Qadence provides a flexible parameter system built on top of Sympy. Parameters can be of different types:
- Fixed parameter: a constant with a fixed, non-trainable value (e.g.
$\dfrac{\pi}{2}$ ). - Variational parameter: a trainable parameter which will be automatically picked up by the optimizer.
- Feature parameter: a non-trainable parameter which can be used to pass input values.
Passing fixed parameters to blocks can be done by simply passing a Python numeric type or a torch.Tensor
.
import torch
from qadence import RX, run, PI
wf = run(RX(0, torch.tensor(PI)))
print(f"{wf = }") # markdown-exec: hide
wf = run(RX(0, PI))
print(f"{wf = }") # markdown-exec: hide
To parametrize a block a VariationalParameter
instance is required. In most cases Qadence also accepts a Python string, which will be used to automatically initialize a VariationalParameter
:
from qadence import RX, run, VariationalParameter
block = RX(0, VariationalParameter("theta"))
block = RX(0, "theta") # Equivalent
wf = run(block)
print(f"{wf = }") # markdown-exec: hide
By calling run
, a random value for "theta"
is initialized at execution. In a QuantumModel
, variational parameters are stored in the underlying model parameter dictionary.
A FeatureParameter
type can also be used. It requires an input value or a batch of values. In most cases, Qadence accepts a values
dictionary to set the input of feature parameters.
from torch import tensor
from qadence import RX, PI, run, FeatureParameter
block = RX(0, FeatureParameter("phi"))
wf = run(block, values = {"phi": tensor([PI, PI/2])})
print(f"{wf = }") # markdown-exec: hide
Since a batch of input values was passed, the run
function returns a batch of output states. Note that FeatureParameter("x")
and VariationalParameter("x")
are simply aliases for Parameter("x", trainable = False)
and Parameter("x", trainable = True)
.
The integration with Sympy becomes useful when one wishes to write arbitrary parameter compositions. Parameters can also be used as scaling coefficients in the block system, which is essential when defining arbitrary analog operations.
from torch import tensor
from qadence import RX, Z, HamEvo, PI
from qadence import VariationalParameter, FeatureParameter, run
from sympy import sin
theta, phi = VariationalParameter("theta"), FeatureParameter("phi")
# Arbitrary parameter composition
expr = PI * sin(theta + phi)
# Use as unitary gate arguments
gate = RX(0, expr)
# Or as scaling coefficients for Hermitian operators
h_op = expr * (Z(0) @ Z(1))
wf = run(gate * HamEvo(h_op, phi), values = {"phi": tensor(PI)})
print(f"{wf = }") # markdown-exec: hide
Parameters are uniquely defined by their name and redundancy is allowed in composite blocks to assign the same value to different blocks. This is useful, for example, when defining layers of rotation gates typically used as feature maps.
from torch import tensor
from qadence import RY, PI, run, kron, FeatureParameter
n_qubits = 3
param = FeatureParameter("phi")
block = kron(RY(i, (i+1) * param) for i in range(n_qubits))
wf = run(block, values = {"phi": tensor(PI)})
print(f"{wf = }") # markdown-exec: hide
Let's look at a final example of an arbitrary composition of digital and analog parameterized blocks:
import sympy
from qadence import RX, RY, RZ, CNOT, CPHASE, Z, HamEvo
from qadence import run, chain, add, kron, FeatureParameter, VariationalParameter, PI
n_qubits = 3
phi = FeatureParameter("Φ")
theta = VariationalParameter("θ")
rotation_block = kron(
RX(0, phi/theta),
RY(1, theta*2),
RZ(2, sympy.cos(phi))
)
digital_entangler = CNOT(0, 1) * CPHASE(1, 2, PI)
hamiltonian = add(theta * (Z(i) @ Z(i+1)) for i in range(n_qubits-1))
analog_evo = HamEvo(hamiltonian, phi)
program = chain(rotation_block, digital_entangler, analog_evo)
from qadence.draw import html_string # markdown-exec: hide
print(html_string(program)) # markdown-exec: hide
Please note the different colors for the parametrization with different types. The default palette assigns blue for VariationalParameter
, green for FeatureParameter
, orange for numeric values, and shaded red for non-parametric gates.