Introduction to TensorFlow

In this tutorial the steps needed to clean a dataset and prepare it for modeling using the machine learning library TensorFlow. The tutorial uses the Wine dataset from the UCI Machine Learning Repository.

Prerequisites

This tutorial includes several machine learning terms. To get a good mathematical understanding of these concepts, please read the Math Primer.

Installation Notes

There are a few packages you will need in order to run this tutorial. We recommend installing the miniconda environment, which makes the installation process quite easy. Please see the README file for this mornings session for instructions on how to install miniconda.

In order to run this tutorial, you will need the following Python packages:

• numpy 1.11 or later
• pandas 0.18 or later
• matplotlib 1.5 or later
• seaborn 0.7 or later
• scikit-learn 0.17.1 or later
• six 1.10.0 or later
• jupyter
• tensorflow 0.8.0

The first seven packages can be installed with the following command:

pip install seaborn scikit-learn jupyter

Alternatively if you are using conda you can do:

conda install seaborn scikit-learn jupyter

For TensorFlow, the installation depends on your environment. Below are installation instructions. For detailed instuctions, please see the TensorFlow Download and Setup page.

Note, skflow is now part of the TensorFlow library. Once you have installed TensorFlow, you can load skflow with the following command:

import tensorflow.contrib.learn as skflow

For detailed instructions about skflow, please read Skflow Readme.

NOTE:

What's better to use? The virtual environment or normal pip installation?

Playing With Outliers

I have added a fun interactive application using the Python visualization library called Bokeh. The app allows you to pick features from the wine data set and define an outlier threshold to explore how this affects the data. The application source code is in the playing_with_outliers directory and is called main.py. To run this application, you will need to install bokeh:

pip install bokeh

Then, to run the application, download the playing_with_outliers directory and its contents. Then, in the directory where you downloaded it, run:

bokeh serve --show playing_with_outliers

The application will open in your browser.