Customize outputs in batch deployments
APPLIES TO: Azure CLI ml extension v2 (current) Python SDK azure-ai-ml v2 (current)
This guide explains how to create deployments that generate custom outputs and files. Sometimes you need more control over what's written as output from batch inference jobs. These cases include the following situations:
- You need to control how predictions are written in the output. For instance, you want to append the prediction to the original data if the data is tabular.
- You need to write your predictions in a different file format than the one supported out-of-the-box by batch deployments.
- Your model is a generative model that can't write the output in a tabular format. For instance, models that produce images as outputs.
- Your model produces multiple tabular files instead of a single one. For example, models that perform forecasting by considering multiple scenarios.
Batch deployments allow you to take control of the output of the jobs by letting you write directly to the output of the batch deployment job. In this tutorial, you learn how to deploy a model to perform batch inference and write the outputs in parquet format by appending the predictions to the original input data.
About this sample
This example shows how you can deploy a model to perform batch inference and customize how your predictions are written in the output. The model is based on the UCI Heart Disease dataset. The database contains 76 attributes, but this example uses a subset of 14 of them. The model tries to predict the presence of heart disease in a patient. It's integer valued from 0 (no presence) to 1 (presence).
The model was trained using an XGBBoost
classifier and all the required preprocessing was packaged as a scikit-learn
pipeline, making this model an end-to-end pipeline that goes from raw data to predictions.
The example in this article is based on code samples contained in the azureml-examples repository. To run the commands locally without having to copy/paste YAML and other files, first clone the repo and then change directories to the folder:
git clone https://github.com/Azure/azureml-examples --depth 1
cd azureml-examples/cli
The files for this example are in:
cd endpoints/batch/deploy-models/custom-outputs-parquet
Follow along in a Jupyter notebook
There's a Jupyter notebook that you can use to follow this example. In the cloned repository, open the notebook called custom-output-batch.ipynb.
Prerequisites
An Azure subscription. If you don't have an Azure subscription, create a Trial before you begin. Try the Azure Machine Learning.
An Azure Machine Learning workspace. To create a workspace, see Manage Azure Machine Learning workspaces.
Ensure that you have the following permissions in the Machine Learning workspace:
- Create or manage batch endpoints and deployments: Use an Owner, Contributor, or Custom role that allows
Microsoft.MachineLearningServices/workspaces/batchEndpoints/*
. - Create Azure Resource Manager deployments in the workspace resource group: Use an Owner, Contributor, or Custom role that allows
Microsoft.Resources/deployments/write
in the resource group where the workspace is deployed.
- Create or manage batch endpoints and deployments: Use an Owner, Contributor, or Custom role that allows
Install the following software to work with Machine Learning:
Run the following command to install the Azure CLI and the
ml
extension for Azure Machine Learning:az extension add -n ml
Pipeline component deployments for Batch Endpoints are introduced in version 2.7 of the
ml
extension for the Azure CLI. Use theaz extension update --name ml
command to get the latest version.
Connect to your workspace
The workspace is the top-level resource for Machine Learning. It provides a centralized place to work with all artifacts you create when you use Machine Learning. In this section, you connect to the workspace where you perform your deployment tasks.
In the following command, enter the values for your subscription ID, workspace, location, and resource group:
az account set --subscription <subscription>
az configure --defaults workspace=<workspace> group=<resource-group> location=<location>
Create a batch deployment with a custom output
In this example, you create a deployment that can write directly to the output folder of the batch deployment job. The deployment uses this feature to write custom parquet files.
Register the model
You can only deploy registered models using a batch endpoint. In this case, you already have a local copy of the model in the repository, so you only need to publish the model to the registry in the workspace. You can skip this step if the model you're trying to deploy is already registered.
