Hacker News Reader

Comment by rezonant | Hacker News Reader

I believe the purpose is for devs who are deploying to Linux, their toolchain may not fully work in Windows, but they want to have a similar dev/test env within Windows... pretty much the whole point of Windows Subsystem for Linux (WSL). It's not that the frameworks don't work on Windows, it's the deployment tech, like Docker, Ansible, their build scripts etc etc. Some users may be doing something totally custom on top of the GPU without a framework layer but that is probably very few users. I'd have to guess though that those users would be the type of users that would get MS to go through the effort though...

View on HN

Do people actually use Docker and Ansible a lot for ML???

Doing something custom on top of the GPU is also not much different on Windows, than Linux. CUDA is basically the same. OpenCL and Vulkan are available too.

I'd like to hear perspective of a person, who actually does ML specifically on Linux for some reason.

Well CUDA is only one thing. Especially in the python environment it's only a matter for time before you run in some wierd dependency issue that is Windows only. For example, getting XGBoost completely up and running on Windows requires you to either build it yourself or download a .dll from a university link. Installing it on Linux is just a proper pip install.

Also windows not having a build in C compiler makes you dependent on the horribly convoluted Visual Basic stack that seems to have a lot of dependencies for some python ML libraries. Docker makes it a lot better to run and I almost always deploy in a docker container because the ML modules I deliver are often interacted with as a black box with a REST API on top.

It looks like Anaconda supports XGBoost on Windows.

You might be right about C compiler. But something itched when you mentioned Docker. Could getting Windows SDK installed be harder, than installing Docker?

I just use chocolatey for that > choco install windows-sdk-10.0

That being said I had to avoid installing VS 2019 for quite awhile because Node.js native module build chain couldn't work with it. There are complexities

I do all of my development on linux, if I can, but to be honest the GPU support is generally better on windows because that seems to be the main platform AMD and NVIDIA target - though linux support is not too bad. GPU support is the only potential benefit in using windows that I can think of though. Everything from package management, to build tools, FOSS support, community, troubleshooting, etc. is generally better on linux.

Docker at this point is InstallShield infrastructure for startup production quality freewares

It is being used both for training and inference in order to package dependencies and scale up.

Just curious: what kind of models are you training, that require scale-up using Docker/Ansible/Kubernetes?

Things like filename limits and command length limits seem to get hit on windows way more frequently, making all code more fragile in general, and a million other little things.

Even popular libraries like zeromq don't support namedpipes on windows because of how complicated they are and how different to everywhere else.

Just determining what visual studio version is installed seems to trip up projects all the time.

Mostly doing several kinds of NLP: My actual setup is a Windows Laptop to SSH into Linux machines w/ tmux session. However, I really appreciate WSL for working offline, etc.

My main reason: It is the most convenient way to have Unix tools (grep/sort/cut/sed/less/...) and bash available. Cygwin always was a pain, MinGW / GitBash felt much better, but ultimately WSL just feels best.

These tools are incredibly valuable to my workflow. Sure, stuff like pandas can be nice for small datasets, and some data sits in some DB/Kafka/distributed system. But there have been countless cases where unix tools allowed me to take xxGB zpfiles of text and do basic examination or even build baseline models within a few hours.

Sure, there always are alternatives to use these tools and there are many equivalents. But I would always prefer WSL + conda for Linux to a typical "Windows Conda" installation with that weird GUI and the need to install so many different applications to even just look into the first or last few lines of a huge textfile.

EDIT: That said, of course I can/could always just run a juypter notebook under windows using windows cuda + GPU and share files with a WSL bash where I do my modifications. But again, everything within the same systems just feels better (ipython shell magic, no worries about if paths to the same file are really identical, etc) and while this is by no means a game-changer, it is just nicer that way.

The reason to use Linux is to follow the crowd. I use Nvidia's docker container because the plurality of devs use it. Over the course of my career I've found that well-trodden paths tend to have a _lot_ fewer bugs thanks to other people finding them first, and when I do get some

https://github.com/pytorch/pytorch/issues/37790

weird

https://github.com/pytorch/pytorch/issues/32575

bug

https://github.com/pytorch/pytorch/issues/25301

I don't have to spend time explaining or justifying or isolating my setup.

Conversely though, my work is itself off the beaten path enough that I'm likely to run into weird bugs. If I was pushing images through a CNN, that'd be well-trodden enough on every platform that I'd be a lot less fussed about which particular platform I use.

Reply on news.ycombinator.com