I was taken aback when looking at Turing for Bayesian modelling that the distributions were just the standard distributions found in the Distributions package! In Python, every Bayesian framework has its own implementation of everything from distributions to log probabilities, but it all composes in Julia.

> It's small and beautiful, but it needs more manpower to compete against Python.

Agreed. Docs of major projects are still incomplete or non-existent, there's a lot of projects that have been nigh abandoned (cough naive bayes cough), and the composability of Julia coupled without strong leadership leads to an ecosystem of overlapping functionality. Still, I hope by leveraging PyCall, Julia can overcome some of these issues.

You mean that, in Julia, we the users have to "compose" our own implementations of models (e.g. log probabilities), as opposed to using the already-made ones in Python?

(The same is true if you use something that's built on top of Stan or JAGS or BUGS or something else.)

In Julia's Turing.jl, everything is built around data structures that are first-class parts of Julia, so there's no need to have special Turing.jl versions of, say, probability distributions.

In Julia, if the DataFrame library is missing something I can just loop like an array and have a method that works just as well as if the library provided it, the CSV, DB and table processing libraries all use the same conventions so if library "A" solves my issue I'm not forced to use the same library "A" serialization method. Basically instead of twisting the logic to what I'm given, I just fill in the blanks. Sure Julia has way more blanks, but Julia also has half of the age of some of the Python's library I use, it's more about the maturity of the community than anything to do with the design choices of the language and I can only hope it gets better and better with time.

now python has like 5 identical numpy-like API re-implemented in multiple framework. That's where resources is wasted (unnecessarily)

If all you do is ML/numerical processing, against data that's already been cleaned up, I bet it's really great tho.

But I agree with you that in the short term it is a replacement for Fortran, Matlab and R. But a new language has to start somewhere. Attacking the niche filled by these language is natural as these are old, dated and ripe for disruption. The key is to gain some critical momentum and user base. That will carry the language to other domains.

I've been using Python for ML for the last 3 years and I've never felt this way. It might be that I'm not all about the hip new languages, but I don't really see the benefit of making Python more ML/Haskell-ish.

The ML use case for Python is roughly as follows: you design, train, and evaluate models. Then, if something is decent enough to use in production, you switch over your application to load and use that instead. I don't really see where Haskell or any language from the ML family can improve in that process.

Sure, the code that you used to implement your model may improve slightly, but I don't see that code improving significantly. The fruit of your labor is usually a protobuf file (encoding the TensorFlow graph) or whatever your framework uses to encode the model you built. The code actually surrounding it is very minimal for most use cases.

> In Julia, things are really small and composable. For example, you have a probabilistic programming library like Turing and a differentiable programming one like Flux, and it's trivial to implement some Bayesian neural networks.

There's nothing stopping you from composing things in Python. But it's simply not a goal of 99% of ML libraries to be composable with other libraries. You're probably never gonna run TensorFlow and PyTorch in the same application (trust me, I've tried, it was a nightmare) and I don't see why you would compose a TensorFlow model with a PyTorch model without incurring tons of overhead in your application around gluing these two things.

There is, and you pointed it out yourself:

> I don't see why you would compose a TensorFlow model with a PyTorch model without incurring tons of overhead in your application around gluing these two things.

This is where Julia is profoundly different. Reusing stuff from different libraries is trivial compared to Python. You could take an activation function made for one libraries use it without any modification or wrapping in another ML library.

These is where Julia will win long term. If you look at the ML libraries in Julia they are tiny. The Python side of things require gargantuan effort because things are not composable. The wheel is reinvented over and over again.

In Julia you can compose a custom distribution, with a bayesian model with an ODE with a neural network with unit number types with custom Julia written CUDA kernels and multithreading.

Edit: That are not designed specifically to work with each other

Can python even hope to do a fraction of that, still be fast and differentiate through everything?

[0]: https://www.tensorflow.org/guide/create_op

[1]: https://pytorch.org/tutorials/advanced/cpp_extension.html

Even then, you're not going to have fast custom types to be used on the GPU without dropping into C++

I don't understand why it's not possible. You're asking if it's possible in the language. I don't see anything stopping you from, as you say, writing your own framework, or simply doing it in TF. Perhaps my ignorance of Julia is showing :)

But there really is a profound difference. The thing is that Python is really just a language to rearrange networks of nodes written in C++. A node written for one ML framework will be incompatible with a node written in another language. If you want them to work together you need to spend a lot of effort gluing them together, and most likely the performance benefits are lost.

