Machine Learning on Cloud
According to Microsoft: Machine learning is a technique that allows computers to use existing data to forecast future behaviors, outcomes, and trends.In this post I will explain how to use Azure Machine Learning Studio as an integrated, end-to-end data science and advanced analytics solution, enabling data scientists to prepare data, develop experiments, and deploy models at cloud scale.
How Does it Work?
(1) Azure Machine Learning Studio is a cloud service in Azure that enables you to experiment with data and create predictive models.
(2) An Experimental Dataset is taken and a workflow of transformations to apply to the data are defined.
(3) Different models are trained and tested to create a predictive experiment model with required accuracy.
(4) The predictive experiment encapsulates the machine learning model and the associated data transformations, ready to be used with new data.
(5) Your predictive experiment can then be published as a web service, which client applications and processes can call to generate predicted values.
What is it Based on?
Azure Machine Learning is built on top of:
- Jupyter Notebook
- Apache Spark
- Docker
- Kubernetes
- Python
- Conda
- Microsoft Machine Learning Library for Apache Spark
- Microsoft Cognitive Toolkit.
Azure ML is compatible with scikit-learn, TensorFlow and Spark ML
How the Components Work Together?
- Python and Jupyter Notebook together provide the basic scripting capabilities in the Azure ML Studio.
- Conda acts as a package manager for compatible python packages like sckit-learn, matplotlib, TensorFlow, etc…
- Docker and Kubernetes provide the underlying environment for the webservice deployment.
- Apache Spark and the Dependant Libraries help in Data gathering from Big Data Jobs which can be linked with the Azure ML Studio.
- Apart from these underlying components Azure ML Studio can further be connected to services such as Azure Storage, Azure SQL Storage, etc.. facilitating a single environment for the whole Data Science life-cycle.
Other ML Tools
In addition to Azure Machine Learning Studio, there are a variety of options at Microsoft to build, deploy, and manage your machine learning models.
- Machine Learning Service
- SQL Server Machine Learning Services
- Microsoft Machine Learning Server
- Azure Data Science Virtual Machine
- Spark MLLib in HDInsight
- Batch AI Training Service
- Microsoft Cognitive Toolkit (CNTK)
- Azure Cognitive Services
Machine Learning Service
Closely related to Machine Learning Studio, Machine Learning Service is an offering in Azure from Microsoft which provides the utility similar to Azure ML Studio.It is currently in preview and provides the environment with much better support to open frameworks in python such as Tensorflow, sci-kit learn etc.It supports Jupyter Notebooks, Visual Studio Code Tools for AI, Azure Batch AI and Containerised Deployment.
ML Service vs Azure ML Studio
When compared with the Machine Learning Service, Azure Machine Learning Studio provides pre-configured algorithms and a drag-drop interface for ML.Azure ML Studio is recommended for beginners to Experiment with ML modelling and deployment and for small-scale, low complexity business problems.However, Machine Learning Service provides a professional and full-fledged environment for solving complex business problems and handle global scale cloud deployment of services.
As you move forward with the , you will learn How to interact with Azure ML on the web.Use Azure ML functionality to clean data, create ML models and deploy them as web services.Consume the web services in real-time client applications.Re-train models using Azure Big data factory.Using Azure ML for IoT and Big data Applications.
Azure ML Platform
This helps understand using Azure ML platform on the web to generate predictions by creating and deploying ML models.
The workflow followed to generate web services from datasets is as follows:
Train a Machine Learning Model with training Experiment.
Create a Predictive experiment from training experiment encapsulating data transformations and training model.
Use the Predictive experiment to create a web service to generate predictions using API-endpoint and API key.
Cortana Intelligence Gallery
The Cortana Intelligence Gallery is a collection of resources people have shared on the Azure ML platform.
It includes various tutorials, datasets and experiments.
Experiments can be copied from the gallery into a workspace and can be explored/modified for a better understanding.
We can also share our work on experiments and web services for others to learn and explore.
Move to the next topic to start creating and experimenting on ML models.
