Update QNGOptimizer to handle deprecations

[antalszava]

Context

The QNGOptimizer raises deprecation warnings because (see related demo):

• inputs have to explicitly specify requires_grad=True;
• the keyword argument diag_approx is deprecated.

Changes

• Switches to using PennyLane NumPy for the optimizer
• Switches how the optimizer uses qml.metric_tensor
• Adds a deprecation warning when diag_approx is specified with QNGOptimizer

[github-actions]

Hello. You may have forgotten to update the changelog!
Please edit doc/releases/changelog-dev.md with:

• A one-to-two sentence description of the change. You may include a small working example for new features.
• A link back to this PR.
• Your name (or GitHub username) in the contributors section.

[codecov]

Codecov Report

❗ No coverage uploaded for pull request base (v0.19.0-rc0@60501d6). Click here to learn what that means.
The diff coverage is n/a.

@@              Coverage Diff               @@
##             v0.19.0-rc0    #1834   +/-   ##
==============================================
  Coverage               ?   98.96%           
==============================================
  Files                  ?      218           
  Lines                  ?    16372           
  Branches               ?        0           
==============================================
  Hits                   ?    16203           
  Misses                 ?      169           
  Partials               ?        0           

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 60501d6...aa37e5e. Read the comment docs.

[josh146]

Thanks Antal for following up on this one! 💪 💯

I have a quick question though before approval.

I tried running the following script on master, and received no warning:

import pennylane as qml
from pennylane import numpy as np

dev = qml.device("default.qubit", wires=1)

@qml.qnode(dev)
def circuit(params):
    qml.RX(params[0], wires=0)
    qml.RY(params[1], wires=0)
    return qml.expval(qml.PauliX(0) + qml.PauliZ(0))

theta = np.array([0.011, 0.012], requires_grad=True)

opt = qml.QNGOptimizer(0.01)
new_params = opt.step(circuit, theta)

res = opt.metric_tensor
exp = np.diag([0.25, (np.cos(theta[0]) ** 2) / 4])
assert np.allclose(res, exp)

This is based on the test you modified here.

After having a look at the demo you linked, I think the issue is with the demo, not the QNGOptimizer! If you have a look at the demo, it is importing vanilla NumPy, and not setting requires_grad=True 🙂

[josh146]

@antalszava do you think we should update the QNG optimizer to match the change in the metric tensor?

That is,

class QNGOptimizer:

    def __init__(self, stepsize=0.01, approx="diag", lam=0):

?

I had forgotten the QNG optimizer also used this old terminology! I think it makes sense to match the metric tensor here, that way we can simply have

Suggested change

	if self.diag_approx:
	metric_tensor_fn = qml.metric_tensor(qnode, approx="diag")
	else:
	metric_tensor_fn = qml.metric_tensor(qnode)
	metric_tensor_fn = qml.metric_tensor(qnode, approx=self.approx)

[josh146]

@antalszava nevermind, I see that an updated version of the demo does use the pennylane numpy! https://4409-214003948-gh.circle-artifacts.com/0/_build/html/demos/tutorial_quantum_natural_gradient.html I am stumped as to why I cannot reproduce it locally 🤔

[antalszava]

Hi @josh146 thank you 🙂 The warnings related to the requires_grad=True casting seem to arise if we step the optimizer once more:

import pennylane as qml
from pennylane import numpy as np

dev = qml.device("default.qubit", wires=1)

@qml.qnode(dev)
def circuit(params):
    qml.RX(params[0], wires=0)
    qml.RY(params[1], wires=0)
    return qml.expval(qml.PauliX(0) + qml.PauliZ(0))

theta = np.array([0.011, 0.012], requires_grad=True)

opt = qml.QNGOptimizer(0.01)
new_params = opt.step(circuit, theta)

res = opt.metric_tensor
exp = np.diag([0.25, (np.cos(theta[0]) ** 2) / 4])
assert np.allclose(res, exp)

theta = new_params
new_params = opt.step(circuit, theta)

[antalszava]

@josh146 I've updated the QNGOptimizer to raise a warning if diag_approx is set and added a new approx keyword. There will be one relevant demo to update.

Perhaps investigating more how the QNGOptimizer could be made compatible with QNodes that take varying input parameter shapes (e.g., circuit(x, y) such that x = np.array(0.2), y = np.array([0.1, 0.2])) could come as a separate PR. Likely the logic of apply_grad from the gradient descent optimizer can be ported over.

[josh146]

Agree. Thanks Antal!

[josh146]

💯