OFA
OFA copied to clipboard
Feature/vqa
Add unconstrained training for VQA, which does not need a pre-defined candidate answer set for both finetuning and inference. In this case, evaluation inference mode must be beamsearch rather than allcand.