SimCLR: Contrastive Learning of Visual Representations