A minimal example of reverse-mode automatic differentiation
imitating the PyTorch interface. We create a bunch of Variable
objects, do simple operations with them and then and call
.backward() to compute the gradient components, accessible via
.grad:
x1 = Variable(1.0)
x2 = Variable(2.0)
a1 = x1*x2
a2 = x1/x2
b1 = a1/a2
b2 = a1*a2
y = b1-b2
y.backward()
# -2.0 4.0
print x1.grad, x2.grad