Self-supervised Pre-training with Hard Examples Improves Visual Representations