Deep Contrastive Representation Learning With Self-Distillation