patrickkidger.bsky.social
I do SciML + open source!
🧪 ML+proteins @ http://Cradle.bio
📚 Neural ODEs: http://arxiv.org/abs/2202.02435
🤖 JAX ecosystem: http://github.com/patrick-kidger
🧑💻 Prev. Google, Oxford
📍 Zürich, Switzerland
65 posts
4,482 followers
654 following
Regular Contributor
Active Commenter
comment in response to
post
...like `pdb`, but with a snappier interface, fewer commands overall, with a focus on a clean REPL/TUI experience to navigate the stack.
Now used by me... and adopted by a host of my coworkers at Cradle! 3/3
comment in response to
post
This library brought to you originally by my need to debug re-raised exceptions in some of the libraries I'm using. Then in my usual fashion I kept adding features, and now we have a whole library 😄
If like me your debugging experience has been a lot of `pdb`, then this library is for you... 2/3
comment in response to
post
Me please! :) I work on bioML. Google scholar mostly reflecting past work in broader sciML:
scholar.google.co.uk/citations?us...
comment in response to
post
Agreed! I have no idea how that works. Is it a thing I submit to or a thing they pick up?
comment in response to
post
Hehe! I originally intended to include the rest of bioML and was then like 'this post is becoming way too long' 😄
comment in response to
post
A huge thanks to Owl, Johanna Haffner, and Jonathan Ziegler for their feedback on the article! I think we ended up with something great. (If a little intimidating.)
Also as timing would have it, also check out @ncfrey.bsky.social's post from yesterday!
ncfrey.substack.com/p/getting-st...
2/2
comment in response to
post
And merged! 😃
comment in response to
post
Shall we add this to the Eqx ecosystem page? I really like it.
docs.kidger.site/equinox/awes...
comment in response to
post
I also went back and took at look at one of the very early READMEs. I think we've stayed true to these goals!
(Also included: Equinox started out as a 'tech demo'. I had to procrastinate from real work somehow, after all. 😄)
comment in response to
post
...the results are now back, and it turns out that *all* of them would have the won the competition!
+ now with bonus analysis on expression, immunogenicity/humanness! 🎉
[2/2]
comment in response to
post
As for this, I'm pretty sure this exists in the main JAX repo! There's a cuDNN implementation here: docs.jax.dev/en/latest/_a...
And there's a bunch of goodies hiding out here:
github.com/jax-ml/jax/t...
(Not super sure what the public API for those is meant to be though lol.)
comment in response to
post
You may be surprised to know that... I have no idea :p We actually use PyTorch at Cradle! This implementation was something I wrote for the sake of an interviewee who was more familiar with JAX. :)
FWIW it's possible to train protein models pretty efficiently: www.biorxiv.org/content/10.1...
comment in response to
post
If they can fix its issues then yes.
But realistically that's a big lift. :)
comment in response to
post
... whilst my professional work has largely focused on things like protein language models! So I'm very happy to doing a better job bridging these two halves of my work. 🌟
I hope this will be a useful starting point for the JAX users of bioML :)
[2/2]
comment in response to
post
Nope, I think it's down :)
I suspect it's just a transitory issue with my hosting provider -- this happened once before and it was resolved in a few hours. (Let's wait and see.)
comment in response to
post
Basically, we want quantitative bio PhDs who care about translational/applied research: bridging the details of real-world protein design projects and the capabilities of ML.
DMs open for any questions and link below! 💥⭐️
jobs.ashbyhq.com/cradlebio/45...
[2/2]
comment in response to
post
Would recommend github.com/patrick-kidg... as an updated+better version!
(Even for pytorch, despite the name.)
comment in response to
post
This is awesome!
And it uses Equinox as well 😁💯
comment in response to
post
Time to write a JAX implementation? :)
comment in response to
post
Yes please! :)
comment in response to
post
Can I be added please? :)
comment in response to
post
Me please! :)
comment in response to
post
Interesting!
I might have to try this out :)
comment in response to
post
I think yes if hatch is used to manage the environments. But I think as a practical matter uv will be taking that crown soon. If they add a way to do something similar then I think that would work :)
comment in response to
post
I have, very occasionally, been known to get mildly disgruntled.
If that's not enough justification for inclusion then I promise to try harder to be grumpy in the future. 😈👿
comment in response to
post
Would love to be added to the AI pack please :)
comment in response to
post
I'm afraid not -- because even the notion of a pytree is something that is baked into JAX itself (`jax.tree_util.register_tree_blahblahetc`)
But if you have your own JAX-independent pytree-like library then maybe you can use Equinox as inspiration to build something similar :)
comment in response to
post
If you're referring to lockfiles, then I deliberately leave this out: when installed then folks often use the latest versions of deps / transitive deps, and this helps to stay on top of when breaking changes occur.
2/2
comment in response to
post
So just running `uv sync` will actually work already. (Modulo the fact that Python 3.13 just came out and not all of the dependencies support it yet.)
1/2
comment in response to
post
All of that said -- this is a pretty nitty detail. I can definitely see 'just use a src directory' as being a simpler best practice when recommending templates to others. Maybe I should do that just for this sake, actually. It's useful to have existing projects as exemplars like this.
2/2
comment in response to
post
Right! :) So to guard against exactly this (and any other similar problems, like non-py files not getting packaged), my CD setup actually builds an {s,b}dist, installs from that, switches to an empty temporary directory, and then runs the tests.
So we're actually covered from that problem!
1/2
comment in response to
post
file. Do you know of any other reason to prefer the `src` approach?
These days I do use uv :) But that's about handling environments+installation, which is not the same as libraries. (These often get conflated, as Poetry handled both.) What did you have in mind?
2/2