shuffle is only needed to introduce randomness, if each time the same unshuffled batch is used to train, the model may overfit.