Level2 memory

[eggerdj]

Summary

This PR fixes #800

Details and comments

The fix is done by adding proper memory marginalization for level 2. We distinguish between level 1 and level 2 using list vs string. This code can be tested on hardware by running the following:

from qiskit_experiments.library import FineXAmplitude
from qiskit_experiments.framework import ParallelExperiment
from qiskit import IBMQ

backend = ...

experiments = []
for qubit in [0, 1]:
    fine_amp = FineXAmplitude(qubit, backend)
    fine_amp.enable_restless()
    experiments.append(fine_amp)
    
parallel_exp = ParallelExperiment(experiments, backend)
parallel_exp.set_run_options(**experiments[0].run_options.__dict__)

exp_data = parallel_exp.run()

With the changes in this PR we get the following output:

exp_data.component_experiment_data(0).figure(0)

and

exp_data.component_experiment_data(1).figure(0)

[chriseclectic]

Some suggestions for cleaning up code a little bit

[chriseclectic]

Since ParallelExperiments always concatenate clbits for sub-experiments you are garuenteed that the composite_clbits are a continuous range. In this case you can simplify this processing like:

clbits = composite_bits[i]
idx = slice(clbits[0], clbits[-1] + 1)
marginalized = [format_counts_memory(mem, header)[idx] for mem in datum["memory"]]

If performance becomes an issue for large numbers of shots (which im sure it will at some point) we should probably update or replace the format_counts_memory into a more efficient vectorized function that can go straight from list/array hex str -> list/array marginal counts

[eggerdj]

Thanks. See here: 871400e note that you still need the [::-1] to conform to Qiskit's bit ordering. Unless the composite_bits already account for this which I doubt? Note that before 871400e I realized that we were calling format_counts_memory a total of n_shots x n_subexperiments times. This is inefficient and by pulling this out of the loop over for i, index in enumerate(metadata["composite_index"]) we only need to call it n_shots times which is much better.

[chriseclectic]

Is it possible to add a test for level-1 marginalization?

[eggerdj]

See here: 5b3eee0 This also required adding a fix to be able to handle single-shot memory vs averaged memory.

[chriseclectic]

I forgot about the reversed bit indexing in my slice example. I think you can remove the need to reverse and re-reverse with ::-1 if you build the slice with that already in mind. I tried some quick tests and I think this is correct:

idx = slice(-1 - composite_qubits[i][-1], -composite_qubits[i][0] or None)
sub_data["memory"] = [shot[idx] for shot in f_memory]

The None case is needed to handle when a qubit[0] = 0

[eggerdj]

That's better thanks. See here: f9eefda

[chriseclectic]

Looks good, its just missing a release note now saying what the "bug fix" was.