Information Maximization Clustering via Multi-View Self-Labelling

Introduction

**This is an implementation code written in Python (version 3.6.9) of IMC-SwAV based on manuscript paper **

Performance

The reported performance of our proposed model is based on ResNet18 encoder architecture.
We train our IMC-SwAV for 15 independent runs on training set and we report the result of testing set only.

Average Performance

Table of average performance and the corresponding standard deviation.

Dataset	Acc	NMI	ARI
CIFAR-10	89.1 (±0.5)	81.1 (±0.7)	79.0 (±1.0)
CIFAR-100-20	49.0 (±1.8)	50.3 (±1.2)	33.7 (±1.3)
STL10	83.1 (±1.0)	72.9 (±0.9)	68.5 (±1.4)
Tiny-Imagenet	27.9 (±0.3)	48.5 (±2.0)	14.3 (±2.1)

Best Performance

Below table reports the best recorded performance from our model.

Dataset	Acc	NMI	ARI
CIFAR-10	89.7	81.8	80.0
CIFAR-100-20	51.9	52.7	36.1
STL10	85.3	74.7	71.6
Tiny-Imagenet	28.2	52.6	14.6

Below, we report separate the result of our proposed IMC-SwAV in CIFAR-100 experiment (100 class)

Dataset	Top-1 ACC	Top-5 ACC	NMI	ARI
CIFAR-100	45.1	67.5	60.8	30.7

Usage

All hyper parameters apply across all datasets (default setup/experiment) in the submission document as following:

Settings related with the multi-crop
--nmb_crops 2 4
--max_scale_crops 1. 0.4
--min_scale_crops 0.2 0.08

Settings related with SwAV
--tau 0.1
--eps 0.05
--project_dim 128
--prototypes 1000

Settings related to the training
--learning_rate 0.0005
--warmup 500
--l2_w 1e-5

Settings related to the dataset
--path ROOT_DIRECTORY_OF_THE_DATASET (the path folder of the dataset)

To run any of the code, the directory path of the dataset is required
otherwise it will automatically download to './dataset'

CIFAR-10

To run the training code.

python train.py --dataset C10 --path ./dataset --size_crops 28 18 \
      --output_dir ./c10 --cpt_name c10.cpt 

CIFAR-100/20

To run the training code.

python train.py --dataset C20 --path ./dataset --size_crops 28 18  \
      --output_dir ./c20 --cpt_name c20.cpt

STL10

To run the training code.

python train.py --dataset STL10 --path ./dataset --size_crops 76 52 \
      --output_dir ./stl10 --cpt_name stl10.cpt --path ./dataset

CIFAR100

To run the training code.

python train.py --dataset C100 --path ./dataset --size_crops 28 18  --batch_size 512 \
      --output_dir ./c100 --cpt_name c100.cpt

Tiny-Imagenet

To run the training code.

python train.py --dataset tiny --path ./dataset --size_crops 56 36 --batch_size 512 \
      --output_dir ./tiny --cpt_name tiny.cpt

The evaluation of the model.

Example evaluation on CIFAR10/100-20/100:

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

python test.py --dataset c10 --path ./dataset --size_crops 28 18 --cpt_load_path ./c10/imc_swav.cpt

python test.py --dataset c20 --path ./dataset --size_crops 28 18 --cpt_load_path ./c20/imc_swav.cpt

python test.py --dataset c100 --path ./dataset --size_crops 28 18 --cpt_load_path ./c100/imc_swav.cpt

Example evaluation on STL10:

python test.py --dataset STL10 --path ./dataset --size_crops 76 52 --cpt_load_path ./stl10/imc_swav.cpt

Example evaluation on Tiny-Imagenet:

python test.py --dataset tiny --path ./dataset --size_crops 56 36 --cpt_load_path ./tiny/tiny.cpt

Notes

• During the training, each epoch reports the model's performance on test (validation set) and the training set (performance on training set is based on transformed instances).

• 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 encoder and prototypes.

• All tests have been performed in Cuda version 10.1.

Acknowledgement for reference repos

• AMDIM
• SwAV

Citation

@misc{ntelemis2021information,
      title={Information Maximization Clustering via Multi-View Self-Labelling}, 
      author={Foivos Ntelemis and Yaochu Jin and Spencer A. Thomas},
      year={2021},
      eprint={2103.07368},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}