Question about inplace operation (RuntimeError: one of the variables (...))

autograd
#1

The following code snippet produces the RuntimeError “one of the variables needed for gradient computation has been modified by an inplace operation”

for i in range(n):
arr[i] = torch.cross(vec[i], vec[i+1])
arr[i] = arr[i] / torch.norm(arr[i])

Whereas if I refactor it to the following:

for i in range(n):
tmp = torch.cross(vec[i], vec[i+1])
arr[i] = tmp / torch.norm(tmp)

It runs without any issue. My question is, is the former example an inplace operation? Does reassigning an array not create a new variable?

#2

if arr is a torch tensor, then assiging into arr[i] will copy the data inplace in the tensor.
So yes this is an inplace operation.