Better Together:
Auto-train a time-series
forecast model with
AML & ADB
Name: Moe Steller
Title: Sr. Cloud Solution Architect, AI/ML
Introduction
Moe Steller
Microsoft - WW Commercial Business
Sr. Cloud Solution Architect, AI/ML
Focus Areas:
NLP / Deep Learning / Time-Series
Scalable AI / AIOps
RAI / Ethics
Industry Focus:
Healthcare, Finance, Supply Chain, Insurance, HLS
Feedback
Your feedback is important to us.
Don’t forget to rate and review the sessions.
Better Together:
Auto-train a time-
series forecast
model with
AML & ADB
Agenda
§ Time Series Forecasting
§ Product Preview
§ Advanced Parameters
§ Better Together
Architecture
§ Demo
Time Series Forecasting
Time Series Forecasting
Multivariate DNN
Univariate
Time Series Forecasting: Algorithm Support
• Elastic Net
• Light GBM
• Gradient Boosting
• Decision Tree
• K Nearest Neighbors
• LARS Lasso
• Stochastic Gradient Descent
• Random Forest
• Extremely Randomized Trees
• XGBoost
• Online Gradient Descent
Regressor
• Fast Linear Regressor
• Elastic Net
• Light GBM
• Gradient Boosting
• Decision Tree
• K Nearest Neighbors
• LARS Lasso
• Stochastic Gradient Descent
• Random Forest
• Extremely Randomized Trees
• XGBoost
• Auto-ARIMA
• Prophet
• ForecastTCN
• Regression
• Logistic Regression
• Light GBM
• Gradient Boosting
• Decision Tree
• K Nearest Neighbors
• Linear SVC
• Support Vector Classification (SVC)
• Random Forest
• Extremely Randomized Trees
• XGBoost
• Averaged Perceptron Classifier
• Naive Bayes
• Stochastic Gradient Descent
• Linear SVM Classifier*
• Classification
• Time Series Forecasting
Time Series Forecasting: Algorithm Support
• Elastic Net
• Light GBM
• Gradient Boosting
• Decision Tree
• K Nearest Neighbors
• LARS Lasso
• Stochastic Gradient Descent
• Random Forest
• Extremely Randomized Trees
• XGBoost
• Online Gradient Descent
Regressor
• Fast Linear Regressor
• Elastic Net
• Light GBM
• Gradient Boosting
• Decision Tree
• K Nearest Neighbors
• LARS Lasso
• Stochastic Gradient Descent
• Random Forest
• Extremely Randomized Trees
• XGBoost
• Auto-ARIMA
• Prophet
• ForecastTCN
• Regression
• Logistic Regression
• Light GBM
• Gradient Boosting
• Decision Tree
• K Nearest Neighbors
• Linear SVC
• Support Vector Classification (SVC)
• Random Forest
• Extremely Randomized Trees
• XGBoost
• Averaged Perceptron Classifier
• Naive Bayes
• Stochastic Gradient Descent
• Linear SVM Classifier*
• Classification
• Time Series Forecasting
Time Series Forecasting: Product Preview
Models Description Benefits
Prophet (Preview)
Prophet works best with time series that have strong
seasonal effects and several seasons of historical data.
To leverage this model, install it locally using pip install
fbprophet.
Accurate & fast, robust to outliers,
missing data, and dramatic changes in
your time series.
Auto-ARIMA (Preview)
Auto-Regressive Integrated Moving Average (ARIMA)
performs best, when the data is stationary. This means that
its statistical properties like the mean and variance are
constant over the entire set. For example, if you flip a coin,
then the probability of you getting heads is 50%, regardless
if you flip today, tomorrow or next year.
Great for univariate series, since the past
values are used to predict the future
values.
ForecastTCN
ForecastTCN is a neural network model designed to tackle
the most demanding forecasting tasks, capturing nonlinear
local and global trends in your data as well as relationships
between time series.
Capable of leveraging complex trends in
your data and readily scales to the largest
of datasets.
