#Recognizing Handwritten Digits Using K-Means Clustering

By: Olivia Angiuli, Martin Reindl, Ty Rocca, Wilder Wohns

#####Required Packages

• python image library (PIL)
  • pip install PIL
• numpy, scipy, scikit-learn:
  • pip install -U numpy scipy scikit-learn
• django version 1.7:
  • pip install django
• iPython Notebook
  • pip install "ipython[notebook]"
• CVXOPT
  • pip install cvxopt --user” (or follow http://cvxopt.org/install/)

To run our code you need to first be in the /code folder of our project. Onece there, our code takes up to 4 command line arguments. The format for this is:

• python main_cluster.py {k} {method} {init_type} {prop}

  • k is the number of clusters
  • method is “means”, “medians”, or “medoids”
  • init_type is “random” or “kplusplus”
  • prop is a number from 0 to 100 which is the percentage of training data to train on

• Running the code in nohup

  • Running the code as a nohuped process is nice becasue when using the full dataset clustering can take up to 4 hours. This no hup format will still give you all of the images and the printed out put will be saved in the *.out folder
  • nohup python main_cluster.py {k} {method} {init_type} {prop} > {method}_{init_type}_clusters{k}.out&

You also have the option to explore our code in an iPython Notebook. The iPython Notebook is a great top level interface for interacting with our code, as well as seeing all of our plots, and a bit of how we process the image in the django app. You can view a static version of our notebook here or you can run it from the code folder with the command: ipython notebook. Within the notebook the main function will run the code much like the one above. Main take 3 arguments (k, method, init_type). Method is by default "means" and initialization type is by default "kplusplus". You can change "k" at the top of the page to change how much data is loaded.

Our Django app is another fun bonus part of the our project. The django app allows us to run a front end interface where our users can see our clusters get put to the test. The bulk of the code is in code/recognize_digits/views.py. This is where all of the backend image processing happens. The font end javascript is stored in code/static/js/homepage.js. This makes the interactive canvas.

• To run the Django app:
  • cd into \code
  • run python manage.py runserver 8000 to run the app on your computer
    • If it is your first time running the app you might need to run python manage.py make migrations
  • Then go to http://localhost:8000/ where the app should now be running
  • Draw a digit and click predict!

• To run the optimization function (which will take a very long time to run since it clusters 102 different ways type python optimize.py and it will run the optimization of the code.

In ../results you will find the outputs from running all possible clusters. Each file is a different method/initialization/k value. The files contain pictures of images from each cluster as well as the clusters means. Each file also contains a .json output which holds the results from running that trial. The json is formatted

{ "k" : k, 
    "prediction_accuracy" : predictions[1], #c ontains prediction percent for outputted clusters
    "cluster_means" : cluster_set, # contains cluster means
    "cluster_stats" : final,    # the results from clustering, this give you the breakdown and purity of each                                        # cluster for each number
    "clustering_time" : clustering_time, # time to cluster in seconds
    "testing_time" : testing_time # time to process the test data and assign clusters
}

**Comparing With SciKit-Learn: ** You can also compare our clusters to those of scikitlearn. The command for that is python kmeans_scikitlearn.py 10

Comparing with Neural Network: You can also compare our algorithim with this implementation of a neural network (we followed this guide) The neural network is executed from the /code/neural_network/ directory. To run execute this command type: python main_neural_network.py. You can also type ipython notebook from this directory and see the iPython notebook version of the code.