question

LeoSabato-0219 avatar image
0 Votes"
LeoSabato-0219 asked LeoSabato-0219 commented

Real-time enpoint and AutoML models

Hi Azure community,

I am experiencing a probem with Real-time endopoint when trying to deploy models trained using AutoML.

Since deploying AutoML models directly from the portal has some problems due to recent changes in azureml-defaults==1.33.0 (see here: https://docs.microsoft.com/en-us/answers/questions/511349/auto-model-deployment-in-container-instance-no-mod.html), I found a work around and managed to deploy AutoML models to real-time-endpoint using the script in attachment (123179-automl-endpoint-deploy-workaround.txt).

All it does is to create a conda environment from a yml file (123353-conda-env-v-1-0-0.txt) to solve the issues introduced with azureml-defaults==1.33.0 and use the same entry scripts that are available after AutoML training.
This works in most cases but for ExtremeRandomTrees models. In these cases I get the following error message:

Failure while loading azureml_run_type_providers. Failed to load entrypoint automl = azureml.train.automl.run:AutoMLRun._from_run_dto with exception cannot import name 'RunType'.

Do you have any idea what the is problem here? Is there any particular package I should add to the conda enviroment?
I cannot find anything similar on the web, so I decided to ask here.

Also, these real-time-endopoints got stuck in transitioning state and I cannot simply remove them from the Portal. How could I remove them instead?

Any idea and help would be very much appreciated thanks.





azure-machine-learningazure-container-instances
· 6
5 |1600 characters needed characters left characters exceeded

Up to 10 attachments (including images) can be used with a maximum of 3.0 MiB each and 30.0 MiB total.

Hi @LeoSabato-0219, I am following up on these issues internally and our team is working to fix this soon. I will post an update as soon as we have one.
In the meantime, you could follow the steps in this thread to get your endpoint deleted as some users have already reported the same.


1 Vote 1 ·

Thanks @romungi-MSFT, looking forward to the hear back from you guys.
I am just amazed by how good AutoML is for fast MVP development (and sometimes not only MVPs!!).

Best,
Sabato

0 Votes 0 ·

@LeoSabato-0219 Is it possible to provide more details of your experiments? That is, are they related to forecasting, DNN or other scenarios? This would help to check us possible scenarios that could be broken in AutoML currently.

0 Votes 0 ·
Show more comments

0 Answers