DifferentiationInterface

[gdalle]

Hi @TimSiebert1!

I stumbled upon your JuliaCon abstract and it made me very excited about your package! I'm the lead developer (with @adrhill) of DifferentiationInterface.jl, a common standard for calling autodiff backends in Julia. If you want to work with us, we'd be thrilled to try and add ADOLC.jl to the list of supported libraries. As long as you can provide functions for pushforwards (JVPs) and pullbacks (VJPs), that's enough to get started!

What's in it for you?

• Access to a wide potential user base. DifferentiationInterface.jl will soon be a dependency of a fairly large number of packages, including the whole SciML ecosystem. It's easier for users to scitch to ADOLC.jl if they don't have to learn a new syntax or dive into your docs.
• Robust testing. DifferentiationInterface.jl is the ultimate autodiff bugfinder: we have uncovered errors or suboptimalities in nearly every autodiff library we support, and we don't plan to stop.
• Easy benchmarking. DifferentiationInterfaceTest.jl, the companion package, is designed to allow seamless comparison between backends. Wanna see if ADOLC.jl beats the competition? That's the way to go.
• Sparse autodiff. With SparseConnectivityTracer.jl and SparseMatrixColorings.jl, we're taking cues from the work of Andrea Walther and giving Julia a sparse autodiff ecosystem that is unique among high-level scientific programming languages. We'd love to see how it compares to the well-tested C counterparts.

So what do you say?

[TimSiebert1]

Hey @gdalle

This sounds great. It would be great to work with you. The first version of ADOLC.jl is close to the final stage and will be (hopefully) ready for registration in a week. After that, I plan to add and supply the functions for your package.

[gdalle]

Awesome, can't wait!
At the moment it's probably easier to host all the library bindings in the DifferentiationInterface repo (as package extensions) for unified testing. So whenever you're ready and the package is registered, feel free start a PR and I'll help you get it ready!

[TimSiebert1]

Great! Thank you very much. I'll let you know.

[j-fu]

If you are interested: @TimSiebert1 will give a talk on his work on Thu, June 6 14:00 CEST. It will be hybrid, we could send a zoom link, but @adrhill, you could join us at WIAS.

[gdalle]

Yeah I'd be glad to attend remotely!

[adrhill]

Same here, I'd love to join remotely!

[j-fu]

Ok will send the link as PM on zulip.

[TimSiebert1]

Oh, cool that you want to join!

The talk will be about AD, especially the univariate Taylor polynomial method. However, there will also be a brief introduction to ADOLC.jl.

[gdalle]

By the way if you need a hand with Julia package infrastructure (tests, docs, github CI, what dependency goes where), let me know, I can lend a hand.

[TimSiebert1]

I would gladly accept that, and that might be quite good. I feel there is a lot to learn, and I don't use the best practices.

[gdalle]

We can move to a private channel if you want. Zulip, Slack, Discourse? I have the same username on all of those.

[gdalle]

In the meantime you might be interested in this blog @adrhill and I wrote: https://modernjuliaworkflows.github.io/

[TimSiebert1]

I wrote to you on Zulip and will check out the blog.