MSVQ: Self-Supervised Learning with Multiple Sample Views and Queues

Updata(12,Nov, 2023)

• MSVQ has been accepted in KBS(Knowledge-Based Systems)!! Please check out our paper on arXiv.

This is the official Pytorch implementation of the MSVQ [MSVQ: Self-Supervised Learning with Multiple Sample Views and Queues] paper, if you find this useful, feel free to give us a Star 🌟✨.

Requirements

To install requirements:

 # name: d2lpy39
 conda env create -f environment.yml

Training

(You need to create the directory './stdout', you can also omit '>stdout/*' so that you can run these commands directly.) To train the model(s) in the paper, run those commands:

nohup python main.py --name msvq --m1 0.99  --m2 0.95 --dataset cifar10       --tem 0.04 --weak --gpuid 0 --queue_size 4096 --logdir cifar10_00 >stdout/cifar10_00 2>&1 &
nohup python main.py --name msvq --m1 0.99  --m2 0.93 --dataset cifar100      --tem 0.03 --weak --gpuid 0 --queue_size 4096 --logdir cifar100_00 >stdout/cifar100_00 2>&1 &
nohup python main.py --name msvq --m1 0.996 --m2 0.99 --dataset stl10         --tem 0.04 --weak --gpuid 0 --queue_size 16384 --logdir stl10_00 >stdout/stl10_00 2>&1 &
nohup python main.py --name msvq --m1 0.996 --m2 0.99 --dataset tinyimagenet  --tem 0.04 --weak --gpuid 0 --queue_size 16384 --logdir tinyimagenet_00 >stdout/tinyimagenet_00 2>&1 &

Evaluation

To evaluate our model on CIFAR10/100, STL10 and Tiny-imagenet, run:

nohup python linear_eval.py --name msvq --dataset cifar10       --gpuid 0  --logdir cifar10_00 >stdout/cifar10_00_01 2>&1 &
nohup python linear_eval.py --name msvq --dataset cifar100      --gpuid 0  --logdir cifar100_00 >stdout/cifar100_00_01 2>&1 &
nohup python linear_eval.py --name msvq --dataset stl10         --gpuid 0  --logdir stl10_00 >stdout/stl10_00_01 2>&1 &
nohup python linear_eval.py --name msvq --dataset tinyimagenet  --gpuid 0  --logdir tinyimagenet_00 >stdout/tinyimagenet_00_01 2>&1 &

Pre-trained Models

You can download pretrained models here:

• this link trained on four datasets.
• Download and place in the "./checkpoints" directory

Results

Our model achieves the following performance:

Image Classification on four datasets

Citation

If you find this repo useful for your research, please consider citing the paper

@article{PENG2024111197,
title = {MSVQ: Self-supervised learning with multiple sample views and queues},
journal = {Knowledge-Based Systems},
volume = {283},
pages = {111197},
year = {2024},
issn = {0950-7051},
doi = {https://doi.org/10.1016/j.knosys.2023.111197},
url = {https://www.sciencedirect.com/science/article/pii/S0950705123009474},
author = {Chen Peng and Xianzhong Long and Yun Li}

Contributors and Contact

📋 If there are any questions, feel free to contact with the authors.