How would you diagnose and address a deep learning model that trains well but performs poorly in production?