Time Series Forecasting: Advanced Parameters
enable_stack_ensable
Two-layer implementation: first layer has the same models as the voting ensemble,
second layer model finding optimal combination of the models from the first layer
enable_voting_ensable Voting implements soft-voting which uses weighted averages
enable_onnx_compatible_models Get a pre-trained or generated ONNX model added
spark_context Used inside Azure Databricks/Spark environment
featurization Detected column type preprocessing/featurization
enable_early_stopping If the score is not improving in the short term
seasonality Set time series seasonality
https://docs.microsoft.com/en-us/python/api/azureml-train-automl-client/azureml.train.automl.automlconfig.automlconfig?preserve-view=true&view=azure-ml-py
Time Series Forecasting: BYOC
Hyperparameter
Databricks Runtime for
Machine Learning
Supported VM series Restrictions
compute_target
Databricks Runtime 8.1 ML D None
to Dv2 None
Databricks Runtime 5.5 LTS ML Dv3 None
+ GPU Clusters
DSv2 None
DSv3 None
FSv2 None
HBv2 Requires Approval
HCS Requires Approval
M Requires Approval
NC None
NCsv2 Requires Approval
NCsv3 Requires Approval
NDs Requires Approval
NDv2 Requires Approval
NV None
NVv3 Requires Approval
Better Together Architecture
Better Together Architecture: ADB + AML
Enterprise Data
Third Party Data
Model Data
Better Together Architecture: ADB + AML
Staging
ADLS v2
Azure Synapse
Enterprise Data
Third Party Data
Loading
Model Data Data Workload
Better Together Architecture: ADB + AML
Staging
Pre-model
Pipeline
ADLS v2
AML
Azure Synapse
AI Lifecycle
Training &
Deployment
ADB
Enterprise Data
Third Party Data
Loading
Model Data AI/ML
Data Workload
Better Together Architecture: ADB + AML
Staging
Pre-model
Pipeline
ADLS v2
AML
Azure Synapse
AI Lifecycle
Training &
Deployment
Post-model
Pipeline
ADB
Enterprise Data
Third Party Data
Loading
Azure Synapse
Model Data Analytical Workload
AI/ML
Data Workload
Better Together Architecture: ADB + AML
Staging
Pre-model
Pipeline
ADLS v2
AML
Azure Synapse
PowerBI
Web Application
AI Lifecycle
Training &
Deployment
Post-model
Pipeline
ADB
Enterprise Data
Third Party Data
Loading
Azure Synapse
Model Data Analytical Workload
AI/ML
Data Workload Front-End
Azure AI: Many-Model Forecasting
Examples:
• Energy and utility companies: Predictive maintenancemodelsforthousands of oil
wells
• Retail organizations: Workforce optimization models for thousands of stores /
Price optimization models
• Restaurant chains:Demand forecasting models across thousands ofrestaurants
• Banks and financial institutes: Models for cash replenishmentfor ATM Machine
• Enterprises: revenue forecasting modelsat each division level
• Document management companies: Text analytics and legal document search
models per each state
Automate Production: All the Ops
MLOps DevOps
DataOps
Auto-train a time-series forecast model
Demo:
Prepare data for time series modeling.
Configure specific time-series parameters in an AutoMLConfig object.
Run inference with time-series data.
DEMO DEMO DEMO
Time Series Forecasting: Next Steps
Automation of model development process and finding the best performing model:
MLFlow: Manages the end-to-end model lifecycle, including tracking experimental runs,
deploying and sharing models, and maintaining a centralized model registry
Hyperopt: Augmented with the SparkTrials class, automates and distributes ML model
parameter tuning
Time Series Forecasting: Next Steps
Automation of model development process and finding the best performing model:
MLFlow: Manages the end-to-end model lifecycle, including tracking experimental runs,
deploying and sharing models, and maintaining a centralized model registry
Hyperopt: Augmented with the SparkTrials class, automates and distributes ML model
parameter tuning
Operationalization of Inference and Deployment:
GPU: Deploy a model on Azure Kubernetes Service (AKS) providing a GPU resource that is
used by the model for inference (highly parallelizable computation)
Docker Image: Docker manages your dependencies, maintain tighter control over
component versions or save time during deployment
Auto-Train a Time-Series Forecast Model With AML + ADB

Auto-Train a Time-Series Forecast Model With AML + ADB

  • 1.
