This repo includes the code for the ICLR 2022 paper Chaos is a Ladder: A New Theoretical Understanding of Contrastive Learning via Augmentation Overlap.
The environment of our codes can be found in requirement.txt.
ARC (Average Relative Confusion) is a new unsupervised evaluation metric for contrastive learning proposed in this paper, which estimates the degree of ''chaos'' of augmented samples by calculating the ratio of samples whose nearest neighbour is from the same anchor. We find that it's strongly correlated to the downstream performance of contrastive learning process. We think that it may be helpful in model selection and some other tasks to improve the performance of contrastive learning.
To calculate ACR and ARC with trained models, run evaluate_ARC.py with following commands
./ARC_test.sh
The code for ACR and ARC calculation can be found in ARC.py.
We observe the relation between ARC and downstream performnace of SimCLR with RandomResizedCrop of different augmentation strength in the following figure
The code of the synthetic dataset can be found in synthetic_data.ipynb.
If you find our code useful, please cite
@inproceedings{
wang2022chaos,
title={Chaos is a Ladder: A New Understanding of Contrastive Learning},
author={Yifei Wang and Qi Zhang and Yisen Wang and Jiansheng Yang and Zhouchen Lin},
booktitle={International Conference on Learning Representations},
year={2022},
url={https://openreview.net/forum?id=ECvgmYVyeUz}
}
This baseline of this repo mainly borrows from SimCLR.