PyTorch: Where we are headed and why it looks a lot like Julia (but not exactly)

I happen to know that pytorch is a pain to maintain in packaging systems. It has a complex build system, many non-python dependencies, and massive build times. I don't know how this factors into the dissatisfaction with a C++ implementation, but I wouldn't be surprised if it were a factor.

In other words, python binary wheels are harder to maintain than source-only python packages. And pytorch uses more than a few. I can't imagine Julia makes the problem much simpler. The main pain point is probably the lack of standard, multi-environment packaging solutions for natively compiled code.

I don't know what it would take for this sort of pain point to improve significantly. Some standards around how C, C++, and Fortran projects are packaged would help. This would allow projects to build on top of existing natively compiled tech a lot better. Maybe the biggest reason those languages don't have the same "ecosystem" as python is utter lack of packaging standardization.

I've used Julia for quite a few years now. It's biggest flaws in my opinion are basically cultural and not technological. It's been adopted mostly by serious domain experts rather then typical software engineers and more 'normal' people. I don't know say junior or senior scientists. This has lead to amazing results but also has it's own detriments.

Some portions of the ecosystem are rock solid, especially the parts where JuliaComputing makes money from consulting(not all but some). Other parts are beds of sand/permanent research projects. The median experience is usually someone points you to a package and it doesn't really do what you hoped it would so you end up adapting it and rolling your own solution to a problem. Maybe you try to make a PR and it gets rejected because of "not invented here"/academia mindsets, either way you made a fix and your code works for you.

What makes this barrier hard to overcome for adoption is: trust, and blind spots. People who aren't experts in a casual work area (maybe computer vision) realize they can't use a tool to do something `basic` and run away to easier ecosystems(R/Python). People who are experts in other areas, check credentials of packages see that an ivy league lead researcher made it and assumes it's great and usable for a general audience. So you'll get a lot of "there's a package for that" but when you go to use it you might find the package is barren for common and anticipatable use cases in industry (or even hobbies).

This makes Julia best positioned as a research tool, or as a teaching tool. Unfortunately - where Julia actually shines is as a practical tool for accomplishing tasks very quickly and cleanly. So there's this uncomfortable mismatch between what Julia could be and what it's being used for today. (yes Julia can do both not arguing against it). The focus on getting headlines far outsurpasses stable useful stuff. Infact, very often after a paper gets published using Julia, a packages syntax will completely change - so no one really benefits except for the person who made the package.

Interestingly, 1 person(with some help of course) fleshed out the majority of the ecosystems need for interchange format support(JSON), database connections, etc. It's not like that person is jobless spending all their days doing it - it was a manageable task for a single smart person to kick off and work hard to accomplish. Why? Because Julia is amazing for quickly developing world class software. That is also kind of its detriment right now.

Because its so easy to create these amazing packages you'll find that a lot of packages have become deprecated or are undocumented. Some researcher just needed a 1 off really quickly to graduate, maybe the base language(or other parts of the ecosystem) changed many times since its release. Furthermore, if you try to revitalize one of these packages you'll sometimes find a rats nest of brilliance. The code is written very intelligently, but unpacking the design decisions to maintain world class performance can be prickly at best.

One of Julia's strengths is it's easy/clean to write fast enough code. One of its downsides is, this attracts people who focus on shaving nanoseconds from a runtime (sometimes needlessly) at the expense of (sometimes) intense code complexity. Performance is important, but, stable and correct features/capabilities mean more to the average person. After-all, this is why people use, pay for, hire for: Matlab, Python and R in the first place - right?

Most people don't want to have to figure out which ANOVA package they should use. Or find out in a bad way some weird bug in one of them and be forced to switch. Meanwhile in R: aov(...).

Do I blame Torch for not using Julia? No. Should they consider using it? Yes, absolutely. Does Julia's cultural issue need attention before risking Python(or anything else) reinventing a flavor of Julia that's more widely used for stability reasons alone - in my opinion, yes (see numba, pyjion, etc). Still love the language, because technologically it's sound, but there are blemishes. I'd chalk it up to growing pains.

I think Python is used by a lot of people now because elementary schools and universities teaching it, easy to start with and you can gradually convert to more advanced (compact) language, you can prototype easily without having to remember a lot of syntax, and you come a long way without an IDE (+ Jupyter Notebook is great).

I love that Torch is converted to Python. I think it’s more importantly with a large ecosystem than the most efficient language.

I'm glad you qualified this with 2/5/whatever. Usually people say #2 in everything, which is of course wrong.

Python more like #100 in terms of speed, #100 in terms of correctness, #100 in terms of sound abstractions, #10 in terms of readability for large programs. Its real strengths are quick hacks and a decent C-API.

The community is smug, conceited, does not value correctness and in general is intoxicated by Python's undeserved success. Many posers and incompetent people.

Try using Go for any serious math project, then do the same using Julia. Report back as to how both approaches went :P. From someone who uses both languages for very different tasks regularly, I would never try to write Torch from scratch in Go. I can't envision a way for it not to be a serious maintenance or performance disaster. Maybe that's a lack of my own creativity, but I'd much sooner use C++ rather then write any large portion of it in Go. If only for template generics...