Take a look at Julia's Flux.jl. It has a really nice API and is quite intuitive to use at a low level, and has higher level components as well.

Julia has a fast maturing data wrangling super-project (Queryverse).

I'm guessing you'd have (slightly) fewer deployment options with Julia though.

100% agree, and there are a number of efforts in the space. mlpack (https://www.github.com/mlpack/mlpack/), Shogun (https://www.shogun-toolbox.org/), and Shark (https://www.shark-ml.org/) are three that have been around for over a decade now. They're a little niche because C++ is not that popular for data science, but they are generally pretty fast (especially mlpack, which focuses on speed).

Absolutely. In 2010 I used FANN for my Msc thesis research, and found it pretty easy to make my own little training sim for stock price data on top of it. Haven't done any ML work since, but I always scratched my head over how Python became the most popular language for this domain.

stock price ml? why aren’t you rich?

Who says they're not...

In general, I think languages with static typing are preferable.

C++ seems ideal for me right now because it is the only other language with a somewhat mature stack (perhaps Julia as well, but I haven't played too much around with that).

Julia is dynamically typed FYI. However, I find that usually when people say they want a statically typed language, they don't actually want static typing (which is just a restriction of language semantics), but instead they want a language with a powerful type system that does work for them.

In this case, julia absolutely is worth checking out. It does static analysis on it's intermediate representation to automatically identify and isolate statically inferrable regions of programs and then stitches them together if it ever encounters any dynamism.

Julia's type system is extremely powerful and expressive and can do a lot of things that are incredibly difficult in fully static languages precisely because it does allow dynamism.

In my opinion, no, it makes no sense to write ML in C++ :

- Python allows for higher level description of algorithms, which means researchers can focus more on the ML stuff and less on low level details.

- There is no performance gain in going from Python to C++, because in both cases the models are compiled to specific binary formats to be executed on dedicated hardwares. TensorFlow enables accelerators not only for training, but also for data transformations and preprocessing.

You will see Modern C++ more like Python these days. It has really neat stuff. A comment is not enough to describe this.

> There is no performance gain in going from Python to C++

The backend of TF/PyTorch is written in C++ anyway, so the more complex the model, the less time it needs to spend in the glue code (frontend) that is written in Python. Therefore, rewriting complex models in full C++, for example by using TF/PyTorch C++ API, probably won't much improve the performance.

In this paper the author rewrites some ML models in Rust using tch-rs (Rust binding for PyTorch C++ API) and finds the performance not that much better (even some models perform worse):

https://www.aclweb.org/anthology/2020.nlposs-1.4/

yes. especially with AI on the edge.