ASzot/ClusterCNN: Using K-means Clustering For Unsupervised CNN ...

ClusterCNN

Summary

Use k-means clustering from the input data set to partially train the network by obtain weights for the convolutional filters. Then compare the impact stochastic gradient descent has on this already partially trained network.

Motivation

Training Convolutional Neural Networks (CNNs) has typically taken vast amounts of labeled data. Obtaining this labeled data is extremely difficult often requiring countless hours of manual annotation. This experiment serves to explore if clustering techniques can be used to automatically set the weights of a CNN. The motivation behind this experiment is viewing convolution filter weights as anchor vectors as described in this paper.

Requirements

• Python 3.6
• numpy
• scipy
• matplotlib
• sklearn
• Theano
• Keras
• MulticoreTSNE
• SphereCluster

Installation

I recommend using Anaconda to manage packages and Python versions. You can find the installer for Anaconda here.

• Create Anaconda environment
  • conda create -n yourenvname python=3 anaconda
  • This will initialize a conda environment with all of the default packages many of which are needed for this project.
• Activate the conda environment
  • source activate yourenvname
• Install additional packages
  • Unfortunately installing Theano and Keras through the conda install command does not work.
  • pip install Theano
  • pip install Keras Be sure to install Keras < version 2.0. The code will not work with version 2.
  • Install the MutlicoreTSNE package:
    • git clone https://github.com/DmitryUlyanov/Multicore-TSNE.git
    • cd Multicore-TSNE
    • pip install -r requirements.txt
    • python setup.py install
    • Go to MulticoreTSNE GitHub to see further instructions on how to install the MulticoreTSNE Python package.
  • SphereCluster:
    • git clone https://github.com/clara-labs/spherecluster.git
    • cd spherecluster
    • python setup.py install
    • Go to SphereCluster GitHub for more information.
  • Upgrade to the most recent version of sklearn to get the IsolationForrest:
    • pip install sklearn --upgrade
• Deactivate your conda environment
  • deactivate yourenvname

K-Means

The effectiveness of the clustering can be determined via the silhouette score.

0.71-1.0 A strong structure has been found

0.51-0.70 A reasonable structure has been found

0.26-0.50 The structure is weak and could be artificial

< 0.25 No substantial structure has been found