Learning fair representation with a parametric integral probability metric