MXM-CLR: A Unified Framework for Contrastive Learning of Multifold Cross-Modal Representations.