Disentangling the Effects of Data Augmentation and Format Transform in Self-Supervised Learning of Image Representations