- Demos/
- Quantum Machine Learning/
Quantum generative adversarial networks with Cirq + TensorFlow
Quantum generative adversarial networks with Cirq + TensorFlow
Published: October 10, 2019. Last updated: October 14, 2025.
Warning
This demo uses TensorFlow, which is a deprecated interface with PennyLane v0.42. Interfacing with TensorFlow will no longer be supported with PennyLane v0.43 and higher. Consider switching to a different machine learning interface with PennyLane, like PyTorch or JAX.
This demo constructs a Quantum Generative Adversarial Network (QGAN) (Lloyd and Weedbrook (2018), Dallaire-Demers and Killoran (2018)) using two subcircuits, a generator and a discriminator. The generator attempts to generate synthetic quantum data to match a pattern of “real” data, while the discriminator tries to discern real data from fake data (see image below). The gradient of the discriminator’s output provides a training signal for the generator to improve its fake generated data.
Using Cirq + TensorFlow
PennyLane allows us to mix and match quantum devices and classical machine learning software. For this demo, we will link together Google’s Cirq and TensorFlow libraries.
We begin by importing PennyLane, NumPy, and TensorFlow.
import numpy as np
import pennylane as qml
import tensorflow as tf
We also declare a 3-qubit simulator device running in Cirq.
dev = qml.device('cirq.simulator', wires=3)
Generator and Discriminator
In classical GANs, the starting point is to draw samples either from some “real data” distribution, or from the generator, and feed them to the discriminator. In this QGAN example, we will use a quantum circuit to generate the real data.
For this simple example, our real data will be a qubit that has been rotated (from the starting state \(\left|0\right\rangle\)) to some arbitrary, but fixed, state.
def real(angles, **kwargs):
qml.Hadamard(wires=0)
qml.Rot(*angles, wires=0)
For the generator and discriminator, we will choose the same basic circuit structure, but acting on different wires.
Both the real data circuit and the generator will output on wire 0, which will be connected as an input to the discriminator. Wire 1 is provided as a workspace for the generator, while the discriminator’s output will be on wire 2.
def generator(w, **kwargs):
qml.Hadamard(wires=0)
qml.RX(w[0], wires=0)
qml.RX(w[1], wires=1)
qml.RY(w[2], wires=0)
qml.RY(w[3], wires=1)
qml.RZ(w[4], wires=0)
qml.RZ(w[5], wires=1)
qml.CNOT(wires=[0, 1])
qml.RX(w[6], wires=0)
qml.RY(w[7], wires=0)
qml.RZ(w[8], wires=0)
def discriminator(w):
qml.Hadamard(wires=0)
qml.RX(w[0], wires=0)
qml.RX(w[1], wires=2)
qml.RY(w[2], wires=0)
qml.RY(w[3], wires=2)
qml.RZ(w[4], wires=0)
qml.RZ(w[5], wires=2)
qml.CNOT(wires=[0, 2])
qml.RX(w[6], wires=2)
qml.RY(w[7], wires=2)
qml.RZ(w[8], wires=2)
We create two QNodes. One where the real data source is wired up to the discriminator, and one where the generator is connected to the discriminator.
@qml.qnode(dev)
def real_disc_circuit(phi, theta, omega, disc_weights):
real([phi, theta, omega])
discriminator(disc_weights)
return qml.expval(qml.PauliZ(2))
@qml.qnode(dev)
def gen_disc_circuit(gen_weights, disc_weights):
generator(gen_weights)
discriminator(disc_weights)
return qml.expval(qml.PauliZ(2))
QGAN cost functions
There are two cost functions of interest, corresponding to the two stages of QGAN training. These cost functions are built from two pieces: the first piece is the probability that the discriminator correctly classifies real data as real. The second piece is the probability that the discriminator classifies fake data (i.e., a state prepared by the generator) as real.
The discriminator is trained to maximize the probability of correctly classifying real data, while minimizing the probability of mistakenly classifying fake data.
