When loading a checkpoint the last iteration is repeated which cause duplicate training data. This also saves the model again, wasting space