TensorFlow Quantum (TFQ) is designed for the problems of NISQ-era quantum machine learning. It brings quantum computing primitives—like building quantum circuits—to the TensorFlow ecosystem. Models and operations built with TensorFlow use these primitives to create powerful quantum-classical hybrid systems.
Using TFQ, researchers can construct a TensorFlow graph using a quantum dataset,
a quantum model, and classical control parameters. These are all represented as
tensors in a single computational graph. The outcome of quantum
measurements—leading to classical probabilistic events—is obtained by TensorFlow
ops. Training is done with the standard
Keras API. The
module allows researchers to experiment with new and interesting quantum
Cirq is a quantum programming framework from Google. It provides all of the basic operations—such as qubits, gates, circuits, and measurement—to create, modify and invoke quantum circuits on a quantum computer, or a simulated quantum computer. TensorFlow Quantum uses these Cirq primitives to extend TensorFlow for batch computation, model building, and gradient computation. To be effective with TensorFlow Quantum, it’s a good idea to be effective with Cirq.
TensorFlow Quantum primitives
TensorFlow Quantum implements the components needed to integrate TensorFlow with quantum computing hardware. To that end, TFQ introduces two datatype primitives:
- Quantum circuit: This represents
quantum circuits (
cirq.Circuit) within TensorFlow. Create batches of circuits of varying size, similar to batches of different real-valued datapoints.
- Pauli sum: Represent linear combinations of tensor products of Pauli
operators defined in Cirq (
cirq.PauliSum). Like circuits, create batches of operators of varying size.
Using the quantum circuit primitives within a
tf.Tensor, TensorFlow Quantum
implements ops that process these circuits and produce meaningful outputs.
The TensorFlow ops are written in optimized C++. These ops sample from circuits, calculate expectation values, and output the state produced by the given circuits. Writing ops that are flexible and performant has some challenges:
- Circuits are not the same size. For simulated circuits, you are unable to
create static operations (like
tf.add) and then substitute different numbers for circuits of different sizes. These ops must allow for dynamic sizes that the statically sized TensorFlow compute graph doesn't allow.
- Quantum data can induce a different circuit structure altogether. This is another reason to support dynamic sizes in the TFQ ops. Quantum data can represent a structural change to the underlying quantum state that is represented by modifications to the original circuit. As new datapoints are swapped in and out at runtime, the TensorFlow compute graph can not be modified after it is built, so support for these varying structures is required.
cirq.Circuitsare similar to compute graphs in that they are a series of operations—and some might contain symbols/placeholders. It is important to make this as compatible with TensorFlow as possible.
For performance reasons, Eigen (the C++ library used in many TensorFlow ops) is
not well suited for quantum circuit simulation. Instead, the circuit simulators
used in the
quantum supremacy experiment
are used as verifiers and extended as the foundation of TFQ ops (all written
with AVX2 and SSE instructions). Ops with identical functional signatures were
created that use a physical quantum computer. Switching between a simulated and
physical quantum computer is as easy as changing a single line of code. These
ops are located in the
TensorFlow Quantum layers expose sampling, expectation, and state calculation to
developers using the
tf.keras.layers.Layer interface. It's convenient to
create a circuit layer for classical control parameters or for readout
operations. Additionally, you can create a layer with a high degree of
complexity supporting batch circuit, batch control parameter value, and perform
batch readout operations. See
tfq.layers.Sample for an example.
Unlike many TensorFlow operations, observables in quantum circuits do not have formulas for gradients that are relatively easy to calculate. This is because a classical computer can only read samples from the circuits that are run on a quantum computer.
To solve this problem, the
tfq.differentiators module provides several
standard differentiation techniques. Users can also define their own method to
compute gradients—in both the “real world” setting of sample-based expectation
calculation, and the analytic exact world. Methods like finite difference are
often the fastest (wall clock time) in an analytic/exact environment. While
slower (wall clock time), more practical methods like
parameter shift or
are often more effective. A
tfq.differentiators.Differentiator is instantiated
and attached to an existing op with
generate_differentiable_op, or passed to
the constructor of
To implement a custom differentiator, inherit from the
tfq.differentiators.Differentiator class. To define a gradient operation for
sampling or state vector calculation, use
As the field of quantum computing grows, more quantum data and model
combinations will arise, making structured comparison more difficult. The
tfq.datasets module is used as the data source for quantum machine learning
tasks. It ensures structured comparisons for the model and performance.
It is hoped that with large community contributions, the
will grow to enable research that is more transparent and reproducible.
Carefully curated problems in: quantum control, fermionic simulation,
classification near phase transitions, quantum sensing, etc are all great
candidates for addition to
tfq.datasets. To propose a new dataset open
a GitHub issue.