qiboml.differentiation

class qiboml.operations.differentiation.Differentiation(circuit: Circuit, decoding: QuantumDecoding)

Bases: ABC

Abstract differentiator object.

circuit: Circuit

decoding: QuantumDecoding

abstract evaluate(parameters: ndarray[tuple[int, ...], dtype[_ScalarType_co]], wrt_inputs: bool = False)

Evaluate the gradient of the quantum circuit w.r.t its parameters, i.e. its rotation angles.

property backend

property non_trainable_gates

nparams(wrt_inputs)

_abc_impl = <_abc._abc_data object>

class qiboml.operations.differentiation.PSR(circuit: Circuit, decoding: QuantumDecoding)

Bases: Differentiation

The Parameter Shift Rule differentiator. Especially useful for non analytical derivative calculation which, thus, makes it hardware compatible.

evaluate(parameters: ndarray[tuple[int, ...], dtype[_ScalarType_co]], wrt_inputs: bool = False)

Evaluate the gradient of the quantum circuit w.r.t its parameters, i.e. its rotation angles. :param parameters: the parameters at which to evaluate the model, and thus the derivative. :type parameters: List[ndarray] :param wrt_inputs: whether to calculate the derivative with respect to, also, inputs (i.e. encoding angles) :type wrt_inputs: bool

or not, by default False. :returns: the calculated jacobian. :rtype: (ndarray)

one_parameter_shift(parameters: ndarray[tuple[int, ...], dtype[_ScalarType_co]], parameter_index: int, wrt_inputs: bool = False) → Tuple[Circuit, Circuit, float]

Compute one derivative of the decoding strategy w.r.t. a target parameter.

static shift_parameter(parameters, i, epsilon, backend)

_abc_impl = <_abc._abc_data object>

property backend

property non_trainable_gates

nparams(wrt_inputs)

circuit: Circuit

decoding: QuantumDecoding

class qiboml.operations.differentiation.Adjoint(circuit: Circuit, decoding: QuantumDecoding)

Bases: Differentiation

Adjoint differentiation following Algorithm 1 from https://arxiv.org/abs/2009.02823. Only qiboml.models.decoding.Expectation. as decoding is supported and all parametrized_gates must have a gradient method returning the gradient of the single gate.

evaluate(parameters: ndarray[tuple[int, ...], dtype[_ScalarType_co]], wrt_inputs: bool = False)

Evaluate the gradient of the quantum circuit w.r.t its parameters, i.e. its rotation angles. :param parameters: the parameters at which to evaluate the model, and thus the derivative. :type parameters: List[ndarray] :param wrt_inputs: whether to calculate the derivate with respect to inputs or not, by default False. :type wrt_inputs: bool

Returns:

the calculated gradients.

Return type:

(ndarray)

_abc_impl = <_abc._abc_data object>

property backend

property non_trainable_gates

nparams(wrt_inputs)

circuit: Circuit

decoding: QuantumDecoding

class qiboml.operations.differentiation.Jax(circuit: Circuit, decoding: QuantumDecoding)

Bases: Differentiation

static _run(circuit, decoding, *parameters)

static _run_with_inputs(circuit, decoding, *parameters)

property circuit

property decoding

_abc_impl = <_abc._abc_data object>

_cast_non_trainable_parameters(src_backend, tgt_backend)

property backend

property non_trainable_gates

nparams(wrt_inputs)

evaluate(parameters, wrt_inputs: bool = False)

Evaluate the jacobian of the internal quantum model (circuit + decoding) w.r.t to its parameters, i.e. the parameterized gates in the circuit. :param parameters: the parameters at which to evaluate the circuit, and thus the derivatives. :type parameters: list[ndarray] :param wrt_inputs: whether to calculate the derivative with respect to, also, inputs (i.e. encoding angles) :type wrt_inputs: bool

or not, by default False. :returns: the calculated jacobian. :rtype: (ndarray)