Optimization & machine learning tutorial

Section author: Nathan Killoran <nathan@xanadu.ai>


This is a more advanced tutorial for users who already have an understanding of Strawberry Fields, e.g., those who have completed the initial teleportation tutorial. Some basic knowledge of Tensorflow is also helpful.

In this tutorial, we show how the user can carry out optimization and machine learning on quantum circuits in Strawberry Fields. This functionality is provided via the Tensorflow simulator backend. By leveraging Tensorflow, we have access to a number of additional funtionalities, including GPU integration, automatic gradient computation, built-in optimization algorithms, and other machine learning tools.

Basic functionality

As usual, we can initialize a Strawberry Fields engine using strawberryfields.Engine().

import strawberryfields as sf
from strawberryfields.ops import *

eng, q = sf.Engine(2)

Replacing numbers with Tensors

When a circuit contains only numerical parameters, the Tensorflow simulator backend works the same as the other backends. However, with the Tensorflow backend, we have the additional option to use Tensorflow objects (e.g., tf.Variable, tf.constant, tf.placeholder, or tf.Tensor) for the parameters of Blackbird states, gates, and measurements.

import tensorflow as tf
alpha = tf.Variable(0.5)
theta_bs = tf.constant(0.0)
phi_bs = tf.sigmoid(0.0) # this will be a tf.Tensor object
phi = tf.placeholder(tf.float32)

with eng:
    # States
    Coherent(alpha)            | q[0]

    # Gates
    BSgate(theta_bs, phi_bs)   | (q[0], q[1])

    # Measurements
    MeasureHomodyne(phi)       | q[0]

To run a Strawberry Fields simulation with the Tensorflow backend, we need to specify 'tf' as the backend argument when calling eng.run(). However, directly evaluating a circuit which contains Tensorflow objects using eng.run() will produce errors. The reason for this is that eng.run() tries, by default, to numerically evaluate any measurement result. But Tensorflow requires several extra ingredients to do this:

  1. Numerical computations must be carried out using a tf.Session.
  2. All tf.Variable objects must be initialized within this tf.Session (the initial values are supplied when creating the variables).
  3. Numerical values must be provided for any tf.placeholder objects using a feed dictionary (feed_dict).

To properly evaluate measurement results, we must therefore do a little more work:

sess = tf.Session()
feed_dict = {phi:0.0}
state = eng.run('tf', cutoff_dim=7, session=sess, feed_dict=feed_dict)

This code will execute without error, and both the output state and the register q will contain numeric values based on the given value for the angle phi. We can select measurement results at other angles by supplying different values for phi in feed_dict.


When being used as a numerical simulator (similar to the other backends), the Tensorflow backend creates temporary sessions in order to evaluate measurement results numerically.

Symbolic computation

Supplying a Session and feed_dict to eng.run() is okay for checking one or two numerical values. However, each call of eng.run() will create additional redundant nodes in the underlying Tensorflow computational graph. A better method is to make the single call eng.run(eval=False). This will carry out the computation symbolically but not numerically. The final state and the register q will both instead contain unevaluted Tensors. These Tensors can be evaluated numerically by running the tf.Session and supplying the desired values for any placeholders:

eng, q = sf.Engine(2)
with eng:
    Dgate(alpha)         | q[0]
    MeasureHomodyne(phi) | q[0]

state = eng.run('tf', cutoff_dim=7, eval=False)
state_density_matrix = state.dm()
homodyne_meas = q[0].val
dm_x, meas_x = sess.run([state_density_matrix, homodyne_meas], feed_dict={phi: 0.0})

Processing data

The parameters for Blackbird states, gates, and measurements may be more complex than just raw data or machine learning weights. These can themselves be the outputs from some learnable function like a neural network:

input_ = tf.placeholder(tf.float32, shape=(2,1))
weights = tf.Variable([[0.1,0.1]])
bias = tf.Variable(0.0)
NN = tf.sigmoid(tf.matmul(weights, input_) + bias)
NNDgate = Dgate(NN)

We can also use the strawberryfields.convert() decorator to allow arbitrary processing of measurement results with the Tensorflow backend [1].

def sigmoid(x):
    return tf.sigmoid(x)

with eng:
    MeasureX             | q[0]
    Dgate(sigmoid(q[0])) | q[1]

Working with batches

It is common in machine learning to process data in batches. Strawberry Fields supports both unbatched and batched data when using the Tensorflow backend. Unbatched operation is the default behaviour (shown above). To enable batched operation, you should provide an extra batch_size argument [2] when calling eng.run(), e.g.,

# run simulation in batched-processing mode
batch_size = 3
eng, q = sf.Engine(2)

with eng:
    Dgate(tf.Variable([0.1] * batch_size)) | q[0]

state = eng.run('tf', cutoff_dim=7, eval=False, batch_size=batch_size)


The batch size should be static, i.e., not changing over the course of a computation.

