Lost in Azure Machine Learning, part I

I have been working with ML Studio for a while now, 18 months to be precise. Not that long of course, but enough to get a good idea of what the product is all about. I know there are people who do not like it and some are huge fans of the tool. I am a bit in the middle, see my article here. One of the things that bothered me the most was the lack of clear strategy where MS was heading to in their tool offering. If you look at my early articles (beginning of 2016) it seemed like there were many products available, but when going deeper into it it was not always clear what was offered or the products were not fully ready yet or they were basically an “MS port” from the open source world on a pre-configured Azure VM.

Not so with ML Studio, since the it was always clear what this product was meant to do. And it was quite useful.

Recently (end September 2017) Microsoft made some interesting changes in their ML offering, putting it all under a single umbrella called … “Machine Learning” or is it “Machine Learning Studio” 😉MLStudio_or_not

Anyway, there are 3 different products now in the ML catalog:

  1. Azure Machine Learning Services
  2. Azure Machine Learning Studio
  3. Data Science Virtual Machine

At first sight it ML Studio and the DS VMs are the same as before. I haven’t examined them in detail, but at first sight these stayed the same. At, least my older experiments still run, so I don’t expect major changes. I will probably go deeper into this at a later time.

The new kid on the block is the Azure Machine Learning Services. I will go into more detail in a next article but here already a short overview:

  1. Experimentation service: the idea behind this is that you build your model (prototype they call it) on your desktop (!) and then run it in a cloud based environment.
  2. Model management: this a mechanism to deploy the models you created and optimized to Docker containers.
  3. Workbench: is a desktop application that focuses on data preparation, a very important and labor intensive phase in data science.
  4. Visual Studio Code Tools for AI.
  5. MMLSpark.

I am really curious for the Experimentation Service and the Model Management offering.  Not sure how this will all work together with AML Studio:

  • Will we be able to port AML Studio models to and from the Experimentation and Model Management,
  • How far will the git integration go? Will we be able to version/fork experiments?
  • How about security and accessing our models ?

Expect my findings soon…