Source code for qailab.qlauncher.passes.backward

 1""" Backward pass algorithm implementation in quantum_launcher. """
 2from collections.abc import Callable
 3from typing import Any, Literal
 4from quantum_launcher.base.base import Backend, Problem, Result
 5from quantum_launcher.routines.qiskit_routines import QiskitBackend
 6from qailab.qlauncher.passes.forward import ForwardPass
 7from qailab.gradient.gradient_calculation import calculate_jacobian
 8
 9

[docs]
10class BackwardPass(ForwardPass):
11    """ Backward Pass, calculates two jacobian matrices: w.r.t. to input and w.r.t. weights"""
12
13    def __init__(self, gradient_method: Literal['param_shift', 'spsa', 'lin_comb'] = 'param_shift', shots: int = 1024) -> None:
14        self.gradient_method = gradient_method
15        self.shots = shots
16        super().__init__()
17

[docs]
18    def run(self, problem: Problem, backend: Backend, formatter: Callable[..., Any] | None = None) -> Result:
19        if formatter is None:
20            raise ValueError('Formatter for Backward pass not found!')
21        if not isinstance(backend, QiskitBackend):
22            raise ValueError('Wrong sampler given into')
23
24        format_result = formatter(problem)[0]
25        if not isinstance(format_result, tuple):
26            raise ValueError("Don't use autobind")
27
28        circuit, params = format_result
29
30        input_params = {x: params[x] for x in params if x.name.startswith('input')}
31        weight_params = {x: params[x] for x in params if x.name.startswith('weight')}
32
33        # All other params must be assigned.
34        input_jacobian = calculate_jacobian(
35            circuit.assign_parameters(weight_params),
36            input_params,
37            backend,
38            self.gradient_method,
39            self.shots
40        )
41        weight_jacobian = calculate_jacobian(
42            circuit.assign_parameters(input_params),
43            weight_params,
44            backend,
45            self.gradient_method,
46            self.shots
47        )
48
49        return Result('', 0, '', 0, {}, {}, self.shots, 0, 0, {'input': input_jacobian, 'weight': weight_jacobian})