Dynamic Channel Selection in Self-Supervised Learning

This is an official Pytorch based implementation of Dynamic Channel Selection in Self-Supervised Learning accepted in IMVIP 2022. Code from channel gating is derived from DGNet while we train the self-supervised approach based on Solo-learn library. We use SimSiam as our self-supervised objective.

Getting Started

Requirements

The main requirements of this work are:

• Python 3.8
• PyTorch == 1.10.0
• Torchvision == 0.11.1
• CUDA 10.2

We recommand using conda env to setup the experimental environments.

Install other requirements

pip install -r requirements.txt

# Clone repo
git clone https://github.com/KrishnaTarun/SSL_DGC.git
cd ./SSL_DGC

Pre-Training

# ImageNet-100
bash bash_files/pretrain/imagent100/simsiam.sh

# CIFAR-100
bash bash_files/pretrain/cifar100/simsiam.sh

KNN - Evaluation

# ImageNet-100
bash bash_files/knn/imagent100/knn.sh

Please consider citing the following paper if you find this work useful for your research.

@inproceedings{Krishna2022DynamicCS,
 title={Dynamic Channel Selection in Self-Supervised Learning},
 author={Tarun Krishna and Ayush Rai and Yasser Abdelaziz Dahou Djilali and Alan F. Smeaton and Kevin McGuinness and Noel E. O'Connor},
 year={2022}
 }