[3jckd]

It's interesting how it's going to play out. On one hand side, Swift is a pleasant language to work with (despite its infancy). But on the other, having a Tensorflow API doesn't suddenly give it a bunch of libraries for statistics, comp. vision, modeling, visualisation, etc. that Python/R/Julia coughMATLABcough have.

Nowadays, it's difficult enough to convince people to drop e.g. MATLAB for R or Python for Julia (let's assume that there's some merit to it), despite them having excellent counterparts for almost everything. Swift's success in this domain depends solely on the adoption by developers/researchers/engineers. Unless they're just going to mostly use it internally (as Google is known to).

Which brings me to the last point - why on Earth would they pick Swift (apart from Christ Lattner being involved) when Julia was on the table? It ticks all their boxes and has more mature ecosystem for all things "data". Provided rationale is hardly convincing.

[aschampion]

Yeah, I don't buy the justification versus Julia because of community size either, given most of Swift's community has little to do with data science. The document even says as much, contradicting that rationale, later on.

As someone who uses TF heavily, I would be much more excited about this project if they'd chosen Julia. Swift's tooling isn't great, and I already have a foot in one language with an immature data science ecosystem (Rust).

[tomjakubowski]

How is Rust-Julia interop?

[celrod]

Julia's `ccall` is great in terms of overhead[0], so calling Rust shared libraries is not a problem. On the Rust side, it took me a while to figure out passing in pointers, and then constructing slices via from_raw_parts[_mut], so that I can transition to safe Rust. Perhaps that is obvious to more experienced Rust programmers, but I was left with the impression that receiving pointers and crunching numbers is not yet a common application for Rust (unlike C, C++, or Fortran). Meaning there is not a lot of introductory material coming from that angle at the moment. Additionally, to get good vectorization seems to require nightly and a fastfloat[1] library. In particular, you'd want associative math / fp-contract for fma and SIMD instructions, and perhaps fno-math-errno to turn off branches in functions like sqrt.

I imagine calling Rust from Julia will be much more common than calling Julia from Rust. I know approximately nothing about this, but there are plenty of questions about embedding Julia into C/C++[2][3]. May be similar for Rust.

[0] https://github.com/dyu/ffi-overhead [1] https://github.com/robsmith11/fastfloat [2] https://discourse.julialang.org/t/support-with-embedding-jul... [3] https://discourse.julialang.org/t/api-reference-for-julia-em...

[ychen306]

I will say this in the risk of talking out of my ass as I have no experience in either language :). Having a statically typed language greatly simplifies the tooling because static analysis is much easier; graph program extraction involves one such analysis. When you have to deploy the trained model in production one would hope not to use Python or Julia.

I'd like to add that, with my limited experience in prototyping some of my ML models, having a static checker to check that your tensors have the right shape is much better than having to run your code.

[dnautics]

what's the problem with deploying julia in production in inference? Some occasional piece of data that looks wrong in an unanticipated way causes a runtime type fault? People deploy high uptime websites with django - how do they do it? Well you use kubernetes (or, gasp, systemd) and have restart and load balancing logic. Even if you were typecheck-compiled, you can't guarantee some other developer logic or system error, or an errant bit flip from a cosmic ray, won't take your setup down. Static checker doesn't really matter. If you're at the point where you're ready to deploy, you're probably good for at least 95-99% of the data you'll ingest. The rest of the gap can be closed using rolling update.

[shele]

On the other hand, Julia can do the right thing dynamically. Your matrix happens to be symmetric? Julia will chose an appropriate factorisation and will propagate that knowledge through dynamic dispatch.

[skohan]

"Automagical" solutions tend to work until they don't. I would much rather have my solution be provably correct at compilation time than to depend on high-overhead runtime systems.

[dnautics]

You don't know what you're talking about.

[skohan]

Tell me where I'm wrong

[ychen306]

Which you don't need most of the time in production because you can statically determine which operation needs to happen without the overhead of dynamically choosing the operation.

[skohan]

Just because you can deploy production code without static checking does not make it a good idea. And to me if you need to rely on an external process to restart your application for NPE's etc. it is a sign your application is not that robust.

It's possible to write production code in brainfuck if that's what you really want to do. Statically checked code is easier to implement correctly, easier to modify, and easier to maintain.

[dnautics]

> Statically checked code is easier to implement correctly, easier to modify, and easier to maintain.

If that were true literally no one would program in python. Static checking is not the end all to uptime and stability. I wrote an elixir program in three days that served as a testbench for a senior's go program (which took him six months to write). This senior believes in static typechecking for everything and doesn't write unit tests. The testbench handles thousands of parallel async requests without a hiccup and even survives operating system resource exhaustion, where the go program falls over and panics.

Erlang is not statically checked (there is a static typechecker, but it's not fully typed). I promise you a well written erlang program has much higher stability than a well written go program. There is a reason why kubernetes exists, after all.

