Embedding And Clustering Your Data Can Improve Contrastive Pretraining