Question - the best loss for learning function depending on angle and radius

[serhiy-yevtushenko]

Hi. Could you please advice, what type of loss suite better, if estimated function returns angular value, but real loss (in which one is interested) depends as well on the distance from coordinates start point (0, 0) (i,e., functions returns angle, but practicall loss is proportional to radius*sin(angular value), and requires as well expensive computation (although less expensive that original function computation))?