Unsupervised-classification-of-datasets

10-701 Course Project

Overview

In this project, we studies the problem of unsupervised visual learning. We investigate both classical and state of the art unsupervsied learning methods and compare their performance. To analyze these methods, we choose the task of image classifcation. We run our experiments on the popular MNIST dataset, and also extend our experiments to a self-generated 3D shape dataset.

Folder Structure

.
├── Capsules                   # Folder contains experiments about Stacked Capsule Autoencoders
├── MoCo                       # Folder contains experiments about MoCo (Momentum Contrast)
├── Autoencoder                # Folder contains experiments about Autoencoder    
├── InfoGAN                    # Folder contains experiments about InfoGAN
├── scripts                    # Scripts to run different baseline methods such as GMM & K-Means
└── README.md

Environment

To install the repective environment, refer to Capsules, MoCo, InfoGAN, and Autoencoder for installation details.

Dataset

Our experiments are done on MNIST & self-rendered shapenet dataset. The shapeNet dataset can be downloaded here.

Run baseline methods (GMM, K-means)

To run baseline methods, use the following commands:

python scripts/gmm_clustering.py --dataset mnist

To run shapeNet tests, remember to change the the dataset directory to your extracted dataset directory.

python scripts/gmm_clustering.py --dataset shapenet

python scripts/kmeans_clustering.py --dataset mnist

Run MoCo, Capsule, VAE experiments

Please refer to Capsules, MoCo, InfoGAN, and Autoencoder for training and testing details.