marco_z
marco_z on Hacker News
Through tears and mud, how else?
Posted by
marco_z on Thursday December 18, 2025 @02:43AM
from the show-hn dept.
https://github.com/ocramz/lightning-extra
After building ML systems in various organizations, I collected some useful kit I wrote into a single library.What you can do with this:* Save (and retrieve) model checkpoints (optionally with a content-addressable naming scheme) on blob storage* Load datasets incrementally from blob storage into Pytorch, using a local disk cache* store your training metrics into SQLiteDesign principles :*...
E.g the pieces from the rows you eliminate reappear at random at the top row.
probably one needs something like lazy evaluation, so that things are only computed if needed.
Boumal was advised by Absil IIRC :) And in fact you can see this in his more modern presentation of the material.
Good pointer! I did know about CVXPY but haven't used it yet. Though in requiring everything to be (log-)convex it constrains the user programs quite a bit, but I imagine it would have worked in this case because the objective is linear and the feasible set convex. I won't deny the role accident and force of habit had in picking this particular implementation path; I used Pytorch because it's my daily driver :D and because I had found 'mctorch' which provided the Riemann SGD logic.
It's incredible how many algorithms can be made differentiable end-to-end.
E.g. in https://arxiv.org/abs/1905.11885 "Differentiable ranks and sorting using optimal transport" they show how a relaxation of an array sorting procedure leads to differentiable rank statistics (quantiles etc.). The underlying theory is quite close to what I show in the blog post.
In DETR they don't actually use a differentiable Hungarian algorithm, it's only used to score their neural predictions outside of the training loop IIUC.
Yes, the problem can now be solved efficiently with a scikit one-liner, but the aim of this post was more speculative rather than beating the benchmark :)
Author here! Thank you f1shy for sharing.
Sugoi! I imagine this approach would be very useful in learning any other language, too.
An interesting connection between combinatorial optimization and differential geometry. And some Pytorch too, because when all you have is a hammer..
I built most of this one year ago, but only now managed to write it up. I enjoyed the change of pace of a pure speculative experiment (with some deep connections to a bunch of interesting theory and applications), and I hope you will enjoy reading about it too.
Hi there, while the word list will help, I think plugging an LLM to a user-facing form is still a very, very bad idea, especially if this is meant for children.
> you could ship logs into Splork, then let it ingest that data into FlorbHut, and then use a frontend like Glorply to search logs
Seriously though, is there a single OSS product that does all of this ? Like, for a small multitenant app (i.e not "web-scale"), and that doesn't force one to get a degree in observability just to get stuff done.