Parameters supplied to a circuit in batch-mode operation can either be scalars or vectors (of length batch_size). Scalars are automatically broadcast over the batch dimension.

alpha = tf.Variable([0.5] * batch_size)
theta = tf.constant(0.0)
phi = tf.Variable([0.1, 0.33, 0.5])

Measurement results will be returned as Tensors with shape (batch_size,). We can picture batch-mode operation as simulating multiple circuit configurations at the same time. Combined with appropriate parallelized hardware like GPUs, this can result in significant speedups compared to serial evaluation.

Example: variational quantum circuit optimization

A key element of machine learning is optimization. We can use Tensorflow’s automatic differentiation tools to optimize the parameters of variational quantum circuits. In this approach, we fix a circuit architecture where the states, gates, and/or measurements may have learnable parameters associated with them. We then define a loss function based on the output state of this circuit.


The state representation in the simulator can change from a ket (pure) to a density matrix (mixed) if we use certain operations (e.g., state preparations). We can check state.is_pure to determine which representation is being used.

In the example below, we optimize a Dgate to produce an output with the largest overlap with the Fock state \(n=1\).

#!/usr/bin/env python3
import strawberryfields as sf
from strawberryfields.ops import *
import tensorflow as tf

eng, q = sf.Engine(1)

alpha = tf.Variable(0.1)
with eng:
    Dgate(alpha) | q[0]
state = eng.run('tf', cutoff_dim=7, eval=False)

# loss is probability for the Fock state n=1
prob = state.fock_prob([1])
loss = -prob  # negative sign to maximize prob

# Set up optimization
optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.1)
minimize_op = optimizer.minimize(loss)

# Create Tensorflow Session and initialize variables
sess = tf.Session()

# Carry out optimization
for step in range(50):
    prob_val, _ = sess.run([prob, minimize_op])
    print("Value at step {}: {}".format(step, prob_val))


This example program is included with Strawberry Fields as examples/optimization.py.

After 50 or so iterations, the optimization should converge to the optimal probability value of \(e^{-1}\approx 0.3678\) at a parameter of \(\alpha=1\) [3].


When optimizing circuits which contains energy-changing operations (displacement, squeezing, etc.), one should be careful to monitor that the state does not leak out of the given truncation level. This can be accomplished by regularizing or using tf.clip_by_value on the relevant parameters.

Tensorflow supports a large set of mathematical and machine learning operations which can be applied to a circuit’s output state to enable further processing. Examples include tf.norm, tf.self_adjoint_eig, tf.svd, and tf.inv [4].

Exercise: Hong-Ou-Mandel Effect

Use the optimization methods outlined above to find the famous Hong-Ou-Mandel effect, where photons bunch together in the same mode. Your circuit should contain two modes, each with a single-photon input state, and a beamsplitter with variable parameters. By optimizing the beamsplitter parameters, minimize the probability of the \(|1,1\rangle\) Fock-basis element of the output state.


[1]Note that certain operations – in particular, measurements – may not have gradients defined within Tensorflow. When optimizing via gradient descent, we must be careful to define a circuit which is end-to-end differentiable. The gradient support of Tensorflow is constantly growing; users are recommended to check the Tensorflow docs and the Tensorflow github page for the latest implementation details.
[2]Note that batch_size should not be set to 1. Instead, use batch_size=None, or just omit the batch_size argument.
[3]In this tutorial, we have applied classical machine learning tools to learn a quantum optical circuit. Of course, there are many other possibilities for combining machine learning and quantum computing, e.g., using quantum algorithms to speed up machine learning subroutines, or fully quantum learning on unprocessed quantum data.
[4]Remember that it might be necessary to reshape a ket or density matrix before using some of these functions.