pretrain: Data used when pre-training on VQAv2
image_features: Image features can be used by Lxmert and VisualBert during both pre-training, fine-tuning and inference.
finetune: Data (e.g., knowledge, captions) used when training on OK-VQA
validate: Data used during inference on validate sets of OK-VQA