SPICE: Semantic Pseudo-labeling for Image Clustering

By Chuang Niu, Hongming Shan, and Ge Wang

This is a Pytorch implementation of the paper.

• SOTA on 5 benchmarks. Please refer to Papers With Code for Image Clustering

Installation

Please refer to requirement.txt for all required packages. Assuming Anaconda with python 3.7, a step-by-step example for installing this project is as follows:

conda install pytorch==1.6.0 torchvision==0.7.0 cudatoolkit=10.1 -c pytorch
conda install -c conda-forge addict tensorboard python-lmdb
conda install matplotlib scipy scikit-learn pillow

Then, clone this repo

git clone https://github.com/niuchuangnn/SPICE.git
cd SPICE

Data

Prepare datasets of interest as described in dataset.md.

Training

Read the training tutorial for details.

Evaluation

Evaluation of SPICE-Self:

python tools/eval_self.py --config-file configs/stl10/eval.py --weight PATH/TO/MODEL --all 1

Evaluation of SPICE-Semi:

python tools/eval_semi.py --load_path PATH/TO/MODEL --net WideResNet --widen_factor 2 --data_dir PATH/TO/DATA --dataset cifar10 --all 1 

Read the evaluation tutorial for more descriptions about the evaluation and the visualization of learned clusters.

Model Zoo

All trained models in our paper are available as follows.

Dataset	Version	ACC	NMI	ARI	Model link
STL10	SPICE-Self	91.0	82.0	81.5	Model
	SPICE	93.8	87.2	87.0	Model
	SPICE-Self*	89.9	80.9	79.7	Model
	SPICE*	92.9	86.0	85.3	Model
CIFAR10	SPICE-Self	83.8	73.4	70.5	Model
	SPICE	92.6	86.5	85.2	Model
	SPICE-Self*	84.9	74.5	71.8	Model
	SPICE*	91.7	85.8	83.6	Model
CIFAR100	SPICE-Self	46.8	44.8	29.4	Model
	SPICE	53.8	56.7	38.7	Model
	SPICE-Self*	48.0	45.0	30.8	Model
	SPICE*	58.4	58.3	42.2	Model
ImageNet-10 (ImageNet Pretrained)	SPICE-Self	96.9	92.7	93.3	Model
	SPICE	96.7	91.7	92.9	Model
ImageNet-Dog (ImageNet Pretrained)	SPICE-Self	54.6	49.8	36.2	Model
	SPICE	55.4	50.4	34.3	Model
ImageNet-10	SPICE-Self	92.1	82.8	83.6	Model
	SPICE	95.9	90.2	91.2	Model
ImageNet-Dog	SPICE-Self	64.6	57.2	47.9	Model
	SPICE	67.5	62.7	52.6	Model
TinyImageNet	SPICE-Self	30.5	44.9	16.3	Model
	SPICE-Self*	29.2	52.5	14.5	Model

More models based on ResNet18 for both SPICE-Self* and SPICE-Semi*.

Dataset	Version	ACC	NMI	ARI	Model link
STL10	SPICE-Self*	86.2	75.6	73.2	Model
	SPICE*	92.0	85.2	83.6	Model
CIFAR10	SPICE-Self*	84.5	73.9	70.9	Model
	SPICE*	91.8	85.0	83.6	Model
CIFAR100	SPICE-Self*	46.8	45.7	32.1	Model
	SPICE*	53.5	56.5	40.4	Model

Acknowledgement for reference repos

• GATCluster
• MOCO
• SCAN
• FixMatch
• IIC

Citation

@misc{niu2021spice,
      title={SPICE: Semantic Pseudo-labeling for Image Clustering}, 
      author={Chuang Niu and Ge Wang},
      year={2021},
      eprint={2103.09382},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}