MC-SSL: Towards Multi-Concept Self-Supervised Learning

This repository contains the official PyTorch self-supervised pretraining, finetuning, and evaluation codes for

MCSSL: Towards Multi-Concept Self-Supervised Learning.

Main Architecture

Visualization of Self-supervised Clustering

Self-supervised pre-training

python -m torch.distributed.launch --nproc_per_node=8 --use_env main_MCSSL.py --batch_size 64 --epochs 800 --data_location 'path/to/imageNet/trainingimgs'

Architecture	# paramters	Finetuning Accuracy	download
ViT-S/16	22M	82.4 %	checkpoint
ViT-B/16	85M	84.0 %	checkpoint

Finetuning

We rely on the finetuning strategy of Deit

Acknowledgement

This repository is built in top of the SiT and the DINO repository.

Citation

If you use this code for a paper, please cite:

@article{atito2021mc,

  title={MC-SSL0. 0: towards multi-concept self-supervised learning},

  author={Atito, Sara and Awais, Muhammad and Farooq, Ammarah and Feng, Zhenhua and Kittler, Josef},

  journal={arXiv preprint arXiv:2111.15340},

  year={2021}

}