Way to tell how PennyLane will differentiate a QNode

PennyLane Help
1

In PL v0.21.0 I could, given a QNode circuit and a device dev do for example

circuit.get_best_method(dev, circuit.interface)[-1]['method'] == 'numeric'

to test whether computing the jacobian on that device would be done with the finite difference method.

I understand that this was never officially supported UI and indeed in the current master this no longer works as get_best_method() no longer returns a “human readable” description of the method as its last argument.

Is there a more official way to find out how PL will differentiate a QNode on a device?

2

Hey @cvjjm! Ah, this is a consequence of us moving away from ‘string’ identifiers, and towards using gradient transform functions directly.

Currently, the only change is that:

  • the "analytic" string is now represented by either qml.gradients.param_shift or qml.gradients.param_shift_cv depending on the device, and

  • the "numeric" string is now represented by qml.gradients.finite_diff.

Here is a quick function that will replicate the behaviour of PL v0.21:

def get_best_method(dev, interface):
    transform = qml.QNode.get_best_method(dev, interface)[0]

    if transform is qml.gradients.finite_diff:
        return "numeric"

    if transform in (qml.gradients.param_shift, qml.gradients.param_shift_cv):
        return "analytic"

    # only other options by this point are "backprop" or "device"
    return transform

Is there a more official way to find out how PL will differentiate a QNode on a device?

Not at the moment, but this is good feedback and something we will look at adding (and definitely useful!).

Regarding the heuristics for determining the ‘best’ gradient method; this is functionality we are planning to evolve further down the line, and may even be moved out of the QNode initialization at some point; since we have identified several places where knowing the circuit structure itself — which is not known at QNode initialization, only at execution — can result in better heuristics.

So we may factor out these static methods into utility functions, and make them ‘public’.