Towards More Sample Efficiency in Reinforcement Learning with Data Augmentation