    Better Together: Auto-train atime-series forecast model with AML & ADB Name: Moe Steller Title: Sr. Cloud Solution Architect, AI/ML
  • 2.
    Introduction Moe Steller Microsoft -WW Commercial Business Sr. Cloud Solution Architect, AI/ML Focus Areas: NLP / Deep Learning / Time-Series Scalable AI / AIOps RAI / Ethics Industry Focus: Healthcare, Finance, Supply Chain, Insurance, HLS
  • 3.
    Feedback Your feedback isimportant to us. Don’t forget to rate and review the sessions.
  • 4.
    Better Together: Auto-train atime- series forecast model with AML & ADB
  • 5.
    Agenda § Time SeriesForecasting § Product Preview § Advanced Parameters § Better Together Architecture § Demo
  • 6.
  • 7.
  • 8.
    Time Series Forecasting:Algorithm Support • Elastic Net • Light GBM • Gradient Boosting • Decision Tree • K Nearest Neighbors • LARS Lasso • Stochastic Gradient Descent • Random Forest • Extremely Randomized Trees • XGBoost • Online Gradient Descent Regressor • Fast Linear Regressor • Elastic Net • Light GBM • Gradient Boosting • Decision Tree • K Nearest Neighbors • LARS Lasso • Stochastic Gradient Descent • Random Forest • Extremely Randomized Trees • XGBoost • Auto-ARIMA • Prophet • ForecastTCN • Regression • Logistic Regression • Light GBM • Gradient Boosting • Decision Tree • K Nearest Neighbors • Linear SVC • Support Vector Classification (SVC) • Random Forest • Extremely Randomized Trees • XGBoost • Averaged Perceptron Classifier • Naive Bayes • Stochastic Gradient Descent • Linear SVM Classifier* • Classification • Time Series Forecasting
  • 9.
    Time Series Forecasting:Algorithm Support • Elastic Net • Light GBM • Gradient Boosting • Decision Tree • K Nearest Neighbors • LARS Lasso • Stochastic Gradient Descent • Random Forest • Extremely Randomized Trees • XGBoost • Online Gradient Descent Regressor • Fast Linear Regressor • Elastic Net • Light GBM • Gradient Boosting • Decision Tree • K Nearest Neighbors • LARS Lasso • Stochastic Gradient Descent • Random Forest • Extremely Randomized Trees • XGBoost • Auto-ARIMA • Prophet • ForecastTCN • Regression • Logistic Regression • Light GBM • Gradient Boosting • Decision Tree • K Nearest Neighbors • Linear SVC • Support Vector Classification (SVC) • Random Forest • Extremely Randomized Trees • XGBoost • Averaged Perceptron Classifier • Naive Bayes • Stochastic Gradient Descent • Linear SVM Classifier* • Classification • Time Series Forecasting
  • 10.
    Time Series Forecasting:Product Preview Models Description Benefits Prophet (Preview) Prophet works best with time series that have strong seasonal effects and several seasons of historical data. To leverage this model, install it locally using pip install fbprophet. Accurate & fast, robust to outliers, missing data, and dramatic changes in your time series. Auto-ARIMA (Preview) Auto-Regressive Integrated Moving Average (ARIMA) performs best, when the data is stationary. This means that its statistical properties like the mean and variance are constant over the entire set. For example, if you flip a coin, then the probability of you getting heads is 50%, regardless if you flip today, tomorrow or next year. Great for univariate series, since the past values are used to predict the future values. ForecastTCN ForecastTCN is a neural network model designed to tackle the most demanding forecasting tasks, capturing nonlinear local and global trends in your data as well as relationships between time series. Capable of leveraging complex trends in your data and readily scales to the largest of datasets.
  • 11.
