How to enable ICE and use What-If on Databricks?

[FMalnegro]

Hi!

Been trying to explore the framework and wanted to try out and further understand perturbations and what-ifs. But upon rendering the ExplanationDashboard, there's a note below stating that an operationalized model is needed to make predictions on hypothetical datapoints for what-ifs. How do we proceed from here?

Thanks!

[imatiach-msft]

@FMalnegro what platform are you trying this on? Is this on a VM or on local compute? There are some issues with what if analysis on remote compute because the browser needs to be able to send POST requests to the remote compute which runs the flask service. We've been trying to resolve this on different platforms one at a time.

[FMalnegro]

Hi, @imatiach-msft

Thanks for your response. I'm on Databricks 7.0 ML. Are there workarounds for this type of setup as of the moment?

[imatiach-msft]

@FMalnegro unfortunately Databricks is the most difficult environment of all, because the driver/worker nodes are under their own virtual network, and for WhatIf analysis we basically need to call predict on the model using the python code. I created an issue on databricks forum several days ago:
https://forums.databricks.com/questions/42981/call-flask-app-on-azure-databricks.html
As of right now I'm not sure this is even possible in Databricks.

[FMalnegro]

Thanks for the insight, @imatiach-msft

Your description does clarify how difficult it is to make this work in Databricks. In the meantime, I will also follow the link that you have shared for any updates on the matter.

[darrenyu1111]

@imatiach-msft in azure ML service - notebook(jupyter) on a computer instance(VM STANDARD_DS3_V2 ) it seems doesn't work.

[imatiach-msft]

@darrenyu1111 this should have been just fixed with the latest release 0.14.3 (what if and ICE plots on compute instance in AzureML)

[darrenyu1111]

@imatiach-msft , thanks. it works.