Brilliant idea. That would make it useful for Keras, too, using decorated callbacks during training