This browser is no longer supported.
Upgrade to Microsoft Edge to take advantage of the latest features, security updates, and technical support.
![](data:image/svg+xml, %3Csvg xmlns='http://www.w3.org/2000/svg' height='64' class='font-weight-bold' style='font: 600 30.11764705882353px "SegoeUI", Arial' width='64'%3E%3Ccircle fill='hsl(35.2, 52%, 13%)' cx='32' cy='32' r='32' /%3E%3Ctext x='50%25' y='55%25' dominant-baseline='middle' text-anchor='middle' fill='%23FFF' %3ELM%3C/text%3E%3C/svg%3E)
I am trying to deploy an endpoint from the machine learning studio, but all endpoints get stuck in the transitioning state.
When looking at the container's activity log, I can see the following operations took place:
And if I select the top level failed action, I can see this error message:
However, I should have the usual permissions for the group I'm deploying in, as I've created other resources in this group before.
Am I missing a different permission which would not be needed for other resources?
Resource group: intrglmpdev00002
Subscription Id: 931c1c11-140f-4489-a457-6f4b22023b26
Workspace: LimburgsMooisteML1
While searching for a solution earlier I found this thread; https://learn.microsoft.com/en-us/answers/questions/39341/azure-ml-endpoint-stuck-in-transitioning-state.html I have also sent an email to microsoft as specified there.
EDIT: I also just noticed that the description that I entered for testdeployment3 does not show up in the endpoint specifications.
@Louis Machiels This error indicates that an audit has run after the previous step i.e after checking your container registry credentials and it seems to have failed the vulnerability assessment. Does the same error occur for other registries too? The policy must have been setup on the policy page of azure portal. Could you check if this was setup by your administrator recently?
I'm not sure if I can try other registries, this is the one it made by default when I tried to deploy an endpoint in Azure ML. I can't find an option to specify a different container registry, should there be one?
@Louis Machiels This could be any container registry. I think this may be mentioned in your environment.yml file. Do you have any documentation of the steps you are following to review the files that are uploaded?
My environment.yaml only contains the libraries that my software depends on, such as PyTorch. All the files I've uploaded are other python scripts with utility functions. The score.py file I've uploaded to handle the main work is a file I adapted and tested myself.
I'm not sure if I can inspect anything about the endpoints I'm deploying, the container registry seems to be empty.
![](data:image/svg+xml, %3Csvg xmlns='http://www.w3.org/2000/svg' height='64' class='font-weight-bold' style='font: 600 30.11764705882353px "SegoeUI", Arial' width='64'%3E%3Ccircle fill='hsl(0, 5%, 10%)' cx='32' cy='32' r='32' /%3E%3Ctext x='50%25' y='55%25' dominant-baseline='middle' text-anchor='middle' fill='%23FFF' %3EFC%3C/text%3E%3C/svg%3E)
Found below error in the logs:
Status Code: 400 BadRequest Reason Phrase: Python interpreter version must be specified in the Conda file.
Please try specifying python version in your Conda file, something like:
dependencies:
Thank you for the answer! Is there any place I can view the logs? I have tried searching Application Insights, but it doesn't seem to show any logs for me.
My previous 3 deployments show no Deployment Logs in the ml studio:
I added a python=3.8.5 requirement to the environment per your suggestion.
This seems to have made some progress, as I now see a Deployment Log entry:
(RequestId = 333db2d3-f01e-005d-2cc7-bcd5f3000000 )
I would also like to ask what I could do to solve this error. If I understand it correctly it's a problem with an SAS signature, but I expect Azure to handle authentication among Azure services that would be needed.
I found the logs in blob storage under a storage account container, azureML/ImageLogs/<guid>
The last deployment which suffered the SAS failure had an error for a dependency it couldn't find using conda. I added the proper channel specifications and tried a new deployment. This one resulted in a new log:
ERROR conda.core.link:_execute(502): An error occurred while installing package 'conda-forge::olefile-0.46-pyh9f0ad1d_1'.
FileNotFoundError(2, "No such file or directory: '/azureml-envs/azureml_0d18ae13fb0e9b1a3353c71850ecba54/bin/python3.8'")
I found the following thread which suggests it's a problem between conda 4.5.11 and Azure ML:
https://stackoverflow.com/questions/60636558/error-conda-core-link-execute502-an-error-occurred-while-installing-package
And the logs specify that anaconda 4.5.11 is being used in endpoint deployment.
I'm not sure if I can use pip instead of conda to install my dependencies?
Fixed this error by changing my environment.yml line from
- conda-forge::opencv
to
- pip:
- opencv-python
but now I run into the following error:
I fixed this and another problem by adding gunicorn and flask to my environment.yml, and changing the opencv requirement to opencv-python-headless.
Now I'm getting this error:
I will try to fix this by also adding applicationinsights to my environment.yml
Adding the applicationinsights dependency let me get on to the next error:
So far I've had to add gunicorn, flask, applicationinsights, and now it seems to require azure.ml. I'm not sure if adding all these dependencies to my own environment.yml file is the right way to solve this, as I would expect the endpoint deployment to install these dependencies autonomously. I also haven't been able to find a list of required dependencies, so I have to add one dependency, start a new deployment, wait 20 minutes for the deployment, and then see which new dependency is missing.
@Louis Machiels Do you have an entry for the azureml-sdk in your file?
dependencies:
- pip:
- azureml-sdk
I added azureml-sdk, azureml-defaults, and azureml-telemetry.
Now I run into the following error:
But this error seems to be due to the pytorch library:
https://github.com/ultralytics/yolov5/issues/22
I'm trying to deploy a .pt model using my own score.py script which uses many functions of the ultralytics repository. Sadly their way of saving and loading a model seems to require a specific folder structure, which I don't think I can replicate in the AzureML endpoint. My next idea is to export my .pt model to an ONNX model and try to run inference using that. Alternatively I could look for a tensorflow implementation of Yolov5.