Contextual vision transformers for robust representation learning