__deepcopy__ for DualGraphModule

[andreasfloros]

DualGraphModule inherits deepcopy behavior from GraphModule, which ignores the train_graph and eval_graph.

This PR implements __deepcopy__ for DualGraphModule.

Fixes #8634

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/vision/8708

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit 4e10111 with merge base 22e86bd ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[NicolasHug]

Thank you for the PR @andreasfloros . To be completely transparent here, I was hoping we could find a much more obvious fix for this problem. The current fix in this PR involves importing and relying on various private structures from torch.fx.graph_module, and I'll admit I'm not confident enough in my understanding of that code to be able to land it right now.

Basically, if that code were to break (because e.g. of an upstream change in some internal of torch.fx.graph_module), I wouldn't be able to fix this again.

Do you think we can address the original problem from #8634 in a more obvious way?

[andreasfloros]

@NicolasHug the issue is with the hooks. Other than that, deepcopy can be implemented without private structures.
A more obvious implementation is to copy the entire __dict__:

def __deepcopy__(self, memo):
    res = type(self).__new__(type(self))
    memo[id(self)] = res
    res.__dict__ = copy.deepcopy(self.__dict__, memo)
    return res

However, this is not done in GraphModule and there is a comment regarding bc (https://github.com/pytorch/pytorch/blob/main/torch/fx/graph_module.py#L880).

What do you think?

[NicolasHug]

Oh, I didn't realize the fix was actually mostly taken from the implementation in the base class (you did write that as a comment, I missed it!)
I guess that's fine. Let's hope this file doesn't get updated too often. If it does, we might have to revert the fix, but for now this is hopefully OK.

Thanks a lot for the fix @andreasfloros . I pushed a small non-regression test and I'll merge the PR soon