Hacker News Reader

Comment by komuher | Hacker News Reader

What purpose did you use Julia for? Julia was hyped like 2 years ago but right now i didn't see any new cool lib or even a blog posts and DL libs are lagging hard compared to python DL libs (torch, tf, jax etc.).

View on HN

Julia was hyped, and now is used in production by several well known institutions.

https://juliacomputing.com/case-studies/

Most of those libraries are Python bindings to libraries written in Fortran, C, and C++, whereas on Julia's case many libraries are pure Julia code.

It was hyped and probably are used in even more places (in my previous work i even wrote a full Julia based parsers etc.) But still i didn't see much advantage of writing in Julia now 2 years and even maybe 1 year ago there were few (before numba and other JIT python stuff was working fine). Right now for me its just l'art pour l'ar you almost never get performance like in C/Nim/RUST and you don't even have much good libs in the community (compared to python thier main rival).

As someone who uses TF/PyTorch/JAX and numba for their day-to-day DL work, I've been watching Julia's post-hype developments in this area intently. Julia has probably the most vibrant ML ecosystem of all the non-Python languages (remind me when someone has implemented a transformers library for Arraymancer [1]), and absolutely smokes Python in certain niches (e.g. neural ODEs and more under the SciML [2] umbrella, I'd argue Python is "lagging hard" there). Numba can work for speeding up simple operations, but is rife with limitations that prevent optimizing more idiomatic Python code and is an absolute pain to debug. Likewise, trying to decipher or even catch errors from deep in the C++ bowels of most modern DL frameworks is an exercise of frustration.

Python may still be the best language for non-cutting edge deep learning, but we ought to consider whether we've overfit our algorithms to the limitations of the current system [3].

[1] https://github.com/mratsim/Arraymancer [2] https://sciml.ai/ [3] https://news.ycombinator.com/item?id=20301619

Torch is not Python, PyTorch is. Which also adds to your point. Remember SciLua and Torch? I far prefer Lua (moonscript) to Python, but I feel like Python will always win for DL. I guess Julia for DL will meet the same fate.

Regardless of future mainstream appeal, I don't think Julia will meet the same fate as Lua in the area simply because of how simple it is to create and maintain DL libraries compared to Python. Multiple dispatch, native matrix support and high performance out of the box (plus native and composable GPU operation libraries like CUDA.jl) means anyone can write their own DL library from scratch in a few thousand lines (and there are tons of those already, even if the community converged mostly into Flux, with Knet as a second option).

Reply on news.ycombinator.com