I don’t want to challenge your efforts, but I think that the final outputs could also be computed by the Dot() so that we can avoid the full computation in test or validation.
right?
I am reiterating that the current implementation serves pedagogical purposes.
The assignment is more focused here on sequential vectorization of question and answer model and then create a model to get desired output based on designated hidden layer output. But this can be surely created with dot() product.
I suppose this assignment is more focused on task specificity model.