You can’t train an integer, but you can train a float.
As I already said in a previous answer, you change the classification part of the record reader to be regression.
And then you change the output layer to use LossMSE with Identity activation.
And obviously you will want to apply regression evaluation instead of the normal classification regression.