DCCS

Official PyTorch implementation for ECCV'20 paper: Deep Image Clustering with Category-Style Representation

Coming soon

• A new clustering method which achieves 85.8% clustering accuracy on CIFAR-10 (with 0.8% standard deviations).

Package dependencies

• python >= 3.6
• pytorch == 1.2.0
• torchvision == 0.4.0
• scikit-learn == 0.21.3
• tensorboardX
• matplotlib
• numpy
• scipy

Create the environment with Anaconda

$ conda create -n dccs python=3.6
$ source activate dccs
$ conda install pytorch=1.2.0 torchvision=0.4.0 cudatoolkit=10.0 -c pytorch
$ conda install scikit-learn=0.21.3
$ pip install tensorboardX
$ conda install matplotlib

Prepare datasets

For MNIST, Fashion-MNIST, CIFAR-10 and STL-10, you can download the datasets using torchvision.

For example, you can download CIFAR-10 with

torchvision.datasets.CIFAR10('path/to/dataset', download=True)

For ImageNet-10, you can download ImageNet, select the images of 10 classes listed in './data/imagenet10_classes.txt', and resize the images to 96x96 pixels.

Command to run DCCS

You can run DCCS on MNIST with

$ CUDA_VISIBLE_DEVICES=0 python train.py --dataset-type=MNIST --dataset-path=path/to/dataset --beta-aug=2 

For CIFAR-10, you can use

$ CUDA_VISIBLE_DEVICES=0 python train.py --dataset-type=CIFAR10 --dataset-path=path/to/dataset --beta-aug=4 

Citation

If you are interested in our paper, please cite:

@inproceedings{zhao2020deep,
  title={Deep Image Clustering with Category-Style Representation},
  author={Zhao, Junjie and Lu, Donghuan and Ma, Kai and Zhang, Yu and Zheng, Yefeng},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2020}
}