This repository contains PyTorch evaluation code, training code and pretrained models for SimMatch. Most of the code in this repository is adapt from here.
For details see SimMatch: Semi-supervised Learning with Similarity Matching by Mingkai Zheng, Shan You, Fei Wang, Chen Qian, and Chang Xu
To run the code, you probably need to change the Dataset setting (ImagenetPercentV2 function in dataset/imagenet.py), and Pytorch DDP setting (dist_init function in util/dist_utils.py) for your server environment.
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 pre-trained model.
| Arch | Setting | Epochs | Accuracy | Download | |
|---|---|---|---|---|---|
| SimMatch | ResNet50 | 1% | 400 | 67.2 % | simmatch-1p.pth |
| SimMatch | ResNet50 | 10% | 400 | 74.4 % | simmatch-10p.pth |
If you want to test the pre-trained model, please download the weights from the link above, and move them to the checkpoints folder. The evaluation scripts also have been provided in script/train.sh
If you find that SimMatch interesting and help your research, please consider citing it:
@article{zheng2022simmatch,
title={SimMatch: Semi-supervised Learning with Similarity Matching},
author={Zheng, Mingkai and You, Shan and Huang, Lang and Wang, Fei and Qian, Chen and Xu, Chang},
journal={arXiv preprint arXiv:2203.06915},
year={2022}
}