What is Azure Machine Learning?

Azure machine learning services allow you to create, test, manage, deploy, migrate, or monitor ML models in a scalable cloud-based environment. Azure machine learning services support thousands of open-source packages available in Python such as TensorFlow and Matplotlib. The supported ML tools make it easy to explore, transform, create, and test data models. E.g., Azure Machine Learning for Visual Studio Code and Jupyter notebooks. Azure ML services assist us with automated model generation and tuning to develop efficient and accurate models. The best part about Azure ML service - you can train your model over the local machine and then deploy it on the cloud. Azure offers computing services like Azure Databricks, Azure Machine Learning Compute, and advanced hyperparameter tuning services allowing you to create better models. Once you have created the right model, it’s time to deploy it over containers like Docker which makes it easy to deploy it to Azure Kubernetes Service or Azure Container Instances. You can take care of deployed models and monitor various executions to get the best outcome. Once it is deployed, you will get asynchronous predictions (real-time) on a massive amount of data. The advanced machine learning pipelines make a collaborative environment in all steps, including data preparation for deployment. Azure Machine Learning Service is a fully managed cloud service that is used to train, deploy, and manage Azure machine learning models. In this post, we are going to see what Azure Machine Learning service is and describe the end-to-end workflow for a machine learning service.

Azure machine learning service has the potential to auto-train and autotune a model. The Azure machine learning software development kit (SDK) available for Python and open-source packages allows us to create and train accurate deep learning and ML models in an Azure machine learning service workspace. Various ML components can be accessed through Python packages like Scikit-learn, PyTorch, MXNet, TensorFlow, Microsoft Cognitive Toolkit (CNTK), etc. After creating the model, you need to create a container like Docker which should be tested locally. Once tested successfully, you can deploy it as a web service using Azure Kubernetes service or Azure Container service. Now, the Azure portal or Azure Machine Learning SDK for Python will assist you to manage the deployed web model. It will help you to evaluate model metrics, redeployment of modified versions, and model tracking simultaneously.

The working mechanism

• Create your machine learning training program in Python and then configure a Compute Target.
• Put the program in the computed target for executing it in this environment. At the time of training, the program can be read from datastore or written in datastore. The running records are stored as runs inside the workspace. They are grouped inside experiments.
• Make a query to the experiments. It is done for logged metrics obtained from the present and past executions. If the desired output is not obtained from the metrics then iterate all the steps from the beginning of the first step. And, if the obtained outcome is relevant, register your model. It is done inside the model registry.
• Create a scoring program.
• Create an image. Register this image in the image registry.
• Finally, deploy this image in Azure. It is deployed as a web service.

Workspace

The workspace represents the topmost level centralized resource of the Azure Machine Learning service. It allows you to create all the artifacts for your work. It holds the list of all computing targets used for the training-developed model. It also stores the log of training execution, metrics, snapshots, and outputs. This data will assist you in choosing the best training model for your work.

The model is registered through the workspace. This registered model and scoring script is used for making an image. This image is deployed as a REST-service-based HTTP endpoint. You can use Azure Container Instances, Field-programmable gate array (FPGA), or Azure Kubernetes Service for the deployment. You can also deploy this image over the Azure IoT Edge device where it is deployed as a module.

You can take advantage of multiple workspaces sharing each of them among your team members. You can assign multiple roles to the users such as Owners, Readers, or Contributors.
A newly created workspace automatically creates all useful Azure resources such as:

• Azure storage account: a default datastore.
• Azure Container Registry: a docker registered container, and it is used while training during deployment of your model.
• Azure Key Vault: keeps keys used by computing targets and various other useful information accesses.
• Azure Application Insights: keeps monitoring data from your model.

Model:

A model is the code script that takes input and avails output. While developing a machine learning model, it requires the selection of algorithms, availing data to it, and tuning of hyperparameters. Training involves an iterative process which provides a trained model inheriting what it learned from the training process.

A model is obtained by executing in Azure Machine Learning. If you have created your model by training it outside, don’t worry, you can still use it in Azure Machine Learning. All it requires is registering it in the Azure Machine Learning service workspace.

Model Registry:

A model registry is responsible for keeping records. It was recorded from all the models of your Azure Machine Learning service workspace. You can recognize a model by its version and name. Every time you register a model with the name which has been already being used before, the registry stores it as a new version. The version is increased and the name of the model remains the same.

You can also put additional metadata tags during the registration of your model. It helps you with search it through that tag. Always remember that suppose if a model is getting used by the image then it can’t be deleted.

Image:

The image provides an environment to deploy your model independently. It has all the components required by the model. An image contains a model, application or a script (this script is provided as an input to the model which returns an output), dependencies required by your model or script.

