September 5th, 2013

Replay: Modeling MNIST With RF Hands-on Demo

RSS icon RSS Category: Uncategorized
Mnist table

Last week Spencer put together a great hands on for modeling data using H2O (http://www.meetup.com/H2Omeetup/).  This post is a write-up of the workflow for generating an RF model on MNIST data for those of you who want to walk through the demo again, or maybe missed the live action version.  I’m running through one of our local servers, with an allocation to H2O of 20 gigs.
RF on MNIST data: Spencer used a data set of pre-GPUed MNIST data similar to that provided by Kaggle in a currently running competition.  If you’re interested in some of the different approaches to the MNIST data (including Neural Nets and K Nearest Neighbors) I highly recommend taking a look at http://yann.lecun.com/exdb/mnist/.
Problem: The training data are 60,000 observations of 786 variables, testing data are 10,000 observations. Each independent variable corresponds to one square pixel of  an image. The value given for any variable indicates the level of saturation of the pixel.  Results are given and discussed below. Here is the step by step process for generating these results.

  1. Starting at the drop down menu Data inhale and parse data (both the testing and training sets).
  2. From the Model drop down menu choose **Random Forest**
    Mnist table
  3. Set Ntree = 50, and Features = 200. Leave all other options in default.  Note that H2O automatically ignores all constant columns, so you need not sort through the data summary by hand  to find those variables.Request Rf form
  4. Step 3 generates a model, the confusion matrix shown below is the output of this model.
  5. The model key is at the top of the RF results page; highlight and copy it. From the drop down menu Score, select RF.RFview data key
  6. In the specification page for scoring your RF model enter the .hex  key for your testing data, paste the model key, specify the dependent variable column, and submit.Request RFscore

At this point you have built a model and verified that it works. In practice, the motivation is generally to actually predict an outcome of interest – which you can now do with this same model by returning to the drop down menu Score and selecting Predict. Feeding  Predict data with the same predictors as contained in your training set produces a column of predictions matching each observation.
Results:  In an RF model of 50 trees,  features set to 200, and all other options left in default, H2O produces this confusion matrix.Confusion matrix
Testing the generated RF model on the test set  produces a classification error of 3.28%.Confusion matrix full scoring
So- there you have it. A walkthrough of Spencer’s meetup presentation that you can follow step by step.

Leave a Reply

An Introduction to Time Series Modeling: Traditional Time Series Models and Their Limitations

In the first article in this series, we broke down the preprocessing and feature engineering

December 3, 2021 - by Adam Murphy
Announcing the Fully Managed H2O AI Cloud

The H2O AI Cloud is the leading platform to make and access your own AI

December 1, 2021 - by Michelle Tanco
H2O.ai Tools for a Beginner

Note: this is a community blog post by Shamil Dilshan Prematunga. It was first published

November 30, 2021 - by Jo-Fai Chow
Amazon Redshift Integration for H2O.ai Model Scoring

We consistently work with our partners on innovative ways to use models in production here

November 22, 2021 - by Eric Gudgion
Building Resilient Supply Chains with AI

A global pandemic, a fundamental shift in the demand for goods and services worldwide, and

November 11, 2021 - by Adam Murphy
Introducing the H2O.ai Wildfire Challenge

We are excited to announce our first AI competition for good - H2O.ai Wildfire Challenge. We’ve

November 5, 2021 - by Jo-Fai Chow

Start your 14-day free trial today