Tutorial on the gradients framework (fixes #1390)

View / edit / reply to this conversation on ReviewNB

ElePT commented on 2023-06-01T15:05:57Z ----------------------------------------------------------------

I think that the title "Qiskit Gradient Framework using Primitives" is a bit more fitting, "The Gradient Framework" sounds too generic for me.

Also, I like how concise the tutorial goal is, but I think it could even be made into a single sentence, and include a link to the API ref. for the primitives:

This tutorial demonstrates the use of the qiskit.algorithms.gradients module to evaluate quantum gradients using the Qiskit Primitives

Finally, if it is not longer useful, could you please remove the comment? It is fine to have some comments while drafting the tutorial, but we should avoid them in the final version :)

maxwell04-wq commented on 2023-06-01T19:32:19Z ----------------------------------------------------------------

Incorporated.

review-notebook-app[bot] commented 1 year ago

View / edit / reply to this conversation on ReviewNB

ElePT commented on 2023-06-01T15:05:58Z ----------------------------------------------------------------

I would call this section something like "A quick refresher on Qiskit Primitives".

Qiskit Primitives is a new framework for executing quantum jobs. Instead of binding parameters to quantum circuits, the primitives module uses either a Sampler or an Estimator class to evaluate the quantum circuits. These Primitive classes take the circuits, the observable hamiltonians, and the circuit parameters and return the results of quantum simulations.

And I would suggest a few changes to the sentence above, I know that explaining the primitives is still a bit tricky, but to make it better it is important to be precise describing what the primitives do:

The Qiskit Primitives work as an abstraction level between algorithms and (real/simulated) quantum devices. Instead of having to manually deal with tasks such as parameter binding or circuit transpilation, the primitives module offers a Sampler and an Estimator class that take the circuits, the observable Hamiltonians, and the circuit parameters and return the sampling distribution and the computed expectation values respectively.

Now, for the following section:

Qiskit Primitives provides two classess for evaluating the circuit:

The Estimator class allows to evaluate the expectation values of the observables.

The Sampler class allows to evaluate the quasi-probability distribution of the observables.

I would say:

use qiskit.primitives instead of "Qiskit Primitives"
there is a small typo on classes (double s at the end)
it is important to understand what each primitive does. The estimator calculates expectation values of circuit-observable pairs, while the sampler returns prob. distributions from only circuits (no observables here). I think the description could be changed to:
The Estimator class allows to evaluate expectation values of observables with respect to states prepared by quantum circuits
The Sampler class returns quasi-probability distributions as a result of sampling quantum circuits

maxwell04-wq commented on 2023-06-01T19:37:37Z ----------------------------------------------------------------

Incorporated.

review-notebook-app[bot] commented 1 year ago

View / edit / reply to this conversation on ReviewNB

ElePT commented on 2023-06-01T15:05:59Z ----------------------------------------------------------------

Overall, I really like this section, but to push it to the next level I would restructure it a little bit, and explain the base classes first, include classical gradients and talk about the QGT. This would look like this:

The algorithms.gradients module contains both circuit gradients and circuit metrics. The gradients extend the BaseEstimatorGradient and BaseSamplerGradient base classes. These are abstract classes on top of which different gradient methods have been implemented. The methods currently available in this module are:

Parameter Shift Gradients
Finite Difference Gradients
Linear Combination of Unitaries Gradients
Simultaneous Perturbation Stochastic Approximation (SPSA) Gradients

Additionally, the module offers reverse gradients for efficient classical computations.

The metrics available are based on the notion of the Quantum Geometric Tensor (QGT). There is a BaseQGT class (Estimator-based) on top of which different QGT methods have been implemented:

Linear Comination of Unitaries QGT
Reverse QGT (classical)

As well as a Quantum Fisher Information class (QFI) that is initialized with a reference QGT implementation from the above list.

