如何使用自动微分？¶

变分电路的参数可以使用Tensorflow或Pytorch框架进行优化。 作为深度学习框架，Tensorflow支持自动微分。下面的脚本通过以保真度作为相应的损失函数，优化两个旋转的参数，使电路输出与目标状态匹配。

请注意，如下例所示，旋转角度必须取实数值，以确保旋转门表示幺正算子。 Qibo不提供Tensorflow和Pytorch作为原生后端；必须安装并使用Qiboml来提供这些量子机器学习后端。

In [ ]:

from qibo import Circuit, gates, set_backend
from qibo.quantum_info import infidelity
import qibo

# 设置后端为 "qiboml" 和平台为 "tensorflow"
set_backend(backend="qiboml", platform="tensorflow")

# 获取当前后端并初始化 TensorFlow
backend = qibo.get_backend()
tf = backend.tf

# 优化参数
nepochs = 1000
optimizer = tf.keras.optimizers.Adam()
target_state = tf.ones(4, dtype=tf.complex128) / 2.0

# 定义电路的初始参数
params = tf.Variable(
    tf.random.uniform((2,), dtype=tf.float64)
)

# 定义量子电路的结构
circuit = Circuit(2)
circuit.add(gates.RX(0, params[0]))
circuit.add(gates.RY(1, params[1]))

# 开始优化循环
for _ in range(nepochs):
    with tf.GradientTape() as tape:
        # 设置电路参数
        circuit.set_parameters(params)
        # 计算最终量子态
        final_state = circuit().state()
        # 计算损失函数（不保真度）
        loss = infidelity(final_state, target_state, backend=backend)
    # 计算梯度
    grads = tape.gradient(loss, params)
    # 应用梯度更新参数
    optimizer.apply_gradients(zip([grads], [params]))

from qibo import Circuit, gates, set_backend from qibo.quantum_info import infidelity import qibo # 设置后端为 "qiboml" 和平台为 "tensorflow" set_backend(backend="qiboml", platform="tensorflow") # 获取当前后端并初始化 TensorFlow backend = qibo.get_backend() tf = backend.tf # 优化参数 nepochs = 1000 optimizer = tf.keras.optimizers.Adam() target_state = tf.ones(4, dtype=tf.complex128) / 2.0 # 定义电路的初始参数 params = tf.Variable( tf.random.uniform((2,), dtype=tf.float64) ) # 定义量子电路的结构 circuit = Circuit(2) circuit.add(gates.RX(0, params[0])) circuit.add(gates.RY(1, params[1])) # 开始优化循环 for _ in range(nepochs): with tf.GradientTape() as tape: # 设置电路参数 circuit.set_parameters(params) # 计算最终量子态 final_state = circuit().state() # 计算损失函数（不保真度） loss = infidelity(final_state, target_state, backend=backend) # 计算梯度 grads = tape.gradient(loss, params) # 应用梯度更新参数 optimizer.apply_gradients(zip([grads], [params]))

[Qibo 0.2.21|INFO|2025-10-15 10:53:24]: Using qiboml (tensorflow) backend on /device:CPU:0

In [2]:

print("Final parameters:", params)

print("Final parameters:", params)

Final parameters: <tf.Variable 'Variable:0' shape=(2,) dtype=float64, numpy=array([0.26564624, 1.40318672])>

In [3]:

final_state

final_state

Out[3]:

<tf.Tensor: shape=(4,), dtype=complex128, numpy=
array([0.75721303+0.j        , 0.6396014 +0.j        ,
       0.        -0.10117105j, 0.        -0.08545699j])>

In [4]:

loss

loss

Out[4]:

<tf.Tensor: shape=(), dtype=float64, numpy=0.5035198562285943>

In [5]:

nepochs = 1000
optimizer = tf.keras.optimizers.Adam()
target_state = tf.ones(4, dtype=tf.complex128) / 2.0
params = tf.Variable(tf.random.uniform((2,), dtype=tf.float64))

@tf.function
def optimize(params):
    with tf.GradientTape() as tape:
        circuit = Circuit(2)
        circuit.add(gates.RX(0, theta=params[0]))
        circuit.add(gates.RY(1, theta=params[1]))
        final_state = circuit().state()
        loss = infidelity(final_state, target_state, backend=backend)
    grads = tape.gradient(loss, params)
    optimizer.apply_gradients(zip([grads], [params]))

for _ in range(nepochs):
    optimize(params)

nepochs = 1000 optimizer = tf.keras.optimizers.Adam() target_state = tf.ones(4, dtype=tf.complex128) / 2.0 params = tf.Variable(tf.random.uniform((2,), dtype=tf.float64)) @tf.function def optimize(params): with tf.GradientTape() as tape: circuit = Circuit(2) circuit.add(gates.RX(0, theta=params[0])) circuit.add(gates.RY(1, theta=params[1])) final_state = circuit().state() loss = infidelity(final_state, target_state, backend=backend) grads = tape.gradient(loss, params) optimizer.apply_gradients(zip([grads], [params])) for _ in range(nepochs): optimize(params)

In [6]:

import torch

from qibo import Circuit, gates, set_backend
from qibo.quantum_info.metrics import infidelity

set_backend(backend="qiboml", platform="pytorch")

# Optimization parameters
nepochs = 1000
optimizer = torch.optim.Adam
target_state = torch.ones(4, dtype=torch.complex128) / 2.0

# Define circuit ansatz
params = torch.tensor(
    torch.rand(2, dtype=torch.float64), requires_grad=True
)
circuit = Circuit(2)
circuit.add(gates.RX(0, params[0]))
circuit.add(gates.RY(1, params[1]))

optimizer = optimizer([params])

for _ in range(nepochs):
    optimizer.zero_grad()
    circuit.set_parameters(params)
    final_state = circuit().state()
    loss = infidelity(final_state, target_state)
    loss.backward()
    optimizer.step()

import torch from qibo import Circuit, gates, set_backend from qibo.quantum_info.metrics import infidelity set_backend(backend="qiboml", platform="pytorch") # Optimization parameters nepochs = 1000 optimizer = torch.optim.Adam target_state = torch.ones(4, dtype=torch.complex128) / 2.0 # Define circuit ansatz params = torch.tensor( torch.rand(2, dtype=torch.float64), requires_grad=True ) circuit = Circuit(2) circuit.add(gates.RX(0, params[0])) circuit.add(gates.RY(1, params[1])) optimizer = optimizer([params]) for _ in range(nepochs): optimizer.zero_grad() circuit.set_parameters(params) final_state = circuit().state() loss = infidelity(final_state, target_state) loss.backward() optimizer.step()

[Qibo 0.2.21|INFO|2025-10-15 11:03:56]: Using qiboml (pytorch) backend on cpu
C:\Users\Administrator\AppData\Local\Temp\ipykernel_19616\2161827941.py:14: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.detach().clone() or sourceTensor.detach().clone().requires_grad_(True), rather than torch.tensor(sourceTensor).
  params = torch.tensor(

In [7]:

final_state

final_state

Out[7]:

tensor([0.6389+0.0000j, 0.6061+0.0000j, 0.0000-0.3437j, 0.0000-0.3261j],
       dtype=torch.complex128, grad_fn=<ViewBackward0>)

In [8]:

loss

loss

Out[8]:

tensor(0.5003, dtype=torch.float64, grad_fn=<RsubBackward1>)

In [ ]: