Looks like you’re on the right track, but you might also need to convert pd_sales.advert.values to a Theano variable (e.g. a shared variable or constant). Also, you’ll either need to make normalize a Theano Op, as well, or convert those explicit Numpy operations into their Theano equivalents.
In general, if you use Numpy/Theano array/tensor methods—instead of functions/operators—you can get away with alot more, since their APIs are designed to match (somewhat).