Improving Transformation Invariance in Contrastive Representation Learning