Implemented Operations between two TensorDicts are defined by their position in their respective trees, not their keys

[ludwigwinkler]

Describe the bug

For example, the add function is implemented between two tensordicts.
Yet which of the leaves of two tensordicts are added is determined by their position in the respective tensordicts, not their explicit keys.

To Reproduce

import torch
from tensordict import TensorDict

x0 = TensorDict(
    {
        "y": torch.tensor([1.0, 0.1, -0.1]),
        "x": torch.tensor([-1.0, 0.1, 0.0]),
    }
)

x1 = TensorDict(
    {
        "x": torch.tensor([-1.0, 0.1, 0.0]),
        "y": torch.tensor([2.0, 0.1, -0.1]),
    }
)

print(f"+: ", (x0 + x1).to_dict())
print(f"*: ", (x0 * x1).to_dict())

Output:

+:  {'y': tensor([ 0.0000,  0.2000, -0.1000]), 'x': tensor([ 1.0000,  0.2000, -0.1000])}
*:  {'y': tensor([-1.0000,  0.0100, -0.0000]), 'x': tensor([-2.0000,  0.0100, -0.0000])}

Here the leaves are processed irrespective of their key.
Thus position in the tree trumps the keys.

Additional context

I just became aware of tensordict.apply() which should probably be used instead and implements a tree_map() functionality?

As far as I can tell, maybe the add functionality here should use an items() iterator which sortes the values by their keys?

tensordict/tensordict/base.py

Line 7026 in 6165e77

def add(self, other: TensorDictBase | float, alpha: float | None = None):

[vmoens]

On it