I meant the former (easy, fast, works well very often) but you could do either. If you decide to tune the whole thing, the training cost, time and complexity gr

I meant the former (easy, fast, works well very often) but you could do either. If you decide to tune the whole thing, the training cost, time and complexity grow by a lot, and of course you're destroying a lot of information that exists in the model.