The TF device¶
The strawberryfields.tf
device gives access to Strawberry Field’s tf
state simulator
backend. This simulator device has the following features:
The simulator is written using TensorFlow, so supports classical backpropagation using PennyLane. Simply use
interface="tf"
when creating your QNode.Quantum states are represented in the Fock basis \(\left 0 \right>, \left 1 \right>, \left 2 \right>, \dots, \left \mathrm{D 1} \right>\), where \(D\) is the usergiven value for
cutoff_dim
that limits the dimension of the Hilbert space.The advantage of this representation is that any continuousvariable operation can be represented. However, the simulations are approximations, whose accuracy, the simulation time, and required memory increases with the cutoff dimension.
Warning
It is often useful to keep track of the normalization of a quantum state during optimization, to make sure the circuit does not “learn” to push its parameters into a regime where the simulation is vastly inaccurate.
Usage¶
You can instantiate the TF device in PennyLane as follows:
>>> import pennylane as qml
>>> import tensorflow as tf
>>> dev = qml.device('strawberryfields.tf', wires=2, cutoff_dim=10)
The device can then be used just like other devices for the definition and evaluation of QNodes within PennyLane.
For instance, the following simple example defines a QNode that first displaces the vacuum state,
applies a beamsplitter, and then returns the marginal probability on the first wire. This function
is then converted into a QNode which is placed on the strawberryfields.tf
device:
@qml.qnode(dev, interface="tf")
def circuit(x, theta):
qml.Displacement(x, 0, wires=0)
qml.Beamsplitter(theta, 0, wires=[0, 1])
return qml.probs(wires=0)
We can evaluate the QNode for arbitrary values of the circuit parameters:
>>> x = tf.Variable(1.0)
>>> theta = tf.Variable(0.543)
>>> with tf.GradientTape() as tape:
... res = circuit(x, theta)
>>> print(res)
tf.Tensor(
[[4.8045865e01+0.j 3.5218298e01+0.j 1.2907754e01+0.j 3.1538557e02+0.j
5.7795495e03+0.j 8.4729097e04+0.j 1.0349592e04+0.j 1.0811385e05+0.j
9.6350857e07+0.j 6.1937492e08+0.j]], shape=(1, 10), dtype=complex64)
We can also evaluate the derivative with respect to any parameter(s):
>>> jac = tape.jacobian(res, x)
>>> print(jac)
<tf.Tensor: shape=(1, 10), dtype=float32, numpy=
array([[7.0436597e01, 1.8805575e01, 3.2707882e01, 1.4299491e01,
3.7763387e02, 7.2306832e03, 1.0900890e03, 1.3535164e04,
1.3895189e05, 9.9099987e07]], dtype=float32)>
Note
The qml.state
, qml.sample
and qml.density_matrix
measurements
are not supported on the strawberryfields.tf
device.
The continuousvariable QNodes available via Strawberry Fields can also be combined with qubitbased QNodes and classical nodes to build up a hybrid computational model. Such hybrid models can be optimized using the builtin optimizers provided by PennyLane.
PennyLane CV templates, such as Interferometer()
and CVNeuralNetLayers()
, can also be used:
dev = qml.device("strawberryfields.tf", wires=3, cutoff_dim=5)
@qml.qnode(dev, interface="tf")
def circuit(weights):
for i in range(3):
qml.Squeezing(0.1, 0, wires=i)
qml.templates.Interferometer(
theta=weights[0],
phi=weights[1],
varphi=weights[2],
wires=[0, 1, 2],
mesh="rectangular",
)
return qml.probs(wires=0)
Once defined, we can now use this QNode within any TensorFlow computation:
>>> weights = qml.init.interferometer_all(n_wires=3)
>>> weights = [tf.convert_to_tensor(w) for w in weights]
>>> with tf.GradientTape() as tape:
... tape.watch(weights)
... res = circuit(weights)
>>> grad = tape.gradient(res, weights)
[<tf.Tensor: shape=(3,), dtype=float64, numpy=array([4.93799348e07, 5.99637985e07, 8.90550478e09])>,
<tf.Tensor: shape=(3,), dtype=float64, numpy=array([2.09796852e07, 1.01452002e08, 4.34359642e08])>,
<tf.Tensor: shape=(3,), dtype=float64, numpy=array([ 8.36735126e10, 1.21872290e10, 1.81160686e09])>]
Note
The strawberryfields.tf
device does not support Autograph mode (tf.function
).
Device options¶
The Strawberry Fields TF device accepts additional arguments beyond the PennyLane default device arguments.
cutoff_dim
the Fock basis truncation when applying quantum operations
hbar=2
The convention chosen in the canonical commutation relation \([x, p] = i \hbar\). Default value is \(\hbar=2\).
shots=None
The number of circuit evaluations/random samples used to estimate expectation values of observables. The default value of
None
means that the exact expectation value is returned.
Supported operations¶
The Strawberry Fields TF device supports all continuousvariable (CV) operations and observables provided by PennyLane, including both Gaussian and nonGaussian operations.
Supported operations:
Beamsplitter interaction. 

Prepares a coherent state. 

Controlled addition operation. 

Controlled phase operation. 

CrossKerr interaction. 

Cubic phase shift. 

Prepares a displaced squeezed vacuum state. 

Phase space displacement. 

Prepare subsystems using the given density matrix in the Fock basis. 

Prepares a single Fock state. 

Prepare subsystems using the given ket vector in the Fock basis. 

Prepare subsystems in a given Gaussian state. 

A linear interferometer transforming the bosonic operators according to the unitary matrix \(U\). 

Kerr interaction. 

Quadratic phase shift. 

Phase space rotation. 

Prepares a squeezed vacuum state. 

Phase space squeezing. 

Prepares a thermal state. 

Phase space twomode squeezing. 
Supported observables:
The identity observable \(\I\). 

The photon number observable \(\langle \hat{n}\rangle\). 

The tensor product of the 

The position quadrature observable \(\hat{x}\). 

The momentum quadrature observable \(\hat{p}\). 

The generalized quadrature observable \(\x_\phi = \x cos\phi+\p\sin\phi\). 

An arbitrary secondorder polynomial observable. 