Datasets
An Azure ML experiment requires at least one dataset on which the model is created. The data can be imported directly or into Azure Storage and used for creating a model.
Generally, in the case of Big Data Sources:
Hive or U-SQL jobs are used to clean and prepare for analysis.
A business process stores large volumes of data in a database or data warehouse.
This data can be stored on Azure Storage and can be easily imported to Azure ML Workspace. Data can also be imported from Hive or Azure SQL database.
Further in the topic you will learn different types of data used in experiments and ways to use them
Big Data Sources – Considerations
When working with Big Data sources the following points must be considered:
- Multiple files can be uploaded to the Azure ML workspace but maximum size is less than 2 GBs.
- However, import up to 10 GB of data from other sources is possible.
- If you need to work with even larger volumes of data, statistical sampling techniques must be used to sample ten gigabytes of data for training.
Training Data
Datasets are broadly classified into two types in Azure ML perspective.
Training Data to train the Experiment
Reference Data providing additional information.
TRAINING DATA : The Data to train the Experiment is mandatory data and acts as the start of Dataflow.
Move to the next card to learn about reference data.
Reference Data
REFERENCE DATA is not directly used in the experiment or for training a model.It is only used to provide additional information.
For an example when a web service is published based on the ML model, a service endpoint is created to consume the web service.
Based on the client using the service, the reference data to be used can be varied accordingly to provide customisation and flexibility to the web service.
Data Import – Trivia
In this topic, you have learnt ways to import data from various Azure Data stores and the types of data used. Few other notable points are:
Data can also be imported from On-premises SQL Server or several other online sources using Import module.
Multiple Data formats such as .txt, .csv, .nh.csv, .tsv, .nh.tsv, Excel file, Azure table, Hive table, SQL database table, .RData etc.. are supported.
Data types recognised by ML Studio are String, Integer, Double, Boolean, DateTime and TimeSpan.
For more details refer the following:
Import Data
Prepare Data
Scenarios for advanced analytics.
The Lifecycle
A Data Science solution involves:
– Identifying a Business Problem
– Data Extraction
– Data Cleaning
– Data Transformation
– Exploratory Data Analysis
– Data Visualisation
– Machine Learning Modeling
– Publishing the Prediction Service
Adopting ML Studio
Azure Machine Learning Studio helps through the entire lifecycle of a Data Science solution.
We already had a look at Data Import from various sources, which can be considered as Data Extraction.
Data Cleaning and Transformation are done based on our business problem and the approach we take to provide a solution for it. You can see this in the videos over the following cards.
Data Cleaning and Transformation
Data Cleaning refers to detecting and correcting missing, inaccurate and irrelevant parts of data while Data transformation refers to converting data from one format to another.
Azure ML Studio provides these features through Filters, Scale and Reduce, Manipulation and Sample and Split.
- Filters : Transforms and cleans digital data and can help in Speech Processing
- Manipulation : Cleans missing values, Meta data editing, SQL transforms, etc..
- Scale and Reduce : Normalization, Grouping, Clipping, etc..
- Sample and Split : Partition and Sample data.
Check this, for more info on Data Cleaning and Transformation.
Data Visualisation
Exploratory Data Analysis and Data Visualisation are facilitated by Notebooks in the Azure ML Studio.Data Visualisation is readily available in Azure ML Studio over the right-click of uploaded datasets.
However, Notebooks can be used to further visualise data in a required manner with Python/R scripts adding flexibility and functionality.
Watch the following video to learn using Notebooks for data visualisation.
Machine Learning Modelling
Coming over to devising a Machine Learning Model, Azure ML supports 25 Algorithms.
They are divided into 4 major classes for both Structured and unstructured data.
– Regression
– Classification
– Anomaly Detection
– Clustering
Go through Machine Learning Axioms and other ML courses for model selection and feature selection.
Supervised Learning
Azure ML Platform is equipped with over 20 types of supervised learning methods in-built and ready to use.
