Contrastive learning is now a popular choice for representation learning in various domains, including image and natural language processing. However, contrastive learning for time-series data is relatively limited, due to its unrecognizable, high-dimensional temporal structures. It is still difficult to generate valid augmented views that are semantically accurate, despite the significant research advances in the field of time-series data augmentation. In this work, we survey recent works in time-series contrastive learning and propose a simple augmentation-agnostic technique that can effectively improve the fidelity of the augmented views.