Difference between Azure ML and Azure ML experimen

Question

I am new to Azure ML. I am having some doubts .Could anyone please clarify my doubts listed below.

1. What is the difference between Azure ML service Azure ML experimentation service.
2. What is the difference between Azure ML workbench and Azure ML Studio.
3. I want to use azure ML Experimentation service for building few models and creating web API's. Is it possible to do the same with ML studio.
4. And also ML Experimentation service requires me to have a docker for windows installed for creating web services. Can i create web services without using docker?

Answer 1

I'll do my best to answer these questions and feel free to ask more questions. :)

What is the difference between Azure ML service Azure ML experimentation service?

Essentially, Azure ML Service (I may reference this as Azure ML Studio) uses a drag and drop interface to build out your workflow and test models. Azure ML experimentation is a new offering from the Azure Portal to host them directly in Azure and offer a better way to manage your models. Experimentation will use Azure ML Workbench to build out your models.

What is the difference between Azure ML workbench and Azure ML Studio?

The biggest difference is ML Studio has the drag and drop interface to build the workflow and models, whereas Workbench lets you use Python to programmatically build out your models. Workbench also includes a really nice and powerful way to clean your data from the app. In Studio you have some good modules to clean data, but I don't think it's as powerful as what you can do in Workbench.

EDIT: The Workbench application is deprecated and has been replaced by/upgraded to ML Services. The core functionality is unchanged, though.

I want to use azure ML Experimentation service for building few models and creating web API's. Is it possible to do the same with ML studio?

I would actually say it's much easier to do this in ML Studio. The drag and drop interface is very intuitive and it is only a couple of clicks to create a web API to call your model. I feel, as it is currently at the time of this writing, is more complex to deploy your model and it involves using the Azure CLI.

And also ML Experimentation service requires me to have a docker for windows installed for creating web services. Can I create web services without using docker?

Here I'm not too familiar with the Docker parts of Workbench, but I believe you can create and deploy without using Docker. It will require an Azure Model Management resource, though, I believe.

I hope this helps and, again, feel free to ask more questions.

Answer 2

1. The AML Experimentation is one of our many new ML offerings, including data preparation, experimentation, model management, and operationalization. Workbench is a PREVIEW product that provides a GUI for some of these services. But it is just a installer/wrapper for the CLI that is needed to run. The services are Spark and Python based. Other Python frameworks will work, and you can get a little hacky to call Java/Scala from Python. Not really sure what you mean by an "Azure ML Service", perhaps you are referring to the operationalization service I mentioned above. This will quickly let you create new Python based APIs using Docker containers, and will connect with the model management account to keep track of the linage between your models and your services. All services here are still in preview and may breaking change before GA release.

2. Azure ML Studio is an older product that is perhaps simpler for some(myself an engineer not a data scientist). It offers a drag and drop experience, but is limited in it's data size to about 10G. This product is GA.

3. It is, but you need smaller data sizes, and the job flow is not spark based. I use this to do rapid PoC's. Also you will less control over the scalability of your scoring (batch or real time), because it is PaaS, compared to the newer service which is more IaaS. I would recommend looking at the new service instead of studio for most use cases.

4. The web services are completely based on Docker. Needing docker for experimentation is more about running things locally, which I myself rarely do. But, for the real time service, everything you package is placed into a docker container so it can be deployed to an ACS cluster.