save_image manipulates input tensor inplace

[ptrblck]

Reported in forum: topic by Shisho_Sama.

Reason:
save_image manipulates the input tensor inplace in this line of code.

Code snippet to reproduce:

x = torch.randn(10, 1, 100, 100)
grid = torchvision.utils.make_grid(x)
print(grid.min(), grid.max(), grid.mean())
> tensor(-4.4464) tensor(4.4604) tensor(-0.0003)
torchvision.utils.save_image(grid, 'tmp.png')
print(grid.min(), grid.max(), grid.mean())
> tensor(0.) tensor(255.) tensor(48.1135)

Working on a copy would solve this issue.

Let me know, if you would be interested in a quick fix adding a clone() to the mentioned line of code or if another approach would be preferable.

[pmeier]

This was already fixed in #1244.

[ptrblck]

Thanks for the information and sorry for not being able to find the PR :/