How to change how data is sampled per batch in a tensroflow.data.Datasets dataset?

Question

I want to train an embedding model on the Stanford dogs dataset using tfa.losses.TripletSemiHardLoss(). Problem is that you have to have a really large batch si