jeperez / ml_demo

machine learning demos with mallet, deeplearning4j and surus

Machine Learning Demos

The content of this repository allows to run three machine learning demos.

1. Clone this repo to a directory of your choice
2. Create the docker images for the demos
3. Run the demos

Demo Quick Facts

• The three demos cover different areas of machine learning
  • Natural Language Processing
  • Pattern Recognition
  • Outlier Detection
• Each demo shows a different Java toolkit
  • Mallet
  • Deeplearning4j
  • Standalone Java code and Surus
• Everything is open and freely available
  • Machine learning toolkits
  • Data for examples
  • Related publications

Repository Organisation

For each demo a different docker image setup is provided and the content of the repository is organized accordingly.

 +-- README.md (the file you are now reading)
 |
 +-- image_javadev
 |      +-- Dockerfile (parent dockerfile for java development)
 |
 +-- image_mallet
 |      +-- Dockerfile (setup for Mallet toolkit)
 |      +-- data
 |      |     +-- rt-polarity.* files (movie review data for sentiment analysis example)
 |      +-- papers
 |            +-- sentiment analysis paper
 |
 +-- image_deeplearning4j
 |      +-- Dockerfile (setup for deeplearning4j toolkit)
 |      +-- data
 |      |     +-- nist files (handwritten digits)
 |      +-- models
 |      |     +-- pretrained lenet5 deep neural net models at 1, 5, and 10 epochs of the training
 |      +-- papers
 |      |     +-- deep neural network papers: lenet, imagenet, facenet
 |      +-- dl4j-demo
 |            +-- ... maven project with demo code: LeNetMnistTrainer.java and LeNetMnistTester.java
 |      
 +-- image_outlier
        +-- Dockerfile (setup for outlier detection)
        +-- data
        |     +-- wdbc files (diagnostic breast cancer)
        +-- papers
        |     +-- robust local outlier factor (lof) and robust pca
        +-- outlier-demo
              +-- ... maven project with demo code LofTester.java

This has the benefit that the demos are independent of each other and allows for simple cleanup once the demos are no longer used.

Create Docker Images

docker build --tag javadev:latest image_javadev
docker build --tag mallet:latest image_mallet
docker build --tag dl4j:latest image_deeplearning4j
docker build --tag outlier:latest image_outlier

Sentiment Analysis with Mallet

Start a Mallet demo container.

docker run -it mallet:latest

Inside the container check out the data files

wc -l data/*.txt
head -7 data/rt-polarity.test.txt | tail -4 | cut -c -150

Import the data into the Mallet format and train a Naive Bayes classifier

mallet import-file --input data/rt-polarity.train.txt --output data.mallet.sentiment.train
mallet train-classifier --trainer NaiveBayes --training-portion 0.9 --input data.mallet.sentiment.train --output-classifier model.mallet.naivebayes
ls -l

Use the classifier to determine the sentiment of some new data

mallet classify-file --classifier model.mallet.naivebayes --input data/rt-polarity.test.txt --output - | head -7
head -7 data/rt-polarity.test.txt | cut -c -150

Recognition of Handwritten Digits with Deeplearning4j

Start the deeplearning4j container

docker run -it dl4j:latest

Inside the container check out the data files

ls -l data/*

Training deep neural net takes hours. That's why there are some pretrained nets

ls -l models/*

Evaluate a model. First on 10,000 digits then on a single image

java -cp /usr/share/dl4j-demo/dl4j-demo.jar org.ece16.dl4j.LeNetMnistTester models/lenet.dl4j.epoch10.model
java -cp /usr/share/dl4j-demo/dl4j-demo.jar org.ece16.dl4j.LeNetMnistTester models/lenet.dl4j.epoch10.model 115

You can also train your own net

java -Xmx1024m -cp dl4j-demo.jar org.ece16.dl4j.LeNetMnistTrainer

Detecting Breast Cancer with Custom Outlier Detection Code

Start the outlier container

docker run -it outlier:latest

Inside the container check out the data files

ls -l data/wdbc*

Search for outliers in the wdbc data with a varying value of k and thresold 2.4

java -cp /home/demo/outlier-demo/outlier-demo.jar org.ece16.lof.LofTester data/wdbc.data 2.4

Search for outliers in the wdbc data with k=40 and thresold 2.4

java -cp /home/demo/outlier-demo/outlier-demo.jar org.ece16.lof.LofTester data/wdbc.data 2.4 40