set -e
# <set_variables>
export ENDPOINT_NAME="<YOUR_ENDPOINT_NAME>"
# </set_variables>
# <name_endpoint>
ENDPOINT_NAME="heart-classifier-custom"
# </name_endpoint>
# The following code ensures the created deployment has a unique name
ENDPOINT_SUFIX=$(cat /dev/urandom | tr -dc 'a-zA-Z0-9' | fold -w ${1:-5} | head -n 1)
ENDPOINT_NAME="$ENDPOINT_NAME-$ENDPOINT_SUFIX"
# <register_model>
MODEL_NAME='heart-classifier-sklpipe'
az ml model create --name $MODEL_NAME --type "custom_model" --path "model"
# </register_model>
echo "Creating compute"
# <create_compute>
az ml compute create -n batch-cluster --type amlcompute --min-instances 0 --max-instances 5
# </create_compute>
echo "Creating batch endpoint $ENDPOINT_NAME"
# <create_endpoint>
az ml batch-endpoint create -n $ENDPOINT_NAME -f endpoint.yml
# </create_endpoint>
echo "Showing details of the batch endpoint"
# <query_endpoint>
az ml batch-endpoint show --name $ENDPOINT_NAME
# </query_endpoint>
echo "Creating batch deployment $DEPLOYMENT_NAME for endpoint $ENDPOINT_NAME"
# <create_deployment>
az ml batch-deployment create --file deployment.yml --endpoint-name $ENDPOINT_NAME --set-default
# </create_deployment>
echo "Update the batch deployment as default for the endpoint"
# <set_default_deployment>
DEPLOYMENT_NAME="classifier-xgboost-custom"
az ml batch-endpoint update --name $ENDPOINT_NAME --set defaults.deployment_name=$DEPLOYMENT_NAME
# </set_default_deployment>
echo "Showing details of the batch deployment"
# <query_deployment>
az ml batch-deployment show --name $DEPLOYMENT_NAME --endpoint-name $ENDPOINT_NAME
# </query_deployment>
echo "Invoking batch endpoint"
# <start_batch_scoring_job>
JOB_NAME = $(az ml batch-endpoint invoke --name $ENDPOINT_NAME --input https://azuremlexampledata.blob.core.chinacloudapi.cn/data/heart-disease-uci/data --query name -o tsv)
# </start_batch_scoring_job>
echo "Showing job detail"
# <show_job_in_studio>
az ml job show -n $JOB_NAME --web
# </show_job_in_studio>
echo "List all jobs under the batch deployment"
# <list_all_jobs>
az ml batch-deployment list-jobs --name $DEPLOYMENT_NAME --endpoint-name $ENDPOINT_NAME --query [].name
# </list_all_jobs>
echo "Stream job logs to console"
# <stream_job_logs>
az ml job stream -n $JOB_NAME
# </stream_job_logs>
# <check_job_status>
STATUS=$(az ml job show -n $JOB_NAME --query status -o tsv)
echo $STATUS
if [[ $STATUS == "Completed" ]]
then
echo "Job completed"
elif [[ $STATUS == "Failed" ]]
then
echo "Job failed"
exit 1
else
echo "Job status not failed or completed"
exit 2
fi
# </check_job_status>
echo "Download scores to local path"
# <download_outputs>
az ml job download --name $JOB_NAME --output-name score --download-path ./
# </download_outputs>
echo "Delete resources"
# <delete_endpoint>
az ml batch-endpoint delete --name $ENDPOINT_NAME --yes
# </delete_endpoint>
Create a scoring script
You need to create a scoring script that can read the input data provided by the batch deployment and return the scores of the model. You're also going to write directly to the output folder of the job. In summary, the proposed scoring script does as follows:
- Reads the input data as CSV files.
- Runs an MLflow model
predict
function over the input data. - Appends the predictions to a
pandas.DataFrame
along with the input data. - Writes the data in a file named as the input file, but in
parquet
format.
code/batch_driver.py
import os
import pickle
import glob
import pandas as pd
from pathlib import Path
from typing import List
def init():
global model
global output_path
# AZUREML_MODEL_DIR is an environment variable created during deployment
# It is the path to the model folder
# Please provide your model's folder name if there's one:
output_path = os.environ["AZUREML_BI_OUTPUT_PATH"]
model_path = os.environ["AZUREML_MODEL_DIR"]
model_file = glob.glob(f"{model_path}/*/*.pkl")[-1]
with open(model_file, "rb") as file:
model = pickle.load(file)
def run(mini_batch: List[str]):
for file_path in mini_batch:
data = pd.read_csv(file_path)
pred = model.predict(data)
data["prediction"] = pred
output_file_name = Path(file_path).stem
output_file_path = os.path.join(output_path, output_file_name + ".parquet")
data.to_parquet(output_file_path)
return mini_batch
Remarks:
- Notice how the environment variable
AZUREML_BI_OUTPUT_PATH
is used to get access to the output path of the deployment job. - The
init()
function populates a global variable calledoutput_path
that can be used later to know where to write. - The
run
method returns a list of the processed files. It's required for therun
function to return alist
or apandas.DataFrame
object.
