Automated ML and Cognitive Services in Power BI enables compelling new analytics scenarios – with cognitive services you can perform sentiment analysis, pull key phrases from unstructured text, tag images with their content and more! And with Automated ML you can build your own binary prediction, classification, and regression machine learning models on your historical data to make predictions on new data.
But first you have to turn it on. That’s what this post is about.
Why do I have to turn it on? Why doesn’t everything just work?
One of the goals of premium capacity is to give customers a fine-grained control of exactly what is run in each capacity. Each capacity has dedicated resources assigned, is isolated from others, and can only run generally available workloads by default. So if there is a mission critical app running in a capacity, administrators are able to constrain other functionality to get the performance they need. For workloads in preview to run in a capacity, the capacity administrator needs to enable them and configure the maximum memory percentage available to it.
Power BI AI is a capacity workload and must be enabled before the features can be used. You can learn more how Power BI capacities function here. Note that this capacity workload includes cognitive services and Automated ML. All other AI features such as Key Influencers and Quick Insights do not require a premium subscription.
You must be the Power BI Administrator or capacity administrator enable the Power BI AI capacity workload. Select the gear in the top right panel and click on “Admin Portal”:
Then click on the Capacity Settings on the left-hand panel on the subsequent page:
Click on the capacity where the workload is to be enabled. In the page above, it is “AICapacity1”:
Note that you can assign workspaces to this capacity after it is configured the way you want.
Then click “Workloads” under “More Options” and turn both Dataflows and AI (Preview) to On and set the maximum memory:
Automated ML performs numerous AI calculations in memory. While the memory needed depends on data size and content, it is a good idea to give a liberal amount of memory for the AI workload. Our own testing was on 5GB max memory or higher (or 100% on an A2 node).
The impact of enabling each workload is that capacity is now shared – memory used for one workload may not be available for others.
After this is applied, cognitive services and automated ML will be available in all of the workspaces in the selected premium capacity.