komuher

Nim have Arraymancer https://github.com/mratsim/Arraymancer sure it isn't easy but still most of top dogs have some libs supporting BLAS/CUDA etc.

Ok let me get it here cause there are a lot of misunderstanding my post on top.

Swift have built in simd same about Rust https://github.com/apple/swift-evolution/blob/master/proposa...

Nim have arraymancer https://github.com/mratsim/Arraymancer

@inbounds it is just compiler option u can use it in any language with LLVM backend i would be suprised if Julia will be faster then Rust/Swift or Nim in this regard. But True about GO in that particular case.

Ok my fault i was writing from perspective of ML engineer not reasercher (I'm using Julia for 1.5 year now and my bois reaserchers prefer pure julia solutions cause its easier to write u can use symbols and not using OOP etc.)

But for production ready models PyTorch and TF is miles ahead first of all: NLP, audio and vision based packages building frameworks, (attention layers, vocoders etc.) then u have option to compile models using XLA and use TPU (about 2/3 times cheaper then gpu for most of our models [audio and nlp])

Next inference performance (dunno about now maybe this change but about ~8 months ago flux was about 15-20% time slower [tested on VGG and Resnet's]then pytorch 1.0 without XLA)

Time to make it to production: Sure maybe writing model from scratch can take a bit longer on PyTorch then Flux (if u not using build in torch layers) but getting in into production is a lot faster, first of all u can compile model (something not possible in Flux) and u can just use it anywhere from Azure and AWS to GCP and Alibaba Cloud make a rest api using Flask/Fast-api etc. or just using ONNX.

Dont get me wrong i love Julia and Flux but there is still a LONG way before most people can even consider using Flux on production enviroment not for reasearch or some MVP stuff.

How much LOC did u guys have at the moment, and how much data are u monitoring? For smaller problems Julia was working perfect for us.

For numeric code sure but its because Julia are using BLAS (or any other instruction for CPU/GPU u give to LLVM). Julia (no BLAS) -> In matmul its on pair with Golang and Swift and a bit slower then RUST and Nim. If u need BLAS then just use lib with that :)

Java is good for big projects but not for Machine Learning purposes (Also im C# fanboy so still will prefer C# over Java :D)

Flux but its not even close to PyTorch or TF in terms of features and performance

Something about Julia from me. I was big Julia fan using it for last 1.5 year and in my company we are using Julia for data preprocessing and parsing hundred of GB's of data. (We also use Python for ML and start to moving julia code to Nim for data prepro) If u consider Julia as ur next language dont use it to any medium to big project outside pure scientific purposes it just isnt ready yet. We have about 3.5k loc in Julia at the moment for our preprocessing service and as much problems as we have with GC in Julia i never saw in any other language: random crashesh, memory leaks, big slowdowns with long running processes and there is a lot of topic on julia forum about that still not solved. Julia is beautiful and easy to use but it is nowhere near stable and/or fast as any other "new" languages (RUST, Nim, Golang, even Swift). And if u consider using Julia for Deep Learning then best performance and stability in this order are: Pytorch, Tensorflow/MXNet/(JAX), Nothing, Flux. Its a long way to go for Flux before it'll be even close in usability considering other flag frameworks.

Kinda biased benchmarks u just took best of crystal and worst of nim from this xd -> https://github.com/kostya/benchmarks