This repository contains PyTorch evaluation code, training code and pretrained models for CoNe.
For details see CoNe: Contrast Your Neighbours for Supervised Image Classification by Mingkai Zheng, Shan You, Lang Huang, Xiu Su, Fei Wang, Chen Qian, Xiaogang Wang, and Chang Xu
To run the code, you probably need to change the Dataset setting (dataset/imagenet.py), and Pytorch DDP setting (util/dist_init.py) for your own server environments.
The distributed training of this code is based on slurm environment, we have provided the training scrips in script/train.sh
We also provide the pretrained model for ResNet50
| Arch | BatchSize | Epochs | Top-1 | Download | |
|---|---|---|---|---|---|
| CoNe | ResNet50 | 1024 | 100 | 78.7 % | 100ep-ResNet50-CoNe.tar |
If you want to test the pretained model, please download the weights from the link above, and move it to the checkpoints folder (create one if you don't have .checkpoints/ directory). The evaluation scripts also have been provided in script/train.sh
If you find that CoNe interesting and help your research, please consider citing it:
@article{zheng2023cone,
title={CoNe: Contrast Your Neighbours for Supervised Image Classification},
author={Zheng, Mingkai and You, Shan and Huang, Lang and Su, Xiu and Wang, Fei and Qian, Chen and Wang, Xiaogang and Xu, Chang},
journal={arXiv preprint arXiv:2308.10761},
year={2023}
}