This is very different in Julia because you are using pure Julia code. What is specialized C++ nodes in Python, are for the most part just plain functions in Julia. There is nothing special about them. That is why they can be used anywhere.

In Python frameworks each of these C++ nodes need to contain information about how auto-differentiation is done. That is why they are tailor made for one specific ML library.

In Julia the whole automatic differentiation stuff is perform code transformations on regular Julia code.

Are you familiar with LISP, and the idea of code as data? That is what we are talking about here. A regular syntax graph of regular Julia code gets manipulated to do autodiff. You are not organizing C++ nodes in a network.

That does sound very different than Python. Admittedly, I haven't used Julia, because I never had to for professional purposes. For personal projects, I use PyTorch, because I find it easy to use, intuitive, and I have a lot of experience with it. Maybe one day I will be forced to use Julia and will appreciate what it brings :)

I don’t much use Julia, but it is a beautiful language. I started a tiny side project for using Julia for non-numerical things like querying SPARQL endpoints, text processing, etc. Julia could be a universal, use for almost everything language.

It's not that it's not possible. Yeah, you could always "write your own framework".

It's that it's not readily available by the language, ready-made to compose, without resorting to external libs on C/C++/Fortran, and so on.

So, "not possible" not in the sense that "it's not possible to write a program that solves the halting problem". More like "it's not possible to write a web app in assembly". Which means, yeah, it's possible, technically, but you wouldn't like it, it would be full of holes, a pain to maintain, and not a good time investment.

I think the reason why people use python is because it’s the only universally-accepted language for coding interviews.

I am not asking anyone to write Python when it isn't part of the job.

https://numba.pydata.org/numba-doc/latest/cuda-reference/ker...

which basically calls NVPTX

https://github.com/numba/numba/blob/ec6fa07c12c3703a52c2b4ac...

which is exactly what CUDA does anyway; CUDA code is actually a frontend for an ISA

https://llvm.org/docs/NVPTXUsage.html

The irony here is that tiny Julia libraries give the same power as much larger Python libraries which require highly specialized and trained developers to evolve and maintain.

With Julia it is much easier to grow the eco system because you can make lots of relatively small packages which can then be combined in almost any way.

The challenge for the Julia community today is really to make people new to the environment aware of this.

I have encountered people who thought Flux couldn't do anything because it was so small. He was so used to the huge monolithic libraries in Python, that it did not occur to him that adding something like an new activation function is literally one line of code. He is used to that requiring adding various C++ nodes. Wrap those up and God knows what more steps you need when you extend something like TensorFlow.

I don't think most people realize how powerful Julia is. They are not used to being able to combine libraries the way you can do in Julia. A big part of the effort for the Julia community will simply be to write more tutorial and introduction which better introduce beginners to these kinds of capabilities. Once you know Julia it is pretty obvious.

But grab people's attention and make them realize Julia could be a solution to their problem, I think one needs a lot of sort of shallow and quick intros to these kinds of things.

^This. As we're seeing here in the discussion thread Python folks don't really realize what they're missing when we're talking about composability of of libraries in Julia - it's kind of difficult to explain the impact on productivity without people actually trying it out. I think Julia and Flux are getting to the point where they're quite useable and comparable with PyTorch. Also some of the Neural ODE stuff and differentiable programming in SciML seems just a lot easier to implement than it would be in Python.

I don't think I realize how powerful Julia is, and it's my main language (and I have written libraries, and contributed to Base).

See the paper.

It allows for the ecosystem to have a combinatorial explosion of possibilities as each domain expert works on their own package.

I can't foresee every application, but the emergence is the point. This is just a crazy example to show the bounds of what's possible.

Julia is not ML-ish. In fact, at first blush, Julia reads very similarly to Python. You even get list comprehensions in Julia!