_maxwell04-wq commented on 2023-06-01T19:40:37Z_ ----------------------------------------------------------------

Incorporated.

review-notebook-app[bot] commented 1 year ago

View / edit / reply to this conversation on ReviewNB

ElePT commented on 2023-06-01T15:06:00Z ----------------------------------------------------------------

I love that you have made a diagram, this shows a lot of effort :) I think we can replace Qiskit Algorithms with Qiskit Primitives at the bottom (as the gradients are built on top of that), and add the Reverse Gradient to the list (I know that I mentioned that a reverse gradient tutorial was optional, we don't have to go into detail, but I think that we should at least name them and show where they fit in the overall structure).

I am not so sure about the "Circuit" in "Circuit Gradients" and "Circuit Metrics", we can maybe drop it.

maxwell04-wq commented on 2023-06-01T20:12:42Z ----------------------------------------------------------------

Incorporated.

review-notebook-app[bot] commented 1 year ago

View / edit / reply to this conversation on ReviewNB

ElePT commented on 2023-06-01T15:06:01Z ----------------------------------------------------------------

We are trying to move away from having import blocks at the beginning of the tutorial, instead, we encourage to add the relevant imports at the top of the cell where they are used. You can see this, for example, in this tutorial https://qiskit.org/ecosystem/ibm-runtime/tutorials/grover_with_sampler.html

maxwell04-wq commented on 2023-06-01T20:28:57Z ----------------------------------------------------------------

Incorporated.

review-notebook-app[bot] commented 1 year ago

View / edit / reply to this conversation on ReviewNB

ElePT commented on 2023-06-01T15:06:02Z ----------------------------------------------------------------

Given that we no longer talk about the difference between First and Second Order gradients, maybe just "Gradients" fits this section better.

Here, and in the following subsections, I think it's important to always be very clear about when we are showing estimator gradients and when we are using sampler gradients. You already do this in the Parameter Shift Gradients section, but in the other sections you don't mention that you are talking about estimator gradients.

maxwell04-wq commented on 2023-06-01T20:30:46Z ----------------------------------------------------------------

I'm adding a note in the Sampler example that all the following examples will be with estimators.

review-notebook-app[bot] commented 1 year ago

View / edit / reply to this conversation on ReviewNB

ElePT commented on 2023-06-01T15:06:03Z ----------------------------------------------------------------

Citing the spsa paperwould be nice.

maxwell04-wq commented on 2023-06-01T20:35:20Z ----------------------------------------------------------------

Done.

review-notebook-app[bot] commented 1 year ago

View / edit / reply to this conversation on ReviewNB

ElePT commented on 2023-06-01T15:06:04Z ----------------------------------------------------------------

similar comment to the one I left above, we should spread out the imports.

maxwell04-wq commented on 2023-06-01T20:37:15Z ----------------------------------------------------------------

Done.

review-notebook-app[bot] commented 1 year ago

View / edit / reply to this conversation on ReviewNB

ElePT commented on 2023-06-01T15:06:06Z ----------------------------------------------------------------

Line #1.    #Make circuit copies for different VQEs

review-notebook-app[bot] commented 1 year ago

View / edit / reply to this conversation on ReviewNB

ElePT commented on 2023-06-01T15:06:07Z ----------------------------------------------------------------

This problem can also be tackled with the Sampler-based SamplingVQE, as this algorithm requires the Hamiltonian to be diagonal (this is not the case for regular VQE). In this case, the algorithm will use a Sampler-based gradient.

review-notebook-app[bot] commented 1 year ago

View / edit / reply to this conversation on ReviewNB

ElePT commented on 2023-06-01T15:06:08Z ----------------------------------------------------------------

Line #3.    vqe = SamplingVQE(sampler=sampler, ansatz=wavefunction_1, optimizer=optimizer)

I think that the point of this example was to show the use of a sampler gradient, right? but no custom gradient is set here

_maxwell04-wq commented on 2023-06-01T20:41:21Z_ ----------------------------------------------------------------

SamplingVQE does not take a gradient as a parameter. Dropping this example.