Azure machine learning has two types of images- FPGA image and Docker image.

An FPGA image is used while deploying a field-programmable gate array in Azure ML while Docker image is used to deploy computer targets such as Azure Kubernetes Service or Azure Container Instances.

Image registry:

An image registry holds a record of images created by your models. It avails additional metadata tags while creating an image. The metadata tags are kept by an image registry. These tags can be used as a query to find your image.

Deployment:

A Deployment instantiates your image into a web service which is further hosted over the cloud. Also, it can instantiate an image into an IoT module for using it in an integrated device deployment.

Web Service:

The web service deployed over the cloud can use either Azure Kubernetes service or Azure Container Instances or FPGAs. You can create a service from your image which keeps your model, program, and other necessary files. The image provides HTTP endpoints for establishing a connection to send or receive requests to the web service and load balancers.

Azure avails valuable application insights so that you can monitor your deployed web service through model telemetry. All you need is to choose this feature. The insights data collected from telemetry are kept inside storage account instances and are only authorized to you. You can also add an automatic scaling feature to scale up and scale down your deployment.

IoT module:

IoT module is a Docker container that keeps your model, program/ application, and various other dependencies. These modules are deployed on edge devices with the help of Azure IoT Edge.

If you’ve selected a monitoring feature then Azure will gather all telemetry information from the models. This model lies inside the Azure IoT Edge module. This data can be easily accessed inside the storage account instance. The execution of the module is taken care of by Azure IoT Edge. The main device which hosts your model is also monitored by it.

Datastore:

A datastore provides a storage abstraction for your Azure account. It leverages the Azure file share and Azure blob container to keep the data. Each workspace keeps a default datastore. Here, you can register other data stores on demand. To access this information, you need the Python SDK API. You can also use Azure Machine Learning CLI to retrieve files from here.

Run:

A run record keeps information like metadata about execution such as timestamp, duration, etc., metrics logged by your program and output files which are automatically collected by the experiment. Run is produced when you provide a program file for training your model. A run contains zero or more child runs. E.g., a top-level run can have two child runs, all of them having their own run as a child.

Experiment:

The experiment contains various runs from a particular program. It is available in a workspace. Every time you provide a run, you have to give an experiment name. All the run data are stored in an experiment. Suppose if you submit a run and you provide an experiment with a name that doesn’t exist, it automatically creates a new experiment with the provided name.

Pipeline :

A machine learning pipeline is responsible for creating and directing workflow involved in machine learning phases. E.g., a pipeline provides data preparation, deployment, training of your model, and inferencing tasks. Each phase in the pipeline contains various steps; all of them run independently in multiple compute targets.

Compute Target:

It is a computing resource that is used while executing your training program or hosting service during deployment.

Compute targets are associated with a workspace. Compute targets apart from the local machine is shared among workspace users.

Managed and unmanaged compute targets:

Managed to compute targets are the ones managed and created by Azure Machine Learning service. They are optimized to handle workloads. Through the Azure portal, Azure Machine Learning SDK, or Azure CLI, you can easily develop a machine learning compute instance in a workspace. Other compute targets must be developed outside of the workspace after that they are attached to it.

The compute targets which are not managed by Azure Machine Learning service are called unmanaged compute targets. They are created outside and attached to the workspace for further use in the Azure Machine Learning service. It requires extra steps for maintaining or enhancing their performance to handle ML workloads.

Run configuration:

It keeps the instruction set to specify your program execution in a compute target. It contains a long set of behavior definitions like which environment to use while building specifications from Python or Conda. It is persistent in a directory which includes the training program.

Training script:

It requires a directory for training scripts and other related files while training your model. You need to provide an experiment name for keeping data collected during training. The whole directory is cloned to the training environment during the training process. The script named by run configuration is executed in the beginning. Finally, a directory snapshot is kept under the workspace of the experiment.

Logging:

The Azure Machine Learning Python SDK is used for collecting log metrics. After execution, you can make queries to identify which run was obtained during deployment.

Snapshot:

When the run is provided, the Azure Machine Learning compresses the directory containing the script. This zip file is then sent to the compute target where it is extracted and executed. Azure Machine Learning keeps this file as a snapshot to avail you with the run record. You can easily download this snapshot which is accessible in a run record.

Activity:

Activity points towards the long-running operations such as creation or deletion of computing targets, execution of a script over the compute target, etc. It provides notifications through SDK which helps you to easily monitor operational progresses.

Conclusion: Microsoft AzureML Studio is a Complete ease-of-use tool to both new & expert data scientist. The use of the tool & the integration with different sources & communities like Python & R makes it a powerful & practical tool. Hence it has the capacity to allow users to work on the related analytics. Having knowledge of the machine learning would fetch you oppurchunities into different positions.