Reproduce the scores in the paper

[roboinlovo]

I read your paper and tried to reproduce the scores shown in the paper.
However, I could not get those scores. Could you give me any idea to reproduce the scores on GQA (val: 73.6, test-dev: 72.1)?

I followed the instruction in this repository and got the score 71.43 (99.86) with test.sh.
Does split 'val' produce the score for val (should be 73.6)?
With split 'test', I got the following error.

Traceback (most recent call last):
  File "evaluate.py", line 159, in <module>
    eval_cfrf_score, _, _, _, bound = evaluate(model, eval_loader, args)
  File "src/FFOE/train.py", line 158, in evaluate
    for i, (v, b, w, e, attr, q, s, a, ans) in enumerate(dataloader):
ValueError: too many values to unpack (expected 9)

I really appreciate it if you could answer my question.

[Dhan800]

I have questions about their results too. In their original paper, there is no mention of integrating LXMERT into their model. In the code base, the model is already LXMERT-integrated. No way to verify the performance of their method before integration.