SemPPL: Predicting pseudo-labels for better contrastive representations