Warning
Take into account that all the batch executors have write access to this path at the same time. This means that you need to account for concurrency. In this case, ensure that each executor writes its own file by using the input file name as the name of the output folder.
Create the endpoint
You now create a batch endpoint named heart-classifier-batch
where the model is deployed.
Decide on the name of the endpoint. The name of the endpoint appears in the URI associated with your endpoint, so batch endpoint names need to be unique within an Azure region. For example, there can be only one batch endpoint with the name
mybatchendpoint
inwestus2
.Configure your batch endpoint.
Create the endpoint:
Create the deployment
Follow the next steps to create a deployment using the previous scoring script:
First, create an environment where the scoring script can be executed:
Create the deployment. Notice that
output_action
is now set toSUMMARY_ONLY
.Note
This example assumes you have a compute cluster with name
batch-cluster
. Change that name accordingly.To create a new deployment under the created endpoint, create a YAML configuration like the following. You can check the full batch endpoint YAML schema for extra properties.
$schema: https://azuremlschemas.azureedge.net/latest/modelBatchDeployment.schema.json endpoint_name: heart-classifier-batch name: classifier-xgboost-custom description: A heart condition classifier based on XGBoost and Scikit-Learn pipelines that append predictions on parquet files. type: model model: azureml:heart-classifier-sklpipe@latest environment: name: batch-mlflow-xgboost image: mcr.microsoft.com/azureml/openmpi4.1.0-ubuntu20.04:latest conda_file: environment/conda.yaml code_configuration: code: code scoring_script: batch_driver.py compute: azureml:batch-cluster resources: instance_count: 2 settings: max_concurrency_per_instance: 2 mini_batch_size: 2 output_action: summary_only retry_settings: max_retries: 3 timeout: 300 error_threshold: -1 logging_level: info
Then, create the deployment with the following command:
DEPLOYMENT_NAME="classifier-xgboost-parquet" az ml batch-deployment create -f endpoint.yml
At this point, our batch endpoint is ready to be used.
Test the deployment
To test your endpoint, use a sample of unlabeled data located in this repository, which can be used with the model. Batch endpoints can only process data that's located in the cloud and is accessible from the Azure Machine Learning workspace. In this example, you upload it to an Azure Machine Learning data store. You're going to create a data asset that can be used to invoke the endpoint for scoring. However, notice that batch endpoints accept data that can be placed in multiple type of locations.
Invoke the endpoint with data from a storage account:
JOB_NAME = $(az ml batch-endpoint invoke --name $ENDPOINT_NAME --deployment-name $DEPLOYMENT_NAME --input azureml:heart-dataset-unlabeled@latest | jq -r '.name')
Note
The utility
jq
might not be installed on every installation. You can get instructions on GitHub.A batch job is started as soon as the command returns. You can monitor the status of the job until it finishes:
Analyze the outputs
The job generates a named output called score
where all the generated files are placed. Since you wrote into the directory directly, one file per each input file, then you can expect to have the same number of files. In this particular example, name the output files the same as the inputs, but they have a parquet extension.
Note
Notice that a file predictions.csv is also included in the output folder. This file contains the summary of the processed files.
You can download the results of the job by using the job name:
To download the predictions, use the following command:
az ml job download --name $JOB_NAME --output-name score --download-path ./
Once the file is downloaded, you can open it using your favorite tool. The following example loads the predictions using Pandas
dataframe.
import pandas as pd
import glob
output_files = glob.glob("named-outputs/score/*.parquet")
score = pd.concat((pd.read_parquet(f) for f in output_files))
The output looks as follows:
age | sex | ... | thal | prediction |
---|---|---|---|---|
63 | 1 | ... | fixed | 0 |
67 | 1 | ... | normal | 1 |
67 | 1 | ... | reversible | 0 |
37 | 1 | ... | normal | 0 |
Clean up resources
Run the following code to delete the batch endpoint and all the underlying deployments. Batch scoring jobs aren't deleted.
az ml batch-endpoint delete --name $ENDPOINT_NAME --yes