> Sure, the code that you used to implement your model may improve slightly, but I don't see that code improving significantly. The fruit of your labor is usually a protobuf file (encoding the TensorFlow graph) or whatever your framework uses to encode the model you built. The code actually surrounding it is very minimal for most use cases.

It's really about breaking down the compositional boundaries here. Upthread, someone talked about how they could play around with Bayesian Nets by simply mixing Flux.jl (the differential programming library) with Turing.jl (the Bayesian programming library). Mixing Tensorflow/PyTorch with PyStan, for example, can be a nightmare. That's also why there's so many implementations of probabilistic programming frameworks (Edward, PyMC3, PyStan, Pyro, etc); they all use different underlying libraries. In Julia they just all compose together.

> I don't see why you would compose a TensorFlow model with a PyTorch model without incurring tons of overhead in your application around gluing these two things

I find the change to be illuminating. Before I started spending more time with Julia, I would often do derivations on paper for more experimental work, and then implement it using TF/PyTorch in Python from scratch, reading other code where I could for some help. In Julia I can import a library and I'm ready to go. It feels just like working with math itself.

Julia also lets you compose long trains of operators together. That also helps when I'm doing long sessions in the REPL exploring data. It lets me define a few functions in a file, import the file, and just pipe (Julia has a pipe operator which is just function application) output between functions to plot data or calculate errors.

Moreover Julia is a lot more performant for REPL work than Python. In Python I'll usually work with a subset of a dataset to get an idea for it, then run a file with code to process the entire dataset. In Julia I can often prototype and run the algorithm in the REPL itself.

I also want to stress that Julia is quite a bit faster for REPL development, both in terms of raw speed

  function julia_blues(bummers...)

The stupidest, (stupidest in the sense that I really wish they didn't bother me, because they're silly things!), things that bother me are:

1. The 'end' keyword. It's everywhere! Loops, if-else statements, function bodies. I mean I understand why it's helpful for parsing and providing more info to the compiler, but honestly, I would've rather we just stuck with curly braces '{}'! At least it's fewer characters to type and it just feels less crowded on screen while reading. It feels like such a petty complaint, but honestly, it feels like I'm writing Pascal or Matlab all over again. Which leads me to my second point.

2. The default choice of 1 based indexing. I'm not going to go into it, because plenty of people before me have beat this dead horse already[1][2][3], but I can't help but be saddened by this choice and its implications. It's important to acknowledge the fact that Julia started as a competitor to Matlab and Octav, so it makes sense from that perspective. However, it could have been a much more general purpose language, with huge performance benefits over popular interpreted languages like Python, JS, and Ruby. It could have been a unifying force in the scientific computing community, bridging the gap between R and Python users with a greenfield approach that would have been a force to be reckoned with. Instead, rightly or not, it's viewed largely as 'just' a matlab replacement.

Now, regardless of whether 1 or 0 based indexing is truly 'better' or the 'end' keyword is no big deal, the reality is that there's a huge demographic of potential users that won't buy into Julia, because it doesn't quite feel as ergonomic as Python/JS/Ruby and won't take it seriously as a general purpose language, because it looks/feels like Matlab and they only use 'real' programming languages. Again, I'm not saying this is right, but it is the reality we're faced with and it just feels like a huge missed opportunity and bums me out.

  end

I share your frustration with the `end` keyword - it's just needlessly verbose, and for something used so frequently it makes sense to be as terse as possible.

I have some similar quibbles with Rust syntax: I know it's a minor issue, but I'm really disappointed that snake_case was adopted. It's just ergonomically inferior to camelCase in every measure. It's one more character to type for every word, and on top of that, on a US keyboard, the underscore character is a pinky key way far away from the center of the keyboard. Making this one of the most frequently typed characters in the language makes no sense.

Keyboard remapping seems like an extreme solution, and I don't want to train my fingers in such a way that when I sit down at a different workstation that doesn't have my .vimrc I can't type rust anymore.

You don't need snake_case for contrast. You can use lowerCamel and UpperCamel to denote the same levels of significance. SCREAMING_SNAKE is fine for constants because they don't get used all that often, but rust maps the hardest-to-type case to the most-frequently-used, which in my opinion is an ergonomic failure.

Yeah someone always wants to capitalize abbreviations/acronyms weirdly, or treat camel case like it was title case in handling short words or...

