Towards a Universal Code Formatter through Machine Learning (2016) | Hacker News Reader

Towards a Universal Code Formatter through Machine Learning (2016)

arxiv.org

74 pointsbreck8 comments

Great to see this picked up on HN - it's really interesting work. Terence is my colleague at USF (in fact he's the guy that got me to go there!) and it was great to see the journey from "could this possibly work" to "holy crap it does work!" over a period of many months. He's known for his work with parsers, of course (i.e. ANTLR), but recently he's been investing heavily in machine learning. This project is a great example of how domain experts can go into whole new areas by building on top of machine learning methods.

For anyone interested in getting into machine learning, I hope you'll consider the MSAN program at USF (https://www.usfca.edu/arts-sciences/graduate-programs/analyt...). Terence is the guy that made this program happen at USF and it's turned into something really special.

Disclaimer: I'm a deep learning researcher and teacher at USF.

One possible angle for improvement of this technology: use a deep learning net to conjure up a different feature vector than the one I handcrafted from language/grammar expertise. I believe this was your idea. :) Glad to have you on board teaching and doing research!

I'm pretty sure this would work really well.

I've trained CharCNN on log files, and it generates really good examples files. To me that shows that even a comparatively simple model can capture syntax rules, so I'd imagine a LSTM would generate really good feature vectors.

It would be spectacular to have a tool like this to enforce, and automatically revise, code style guidelines as languages evolve. Keeping up with something like the C++ Core Guidelines [0] can be a fair amount of work, as you need to respect existing idioms to keep your codebase coherent (a higher priority than respecting the "one true style", in my opinion).

A tool like this would be brilliant if it could automatically generate suggestions for "hybrid styles" that fit your existing codebase, and provide suggestions to migrate existing code patterns to "correct" ones over time.

[0]: https://isocpp.github.io/CppCoreGuidelines/CppCoreGuidelines

First step to using CodeBuff would be getting an ANTLR grammar for C++ or at least a fuzzy version. It's still a prototype but should do pretty well.

Is there a prototype anywhere? I want it!

This seems to be it: https://github.com/antlr/codebuff

Yep, that's it. Somebody has ported from Java to C# as well. Next step is really to convert to use a Random Forest classifier. I'm stuck elsewhere at the moment.

Reply on news.ycombinator.com