SlideShare a Scribd company logo

Design Patterns for Smart Applications

•Download as PPTX, PDF•
•
Sri Ambati
Sri Ambati

Tom Kraljevic walks through building a machine learning web application with an H2O generated POJO model, 09.17.15.

1 of 8
Building a Smarter WebApp with a Generated POJO Model H2O.ai Tom Kraljevic September 17, 2015
Target Audience • JavaScript WebApp developer who wants to incorporate machine learning • More code, less PowerPoint
Building a Smarter Application • Q: What is a smarter application? • A: An app that learns from data [From rules-based to model-based]
Software Pieces • Front-end – Web browser – JavaScript application (run in the browser) • Back-end – Jetty servlet container – H2O-generated model POJO (hosted by servlet container) • Out-of-band – H2O
Overview Step 1: Picking the question you want your model to answer Step 2: Using your data to build a model Step 3: Exporting the generated model as a Java POJO Step 4: Compiling the model Step 5: Hosting the model in a servlet container Step 6: Running the JavaScript app in a browser Step 7: Using a REST API to make predictions Step 8: Incorporating the prediction into your application
Resources • This meetup – https://github.com/h2oai/h2o- meetups/tree/master/2015_09_17_PojoWebApp • H2O-3 Slater Release – http://h2o-release.s3.amazonaws.com/h2o/rel- slater/1/index.html • Generated POJO Model Javadoc – http://h2o-release.s3.amazonaws.com/h2o/rel- slater/1/docs-website/h2o-genmodel/javadoc/index.html • POJO REST API Example – https://github.com/h2oai/h2o-world-2015- training/tree/master/tutorials/pojo_webapp
Demonstration
Next Steps for a Real Use Case • Scoring (judging how good the predictions really are) – Need to get the correct answers from somewhere • Storing predictions (and the correct answers) – Often Hadoop. Seems easy, but this can be real work to organize • Model update frequency – From the stored data above – Need depends on the problem – Hourly, daily, monthly? – Cost of training the model becomes a factor • Hot swapping the model – Separating front-end and back-end makes this easier – Java reflection for in-process hot-swap – Load balancer for servlet container hot-swap

More Related Content

Design Patterns for Smart Applications

  • 1. Building a Smarter WebApp with a Generated POJO Model H2O.ai Tom Kraljevic September 17, 2015
  • 2. Target Audience • JavaScript WebApp developer who wants to incorporate machine learning • More code, less PowerPoint
  • 3. Building a Smarter Application • Q: What is a smarter application? • A: An app that learns from data [From rules-based to model-based]
  • 4. Software Pieces • Front-end – Web browser – JavaScript application (run in the browser) • Back-end – Jetty servlet container – H2O-generated model POJO (hosted by servlet container) • Out-of-band – H2O
  • 5. Overview Step 1: Picking the question you want your model to answer Step 2: Using your data to build a model Step 3: Exporting the generated model as a Java POJO Step 4: Compiling the model Step 5: Hosting the model in a servlet container Step 6: Running the JavaScript app in a browser Step 7: Using a REST API to make predictions Step 8: Incorporating the prediction into your application
  • 6. Resources • This meetup – https://github.com/h2oai/h2o- meetups/tree/master/2015_09_17_PojoWebApp • H2O-3 Slater Release – http://h2o-release.s3.amazonaws.com/h2o/rel- slater/1/index.html • Generated POJO Model Javadoc – http://h2o-release.s3.amazonaws.com/h2o/rel- slater/1/docs-website/h2o-genmodel/javadoc/index.html • POJO REST API Example – https://github.com/h2oai/h2o-world-2015- training/tree/master/tutorials/pojo_webapp
  • 8. Next Steps for a Real Use Case • Scoring (judging how good the predictions really are) – Need to get the correct answers from somewhere • Storing predictions (and the correct answers) – Often Hadoop. Seems easy, but this can be real work to organize • Model update frequency – From the stored data above – Need depends on the problem – Hourly, daily, monthly? – Cost of training the model becomes a factor • Hot swapping the model – Separating front-end and back-end makes this easier – Java reflection for in-process hot-swap – Load balancer for servlet container hot-swap

Editor's Notes

  1. An app that makes predictions using a model to affect program behavior (and subsequently affect user behavior!)