dmm | JuliaCon workshops

You're viewing

dmm's journal
Create a Dreamwidth Account Learn More

Reload page in style: site light

JuliaCon this year is 8 days of workshops (July 20-27), following by 3 days of the main conference (July 28-30).

The list of workshops is here: juliacon.org/2021/workshops/

The most interesting for me is the workshop on the next generation of autodiff systems (July 27), "Diffractor: Next-Gen AD for Julia": pretalx.com/juliacon2021/talk/review/MJ7T9LXLXNDWHPW9WJN7SGTWU7L9QMJW

All workshops start at 10am United States East Coast time and last 3 hours, two parallel workshops each day; if you don't care about interactive real-time aspects, youtube recordings will be available...

Full schedule (UTC time default, can switch to US East Coast): pretalx.com/juliacon2021/schedule/

Flat | Top-Level Comments Only

From:

dmm

I will record my notes from those workshops here.

From:

dmm

So, it's time to overview the workshop (one more day of them left, and we know everything at this point).

1) Diffractor is finally released, thanks to the workshop pushing the authors to do that: https://juliadiff.org/

https://github.com/JuliaDiff/Diffractor.jl

Diffractor workshop itself is going to be cancelled/postponed, because of the speaker's illness: https://twitter.com/KenoFischer/status/1419680777135136771

(The chaotic times we live in.)

I am going to try it (I've spent a lot of time lately working with Zygote and ended up being less than happy so far, to say the least; but it might be that I just need some help from more experienced people).

2) The most non-trivial workshop was on quantum computing. It was conducted by the Amazon Braket people, and while the canonical Amazon Braket API is in Python, there is a full-strength Julia API as well.

https://www.youtube.com/watch?v=tEd6driz37w - that's video recording of the live session (I think even the live chat can be replayed)

https://pretalx.com/juliacon2021/talk/V3N73B/ - that's the description of the workshop

https://pigeonhole.at/QUANTUM/q/2333682 - that's the Q&A for this workshop

Now, speaking of Amazon Braket: https://aws.amazon.com/braket/

You can either work on a simulator (which you can also just run on your home machine), or you can work on a real hardware from one of the three providers at the moment: https://aws.amazon.com/braket/hardware-providers/

I don't understand much about IonQ, but you can work on actual 32-qubit quantum processors from Rigetti: https://aws.amazon.com/braket/hardware-providers/rigetti/

That's not too bad, you can process 4 billion states in parallel, not enough to "take over the world", but enough to do interesting things.

You can also use D-Wave quantum annealers, some of which are pretty formidable at this stage ("Advantage QPU" has more than 5000 qubits; note that quantum annealers are not full-strength quantum computers, but are good for specialized problems; I wonder if people would be open to calling then "semi-quantum computers"): https://aws.amazon.com/braket/hardware-providers/dwave/

Amazon Braket is also actively hiring; they seem to be quite bullish on the future trajectory of this.

Edited Date: 2021-07-26 09:48 pm (UTC)

From:

dmm

Other than that:

1) "GPU programming in Julia" is quite useful, if you want to do GPU programming in Julia. Don't listen to it in advance, but if you are about to start writing software for GPUs, this is a great workshop and will save you tons of time explaining what to do and what to avoid.

2) "Package development in VSCode". You probably want to first master Visual Studio Code and its Julia extension on your own a bit (the introductory part of this workshop is way too fast and abbreviated, it pretends to be self-contained in this sense, but it is not). Once you've done that, it is a tremendously useful workshop, teaches all kinds of tips and tricks for Julia programming inside Visual Studio Code.

Note that GitHub Copilot also works nicely for Julia, not just for Python, although the Copilot is a closed beta at the moment. But it will not remain a closed beta forever, and this is a good enough reason on its own to master Visual Studio Code.

3) "Simulating Big Models in Julia with ModelingToolkit" - very useful tutorial and overview of the state of symbolic computations in Julia, and of multiple orders of magnitudes speed-ups enjoyed by its users, and it was conducted by the brilliant Chris Rackauckas, although it was just a reasonably good workshop (much less impressive than his SciML workshop last year: "The best talk of JuliaCon 2020 was this 4-hour workshop: "Doing Scientific Machine Learning (SciML) With Julia" by Chris Rackauckas: https://www.youtube.com/watch?v=QwVO0Xh2Hbg ")

4) "JuliaReach" - an interesting material, but the presentation was so-so.

5) "Intro to Metaprogramming" - I am sure this will be super-useful for anyone who wants to master Lisp macros and such, but that's still tomorrow, so I am looking forward to it, since my level of mastery of these things is rudimentary and needs improvement.

Edited Date: 2021-07-26 10:14 pm (UTC)

From:

dmm

Now about Diffractor. This is a not-yet-released package (but I presume they can't avoid releasing it within the next few days). It's very mathematical, based on categorical things like "optics": see "Categories of Optics" written by a Wesleyan University graduate student: https://arxiv.org/abs/1809.00738

But I am bumping into Flux/Zygote limitations so much lately that I hope this would solve all those problems.

The author is a rather famous guy who happens to be a co-founder and CTO of Julia Computing:

https://twitter.com/kenofischer/status/1317961750218706946?lang=en

https://github.com/Keno

https://en.wikipedia.org/wiki/Keno_Fischer

Edited Date: 2021-07-17 01:13 am (UTC)

From:

dmm

And this 14-page "terminology.pdf" gist: https://gist.github.com/Keno/4a6507b75288b1fe671e9d1cc683014f

and this video, "Non-local compiler transformations in the presence of dynamic dispatch": https://www.youtube.com/watch?v=mQnSRfseu0c

Now, I am NOT a fan of categorical methods, my intuition works in different ways, but I was torturing myself with Flux/Zygote corner cases so much in recent weeks, that I am really hoping this package would solve those problems and would "just work"...

Edited Date: 2021-07-17 01:50 am (UTC)

From:

dmm

So, this is released now: https://github.com/JuliaDiff/Diffractor.jl

Here is some of my fresh self-torture with Zygote (it might be that I just need some help from more experienced people): the README here https://github.com/anhinga/julia-flux-drafts/tree/main/arxiv-1606-09470-section3/dictionary-experiments

From:

dmm

Let's also start studying the full schedule (the posters are separate, we'll learn more about them later, these are just talks and such): https://pretalx.com/juliacon2021/schedule/