Tools to ride the waves of loss landscapes!
This package includes different methods to compute gradients of expecation values and the Quantum Fisher Information (QFI) of states. It is based on Qiskit.
Gradient methods are located in surfer.gradient and derive the GradientCalculator interface. Each method has a custom constructor but offers a
compute(operator: qiskit.opflow.OperatorBase, circuit: qiskit.QuantumCircuit, values: numpy.ndarray)method to compute the gradients.
See examples/gradient.py for an example.
Supported methods
ReverseGradient: Statevector-based, scales linearly in the number of parameters, see also arXiv:2009.02823.ForwardGradient*: Statevector-based, scales quadratically in the number of parameters.
QFI methods are located in surfer.qfi and derive the QFICalculator interface. Each method has a custom constructor but offers a
compute(circuit: qiskit.QuantumCircuit, values: numpy.ndarray)method to compute the QFI of the state that is obtained by applying circuit to the all zero state.
See examples/qfi.py for an example.
Supported methods
ReverseQFI: Statevector-based, scales quadratically in the number of parameters, see also arXiv:2011.02991.LinearCombination*: Proxy for Qiskit's QFI calculation based on the linear combination of unitaries.StochasticApproximation*: Monte Carlo-style stochastic approximation of the QFI, based on arXiv:2103.09232.
*Note: Only ReverseGradient and ReverseQFI are currently fully functional. The other methods are still work in progress and have some sharp bits. They do not unroll circuits (so you have to pass them in a supported basis), do not support the product rule (so no re-using of parameters!) and determine the order of parameters from the circuit (instead of respecting e.g. the order of your ParameterVector).