2021-CVPR-MvCLN
This repo contains the code and data of the following paper accepted by CVPR 2021
Requirements
pytorch==1.5.0
numpy>=1.18.2
scikit-learn>=0.22.2
munkres>=1.1.2
logging>=0.5.1.2
Configuration
The hyper-parameters, the training options (including the ratiao of positive to negative, aligned proportions and switch time) are defined in the args. part in run.py.
Datasets
The Scene-15 and Reuters-dim10 datasets are placed in "datasets" folder. The NoisyMNIST and Caltech101 datasets could be downloaded from Google cloud or Baidu cloud with password "rqv4".
Usage
After setting the configuration and downloading datasets from the cloud desk, one could run the following code to verify our method on NoisyMNIST-30000 dataset for clustering task.
python run.py --data 3
The expected outputs are as follows:
******** Training begin, use RobustLoss: 1.0*m, use gpu 0, batch_size = 1024, unaligned_prop = 0.5, NetSeed = 64, DivSeed = 249 ********
=======> Train epoch: 0/80
margin = 5
distance: pos. = 2.5, neg. = 2.5, true neg. = 2.5, false neg. = 2.49
loss = 3.41, epoch_time = 12.07 s
******** testing ********
CAR=0.1012, kmeans: acc=0.1791, nmi=0.0435, ari=0.021
******* neg_dist_mean >= 1.0 * margin, start using fine loss at epoch: 3 *******
=======> Train epoch: 10/80
distance: pos. = 0.76, neg. = 5.38, true neg. = 5.83, false neg. = 1.34
loss = 0.09, epoch_time = 15.17 s
******** testing ********
CAR=0.8712, kmeans: acc=0.9462, nmi=0.8705, ari=0.8862
......
=======> Train epoch: 80/80
distance: pos. = 0.25, neg. = 5.34, true neg. = 5.8, false neg. = 1.17
loss = 0.03, epoch_time = 14.18 s
******** testing ********
CAR=0.8753, kmeans: acc=0.9459, nmi=0.8744, ari=0.8859
******** End, training time = 1276.29 s ********
Citation
If you find our work useful in your research, please consider citing:
@inproceedings{yang2021MvCLN,
title={Partially View-aligned Representation Learning with Noise-robust Contrastive Loss},
author={Mouxing Yang, Yunfan Li, Zhenyu Huang, Zitao Liu, Peng Hu, Xi Peng},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month={June},
year={2021}
}
