Using PennyLane

Release news

Development

API

Internals

  1. Docs
  2. qml
  3. qml.transform

qml.transform

transform(tape_transform=None, pass_name=None, *, setup_inputs=None, expand_transform=None, classical_cotransform=None, is_informative=False, final_transform=False, use_argnum_in_expand=False, plxpr_transform=None)[source]

Generalizes a function that transforms tapes to work with additional circuit-like objects such as a QNode.

transform should be applied to a function that transforms tapes. Once validated, the result will be an object that is able to transform PennyLane’s range of circuit-like objects: QuantumTape, quantum function and QNode. A circuit-like object can be transformed either via decoration or by passing it functionally through the created transform.

Parameters:

  • tape_transform (Callable | None) –

    The input quantum transform must be a function that satisfies the following requirements:

    • Accepts a QuantumScript as its first input and returns a sequence of QuantumScript and a processing function.

    • The transform must have the following structure (type hinting is optional): my_tape_transform(tape: qml.tape.QuantumScript, ...) -> tuple[qml.tape.QuantumScriptBatch, qml.typing.PostprocessingFn]

  • pass_name (str | None) – the name of the associated MLIR pass to be applied when Catalyst is used. See Usage Details for more information.

Keyword Arguments:

  • expand_transform=None (Optional[Callable]) – An optional transform that is applied directly before the input transform. It must be a function that satisfies the same requirements as tape_transform.

  • classical_cotransform=None (Optional[Callable]) – A classical co-transform is a function to post-process the classical jacobian and the quantum jacobian and has the signature: my_cotransform(qjac, cjac, tape) -> tensor_like

  • is_informative=False (bool) – Whether or not a transform is informative. If true, the transform is queued at the end of the compile pipeline and the tapes or qnode aren’t executed.

  • final_transform=False (bool) – Whether or not the transform is terminal. If true, the transform is queued at the end of the compile pipeline. is_informative supersedes final_transform.

  • use_argnum_in_expand=False (bool) – Whether to use argnum of the tape to determine trainable parameters during the expansion transform process.

  • plxpr_transform=None (Optional[Callable]) – Function for transforming plxpr. Experimental

Example

First define an input tape transform with the necessary structure defined above. In this example, we copy the tape and sum the results of the execution of the two tapes.

from pennylane.tape import QuantumScript, QuantumScriptBatch
from pennylane.typing import PostprocessingFn

def my_quantum_transform(tape: QuantumScript) -> tuple[QuantumScriptBatch, PostprocessingFn]:
    tape1 = tape
    tape2 = tape.copy()

    def post_processing_fn(results):
        return qml.math.sum(results)

    return [tape1, tape2], post_processing_fn

We want to be able to apply this transform on both a qfunc and a pennylane.QNode and will use transform to achieve this. transform validates the signature of your input quantum transform and makes it capable of transforming qfunc and pennylane.QNode in addition to quantum tapes. Let’s define a circuit as a pennylane.QNode:

dev = qml.device("default.qubit")

@qml.qnode(device=dev)
def qnode_circuit(a):
    qml.Hadamard(wires=0)
    qml.CNOT(wires=[0, 1])
    qml.X(0)
    qml.RZ(a, wires=1)
    return qml.expval(qml.Z(0))

We first apply transform to my_quantum_transform:

>>> dispatched_transform = qml.transform(my_quantum_transform)

Now you can use the dispatched transform directly on a pennylane.QNode.

For pennylane.QNode, the dispatched transform populates the CompilePipeline of your QNode. The transform and its processing function are applied in the execution.

>>> transformed_qnode = dispatched_transform(qnode_circuit)
>>> transformed_qnode
<QNode: device='<default.qubit device at ...>', interface='auto', diff_method='best', shots='Shots(total=None)'>

>>> print(transformed_qnode.compile_pipeline)
CompilePipeline(
  [1] my_quantum_transform()
)

If we apply dispatched_transform a second time to the pennylane.QNode, we would add it to the compile pipeline again and therefore the transform would be applied twice before execution.

>>> transformed_qnode = dispatched_transform(transformed_qnode)
>>> print(transformed_qnode.compile_pipeline)
CompilePipeline(
  [1] my_quantum_transform(),
  [2] my_quantum_transform()
)

When a transformed QNode is executed, the QNode’s compile pipeline is applied to the generated tape and creates a sequence of tapes to be executed. The execution results are then post-processed in the reverse order of the compile pipeline to obtain the final results.

The setup_inputs function will independently applied prior to any application of the transform. This allows for validation of the inputs, separation into positional and keyword arguments, and specification of a call signature and docstring for transforms without a tape definition.

def my_transform_setup(a, b=1, metadata : str = "my_value"):
    "Docstring for my_transform."
    return (a, b), {"metadata": metadata}

my_transform = qml.transform(pass_name="my_pass", setup_inputs=my_transform_setup)

@qml.qnode(qml.device('default.qubit', wires=4))
def circuit():
    return qml.expval(qml.Z(0))

This allows us to perform eager input validation and set default values.

>>> my_transform(circuit)
Traceback (most recent call last):
    ...
TypeError: <transform: my_pass> missing 1 required positional argument: 'a'
>>> new_circuit = my_transform(circuit, a=2)
>>> new_circuit.compile_pipeline[0]
<my_pass(2, 1, metadata=my_value)>

We will also have a docstring and signature. If a tape transform is present, the signature will be determined by that.

