I wish Nvidia would put more work into making sure that its stable software packages work together with other stable software. I have a laptop that's more than 1 year old with tensor cores (RTX 2070), but it's still not supported on stable Windows.

I had to install dev channel MS Windows + subsystem for linux 2 + NVidia CUDA 10.2 (an old version of the NVIDIA driver) to be able to run mixed precision training on my NVIDIA card.

I could try to run Linux, but in that case I may not be able to use the newest games that are created for the same NVIDIA card.

The situation is so bad that Jeremy Howard from Fast.AI suggests people to run training models on the cloud even if they put significant amount of money into having their own NVIDIA cards.

These are problems of third parties, not Nvidia.

Every game that runs on Linux gets full graphics acceleration.

As for Windows, PyTorch AMP allows for mixed precision training on native windows through conda, no apex needed. And of course, the same APIs from Nvidia are available on both.

Thanks, it's great to know that I can use PyTorch AMP for mixed precision training, I was following NVIDIA's documentation, and this wasn't explained.

You could always dual boot.

There’s a lot of words and vendors here I’ve never heard of. Is this a legit rundown of a growing field, or is this Nvidia pitching me all their partners?

The lifecycle diagram is legit.

As for the vendors, Nvidia is just promoting their own services or their partners.

But there aren't any established players yet. It's still a nascent field.

ML Ops is a rapidly growing corner of the AI landscape: https://www.cognilytica.com/2020/04/01/ai-today-podcast-135-...

Thanks - sounds like a good area to google around in, but not to take the blog too seriously yet.

Sounds about right to me.

It's a bit of both. Well, it's definitely an advertising piece but parts of it are somewhat accurate about MLOps.

Working for a devops consultancy and tooling conpany: this is definitely something we have been keeping our eyes on for the last couple of years. My personal opinion is that our local market might be a bit too small to invest heavily in this but I’m glad some of the big names can get benefits out of it.

It looks like the "Data Fixes" section could introduce bias. "Select the right data" reminds me of researchers cherry-picking data for papers.

The purpose of that step is exactly the opposite: to use data that represents (as much as possible) the full generality of the space you seek to model. And on a more practical level, there are datasets that can't be used for commercial purposes but are useful for research (a well known but not the best example would be ImageNet).

"Data Fixes" might also correct for bias. Data often contains errors and inconsistencies, especially if it's the output of some other automation.

"Select the right data" just means don't try to do something silly like predict 2021 housing prices in Ann Arbor using historical data for Pittsburgh from 1980-2007.

for people who work on real production ML. what kind of data pipeline are you using before the MODELING part of it?