[marmaduke]

I’m not insisting I have the only use case, but apart from the examples of traversing language boundaries, I haven’t see a good example of what’s so painful in Numba. What is so challenging that is requires code generation?

[Certhas]

I have a data structure based on which I generate a dynamical behaviour that I want to integrate. So I construct a rhs.

I further want the user of the library be able to pass it new functions that can be integrated into overall dynamical behaviour.

There are different ways to achieve this, the simplest version is with closures. Pass a list of functions, and some parameters and I construct a right hand side function from it. Unfortunately this does not work with numba. What I ended up doing is passing not the function itself but the function text to generate the code of the function to be jited and then eval that. It worked but it was horrible to maintain, and required users to pass function bodies as text witha very specific format.

Now in Julia we will probably eventually transition to a macro based approach, but the simple closure based model just worked.

Previously I had large scale, inhomogeneous right hand side functions that I wanted to jit in numba and that need sparse matrices. So I ended up having to implement sparse matrix algorithms by hand because I can't call scipy.sparse.

Another instance: I implemented a solver for stochastic differential equations with algebraic constraints in numba, partly to be able to use it with numba jited functions and get a complete compiled solver out of it. This already constrained my users to use numba compatible code in their right hand side functions.

In order to get this to work I had to implement a non-linear solver from scratch in numba rather than being able to use scipys excellent set of solvers.

Julia is not a magic silver bullet. Getting the ODE solvers to make full use of sparsity still requires some care and attention. But I simply spend a lot less time on bullshit than before. (so I have more time to spend on HackerNews :P)

I decided to switch over when for one paper I was able to implement a problem using the standard tools and packages available in Julia within half a day. The Python equivalent would have involved using a new library that came with its own DSL, which would have meant rewriting quite a bit of my code to take advantage of it. Easily several days work.

With DifferentialEquations.jl I also could just test half a dozen different numerical algorithms on a problem in a matter of minutes, find out which performed best and use that for MonteCarlo. Saved about a week of computation time on one project alone. That's not a critical amount, nobody cares if the paper comes out a week later or earlier, but it's nice (and I don't waste super computer time). With Python libraries with different DSLs this would have taken considerably longer, and I probably would not have done it. This is the result of having one library and interface rather than a whole bunch, if everyone agreed on scipys ode interface (which just got properly established in scipy 1.0.0) this would be easy in Python as well. But that's also the point that people have been making: Julias design for composition over inheritance makes it convenient to rally around one base package.

I also personally very much like being able to enforce types when I want to. This is a big win for bigger projects for us.

[marmaduke]

> With DifferentialEquations.jl I also could just

yep... I took a look at the DE packages in Julia today, and quite frankly they're much better than the situation in Python, perhaps because of one or more prolific applied mathematicians are making a concerted effort, which is lacking Python? I dunno, but I did recommend my colleagues look at Julia for DEs, for this reason.

That said,

> Pass a list of functions, and some parameters and I construct a right hand side function from it. Unfortunately this does not work with numba.

I'm pretty sure I've done this before with numba, so maybe getting concrete would help, e.g. an Euler step

    def euler(f, dt, nopython=False):
        @numba.jit(nopython=nopython)
        def step(x):
            return x + dt * f(x)

where user can provide regular Python function or a @numba.jit'd function. If a @numba.jit'd function is provided, and nopython=True, this should result in fused machine code. This sort of code gen through nest functions can be done repeatedly for e.g. the time stepping loop.

I've done this for CPU & GPU code for a complex model space (10 neural mass models, 8 coupling functions, 2 integration schemes, N output functions, ...) which, by the above pattern, results in flexible but fast code.

Is this a pattern that captures your use case or not yet?

> implement sparse matrix algorithms by hand because I can't call scipy.sparse.

agreed, this is a surprising omission, which I attribute to not much of the numerical Python community making use of Numba, but could be fixed rapidly.

> constrained my users to use numba compatible code in their right hand side functions

what did you run into that was problematic?

> I had to implement a non-linear solver from scratch in numba rather than being able to use scipys excellent set of solvers

I didn't follow; passing @numba.jit'd functions to scipy is in the Numba guide, so what exactly didn't work?

[Certhas]

This pattern is how I wrote the SDE solver in Python. That works great and is really useful and the reason why I teach closures.

The library we're building now though does something different. Something like this:

  def network_rhs(fs, Network)
    def rhs(y,t)
      y2 = np.dot(Network, y)
      r = empty_like(y)
      for i, f in enumerate(fs):
        r[i] = f(y2[i])
      return r
  return rhs

> what did you run into that was problematic?

For more complex model building the right hand side functions actually make use of fairly complex class hierarchies. That was the major stumbling block. But people also were using dictionaries and other non-numpy data structures and just generally idiomatic Python that is not always supported. Some of that stuff is inherently slow/bad design of course, but it still ended up killing the use of my solver for this project.

They are now rewriting in C++, which is absolutely a great choice for their case (and probably would have been viable for us too if we had had more people with a C/C++ background in the team).

> passing @numba.jit'd functions to scipy is in the Numba guide

I wanted to use scipy.root from numba. Not the other way around.

Now if all of the numerical Python community was standardized on numba, a lot of this would not be an issue. Scipys LowLevelCallable is a great step in the right direction. But fundamentally I don't see how you will ever get the different libraries to play together nicely in a performant way. It would require every API to expose numba jitable functions. Last I checked, the only functions you could call from within numba code were other numba functions and the handful of numpy features the numba authors implemented themselves (I remember waiting for dot and inv support). If I have an algorithm by a student implemented on a networkx graph as a data structure I can't just jit that. In Julia it automatically is.

[marmaduke]

I see what you mean. I’ve done exactly that sort of thing in C with an array of function pointers, but I’m not sure it would work in Numba.

The churn is exhausting but I see the merit of starting over and getting everything done in a fully fledged JITd language.

[Certhas]

As said in another part of the thread, we were in a very good spot to do so. There absolutely are many reasons to not jump in at this time.

[mlthoughts2018]

This just sounds like bad software designto me. You are miswanting something overly generic that’s super not needed, and regardless of implementing in any given language, it sounds like it would benefit hugely from taking a more YAGNI approach to it, restricting its genericity based on likely usage (not intended or imagined usage), and either just manually writing stuff for an exhaustive set of use cases, or code genning just those cases and not allowing or encouraging arbitrary code gen of possible other cases.

I love it when libraries limit what can be done with them and document an extremely specific scope they apply to.

When libraries try to be all things to all people, it’s bad. A sophisticatedcode gen tool that enables library authors to choose to do that is a bad thing, not a good thing.

[Certhas]

You don't know my use case, and you are not right. I have a network of heterogeneous interacting nodes with quite different dynamics on the nodes. I pay great attention to YAGNI, and constantly tell my students and colleagues to cut enerality and work from the specific case outward. But this is just essential complexity of the problem domain. I've spent years implementing the concrete cases, I know what research we couldn't and didn't do because it was to painful to do by hand, and this is the minimum level of generality I can get away with.

I have ideas for a more general library of course, :P But I'm not spending time on them.