>>> my_transform.__doc__
'Docstring for my_transform.'
>>> import inspect
>>> inspect.signature(my_transform)
<Signature (a, b=1, metadata: str = 'my_value')>

We can compose multiple transforms when working in the tape paradigm and apply them to more than one tape. The following example demonstrates how to apply a transform to a batch of tapes.

Example

In this example, we apply sequentially a transform to a tape and another one to a batch of tapes. We then execute the transformed tapes on a device and post-process the results.

import pennylane as qml

H = qml.PauliY(2) @ qml.PauliZ(1) + 0.5 * qml.PauliZ(2) + qml.PauliZ(1)
measurement = [qml.expval(H)]
operations = [qml.Hadamard(0), qml.RX(0.2, 0), qml.RX(0.6, 0), qml.CNOT((0, 1))]
tape = qml.tape.QuantumTape(operations, measurement)

batch1, function1 = qml.transforms.split_non_commuting(tape)
batch2, function2 = qml.transforms.merge_rotations(batch1)

dev = qml.device("default.qubit", wires=3)
result = dev.execute(batch2)

The first split_non_commuting transform splits the original tape, returning a batch of tapes batch1 and a processing function function1. The second merge_rotations transform is applied to the batch of tapes returned by the first transform. It returns a new batch of tapes batch2, each of which has been transformed by the second transform, and a processing function function2.

>>> batch2
(<QuantumTape: wires=[0, 1, 2], params=1>, <QuantumTape: wires=[0, 1, 2], params=1>)

>>> type(function2)
<class 'function'>

We can combine the processing functions to post-process the results of the execution.

>>> function1(function2(result))
np.float64(0.499...)

A dispatched transform is able to handle several PennyLane circuit-like objects:

For each object, the transform will be applied in a different way, but it always preserves the underlying tape-based quantum transform behaviour.

The return of a dispatched transform depends upon which of the above objects is passed as an input:

  • For a QNode input, the underlying transform is added to the QNode’s CompilePipeline and the return is the transformed QNode. For each execution of the pennylane.QNode, it first applies the compile pipeline on the original captured circuit. Then the transformed circuits are executed by a device and finally the post-processing function is applied on the results.

    When experimental program capture is enabled, transforming a QNode returns a new function to which the transform has been added as a higher-order primitive.

  • For a quantum function (callable) input, the transform builds the tape when the quantum function is executed and then applies itself to the tape. The resulting tape is then converted back to a quantum function (callable). It therefore returns a transformed quantum function (Callable). The limitation is that the underlying transform can only return a sequence containing a single tape, because quantum functions only support a single circuit.

    When experimental program capture is enabled, transforming a function (callable) returns a new function to which the transform has been added as a higher-order primitive.

  • For a QuantumScript. It returns a sequence of QuantumScript and a processing function to be applied after execution.

  • For a batch of pennylane.tape.QuantumScript, the quantum transform is mapped across all the tapes. It returns a sequence of QuantumScript and a processing function to be applied after execution. Each tape in the sequence is transformed by the transform.

  • For a Device, the transform is added to the device’s compile pipeline and a transformed pennylane.devices.Device is returned. The transform is added to the end of the device program and will be last in the overall compile pipeline.

If a compilation pass is written in MLIR, using it in a qjit’d workflow requires that it have a transform with a matching pass_name. This ensures that the transform is properly applied as part of the lower-level compilation.

For example, we can create a transform that will apply the cancel-inverses pass, like the in-built qml.transforms.cancel_inverses transform.

my_transform = qml.transform(pass_name="cancel-inverses")

@qml.qjit
@my_transform
@qml.qnode(qml.device('lightning.qubit', wires=4))
def circuit():
    qml.X(0)
    qml.X(0)
    return qml.expval(qml.Z(0))

We can see that the instruction to apply "cancel-inverses" is present in the initial MLIR.

>>> circuit()
Array(1., dtype=float64)
>>> print(circuit.mlir[200:600])
tensor<f64>
}
module @module_circuit {
    module attributes {transform.with_named_sequence} {
    transform.named_sequence @__transform_main(%arg0: !transform.op<"builtin.module">) {
        %0 = transform.apply_registered_pass "cancel-inverses" to %arg0 : (!transform.op<"builtin.module">) -> !transform.op<"builtin.module">
        transform.yield
    }
    }
    func.func public @circui

Transforms can have both tape-based and pass_name-based definitions. For example, the transform below called my_transform has both definitions. In this case, the MLIR pass will take precedence when being qjit’d if only MLIR passes can occur after.

from functools import partial

@partial(qml.transform, pass_name="my-pass-name")
def my_transform(tape):
    return (tape, ), lambda res: res[0]

Note that any transform with only a pass_name definition must occur after any purely tape-based transform, as tape transforms occur prior to lowering to MLIR.

>>> @qml.qjit
... @qml.defer_measurements
... @qml.transform(pass_name="cancel-inverses")
... @qml.qnode(qml.device('lightning.qubit', wires=4))
... def c():
...     qml.X(0)
...     qml.X(0)
...     return qml.expval(qml.Z(0))
...
Traceback (most recent call last):
    ...
ValueError: <cancel-inverses()> without a tape definition occurs before tape transform <defer_measurements()>.
Previous
Next