Why change a tensor's value in a function can change the outer tensor?

1

I found that defining a tensor variable and change its value in a function, outer the function the tensor’s value will change though I did not return anything. So what does happen when I did things above?

from torch import Tensor
a = Tensor([0])
>>> a = tensor(0.)
def b(t):
    t[0]=3
b(a)
print(a)
>>> a=tensor(3.)

It’s may about the scope of a tensor or torch low-level method. However I want someone can give me a complete reason to explain it.

2

The reason for the above scenario is that tensors are mutable objects therefore they are changeable in-place
This implies that when you call b(a), instead of a new local variable a being created in the function scope, a ‘reference’ to a will be made and a[0] will be assigned the value 3.

However if a wasn’t a mutable object (not changeable in-place) the reverse would happen.i.e a new variable a would be created in the function scope and thus not affecting the outer a variable.
For instance, the function below doesn’t change the outer/global a

a = 0
def my_function(b):
    b = 3
    print(b)

my_function(a)
print(a)
2 Likes