Thursday, August 24th, 2017 Meeting

Abstract

This is part 2 of a two part series. Part 1 was presented in early 2016.

In this talk Dan shows you how to build a sophisticated Machine Learning App from the ground up using Tensor Flow (Wrapped by Keras) and packages in Anaconda Python. We start by building a simple API server, with Python package Flask-RESTful, which responds to GET requests containing parameters which affect the behavior of a Keras backend server.

We use Keras-Tensor-Flow to create a Deep Learning model and then generate predictions for any stock symbol of interest. The predictions are affected by features selected by end-user such as various slopes of price-moving-avg and/or date features like Day-of-Week and/or Month-of-Year. Building a Deep Learning model is expensive so we use Postgres to cache each model in case we encounter another identical request in the near future. Additionally we cache all predictions which could be generated from each model. Next in the workflow we use Flask-RESTful to serve predictions to the end-user as JSON documents.

For deployment, we offer two techniques. Technique 1 is simple. Just deploy the app to an EC2 server and configure Flask-RESTful to listen on the public interface. Technique 2 might be more cost-effective. Find some unused hardware in your office which can run Ubuntu 16. Then deploy the app to that server. Next, deploy the Flask-RESTful piece to Heroku which should be free or low cost. Finally, configure both pieces so they communicate using shared data on a Postgres database hosted at Heroku. Each piece will connect to Postgres using Python package SQLAlchemy.

Python is awesome technology and it is perfect for building an elegant Machine Learning Application.

Speaker Bio

Dan Bikle is a graduate of Caltech and works as an independent Data Scientist. He is skilled at extracting knowledge (forecasts mostly) from data. Also, Dan teaches three classes at Santa Clara Adult Education: Hands on Python, Time Series Data Science, and Machine Learning Applications.