Support qiskit2tq to parse Qiskit’s ParameterExpression

[king-p3nguin]

Resolve #263

Details

• I used sympy.lambdify to parse Qiskit's ParameterExpression. Any kind of ParameterExpression can be used as long as Qiskit or Sympy supports the expression. (See following example)
• I will wait until Make TQ Compatible with Qiskit ≥1.0.0 #267 is merged because I have not checked whether it still works with qiskit>=1.0.0.

Example (based on train_state_prep.py)

import torch
import torch.optim as optim

import torchquantum as tq
from torch.optim.lr_scheduler import CosineAnnealingLR

import random
import numpy as np
from qiskit import QuantumCircuit
from qiskit.circuit import ParameterVector, Parameter
from torchquantum.plugin import qiskit2tq

ansatz = QuantumCircuit(2)
ansatz_param = Parameter("Θ") # parameter
ansatz.rx(ansatz_param, 0)
ansatz_param_vector = ParameterVector("φ", 9) # parameter vector
ansatz.u(ansatz_param_vector[0], ansatz_param_vector[1], ansatz_param_vector[2], 0)
ansatz.u(
    ansatz_param_vector[3] + ansatz_param_vector[1], # parameter expression
    ansatz_param_vector[4] * ansatz_param_vector[2],
    ansatz_param_vector[5] / ansatz_param_vector[0],
    1,
)
ansatz.cu(
    np.sin(ansatz_param_vector[6]), # numpy functions
    np.abs(np.sin(ansatz_param_vector[7])), # nested numpy functions
    # complex expression
    np.abs(np.sin(ansatz_param_vector[8])) * np.exp(ansatz_param_vector[1] + ansatz_param_vector[2]),
    0.0,
    0,
    1,
)


def train(target_state, device, model, optimizer):
    model(device)
    result_state = device.get_states_1d()[0]

    # compute the state infidelity
    loss = 1 - torch.dot(result_state, target_state).abs() ** 2

    optimizer.zero_grad()
    loss.backward()
    optimizer.step()
    print(
        f"infidelity (loss): {loss.item()}, \n target state : "
        f"{target_state.detach().cpu().numpy()}, \n "
        f"result state : {result_state.detach().cpu().numpy()}\n"
    )


seed = 42
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)

device = torch.device("cpu")
model = qiskit2tq(ansatz).to(device)
print(model)

n_epochs = 10
optimizer = optim.Adam(model.parameters(), lr=1e-2, weight_decay=0)
scheduler = CosineAnnealingLR(optimizer, T_max=n_epochs)

q_device = tq.QuantumDevice(n_wires=2)
target_state = torch.tensor([0, 1, 0, 0], dtype=torch.complex64)

for epoch in range(1, n_epochs + 1):
    print(f"Epoch {epoch}, LR: {optimizer.param_groups[0]['lr']}")
    train(target_state, q_device, model, optimizer)
    scheduler.step()

Output:

QuantumModuleFromOps(
  (ops): QuantumModuleList(
    (0):  class: RX 
     parameters: Parameter containing:
    tensor([[2.4019]], requires_grad=True) 
     wires: [0] 
     inverse: False
    (1):  class: U3 
     parameters: Parameter containing:
    tensor([[ 2.6075, -0.7360,  2.8859]], requires_grad=True) 
     wires: [0] 
     inverse: False
    (2):  class: U3 
     parameters: Parameter containing:
    tensor([[-1.4243,  1.8295, -0.5866]], requires_grad=True) 
     wires: [1] 
     inverse: False
    (3):  class: CU 
     parameters: Parameter containing:
    tensor([[0.9626, 0.3636, 6.3739, 0.0000]], requires_grad=True) 
     wires: [0, 1] 
     inverse: False
  )
)
Epoch 1, LR: 0.01
infidelity (loss): 0.6692649126052856, 
 target state : [0.+0.j 1.+0.j 0.+0.j 0.+0.j], 
 result state : [-0.0999714 -0.6586203j  -0.5717409 -0.06202715j  0.29171193+0.07556738j
  0.29684204-0.21579707j]

Epoch 2, LR: 0.009755282581475769
infidelity (loss): 0.6621534824371338, 
 target state : [0.+0.j 1.+0.j 0.+0.j 0.+0.j], 
 result state : [-0.09573969-0.6596097j  -0.57720983-0.06837652j  0.28471547+0.07582995j
  0.29304212-0.21264003j]

Epoch 3, LR: 0.009045084971874737
infidelity (loss): 0.6551913022994995, 
 target state : [0.+0.j 1.+0.j 0.+0.j 0.+0.j], 
 result state : [-0.09154983-0.6604437j  -0.58290994-0.07088457j  0.27886063+0.07612759j
  0.28773254-0.21041745j]

Epoch 4, LR: 0.007938926261462366
infidelity (loss): 0.6487162113189697, 
 target state : [0.+0.j 1.+0.j 0.+0.j 0.+0.j], 
 result state : [-0.08761434-0.6611002j  -0.5879854 -0.07454494j  0.2731675 +0.07628112j
  0.28329697-0.20806299j]

Epoch 5, LR: 0.006545084971874737
infidelity (loss): 0.6430189609527588, 
 target state : [0.+0.j 1.+0.j 0.+0.j 0.+0.j], 
 result state : [-0.08412315-0.6615827j  -0.5923966 -0.07776422j  0.26821756+0.07635648j
  0.27936015-0.2060038j ]

Epoch 6, LR: 0.005
infidelity (loss): 0.638312816619873, 
 target state : [0.+0.j 1.+0.j 0.+0.j 0.+0.j], 
 result state : [-0.08122104-0.66191393j -0.5960295 -0.08022463j  0.2642149 +0.07638404j
  0.27601   -0.20435348j]

Epoch 7, LR: 0.003454915028125263
infidelity (loss): 0.6347126364707947, 
 target state : [0.+0.j 1.+0.j 0.+0.j 0.+0.j], 
 result state : [-0.07899112-0.6621257j  -0.5987737 -0.08220325j  0.2611553 +0.07637724j
  0.27347285-0.20307513j]

Epoch 8, LR: 0.0020610737385376348
infidelity (loss): 0.6322224140167236, 
 target state : [0.+0.j 1.+0.j 0.+0.j 0.+0.j], 
 result state : [-0.07744468-0.66225094j -0.60064095-0.08371409j  0.25901985+0.07635573j
  0.2717662 -0.20216379j]

Epoch 9, LR: 0.0009549150281252633
infidelity (loss): 0.6307358741760254, 
 target state : [0.+0.j 1.+0.j 0.+0.j 0.+0.j], 
 result state : [-0.07652031-0.662317j   -0.6017321 -0.08475052j  0.25772044+0.07633403j
  0.2707957 -0.20159304j]

Epoch 10, LR: 0.00024471741852423234
infidelity (loss): 0.6300469636917114, 
 target state : [0.+0.j 1.+0.j 0.+0.j 0.+0.j], 
 result state : [-0.07609182-0.66234535j -0.6022414 -0.08519605j  0.25712723+0.07632397j
  0.27033237-0.20133616j]

[01110011011101010110010001101111]

Hello, could you also add tests for the new functionality?

[king-p3nguin]

Tests are now added.

[01110011011101010110010001101111]

@king-p3nguin Merged in the >= 1.0.0 update; can you take a look into fixing these updates?

[king-p3nguin]

@GenericP3rson Thank you for the update. The tests are working now.