Machine learning as a service
The IT group where I work has been investigating moving more of what we do to the cloud. So, I’ve been checking out the costs and services provided by AWS and Microsoft Azure. There seems to be two big ways to think about moving model training (and other work) to cloud services:
Reproduce the hardware you’re currently using, or better.
Focus on your ultimate output and find a service that makes it.
Reproducing my current hardware resources in the cloud seems to have a higher 3-year cost than just buying my local machines. However, our IT director suggested I think about what my outputs are and instead try to price services that give those. I found this article comparing the main available machine learning as a service options. Utlimately, I’m underwhelmed.
Both Amazon SageMaker and Azure’s Machine Learning Studio sound promising — especially the AutoML aspects available in SageMaker. Maybe I don’t need to think about model selection and hyperparameter tuning. Perhaps AWS can think about it for me. But it looks like there are really only three classifier models in SageMaker — a generalized linear model, XGBoost and NLPs. If you work with images, there are a few more options. If you aren’t doing AutoML, then SageMaker just seems like a jupyter notebook in the cloud. So, I’m back to thinking about what the hardware needs are on the machine running that notebook. Bleh!
For Azure’s Machine Learning Studio, I find the no or low-code aspect cool. But ultimately, I know enough about what I’m doing to want the flexibility of specifying a little more than their GUI allows. So now I’m back to basically wanting Python interface of some kind and thinking about the hardware around that. Bleh!
Ultimately, I think these products are targeted either at people who want machine learning without code (Azure’s Machine Learning Studio) or just want their code in the cloud (AWS’s Sage Maker). Where’s the product for people who can code but want to worry about outputs instead of pricing hardware requirements?