如何明确地将可区分函数传递给T.grad？

Question

请告诉我，如何将各种差异功能传递给T.grad？我想要这样的东西：

x = T.dscalar('x')

ellipic_paraboloid = x ** 2 + y ** 2
hyperbolic_paraboloid = x ** 2 - y ** 2

gradients = theano.function([function_of_xy, x, y], T.grad(function_of_xy, gx, gy))

gradients(ellipic_paraboloid, 1, 1)
gradients(hyperbolic_paraboloid, 1, 1)