Copying Nengo objects

[tbekolay]

From today's meeting, it would be nice to be able to copy Nengo object's cleanly. I think in general we want shallow copies (i.e., don't deepcopy the parameter values); deepcopies aren't necessary since everything we do to make sure we can make multiple simulators also means we don't need to do deepcopies.

Todo:

• Implement __getstate__ __setstate__ (see pickle-config branch; ensure we don't duplicate code with FrozenObject which does some copy stuff)
• Implement .copy method
  • Include an add_to_network arg on the copy method that adds or doesn't add to current Network.context
• Make sure it plays well with Python's copy module

[tbekolay]

Also, related to #941.

[jgosmann]

I started implementing this and it seems to be possible in a quite general manner. The added copy method on NengoObjects can take an add_to_network argument, but neither the Python copy.copy function can nor the unpickling functions. Should we add the object in those cases to the network or shouldn't we? Or maybe only if we are in the context of a network?

Probably we don't want to add the object always to a network (otherwise unpickling outside of a network would not be possible because it would raise an exception).

[hunse]

Yeah, I would say don't add it when unpickling.

Maybe we shouldn't allow copy.copy at all? We could just have a NotImplementedError that tells people to use our copy method. That would make things more straightforward, and I don't think it really makes anything inconvenient, either.

[jgosmann]

I'm using __getstate__ and __setstate__. I have to, to support pickling/unpickling and it gives me copying basically for free. But because of that it is not possible to raise NotImplementedError for the Python copy function. Furthermore, I'm using exactly that function in our copy implementation.

[hunse]

What about something like this?

[jgosmann]

Isn't this using a private method in the copy module?

[hunse]

Yep! Probably a bit strange, but I basically just took out the parts of their code that do copying if there's no __copy__ function. Though we should check if this works in 3.x. Also, maybe better to just write all our own code to do it using __getstate__ and __setstate__, which shouldn't be too hard. I just threw it together quick as a proof of concept.

[jgosmann]

To me it seems to be slightly more effort to disable Python's copy than it is worth. It's actually fully functional, the copy just doesn't get added to the network which might actually not to surprising (and of course in examples we would prefer NengoObject.copy where there is an add_to_network argument).

[hunse]

Yeah, I just don't want people to be confused or surprised by it not getting added to the model. I feel like on the user side, we want the default behaviour to be to add it to the network, since I think that's what users will expect. It seems strange if our copy does it one way and Python's copy does it the other.

[jgosmann]

I finished the PR. Using Python's copy or unpickling in the context of a network will produce a warning that the object was not added to the network. This still allows to use Pythonscopy` but the warning should make it clear that this might not have the intended effect.

[drasmuss]

Copying of Nengo objects is implemented now.