VQA and GQA test splits do not have answers?

[ranery]

Hi, I notice that for VQA and GQA, if we want to test the pretrained model, there are no available ground-truth answers.

if "test" not in self.split:
      answer = entry["answer"]
      labels = answer["labels"]
      scores = answer["scores"]
      if labels is not None:
          target.scatter_(0, labels, scores)

I wonder how you calculate the score on the test set?

[ranery]

Find out that there are EvalAI platforms, close the issue.