Video "PyTorch Fundamentals" - Having a difference of output

[Arcane-WD]

Hello, I have a small doubt in the vid part "PyTorch and NumPy". I create a new ones tensor, then create a variable to hold numpy version of said tensor, then increment tensor using tensor = tensor + 1 there is no data shared b/w tensor and numpy_tensor. But if I use tensor += 1, the data is shared and both show the same output.

[hillions]

From the video, it looks like array = array + 1 creates a new numpy.ndarray, while doing something like array[0] = 9 will not create a new one. (keeping the memory shared) You can check this by looking at the memory location of the array:

array = np.arange(1.0, 8.0)
print(array)
print(f"NumPy array location: {hex(id(array))}")
array[0] = 9
print(array)
print(f"NumPy array location: {hex(id(array))}")
array = array + 1
print(array)
print(f"NumPy array location: {hex(id(array))}")

Gives an output like this:

[1. 2. 3. 4. 5. 6. 7.]
NumPy array location: 0x7fd468e05bd0
[9. 2. 3. 4. 5. 6. 7.]
NumPy array location: 0x7fd468e05bd0
[10.  3.  4.  5.  6.  7.  8.]
NumPy array location: 0x7fd468e11810

I'm guessing that when you create a numpy.ndarray, then a torch.Tensor from it using torch.from_numpy(), then the torch.Tensor still references the old array memory even after you create a new one with array = array + 1. They have become disassociated when you do array = array + 1.

[udaykumarswamy]

Hey @Arcane-WD

here what i found from the documentation of torch.numpy()

torch.Tensor.numpy
Tensor.numpy(*, force=False) → numpy.ndarray
Returns the tensor as a NumPy ndarray.

If force is False (the default), the conversion is performed only if the tensor is on the CPU, does not require grad, does not have its conjugate bit set, and is a dtype and layout that NumPy supports. The returned ndarray and the tensor will share their storage, so changes to the tensor will be reflected in the ndarray and vice versa.

i hope that should resolve your doubt.