Users can also write their own python scripts and embed them into the ML workflow for customised and optimised models using Notebooks and supported ML libraries like sci-kit learn, Tensorflow, etc.
Move to the next cards to check out examples of various Supervised Learning Algorithms along with concepts of Data cleaning and Transformation.
Regression Model
Watch this video to understand how to train a regression model from a sample dataset using Azure ML Experimentation service.
Classification Model
Watch this video to understand how to train a classification model from a sample dataset using Azure ML Experimentation service.
Moving Further
Supervised learning is the widely used ML modelling technique for structured data.
Different algorithms have their own pros and cons. Our requirements must be decided on to select the optimal algorithm.
Refer the following links for a better understanding on Supervised learning methods w.r.t Azure ML platform: Feature Engineering, Algorithm Selection and Evaluating ML Model.
Unsupervised Learning
Unsupervised Learning is used to work on unstructured and un-labelled data.
Clustering is the most commonly used method where similar data is grouped by finding features and grouping them based on their feature set.
Azure ML Platform supports K-Means Clustering for Unsupervised Learning.
K-Means Clustering
Watch the video to learn using K-Means Clustering Algorithm and visualise data after clustering.
Recommenders
Recommenders, as their name suggests, are used in sectors of e-commerce, ads and social platforms etc. They recommend related items of interest based on users previous interaction.
Watch the following video to learn to model a recommender on Azure ML platform.
What’s Next?
In this topic, you learned how to create an ML model from datasets taken and different types of ML models supported on Azure ML Platform.
Move over to next topic to learn to deploy the ML models as web services.
Refer the following links for further info on: K-Means Clustering and Match Box Recommender.
Using the Trained Models
In the previous topic, training different ML models using Azure ML Platform is discussed.
Once the trained models are tested for accuracy and optimised, they need to be deployed for consumption through API.
This topic helps to learn to deploy the trained model as a web service using a predictive experiment.
Move through the cards to find out..
Predictive Experiment
Predictive Experiment is created from successful Training Experiment.
- This defines the workflow of processing web requests to generate predictions.
- It is similarly structured to training experiment but the data transformations like normalisation, are encapsulated into a single block in order to not modify the training experiment.
- Webservice input and output will be the new blocks added to the experiment.
Webservice
A Webservice is to be created from a Predictive Experiment to consume the generated ML model.
Creating a Webservice generates API end-points which can be accessed through Primary and Secondary API-keys.
The API end-points take inputs required by the ML model and given JSON output with predictions.
Webservice Workflow
The following workflow is adopted for building a web service with predictive experiment:
- Create a training experiment and optimize it.
- Generate a predictive experiment from the trained model along with operations like data normalization, encapsulation and transformation.
- Include definitions for web service inputs and outputs that define how client applications will consume the model.
- Modify predictive experiment without affecting training experiment in accordance with client requirements.
- Finally, publish the predictive experiment as a web service that can be accessed through a URL or endpoint.
Web service usage
- Client applications can call the web service, passing in the features and getting the output according to the schema.
- When new training data is gathered, the training experiment can be re-opened and improved by retraining.
- Predictive experiment can be updated accordingly while taking consideration of transforms that encapsulate the trained model and other operations required while preserving other changes of the predictive experiment.
- With updated predictive model the web service can be re-deployed and client applications can continue to access through the same endpoint as before.
Moving Further
For more info, refer to the following links: Preparing Predictive Model and Deploying Web Service.
Now as the basics of creating and deploying a Webservice are understood, Move over to the Next topic to learn more about managing and customising Webservices.
Diving Deep – Web services
This topic covers the
Management
Consumption and
Customizing of Web services according to the client requirements.
You will also learn about metrics and logging regarding the usage of Web services.
Webservices – Factors to consider
After models are created and published as API, we need to make them consumable by various client applications.
Web services from Azure ML Platform are flexible.
- Web services can be made to retrieve reference data from a database.
- Parameters can be passed to retrieve reference data from the required database to service different clients.
- Can be managed and monitored with respect to utilization and activity through logging.
