[BUG] `unravel_key_to_tuple` returns empty tuple when non-`str` elements are present

[LCarmi]

Describe the bug

_unravel_key_to_tuple returns empty tuple when elements that are not tuple or str are present in the key. The expected behavior should be to drop the non-str elements, and return the other elements as a flat tuple.

To Reproduce

>>> import tensordict
>>> ok_key = ("agent", "observation")
>>> ko_key = (None, "observation")
>>> tensordict.utils._unravel_key_to_tuple(ok_key)
('next', 'agent', 'observation')
>>> tensordict.utils._unravel_key_to_tuple(ko_key)
()

Expected behavior

The expected behavior is the path taken when torch.is_compiling() == True

>>> tuple(subk for k in ok_key for subk in tensordict.utils._unravel_key_to_tuple(k))
('next', 'agent', 'observation')
>>> tuple(subk for k in ko_key for subk in tensordict.utils._unravel_key_to_tuple(k))
('next', 'observation')

Reason and Possible fixes

In the file https://github.com/pytorch/tensordict/blob/main/tensordict/csrc/utils.cpp, we can see that

if (_key.size() == 0) {
  return py::make_tuple();
}

which discards the previously accumulated keys in newkey

Checklist

• I have checked that there is no similar issue in the repo (required)
• I have read the documentation (required)
• I have provided a minimal working example to reproduce the bug (required)

[vmoens]

This is the intended behaviour because we want to process only keys that are properly defined NestedKeys and cowardly return an indicator that we didn't find a match for that otherwise.

Happy to reconsider this behaviour. Let me ask this: What would be the goal of having a None in the key? Mixing shape and key indexing I suppose?