Gender Slopes: Counterfactual Fairness for Computer Vision Models by Attribute Manipulation