numpy to PyTorch tensor conversion - memory sharing with respect to '+=' and '=' operators

[AiKoushik]

I was going through numpy to tensor and tensor to numpy conversion, where it was mentioned that memory is not shared between the original and the new tensor/array.

But while performing += for increment, they behaved like they are sharing the same memory. Although the moment I use statements like tensor = tensor + 1 that sharing stops e.g. tensor += 1

Can someone tell me the right topic to read so it doesn't feel like a quirky behaviour?

I am using Google Colab and I'm on CPU.
Code:

import numpy as np

array = np.arange(1, 8)
tensor = torch.from_numpy(array)

array, tensor

array += 1
array, tensor

array = array + 1
array, tensor

array += 1
array, tensor

[sametucuncu]

The "+=" operates in place, so it will change the original numpy array.
But when you use:
array = array + 1
It is same as:
array2 = array + 1
So basically, you are creating a new variable with a different memory address.

You can try these codes and observe how each variable memory address changes.
(Try them in seperate cells.)

import numpy as np
array = np.arange(1, 8)
tensor = torch.from_numpy(array)
# tensor and array share the same memory.

print(f"array: {array}, tensor: {tensor}")
print(f"address of array: {id(array)}, address of tensor: {id(tensor)}")

array += 1 
print(f"array: {array}, tensor: {tensor}") 
print(f"address of array: {id(array)}, address of tensor: {id(tensor)}")
# You will see the id of array same as above.

array = array + 1 
print(f"array: {array}, tensor: {tensor}") 
print(f"address of array: {id(array)}, address of tensor: {id(tensor)}")
# You will see the id of array has changed. Because it is different variable.

array += 1 
print(f"array: {array}, tensor: {tensor}") 
print(f"address of array: {id(array)}, address of tensor: {id(tensor)}")

[AiKoushik]

Thank you. I understood the same from my searching.
So basically the original tensor and the derived ones share the same memory. It's just a mistake in the related video where it says that these use different memory locations.

[shisirkha]

Good explanation of this concept.

I was just scrolling through the discussion page to help others but I'm pretty impressed by your explanation.

Nice.