After some calculation to obtain x 1 y, then appeared dy / dx, and a mapping x → y (y-how obtained by the x-operation).
Obtaining dy / dx need two steps: y.backward (), x.grad, i.e. back-propagation, a gradient is obtained
x → y mapping grad_fn Tensor is a property of the object: y.grad_fn
Note that the back-propagation gradient will accumulate, so it should be cleared before the gradient back-propagation
t.ones = x (2,2 &, requires_grad = True) # tracking all arithmetic operations on the x y = x.sum () # 4. Note that only a scalar y y.grad_fn # y is obtained by a calculation of what , SumBackward0 AT 0x2598370d948 # Y x differential pair, i.e., a gradient of Dy / DX y.backward () # backpropagation Print (x.grad) # output gradient # backpropagation gradient will accumulate, so it should be cleared before the backpropagation gradient y.backward () Print (x.grad) # output gradient y.backward () Print (x.grad) # output gradient # gradient cleared x.grad.zero_ () y.backward () # backpropagation Print (X .grad) # output gradient