    Time Series Forecasting:Advanced Parameters enable_stack_ensable Two-layer implementation: first layer has the same models as the voting ensemble, second layer model finding optimal combination of the models from the first layer enable_voting_ensable Voting implements soft-voting which uses weighted averages enable_onnx_compatible_models Get a pre-trained or generated ONNX model added spark_context Used inside Azure Databricks/Spark environment featurization Detected column type preprocessing/featurization enable_early_stopping If the score is not improving in the short term seasonality Set time series seasonality https://docs.microsoft.com/en-us/python/api/azureml-train-automl-client/azureml.train.automl.automlconfig.automlconfig?preserve-view=true&view=azure-ml-py
  • 12.
    Time Series Forecasting:BYOC Hyperparameter Databricks Runtime for Machine Learning Supported VM series Restrictions compute_target Databricks Runtime 8.1 ML D None to Dv2 None Databricks Runtime 5.5 LTS ML Dv3 None + GPU Clusters DSv2 None DSv3 None FSv2 None HBv2 Requires Approval HCS Requires Approval M Requires Approval NC None NCsv2 Requires Approval NCsv3 Requires Approval NDs Requires Approval NDv2 Requires Approval NV None NVv3 Requires Approval
  • 13.
  • 14.
    Better Together Architecture:ADB + AML Enterprise Data Third Party Data Model Data
  • 15.
    Better Together Architecture:ADB + AML Staging ADLS v2 Azure Synapse Enterprise Data Third Party Data Loading Model Data Data Workload
  • 16.
    Better Together Architecture:ADB + AML Staging Pre-model Pipeline ADLS v2 AML Azure Synapse AI Lifecycle Training & Deployment ADB Enterprise Data Third Party Data Loading Model Data AI/ML Data Workload
  • 17.
    Better Together Architecture:ADB + AML Staging Pre-model Pipeline ADLS v2 AML Azure Synapse AI Lifecycle Training & Deployment Post-model Pipeline ADB Enterprise Data Third Party Data Loading Azure Synapse Model Data Analytical Workload AI/ML Data Workload
  • 18.
    Better Together Architecture:ADB + AML Staging Pre-model Pipeline ADLS v2 AML Azure Synapse PowerBI Web Application AI Lifecycle Training & Deployment Post-model Pipeline ADB Enterprise Data Third Party Data Loading Azure Synapse Model Data Analytical Workload AI/ML Data Workload Front-End
  • 19.
    Azure AI: Many-ModelForecasting Examples: • Energy and utility companies: Predictive maintenancemodelsforthousands of oil wells • Retail organizations: Workforce optimization models for thousands of stores / Price optimization models • Restaurant chains:Demand forecasting models across thousands ofrestaurants • Banks and financial institutes: Models for cash replenishmentfor ATM Machine • Enterprises: revenue forecasting modelsat each division level • Document management companies: Text analytics and legal document search models per each state
  • 20.
    Automate Production: Allthe Ops MLOps DevOps DataOps
  • 21.
    Auto-train a time-seriesforecast model Demo: Prepare data for time series modeling. Configure specific time-series parameters in an AutoMLConfig object. Run inference with time-series data.
  • 22.
  • 23.
    Time Series Forecasting:Next Steps Automation of model development process and finding the best performing model: MLFlow: Manages the end-to-end model lifecycle, including tracking experimental runs, deploying and sharing models, and maintaining a centralized model registry Hyperopt: Augmented with the SparkTrials class, automates and distributes ML model parameter tuning
  • 24.
    Time Series Forecasting:Next Steps Automation of model development process and finding the best performing model: MLFlow: Manages the end-to-end model lifecycle, including tracking experimental runs, deploying and sharing models, and maintaining a centralized model registry Hyperopt: Augmented with the SparkTrials class, automates and distributes ML model parameter tuning Operationalization of Inference and Deployment: GPU: Deploy a model on Azure Kubernetes Service (AKS) providing a GPU resource that is used by the model for inference (highly parallelizable computation) Docker Image: Docker manages your dependencies, maintain tighter control over component versions or save time during deployment