Getting people to be consistent with snake case is much easier, for some reason.

It makes a huge difference, so much so that using camelCase feels like a huge drag.

Qwerty is frankly awful for programming. I encourage everybody to look into alternate layouts.

`end` marks of blocks of code much more clearly than curly braces. It is also requires fewer keyboard taps. To type {} requires holding down four keys in total (shift-[ twice). end is just three key strokes.

But more importantly is saves curly braces for other uses where it is more needed.

Also braces have a nice symmetry, which is also convenient for parsers and tooling to count opens and closes.

I also think "end" is just more visually noisy. Braces are a symbol, so it's easy to filter them out when you're looking at code, but a trail of "end"s takes up more real-estate than is semantically justified.

In julia, f!(types)(parameter) means that f! is a mutating function acting on types which returns a closure which is then called on parameter.

We just write Type{paramemer}

It seems odd to say "we just write" for some arbitrary choice with its own tradeoffs; {} mean a set in Python or a literal array initialiser in C# or a scriptblock in PowerShell or a dfn in Dyalog APL or a JSON dictionary, etc. With only a limited set of whatever symbols happened to end up on a US ASCII keyboard 40+ years ago, there is heavy competition for them, it's why J broke the symmetry of pairs and went for two-character symbols, why PowerShell went for comparators like -gt and -lt, why C# has => which is completely different from <=, why << is not a double application of less-than but is bit shifting ...

To act like there's an objectively better use for {} is to miss how many tradeoffs there are, and how many things people want symbols for out of a standard keyboard.

My point was just that we find Type{parameter} to be nice pleasing syntax, and so reserved the curly braces for that.

Of course it’s not objectively better, that’s a dumb attitude towards syntax.

This is okay in a static language where there are special slots for a type where only a restricted subset of things can happen. You'd know at the parser level that the < and > are referring to the type meaning or the operator meaning.

In julia, types are values and values can live in types, and arbtrary operations can happen in a type.

E.g. I can write Tuple{1 < 2} which just becomes the type Tuple{true}, i.e. a Tuple Type with paramemter True.

Seems to me that "end" is one of the worse available choices; if we're going for keywords, why not make them "endif", "endfor", "endwhile" (or "wend") which makes it explicitly clear what each of:

        end
      end
    end

Re: 0-indexing vs 1-indexing. If you use 0-indexing, you turn off a lot of non-engineering scientific programmers. My personal experience is that 0-indexing is better for more engineering applications, while 1-indexing is better for math. I'm a weirdo in that I don't seem to mind either one though.

I think you mean "non-CS engineers". CS is a minuscule branch of engineering. Plenty of chemical, mechanical, civil (and so on) engineers had their whole education doing maths and programming with 1-indexing.

> I'm a weirdo in that I don't seem to mind either one though.

You are not, as an outsider this is one of the less appealing parts of practical CS, endless bickering about non-substantive issues which most of the time boil down to a matter of personal preference (see also tabs vs spaces, vim vs emacs, react vs vue, golang vs rust and on and on and on...)

https://docs.julialang.org/en/v1/devdocs/offset-arrays/

Subsequent languages don’t have this issue, and when you think about it, it causes a weird disconnect: 1-based indexing corresponds exactly with how you think about and see elements in a collection. 0 based indexing-despite how comfortable one gets with it requires additional arithmetic.

It’s a relic from the past, and unless you’re doing the very specific thing of indexing via offset instead of position, there’s no reason to hang onto this anachronism.

Also this is how you talk normally. You don't talk about the zeroth-column or zeroth-row in daily speech. You talk about first column and first row.

Only reason 0 based indexing make sense to me is because I began programming as a teenagers and was forced to get accustomed to it. But I remember struggling with it. Yes when working with memory, pointers etc it is more elegant. But if you are not, then I think 1-based indexing looks better.

I believe what it really comes down to is that, just as zero-based really is more natural for offsets (zero offset means the first position, great for pointers), one based really is more natural for counting [and _sets_] (when your index is n that means you have counted n elements up until now).

I love experimenting with programming languages but I don’t like recommending languages just as I don’t like recommending movies: I figure that everyone has their own tastes, and that is a good thing.