The generator is trained to maximize the probability that the discriminator accepts fake data as real.
def prob_real_true(disc_weights):
true_disc_output = real_disc_circuit(phi, theta, omega, disc_weights)
# convert to probability
prob_real_true = (true_disc_output + 1) / 2
return prob_real_true
def prob_fake_true(gen_weights, disc_weights):
fake_disc_output = gen_disc_circuit(gen_weights, disc_weights)
# convert to probability
prob_fake_true = (fake_disc_output + 1) / 2
return prob_fake_true
def disc_cost(disc_weights):
cost = prob_fake_true(gen_weights, disc_weights) - prob_real_true(disc_weights)
return cost
def gen_cost(gen_weights):
return -prob_fake_true(gen_weights, disc_weights)
Training the QGAN
We initialize the fixed angles of the “real data” circuit, as well as the initial parameters for both generator and discriminator. These are chosen so that the generator initially prepares a state on wire 0 that is very close to the \(\left| 1 \right\rangle\) state.
phi = np.pi / 6
theta = np.pi / 2
omega = np.pi / 7
np.random.seed(0)
eps = 1e-2
init_gen_weights = np.array([np.pi] + [0] * 8) + \
np.random.normal(scale=eps, size=(9,))
init_disc_weights = np.random.normal(size=(9,))
gen_weights = tf.Variable(init_gen_weights)
disc_weights = tf.Variable(init_disc_weights)
We begin by creating the optimizer:
opt = tf.keras.optimizers.SGD(0.4)
opt.build([disc_weights, gen_weights])
In the first stage of training, we optimize the discriminator while keeping the generator parameters fixed.
cost = lambda: disc_cost(disc_weights)
for step in range(50):
with tf.GradientTape() as tape:
loss_value = cost()
gradients = tape.gradient(loss_value, [disc_weights])
opt.apply_gradients(zip(gradients, [disc_weights]))
if step % 5 == 0:
cost_val = loss_value.numpy()
print("Step {}: cost = {}".format(step, cost_val))
At the discriminator’s optimum, the probability for the discriminator to correctly classify the real data should be close to one.
print("Prob(real classified as real): ", prob_real_true(disc_weights).numpy())
For comparison, we check how the discriminator classifies the generator’s (still unoptimized) fake data:
print("Prob(fake classified as real): ", prob_fake_true(gen_weights, disc_weights).numpy())
In the adversarial game we now have to train the generator to better fool the discriminator. For this demo, we only perform one stage of the game. For more complex models, we would continue training the models in an alternating fashion until we reach the optimum point of the two-player adversarial game.
cost = lambda: gen_cost(gen_weights)
for step in range(50):
with tf.GradientTape() as tape:
loss_value = cost()
gradients = tape.gradient(loss_value, [gen_weights])
opt.apply_gradients(zip(gradients, [gen_weights]))
if step % 5 == 0:
cost_val = loss_value.numpy()
print("Step {}: cost = {}".format(step, cost_val))
At the optimum of the generator, the probability for the discriminator to be fooled should be close to 1.
print("Prob(fake classified as real): ", prob_fake_true(gen_weights, disc_weights).numpy())
At the joint optimum the discriminator cost will be close to zero, indicating that the discriminator assigns equal probability to both real and generated data.
print("Discriminator cost: ", disc_cost(disc_weights).numpy())
The generator has successfully learned how to simulate the real data enough to fool the discriminator.
Let’s conclude by comparing the states of the real data circuit and the generator. We expect the generator to have learned to be in a state that is very close to the one prepared in the real data circuit. An easy way to access the state of the first qubit is through its Bloch sphere representation:
obs = [qml.PauliX(0), qml.PauliY(0), qml.PauliZ(0)]
@qml.qnode(dev)
def bloch_vector_real(angles):
real(angles)
return [qml.expval(o) for o in obs]
@qml.qnode(dev)
def bloch_vector_generator(angles):
generator(angles)
return [qml.expval(o) for o in obs]
print(f"Real Bloch vector: {bloch_vector_real([phi, theta, omega])}")
print(f"Generator Bloch vector: {bloch_vector_generator(gen_weights)}")
About the author
Nathan Killoran
Steering software-driven research and research-driven software at Xanadu
