> are computationally expensive and so not typically done.

how does this expense compare to the training of the model? surely its a vanishing fraction?