Train PyTorch models at scale with Azure Machine Learning

APPLIES TO: Python SDK azure-ai-ml v2 (current)

In this article, you'll learn to train, hyperparameter tune, and deploy a PyTorch model using the Azure Machine Learning (AzureML) Python SDK v2.

You'll use the example scripts in this article to classify chicken and turkey images to build a deep learning neural network (DNN) based on PyTorch's transfer learning tutorial. Transfer learning is a technique that applies knowledge gained from solving one problem to a different but related problem. Transfer learning shortens the training process by requiring less data, time, and compute resources than training from scratch. To learn more about transfer learning, see the deep learning vs machine learning article.

Whether you're training a deep learning PyTorch model from the ground-up or you're bringing an existing model into the cloud, you can use AzureML to scale out open-source training jobs using elastic cloud compute resources. You can build, deploy, version, and monitor production-grade models with AzureML.


To benefit from this article, you'll need to:

  • Access an Azure subscription. If you don't have one already, create a free account.
  • Run the code in this article using either an Azure Machine Learning compute instance or your own Jupyter notebook.
    • Azure Machine Learning compute instance—no downloads or installation necessary
      • Complete the Quickstart: Get started with Azure Machine Learning to create a dedicated notebook server pre-loaded with the SDK and the sample repository.
      • In the samples deep learning folder on the notebook server, find a completed and expanded notebook by navigating to this directory: v2 > sdk > python > jobs > single-step > pytorch > train-hyperparameter-tune-deploy-with-pytorch.
    • Your Jupyter notebook server

You can also find a completed Jupyter Notebook version of this guide on the GitHub samples page.

Before you can run the code in this article to create a GPU cluster, you'll need to request a quota increase for your workspace.

Set up the job

This section sets up the job for training by loading the required Python packages, connecting to a workspace, creating a compute resource to run a command job, and creating an environment to run the job.

Connect to the workspace

First, you'll need to connect to your AzureML workspace. The AzureML workspace is the top-level resource for the service. It provides you with a centralized place to work with all the artifacts you create when you use Azure Machine Learning.

We're using DefaultAzureCredential to get access to the workspace. This credential should be capable of handling most Azure SDK authentication scenarios.

If DefaultAzureCredential doesn't work for you, see azure-identity reference documentation or Set up authentication for more available credentials.

# Handle to the workspace
from import MLClient

# Authentication package
from azure.identity import DefaultAzureCredential

credential = DefaultAzureCredential()

If you prefer to use a browser to sign in and authenticate, you should uncomment the following code and use it instead.

# Handle to the workspace
# from import MLClient

# Authentication package
# from azure.identity import InteractiveBrowserCredential
# credential = InteractiveBrowserCredential()

Next, get a handle to the workspace by providing your Subscription ID, Resource Group name, and workspace name. To find these parameters:

  1. Look for your workspace name in the upper-right corner of the Azure Machine Learning studio toolbar.
  2. Select your workspace name to show your Resource Group and Subscription ID.
  3. Copy the values for Resource Group and Subscription ID into the code.
# Get a handle to the workspace
ml_client = MLClient(

The result of running this script is a workspace handle that you'll use to manage other resources and jobs.


  • Creating MLClient will not connect the client to the workspace. The client initialization is lazy and will wait for the first time it needs to make a call. In this article, this will happen during compute creation.

Create a compute resource to run the job

AzureML needs a compute resource to run a job. This resource can be single or multi-node machines with Linux or Windows OS, or a specific compute fabric like Spark.

In the following example script, we provision a Linux compute cluster. You can see the Azure Machine Learning pricing page for the full list of VM sizes and prices. Since we need a GPU cluster for this example, let's pick a STANDARD_NC6 model and create an AzureML compute.

from import AmlCompute

gpu_compute_taget = "gpu-cluster"

    # let's see if the compute target already exists
    gpu_cluster = ml_client.compute.get(gpu_compute_taget)
        f"You already have a cluster named {gpu_compute_taget}, we'll reuse it as is."

except Exception:
    print("Creating a new gpu compute target...")

    # Let's create the Azure ML compute object with the intended parameters
    gpu_cluster = AmlCompute(
        # Name assigned to the compute cluster
        # Azure ML Compute is the on-demand VM service
        # VM Family
        # Minimum running nodes when there is no job running
        # Nodes in cluster
        # How many seconds will the node running after the job termination
        # Dedicated or LowPriority. The latter is cheaper but there is a chance of job termination

    # Now, we pass the object to MLClient's create_or_update method
    gpu_cluster = ml_client.begin_create_or_update(gpu_cluster).result()

    f"AMLCompute with name {} is created, the compute size is {gpu_cluster.size}"

Create a job environment

To run an AzureML job, you'll need an environment. An AzureML environment encapsulates the dependencies (such as software runtime and libraries) needed to run your machine learning training script on your compute resource. This environment is similar to a Python environment on your local machine.

