Question about the Dropout technique

I Understand that you apply Dropout for every training set in order to get multiple sub-models to train ones network on. However, do you create new set of sub-models for next iteration (apply Dropout for every training set again) or do you keep set reuse the sub-models?