The difference of "=" and "clone" in the computation graph?

autograd
1

Here are two cases use “=” and “clone”:

(1)
x1 = torch.randn(4)
x1.requires_grad()
y1 = x1
z1 = y1.mean().backward()

(2)
x2 = torch.randn(4)
x2.requires_grad()
y2 = x2.clone()
z2 =

Since x1 and x2 can get their grads (which are equal), and my questions are:

  1. what’s the difference in the two cases?
  2. which one is preferred when creating a computation graph?
2

(1) Using = is simply setting y1 to point to the object that x1 points to, so y1 and x1 both points to the same Python object.
(2) Clone made a copy to x2, so even though they are still equal, they aren’t the same object. y1 points to the cloned object while x1 points to the original one.

3

Thanks for the explanation! And which way is better when create a computation graph in your option?

4

It depends on your purpose because they have different behavior. If you are only doing simple calculations, I suggest (1). One thing to note is that because (2) uses clone, (2) uses more memory.