[skohan]

> If that were true literally no one would program in python.

There are many cases in which people don't choose the optimal language. But I would say the size of the python community has more to do with inertia, the breadth of libraries available, and a relatively shallow learning curve than it says about its strengths as a tool for writing good software.

I actually find that Python has some rather serious warts: the whole story around environment/version management is a mess, and the less I have to work with Python in a serious capacity the better.

> I wrote an elixir program ... The testbench handles thousands of parallel async requests without a hiccup and even survives operating system resource exhaustion, where the go program falls over and panics.

Well Erlang is specifically designed for concurrency and stability: if you want to judge your result on those two metrics I hope it is going to perform well.

I never made the claim that static typing is the "end all to uptime and stability" - static typing makes it easier to reason about your code, and to provably eliminate many issues. It's very nice that you implemented a test harness quickly, but come back to me when you've worked on a complex codebase with several other people over an extended period of time.

[ychen306]

The problem comes down to how you use your model, which is all that matters. If there is a mismatch between the language in which your model is embedded in and the language in which your main application/interface is written in, then ideally you want a way to extract the model statically from the dev environment. It's not about type-checking per se, which is nice, but how you get the model out without having to, say, start a python interpreter in your server process just to do the inference (which is why in some aspect Tensorflow is more convenient).

[agibsonccc]

I'd argue in general that outside of python, there's not really much of a focus from google itself. They are largely leaving other language bindings to other people (see what's currently going on with tensorflow 2.0).

Their focus is more on the c bindings and allowing other people to build what they want on top of that.

Other language bindings aren't generally going to be used for anything more than inference. First class actual data science work isn't going to happen in other languages anytime soon (at least outside of julia and R which are at least trying to compete in this niche).

[marmaduke]

Julia doesn't tick the "compile to .o/.h" box. As far as I can tell, the use case for AOT Julia is avoiding package compilation overhead, not delivery of standalone code objects.

_edit_ seems JuliaC does support this sort of thing:

https://juliacomputing.com/blog/2016/02/09/static-julia.html

[StefanKarpinski]

Julia does already support this kind of thing. Moreover with a minuscule fraction of the money that’s being poured into making Swift usable for data science and machine learning, truly top notch support for generating standalone binaries from Julia could readily be developed. Which is kind of frustrating but what can you do?

[celrod]

If you'd like to track the latest developments:

https://github.com/JuliaLang/PackageCompiler.jl

[skohan]

> Tensorflow API doesn't suddenly give it a bunch of libraries for statistics, comp. vision, modeling, visualisation, etc. that Python/R/Julia coughMATLABcough have.

Actually in this case it does. Swift for Tensorflow includes python interop out of the box: https://www.tensorflow.org/swift/api_docs/Global-Variables#/...

The supported use-case would be to do your ML work in Swift, and then call numby etc. from Python.

[cageface]

Swift is a nice language but its reliance on reference counting means you have to work a lot harder to avoid retain cycles than you do in a garbage collected language.

That might have been the right choice for Apple’s uses of Swift where GC pauses affect the user experience but for most other use cases it’s too much of a cognitive burden IMO.

[saagarjha]

Personally, I find that this only really comes up rarely. Most of the time strong references are fine.

[cageface]

My iOS code is loaded with weak/strong ref handling logic. It comes up all the time when using closures in UIKit.

[saagarjha]

UIKit is kind of annoying because it’s really not “idiomatic” Swift: it’s a wrapper around Objective-C (albeit, a very nice one) that happens to bring along with it a bunch of decisions that lead to having to deal with reference lifetimes.

[cageface]

That could be. I'll admit I haven't used Swift at all outside of coding Appkit/UIKit apps.

[dep_b]

iOS Swift is usually the front-end code which means that you're doing a lot of connections to data sources elsewhere. Non-blocking connections always need a weak/strong dance. If you're using Swift for logic most of that is gone.

[newen]

Hah...it's funny how they pretend to give objective rationales for choosing Swift when it's pretty clear the decision was made long before.

[microcolonel]

I feel like GraalVM has a chance to solve some of this at least. I wonder if anyone will make an Octave GraalVM frontend, they already have one for R.

[c-cube]

Isn't Graal an Oracle thing? I don't understand why anyone would want to touch that even with a 10-foot pole.

[microcolonel]

Yeah, I would tend to agree. If Google successfully appeals Oracle's suit against them for implementing Java, I might consider using it in a product, but Oracle makes it legally risky to use any of their products.

[plg]

> why on Earth would they pick Swift

Because of iOS?

[phillipcarter]

iOS isn't really relevant for this. You would certainly deploy a model into an app, but that is likely to be using ONNX: https://medium.com/@alexiscreuzot/building-a-neural-style-tr...