initial_layout or circuit._layout bug

[revilooliver]

Information

• Qiskit Terra version: 'qiskit-terra': '0.11.1',
• Python version: Python 3.7.3
• Operating system: Windows 10

What is the current behavior?

I tried to transpile the same circuit without and with initial_layout but the transpiler is returning an error message:
KeyError: "The item Qubit(QuantumRegister(20, 'q'), 0) does not exist in the Layout"

Steps to reproduce the problem

qr = QuantumRegister(3)
cr = ClassicalRegister(3)
qc = QuantumCircuit(qr, cr)
qc.cx(qr[0], qr[2])
qc.cx(qr[1], qr[2])
qc.measure(qr, cr)

trans_qc = transpile(qc, device)
ini_layout = trans_qc._layout.get_virtual_bits()
trans_qc2 = transpile(trans_qc, backend=device, initial_layout = ini_layout)

What is the expected behavior?

The initial layout is derived from the same circuit so the transpiler should return the same circuit.

Suggested solutions

[1ucian0]

trans_qc in the second transpile is an "embed" circuit. If what you want is to transpile the same circuit without and with initial layout, you should run transpile(qc, backend=device, initial_layout = ini_layout), or am I missing something?

[revilooliver]

Let me clarify it here. I wanted to test whether the transpiled circuit "trans_qc" changes if I transpile it again with the same initial_layout.

I derived the initial layout from the transpiled circuit "trans_qc"
ini_layout = trans_qc._layout.get_virtual_bits()

And transpiled it again:
transpile(trans_qc, backend=device, initial_layout = ini_layout)

But this will raise an error.
Is that because the transpiled circuit "trans_qc" and the original circuit "qc" are in different formats?

[1ucian0]

An "embed circuit" does not have a layout. I think you can reach your goal by running trans_qc2 = transpile(trans_qc, backend=device) without initial_layout. If you print the circuits after each transpilation, you will see the physical layout allocation.

This questions raises many thoughts:

• All the layout selection part should be skipped for embed circuits.
• Should we have to "types" of circuit, embed and not?
  • This would help passes, since some of them run on only embed ones.
• Should the "embed" property be in the circuit object and the transpiler make decisions on it?
  • A temporal fix could be use the internal circuit._layout property as a flag for this.

@kdk @ajavadia ideas?

[revilooliver]

I see. Is there any function/method that I can use to get the physical layout allocation rather than reading the printed circuits? It is difficult to manually check the figures. I thought trans_qc._layout should return the physical layout, but the result doesn't match with the printed circuit.

[1ucian0]

The property trans_qc._layoutreturns something like the following:

Layout({
5: Qubit(QuantumRegister(3, 'q3'), 0),
0: Qubit(QuantumRegister(3, 'q3'), 1),
...
18: Qubit(QuantumRegister(17, 'ancilla'), 15),
19: Qubit(QuantumRegister(17, 'ancilla'), 16)
})

That means that the original qr = QuantumRegister(3) created a quantum register called q3 and qr[0] is being allocated in the physical qubit 5. The ancilla entries were added by the transpiling process, because the backend I used has 20 qubits.

[revilooliver]

Thanks for @1ucian0 's explanation. There are two things that confused me:

1. Based on the naming, trans_qc._layout should be the layout for the transpiled circuit. However, the quantum registers in trans_qc._layout are actually the quantum registers in the original circuit.
   If we call trans_qc.qregs we will get [QuantumRegister(20, 'q')].
   But the trans_qc._layout returns:
   Layout({ 5: Qubit(QuantumRegister(3, 'q3'), 0), 7: Qubit(QuantumRegister(3, 'q3'), 1), ... 19: Qubit(QuantumRegister(17, 'ancilla'), 16) })
   There's a mismatch between the quantum registers names in trans_qc.qregs and trans_qc._layout.

2. The quantum register names in transpiled circuit are too simplified. The quantum registers are named from q[0] to q[n]. But it is difficult for the user to infer that q[0] is actually mapping to the physical qubit 0. Perhaps change the name to "physical_qubit[0]" or explaining it in the documentation.

This is not a bug and I will close it.