The init and apply split is probably the part that clicked for me most. Keeping model structure separate from the actual weights feels much cleaner. Especially when you start adding dropout and batch stats.