Allow backend cal to export instmap

[nkanazawa1989]

Summary

This PR allows BackendCalibrations to override instruction schedule map.
This requires Qiskit/qiskit#6759

Here is the expected usage.

my_calibration = BackendCalibrations(backend)

my_experiment = qiskit_experiments.MyExperiment(**exp_options)
# here we can set instmap-overridden backend from calibration, instead of backend above
result = my_experiment.run(my_calibration.export_backend())

Note that this framework doesn't require special logic to update calibration on experiment end.
Thus this can be applied to any kind of experiment we define.

If we want to apply custom gate

my_calibration = BackendCalibrations(backend)
basis_gates = backend.configurations().basis_gates + ["my_su4_gate"]

my_experiment = qiskit_experiments.MyExperiment(**exp_options)
my_experiment.set_transpiler_options(basis_gates=basis_gates)
result = my_experiment.run(my_calibration.export_backend(basis_gates=basis_gates))

Again, this doesn't require any custom logic. For example, we can run QV experiment with custom gates without modifying the codebase. This will add more values to the calibration module in terms of smooth integration of pulse and algorithm-level experiment.

It is noteworthy that the backend exported by the BackendCalibrations is a standard backend object. This means we can calibrate specific backend, share the experiment with the team, and let other researchers use the custom gate in other qiskit algorithm modules.

Details and comments

• Finalize spec, we have alternative [WIP] Injecting cals in calibration experiment #180
• Merge Add pulse gate pass Qiskit/qiskit#6759 (also we need to update meta package with experiment release)
• Write unittest (for both examples above)
• Write TODO

[eggerdj]

I like this approach. A test or two would be nice.

[eggerdj]

The one thing that this methodology along with the terra PR will not allow us to do is inject calibrations for the custom gates we created in e.g. Rabi and Drag as, e.g., an xp pulse will sometimes be tied to a gate called Rabi and sometimes to a gate called Rp. It would be nice to simplify part of the circuit builders there.

[nkanazawa1989]

Thanks @eggerdj . In experiments that may use custom gate name you always need to

backend = my_cal.export(basis_gates=["Rabi", "measure"])

exp = Rabi(qubit=0)
exp.set_transpile_options(basis_gates=["Rabi", "measure"])  # this can be overridden by self._default_transpile_options
exp.run(backend)

and it might be dynamically updated by self.set_transpile_options depending on configuration. I agree we can simplify the logic to manage gate name.

[eggerdj]

In experiments that may use custom gate name you always need to

backend = my_cal.export(basis_gates=["Rabi", "measure"])

exp = Rabi(qubit=0)
exp.set_transpile_options(basis_gates=["Rabi", "measure"])  # this can be overridden by self._default_transpile_options
exp.run(backend)

and it might be dynamically updated by self.set_transpile_options depending on configuration. I agree we can simplify the logic to manage gate name.

I agree with the need for basis_gates but it will miss the following situation. We have xp and x90p schedules in my_cals. Sometimes we want to run the Rabi experiment on xp and sometimes on x90p. This means we need to attach schedules with names that can change to a gate which is always called Rabi. To me it looks like we would need to do:

backend = my_cal.export(basis_gates=["Rabi", "measure"], schedule_gate_map={"xp": "Rabi"})

exp = Rabi(qubit=0)
exp.set_transpile_options(basis_gates=["Rabi", "measure"])  # this can be overridden by self._default_transpile_options
exp.run(backend)

and to use the x90p we would do

backend = my_cal.export(basis_gates=["Rabi", "measure"], schedule_gate_map={"x90p": "Rabi"})

I don't see how we get around not using something like a schedule_gate_map in the export to backend method. Note that the schedule_gate_map is optional. I.e. if you do not specify it then xp is simply xp in the inst. map. Optionally one could also consider

backend = my_cal.export(basis_gates=["Rabi", "measure"], schedule_gate_map={"x90p": ["Rabi", "x90p"]})

in case we need, for example, x90p as both a Rabi and an x90p.

[yaelbh]

@nkanazawa1989 Is this PR still relevant?

[eggerdj]

Can we keep this open for a bit longer? There are some ideas in here that I may want to consult as I refactor the calibration experiments.