Awesome
Information based Deep Clustering: An experimental study
This code accompanies the paper Information based Deep Clustering: An experimental study.
Abstract:
Recently, two methods have shown outstanding perfor-mance for clustering images and jointly learning the fea-ture representation. The first, called Information Maximiz-ing Self-Augmented Training (IMSAT), maximizes the mu-tual information between input and clusters while using aregularization term based on virtual adversarial examples.The second, named Invariant Information Clustering (IIC),maximizes the mutual information between the clustering ofa sample and its geometrically transformed version. Thesemethods use mutual information in distinct ways and lever-age different kinds of transformations. This work proposesa comprehensive analysis of transformation and losses fordeep clustering, where we compare numerous combinationsof these two components and evaluate how they interactwith one another. Results suggest that mutual informationbetween a sample and its transformed representation leadsto state-of-the art performance for deep clustering, espe-cially when used jointly with geometrical and adversarialtransformations.
How to use:
git clone --recursive https://github.com/jizongFox/DeepClusteringProject.git
If you use conda, you'd prefer using a virtual environment:
conda create -p env python=3.7
conda activate env
Then install backend library:
cd library/deep-clustering-toolbox
python setup install
The installed package support Apex to speed up and reduce the memory usage. Following installation guide
for a faster training.
The main.py is the main entrance of the program and the trainer/* defines the training behaviors.