Azure ML Batch Endpoint: Configuring the output

Demet Sude Saplik 36

Hello,

I created a batch endpoint in Machine Learning Workspace and I would like to configure the output path. I follow the tutorial https://learn.microsoft.com/en-us/azure/machine-learning/how-to-deploy-model-custom-output?view=azureml-api-2&tabs=python and under the part Creating a scoring script, it shows how the environment variable AZUREML_BI_OUTPUT_PATH is used to get the output bath. And in the part Creating the deployment, output_action is set to BatchDeploymentOutputAction.SUMMARY_ONLY which means that the user script will store the output. I run a sample job with this configuration and the results is written somewhere in workspaceblobstore. But I would like to write the results to another Storage Account. My question can I write output to another blob storage or Data Lake Storage? If so how? If this is not possible, can I somehow overwrite the environment variable AZUREML_BI_OUTPUT_PATH so that I can write to a specific place that I want in workspaceblobstore?

Thank you for your time and help.

YutongTie-MSFT 48,156 Reputation points

2023-05-22T23:33:16.9933333+00:00
Hello,

Thanks for reaching out to us, you may consider modify the scoring script to use the appropriate storage account and container.

Here is an example of how to modify the scoring script to write the output to a blob storage account:

import os import json from azure.storage.blob import BlobServiceClient # Get the output path from the environment variable output_path = os.environ.get('AZUREML_BI_OUTPUT_PATH') # Parse the output path to get the storage account name and container name output_path_parts = output_path.split('/') storage_account_name = output_path_parts[2] container_name = output_path_parts[3] # Create a BlobServiceClient object to connect to the storage account connection_string = "DefaultEndpointsProtocol=https;AccountName={};AccountKey=<account_key>;EndpointSuffix=core.windows.net".format(storage_account_name) blob_service_client = BlobServiceClient.from_connection_string(connection_string) # Get a reference to the container container_client = blob_service_client.get_container_client(container_name) # Write the output to a blob in the container output_blob_name = "output.json" output_blob_client = container_client.get_blob_client(output_blob_name) output_blob_client.upload_blob(json.dumps({"result": "Hello World"}), overwrite=True)

In this example, we use the BlobServiceClient class from the azure.storage.blob package to connect to the storage account and container. We get the storage account name and container name from the output path, and then use them to create a connection string for the BlobServiceClient. We then get a reference to the container using the get_container_client method, and write the output to a blob in the container using the upload_blob method.

To use this modified scoring script, you need to set the AZUREML_BI_OUTPUT_PATH environment variable to the path of the output blob in the container. For example, if you want to write the output to a blob named output.json in a container named mycontainer in a storage account named mystorageaccount, you would set the AZUREML_BI_OUTPUT_PATH environment variable to azureml/myservice/output.json

Please let me know how it works. Thanks a lot!
Demet Sude Saplik 36 Reputation points

2023-05-23T11:23:24.34+00:00

Thank you very much for your answer and I have a follow-up. Is it possible to configure the output place in the batch endpoint deployment instead within the scoring script?
Kris Bock 6 Reputation points Microsoft Employee

2023-10-10T11:40:49.5766667+00:00
A bit late to respond, but for anyone else that discovers this question through search:

You can specify the output when you invoke the batch endpoint. See https://learn.microsoft.com/en-us/azure/machine-learning/how-to-access-data-batch-endpoints-jobs?view=azureml-api-2&tabs=sdk#data-outputs.

Two things to note:

Batch Endpoints only support writing outputs in blob Azure Machine Learning data stores.

You need to assign the output as per the AZUREML_BI_OUTPUT_PATH environment variable.

data_path = "batch-jobs/my-unique-path" output = Output(type=AssetTypes.URI_FOLDER, path=f"{default_ds.id}/paths/{data_path}") input="https://azuremlexampledata.blob.core.windows.net/data/heart-disease-uci/data" job = ml_client.batch_endpoints.invoke( endpoint_name=endpoint.name, inputs={ "heart_dataset": input }, outputs={ "score": output } )
Jacob 0 Reputation points

2024-01-11T16:26:05.43+00:00

Thank you @Kris Bock because I think this is exactly what we need.

However, when I use the AssetTypes.URI_FOLDER als output, I get a 400 error saying Job output type must be 'uri_file'.

When I do provide a file, it still tries to mount it as a folder when starting the batch job and it supplies that folder as mount point for AZUREML_BI_OUTPUT_PATH...
Is this a bug? Or am I doing something wrong?

Matt Haythornthwaite 0

If you are using ml_client.batch_endpoints.invoke() to invoke the batch endpoint inferencing job, then you can try using the params_override parameter to set where the output should be stored.

See a code block taken from the docs below:

job = ml_client.batch_endpoints.invoke(
    endpoint_name=endpoint_name,
    input=Input(
        path="https://azuremlexampledata.blob.core.windows.net/data/mnist/sample/",
        type=AssetTypes.URI_FOLDER,
    ),
    params_override=[
        {"output_dataset.datastore_id": f"azureml:{batch_ds.id}"},
        {"output_dataset.path": f"/{endpoint_name}/"},
        {"output_file_name": filename},
    ],
)

1 answer

Sten Schneider 0 Reputation points

2024-04-22T13:51:00.68+00:00

@Kris Bock Thank you for your answer. Is it possible to also register the output as a data asset? I tried to construct the Output object with the "name" parameter (as it works in normal ml pipelines) but here it doesn't seem to have any effect. Is there another way or did I miss something?
Please sign in to rate this answer.

0 comments No comments
Sign in to comment

Share via

Azure ML Batch Endpoint: Configuring the output

1 answer