AzureML allows you to either use a curated (or ready-made) environment or create a custom environment using a Docker image or a Conda configuration. In this article, you'll reuse the curated AzureML environment AzureML-pytorch-1.9-ubuntu18.04-py37-cuda11-gpu. You'll use the latest version of this environment using the @latest directive.

curated_env_name = "AzureML-pytorch-1.9-ubuntu18.04-py37-cuda11-gpu@latest"

Configure and submit your training job

In this section, we'll begin by introducing the data for training. We'll then cover how to run a training job, using a training script that we've provided. You'll learn to build the training job by configuring the command for running the training script. Then, you'll submit the training job to run in AzureML.

Obtain the training data

You'll use data that is stored on a public blob as a zip file. This dataset consists of about 120 training images each for two classes (turkeys and chickens), with 100 validation images for each class. The images are a subset of the Open Images v5 Dataset. We'll download and extract the dataset as part of our training script

Prepare the training script

In this article, we've provided the training script In practice, you should be able to take any custom training script as is and run it with AzureML without having to modify your code.

The provided training script downloads the data, trains a model, and registers the model.

Build the training job

Now that you have all the assets required to run your job, it's time to build it using the AzureML Python SDK v2. For this example, we'll be creating a command.

An AzureML command is a resource that specifies all the details needed to execute your training code in the cloud. These details include the inputs and outputs, type of hardware to use, software to install, and how to run your code. The command contains information to execute a single command.

Configure the command

You'll use the general purpose command to run the training script and perform your desired tasks. Create a Command object to specify the configuration details of your training job.

from import command
from import Input

job = command(
        num_epochs=30, learning_rate=0.001, momentum=0.9, output_dir="./outputs"
    code="./src/",  # location of source code
    command="python --num_epochs ${{inputs.num_epochs}} --output_dir ${{inputs.output_dir}}",
  • The inputs for this command include the number of epochs, learning rate, momentum, and output directory.
  • For the parameter values:
    • provide the compute cluster gpu_compute_target = "gpu-cluster" that you created for running this command;
    • provide the curated environment AzureML-pytorch-1.9-ubuntu18.04-py37-cuda11-gpu that you initialized earlier;
    • configure the command line action itself—in this case, the command is python You can access the inputs and outputs in the command via the ${{ ... }} notation; and
    • configure metadata such as the display name and experiment name; where an experiment is a container for all the iterations one does on a certain project. All the jobs submitted under the same experiment name would be listed next to each other in AzureML studio.

Submit the job

It's now time to submit the job to run in AzureML. This time, you'll use create_or_update on

Once completed, the job will register a model in your workspace (as a result of training) and output a link for viewing the job in AzureML studio.


Azure Machine Learning runs training scripts by copying the entire source directory. If you have sensitive data that you don't want to upload, use a .ignore file or don't include it in the source directory.

What happens during job execution

As the job is executed, it goes through the following stages:

  • Preparing: A docker image is created according to the environment defined. The image is uploaded to the workspace's container registry and cached for later runs. Logs are also streamed to the job history and can be viewed to monitor progress. If a curated environment is specified, the cached image backing that curated environment will be used.

  • Scaling: The cluster attempts to scale up if it requires more nodes to execute the run than are currently available.

  • Running: All scripts in the script folder src are uploaded to the compute target, data stores are mounted or copied, and the script is executed. Outputs from stdout and the ./logs folder are streamed to the job history and can be used to monitor the job.

Tune model hyperparameters

You've trained the model with one set of parameters, let's now see if you can further improve the accuracy of your model. You can tune and optimize your model's hyperparameters using Azure Machine Learning's sweep capabilities.

To tune the model's hyperparameters, define the parameter space in which to search during training. You'll do this by replacing some of the parameters passed to the training job with special inputs from the package.

Since the training script uses a learning rate schedule to decay the learning rate every several epochs, you can tune the initial learning rate and the momentum parameters.

from import Uniform

# we will reuse the command_job created before. we call it as a function so that we can apply inputs
job_for_sweep = job(
    learning_rate=Uniform(min_value=0.0005, max_value=0.005),
    momentum=Uniform(min_value=0.9, max_value=0.99),

Then, you'll configure sweep on the command job, using some sweep-specific parameters, such as the primary metric to watch and the sampling algorithm to use.

In the following code, we use random sampling to try different configuration sets of hyperparameters in an attempt to maximize our primary metric, best_val_acc.

