A Simple Framework for Contrastive Learning of Visual Representations