Depends on the model application? If you have to carry around the graph for inference, then you want it compact. If you just fold the graph into some model, t

Depends on the model application? If you have to carry around the graph for inference, then you want it compact. If you just fold the graph into some model, then it doesnt really matter by inference time