We also define an early termination policy, the BanditPolicy, to terminate poorly performing runs early. The BanditPolicy will terminate any run that doesn't fall within the slack factor of our primary evaluation metric. You will apply this policy every epoch (since we report our best_val_acc metric every epoch and evaluation_interval=1). Notice we will delay the first policy evaluation until after the first 10 epochs (delay_evaluation=10).

from import BanditPolicy

sweep_job = job_for_sweep.sweep(
        slack_factor=0.15, evaluation_interval=1, delay_evaluation=10

Now, you can submit this job as before. This time, you'll be running a sweep job that sweeps over your train job.

returned_sweep_job = ml_client.create_or_update(sweep_job)

# stream the output and wait until the job is finished

# refresh the latest status of the job after streaming
returned_sweep_job =

You can monitor the job by using the studio user interface link that is presented during the job run.

Find the best model

Once all the runs complete, you can find the run that produced the model with the highest accuracy.

from import Model

if returned_sweep_job.status == "Completed":

    # First let us get the run which gave us the best result
    best_run =["best_child_run_id"]

    # lets get the model from this run
    model = Model(
        # the script stores the model as "outputs"
        description="Model created from run.",

        "Sweep job status: {}. Please wait until it completes".format(

Deploy the model as an online endpoint

You can now deploy your model as an online endpoint—that is, as a web service in the Azure cloud.

To deploy a machine learning service, you'll typically need:

  • The model assets that you want to deploy. These assets include the model's file and metadata that you already registered in your training job.
  • Some code to run as a service. The code executes the model on a given input request (an entry script). This entry script receives data submitted to a deployed web service and passes it to the model. After the model processes the data, the script returns the model's response to the client. The script is specific to your model and must understand the data that the model expects and returns. When you use an MLFlow model, AzureML automatically creates this script for you.

For more information about deployment, see Deploy and score a machine learning model with managed online endpoint using Python SDK v2.

Create a new online endpoint

As a first step to deploying your model, you need to create your online endpoint. The endpoint name must be unique in the entire Azure region. For this article, you'll create a unique name using a universally unique identifier (UUID).

import uuid

# Creating a unique name for the endpoint
online_endpoint_name = "aci-birds-endpoint-" + str(uuid.uuid4())[:8]
from import ManagedOnlineEndpoint

# create an online endpoint
endpoint = ManagedOnlineEndpoint(
    description="Classify turkey/chickens using transfer learning with PyTorch",
    tags={"data": "birds", "method": "transfer learning", "framework": "pytorch"},

endpoint = ml_client.begin_create_or_update(endpoint).result()

print(f"Endpoint {} provisioning state: {endpoint.provisioning_state}")

Once you've created the endpoint, you can retrieve it as follows:

endpoint = ml_client.online_endpoints.get(name=online_endpoint_name)

    f'Endpint "{}" with provisioning state "{endpoint.provisioning_state}" is retrieved'

Deploy the model to the endpoint

After you've created the endpoint, you can deploy the model with the entry script. An endpoint can have multiple deployments. Using rules, the endpoint can then direct traffic to these deployments.

In the following code, you'll create a single deployment that handles 100% of the incoming traffic. We've specified an arbitrary color name (aci-blue) for the deployment. You could also use any other name such as aci-green or aci-red for the deployment. The code to deploy the model to the endpoint does the following:

  • deploys the best version of the model that you registered earlier;
  • scores the model, using the file; and
  • uses the curated environment (that you specified earlier) to perform inferencing.
from import (

online_deployment_name = "aci-blue"

# create an online deployment.
blue_deployment = ManagedOnlineDeployment(
    code_configuration=CodeConfiguration(code="./score/", scoring_script=""),

blue_deployment = ml_client.begin_create_or_update(blue_deployment).result()


Expect this deployment to take a bit of time to finish.

Test the deployed model

Now that you've deployed the model to the endpoint, you can predict the output of the deployed model, using the invoke method on the endpoint.

To test the endpoint, let's use a sample image for prediction. First, let's display the image.

# install pillow if PIL cannot imported
%pip install pillow
import json
from PIL import Image
import matplotlib.pyplot as plt

%matplotlib inline

Create a function to format and resize the image.

# install torch and torchvision if needed
%pip install torch
%pip install torchvision

import torch
from torchvision import transforms

def preprocess(image_file):
    """Preprocess the input image."""
    data_transforms = transforms.Compose(
            transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),

    image =
    image = data_transforms(image).float()
    image = torch.tensor(image)
    image = image.unsqueeze(0)
    return image.numpy()

Format the image and convert it to a JSON file.

image_data = preprocess("test_img.jpg")
input_data = json.dumps({"data": image_data.tolist()})
with open("request.json", "w") as outfile:

You can then invoke the endpoint with this JSON and print the result.

# test the blue deployment
result = ml_client.online_endpoints.invoke(


Clean up resources

If you won't be using the endpoint, delete it to stop using the resource. Make sure no other deployments are using the endpoint before you delete it.



Expect this cleanup to take a bit of time to finish.

Next steps

In this article, you trained and registered a deep learning neural network using PyTorch on Azure Machine Learning. You also deployed the model to an online endpoint. See these other articles to learn more about Azure Machine Learning.