You never destandarize input data. Network (or any other machine learning model) won't understand data which is in different scale/space than the one that was used during training.
If you did, however scale the output values, than in an obvious way to have to scale them back in order to obtain "unscaled" results.