Consumption
Web service can be consumed from the published API either in
Request-Response mode or
Batch Mode in asynchronous way.
API endpoints and API Keys are used according to the requirement.
The APIs are built as REST APIs and can be consumed by required client application by passing required parameters as an HTTPS request.
Microsoft Excel along with Azure ML Plug-in can also be used to consume the service.
Parameters
Web Service can be customised to take required parameters to access additional data.
For example, you can specify a database to fetch the data. It will be provided as an additional parameter apart from the required parameters for a prediction.
This helps in customised client consumption such that each client will use the same service for prediction but data output/retrieval is different for each of them.
Monitoring
Azure ML Platform enables Web service management through the Usage Statistics and Logging.
Dashboards provide an overview of the total no.of requests made to the API and their success/fail rate over a selected period of time. They also give the average compute time and latency associated with the API.
Logging can be enabled for a more detailed JSON response files stored automatically on Azure storage providing a detailed report of the request and response.
Moving Further
Refer the following links for more info on Consuming Webservice, Adding Parameters, Managing Webservice and Logging.
Move over to the next topic to learn using these Webservices in Big Data scenarios.
Big Data Scenario
The Webservice API-endpoints can be used as a service by clients for predictions with instant response.
However, if we want to use ML models in scenarios having large sets of data i.e for Big Data, processing must be done in batches at scheduled intervals and in optimal times to reduce latency and promote the asynchronous way of obtaining predictions for our data.
The Azure Data Factory and its pipeline come in to play in these kinds of scenarios.
Azure Data Factory
Azure Data Factory is a cloud service for data movement and transformation and helps orchestrate big data batch processing tasks through defined pipelines in a scheduled way.
Pipeline
- Consists of a sequence of activities, which are performed using the linked services.
- Makes datasets, which define the input and output for each activity.
- Runs activities at scheduled intervals or windows, each operating on the slice of data that’s relevant for that window.
Azure ML in Pipeline
When Big Data batch processing is done by pipeline, predictions can be part of the pipeline with Azure ML being used as a linked service.
Azure ML batch execution activity is used to call a predictive web service from a pipeline.
The input data set is passed to the web service input and, the predicted output from the web service is returned continuing to the next activity in the pipeline.
Other Linked services are data sources, like Azure Storage or Azure SQL Database, and compute services, like Azure HD Insight or Azure Data Link Analytics.
What about Retraining?
In a Big Data Scenario retraining the model might be needed a tad frequently, since the data generated is of huge amounts and may vary over short intervals of time.
We need fresh data to train the ML model to improve the model accuracy with the changing inflow of data.
The retraining can be done manually using Azure ML workspace.
However, considering the amount of data required to be handled while Big Data Processes are ongoing, this is neither efficient nor recommended.
In this situation, Azure Data Factory Pipeline and Azure ML Update Resource Service comes to the rescue to automate the retraining of ML models.
Automating Retraining
To automate the process of retraining, Azure ML provides us a feature to publish training experiment as a retraining web service.
The following activities can be executed in sequence, in the Azure Data Factory Pipeline, to achieve the automation task:
- Azure ML Batch Execution Activity is used to call the retraining web service to generate a new model as a file.
- The model file is passed to an Azure ML Update Resource Activity that updates the scoring experiment replacing the existing model.
Moving Further
Refer the following links for more info on Predictive Pipelines and Update Resource Activity.
As Big Data processing for predictions is looked into, Move over to the next topic to learn process real-time data for predictions using the Webservices of Azure ML Platform.
Streaming Process
A Streaming Process is considered as data processing in real-time.
Azure includes a number of services that you can use to implement a streaming process for real-time data.
An Azure streaming solution consists of inputs, a streaming job, and output.
Input : Are often event hubs or IoT hubs that are used to ingest real-time data at scale.
Streaming Job : Used to process the data. Generally, an Azure Stream Analytics query.
Output : The expected result, could be anything from a database update to a real-time dashboard with analysis.