If the post training is done with specific sampling parameters it would make sense to only use the parameters it was trained with.