Revisiting the Critical Factors of Augmentation-Invariant Representation Learning