drive links for COCOdatasets. Because in china, the dropboxlink is unstable, even I bought a vpn. Sometimes it ended abruptly, but the file is too large, one is 26.3g, the other is approximately 40g.
https://www.dropbox.com/sh/09182lupkagw1ov/AACShSEVClAh6CzbhyIKxmtga?dl=0

I also want to reproduce the results but now I didn't have the data file. Could you help me?

I am eager to get your message.

[alice-cool]

I found the lmdb could produced by the tsv file. Did you have the tsv file? I found peterson github only have smaller file. SO I think the tsv file would not from that github

[alice-cool]

But VQA2.0 features should be the same with the github features from peterson.
https://github.com/peteanderson80/bottom-up-attention#demo
I have all the tsv file from peterson.

[alice-cool]

If the script can convert the peterson features tsv to the format lmdb. If that , I will produce the Imdb file . But which file to produce, by using 10~100 features file or fixed 36 features file?

[alice-cool]

Dear scholar,
I use the pretrained parameters to evaluate the dataset VQA2.0 . But the results is the following. I replace the author', task_util.py with your uploading task_util.py. What is the accuracy ? 0.0?