tensorflow / quantum

Hybrid Quantum-Classical Machine Learning in TensorFlow
https://www.tensorflow.org/quantum
Apache License 2.0
1.77k stars 565 forks source link

Multi-qubit X gate promoted to TOFFOLI gate and could not be serialized by TFQ. #480

Open we-taper opened 3 years ago

we-taper commented 3 years ago

When a multi-qubit X gate is created by the .controlled method, it maybe automatically promoted to cirq.TOFFOLI gate which isn't supported by TFQ.

Minimal example:

import tensorflow_quantum as tfq
qbs = cirq.GridQubit.rect(4, 1)
circuit = cirq.Circuit(
    cirq.X.controlled(num_controls=3).on(*qbs)
)
try:
  layer([circuit])
except Exception as e:
  print(f"Exception:\n{e}")
# Pinrt:
# Exception:
# Cannot serialize op cirq.TOFFOLI(cirq.GridQubit(1, 0), cirq.GridQubit(2, 0), cirq.GridQubit(3, 0)) of type <class 'cirq.ops.three_qubit_gates.CCXPowGate'>

Currently, the work-around is to use cirq.ControlledGate(sub_gate=cirq.X, num_controls=3).on(*qbs) instead. Similar problem happen to cirq.Z as well, which is promoted to cirq.CCZ. Luckily, cirq.Y works because cirq.CCY does not exist.

MichaelBroughton commented 3 years ago

After talking this over with the Cirq people it turns out that this is intended behavior. So we have to turn around and update our serializer functionality to support Toffoli and CCZ.

Danielle-Schuman commented 3 years ago

I am having the same issue, but the workaround @we-taper mentioned does not work for me, since I am getting the same error for the ControlledGate. @we-taper: Are you using the regular TFQ version or the nightly one?

we-taper commented 3 years ago

@Danielle-Schuman I was using the nightly one at the time.

Danielle-Schuman commented 3 years ago

Ok, thank you for the answer, @we-taper :). I was using the regular version, so that might have caused the different behaviour. I meanwhile have found a workaround, though. (I just implemented the Toffoli-gate in a decomposed manner, as described in https://en.wikipedia.org/wiki/Toffoli_gate#/media/File:Qcircuit_ToffolifromCNOT.svg , that worked with the regular TFQ version, too.)

we-taper commented 3 years ago

Thanks! @Danielle-Schuman

Actually, why not working with the nightly-version? The major benefit is that gates will be implemented natively which means faster execution. So far it has been quite smooth for me to use the nightly version (or to be specific tfq-nightly==0.5.0.dev20210218).

lockwo commented 3 years ago

I see there is a pull request under review to fixing this thread, is there plans to merge that soon? It's not breaking since you can use the techniques presented above to solve the problem, but it's a weird interaction.