quantum mixer implementation

[ongspxm]

closes #641

implementing a quantum class neural network from https://qiskit.org/textbook/ch-machine-learning/machine-learning-qiskit-pytorch.html

implementation notes

• currently using QClassic wrapper around the Neural mixer
  • its easier to error out the unit test
  • because qiskit is an optional dependency
• QClassicNet is quite rudimentary
  • returns a quantum modified value as per the tutorial

[ongspxm]

Running in /home/runner/work/lightwood/lightwood
Add input parsed as single string, running 1 git add command.
> Using 'Automated Author <info@mindsdb.com>' as author.
> Using "updating docs" as commit message.
> Running for a PR, the action will use 'feature/qmixer' as ref.
Internal logs
Outputs
Error: Error: There is no tracking information for the current branch.
Please specify which branch you want to merge with.
See git-pull(1) for details.

    git pull <remote> <branch>

If you wish to set tracking information for this branch you can do so with:

    git branch --set-upstream-to=origin/<branch> feature/qmixer

this github action seems to be failing, i ran the command locally but it doesnt seem to help resolve this issue, sorry new to this whole github action thing

[George3d6]

There is already a PR implementing a quantum mixer here: #639

Can you please explain how it's different from it.

I don't mind merging both mixers and/or doing the work to merge them into one and award both of you 10 points for the laptop, but I'd like to understand how this is different given that it came after the original.

[ongspxm]

Yeah, it's not much difference in terms of the quantum neural network structure. The main difference is that this pr is built upon the existing neural mixer to take advantage of the existing batching, and hyper parameter tuning that is in neural. On inspection of both prs, I would say that the other pr spends more effort implenting the pipeline to load and train the network, whereas this effort is geared towards a more generic qiskit supported neural network, with more qiskit features like batched jobs run and all.

…

________________________________ From: George ***@***.***> Sent: Wednesday, October 20, 2021 12:31:30 AM To: mindsdb/lightwood ***@***.***> Cc: Metta Ong ***@***.***>; Author ***@***.***> Subject: Re: [mindsdb/lightwood] quantum mixer implementation (PR #645) There is already a PR implementing a quantum mixer here: #639<#639> Can you please explain how it's different from it. I don't mind merging both mixers and/or doing the work to merge them into one and award both of you 10 points for the laptop, but I'd like to understand how this is different given that it came after the original. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub<#645 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AAK5KRSOHGYXXCAQPB3LY4DUHWMOFANCNFSM5GERJKZA>. Triage notifications on the go with GitHub Mobile for iOS<https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android<https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>.

[mrandri19]

(Author of the other PR here) I have to say that the two PRs came at around the same time. This one is more production-ready since it extends the pre-existing classes rather than creating new ones, hence taking advantage of much of the existing infrastructrue. If I had to choose, I would merge this one.

I have a few comments, stemming from TODOs I have in my PR:

• Since this reuses the Neural class, I am wondering what will happen on a regression problem, especially considering the final softmax. Perhaps this should only support classification.
• This only supports having quantum circuits as the last layer.
• It would be cool to experiment with more complex quantum circuits.

[George3d6]

• Reusing Neural is indeed a good idea assuming that the quantum network doesn't need any custom behaviour, Neural does a bunch of "fancy" logic in terms of early-stopping, switching target encoder flags, lr finding and partially fitting on the dev data, which we found helps a lot in benchmarks.
• Personally I don't see the point of the final softmax, @ongspxm any reason why that was added?
• As for experimenting with "more complex quantum circuits", no idea on that point

At any rate, looking at both PRs more I do like this one best, ultimately I'd want to get the mixers in and actually benchmark them.

@ongspxm any changes you'd like to make on this one or are you happy to merge as is (again, softmax as the last activation looks wrong to me but maybe I'm mistaken, idk), unittest look good so in principle, this works.

@mrandri19 would you want to finish your original PR? If so I can merge that too and the benchmark this both head-to-head and when we have the time I'll merge them into a joint implementation. Alternatively, if you have some ideas that'd build on this PR I can merge this one when @ongspxm is happy with it and then you can PR your changes into the new staging.

[ongspxm]

• yeah, yall right, softmax should be removed. i had a brainfart and saw crossentropyloss, which made me just put it in by habit
• added one last nit, for easy extension to other nets

yeah, i think it should be good to go

[mrandri19]

would you want to finish your original PR? If so I can merge that too and the benchmark this both head-to-head and when we have the time I'll merge them into a joint implementation.

I think it's best to join this one, since in the current state they are doing the same thing and this one does it better.

Alternatively, if you have some ideas that'd build on this PR I can merge this one when @ongspxm is happy with it and then you can PR your changes into the new staging.

I'll do this, thanks!

[mrandri19]

yeah, yall right, softmax should be removed. i had a brainfart and saw crossentropyloss, which made me just put it in by habit

I believe this is correct. The final softmax will be added if needed by _select_criterion:
https://github.com/mindsdb/lightwood/blob/staging/lightwood/mixer/neural.py#L77-L90