Backend argument in experiment constructors

[yaelbh]

BaseExperiment has an optional backend parameter. Do we want to make it mandatory, and remove the backend parameter from the run method?

If no: some experiments don't have a backend parameter in their constructors. Do we want, for the sake of uniformity, to instruct that all the experiments have this parameter? How can we enforce this instruction?

[yaelbh]

Also readout error experiments and tomography experiments.
In the case of tomography, the base class has the backend parameter, but this in practice has no effect since the state tomography and process tomography classes don't.

[wshanks]

Personally, I would prefer that backend were an __init__ argument. It is needed for core functionality of the experiment class and with it optional we have to jump through extra hoops to check if a backend is set or not and substitute default values when they can't be pulled from the backend. There is some case for building an experiment and then running it on different backends but I think the more common workflow is to focus on one backend at a time, and in that case passing the backend at initialization is not a burden.

If we want to preserve the ability to initialize an experiment instance and inspect it without setting up a specific backend, I wonder if we could define a generic backend and set it as the default argument for when a backend is not passed by the user (def __init__(self, qubits, backend=GenericBackend(), ...).

[yaelbh]

Note that, with the generic experiment, we again don't catch cases where experiment developers simply forget backend in their expeiment __init__. Like what seems to be the case with tomography.

[chriseclectic]

@yaelbh readout error and tomography now all take backend kwargs in init since #1019 #998. I personally like having backend as an optional kwarg in all experiments, so you can either set it at init, or you can set it at run. So in principle you can configure an experiment and run it an on multiple backends.

That said it would likely allow some simplifying of the base experiment if backend could only be set at init (you could remove the backend setter and set_backend method). The purpose of the set_backend method in base experiment was so that any functionality that depended on a backend could be specified here, but its still up to person writing an experiment to do validation for if backend is not set where it might be needed (ie in circuit generation). I suppose we could automate that exception handling by making the backend getter raise an exception if its None.

[yaelbh]

In a meeting we decided:

1. We will add the parameter to our existing experiment constructors where it is missing.
2. We will not enforce it to be present in experiment constructors.

[yaelbh]

One of the considerations to not enforcing was related to the fact backends are not serializable.