Image Clustering Using an Augmented Generative Adversarial Network and Information Maximization

Introduction

**This is an implementation code written in Python (version 3.6.9) of a manuscript paper **

Performance

The reported performance of our proposed model is based on custom architecture.
We train our unsupervised proposed method for 5 independent runs on training and testing set and we report the results accordingly.

Best and Average Performance

Below table reports the best and average recorded performance from our model.

Dataset	Best Acc	Aver Acc
MNIST	99.02%	98.85% (±0.14%)
CIFAR-10	70.04%	69.22% (±0.83%)
CIFAR-100/20	32.44%	30.88% (±0.14%)
STL10	58.65 %	74.7% (±1.81%)

Usage

All hyper-parameters of the reported accuracy are stored in 'models.py'. To run the training code.

MNIST

python train.py --dataset mnist --store_path ./models/mnist/mnist.ckpt

CIFAR-10

To run the training code.

python train.py --dataset c10 --store_path ./models/c10/c10.ckpt

CIFAR-100/20

To run the training code.

python train.py --dataset c100 --store_path ./models/c100/c100.ckpt

STL10

To run the training code.

python train.py --dataset stl10 --store_path ./models/stl10/stl10.ckpt

The evaluation of the model.

Example evaluation on MNIST, CIFAR10/100-20 & STL10:

Through the argument '--store_path', the full path of the stored model is parsed.

python test.py --dataset mnist --store_path ./models/mnist/mnist.ckpt

python test.py --dataset c10 --store_path ./models/c10/c10.ckpt

python test.py --dataset c100 --store_path ./models/c100/c100.ckpt

python test.py --dataset stl10 --store_path ./models/stl10/stl10.ckpt

Notes

• The classifier head is trained and evaluated only for labelled set on STL10 dataset. The unlabelled part of STL10 is used only to train the GAN model.

• All tests have been performed in Cuda version 10.1.

Citation

@ARTICLE{9451540,
  author={Ntelemis, Foivos and Jin, Yaochu and Thomas, Spencer A.},
  journal={IEEE Transactions on Neural Networks and Learning Systems}, 
  title={Image Clustering Using an Augmented Generative Adversarial Network and Information Maximization}, 
  year={2021},
  volume={},
  number={},
  pages={1-14},
  doi={10.1109/TNNLS.2021.3085125}}