Capsule networks with non-iterative cluster routing

A pytorch implementation.

Capsule networks with non-iterative cluster routing
Zhihao Zhao, Samuel Cheng Elsevier Neural Networks, Volume 143, November 2021.

Use this bibtex to cite this repository:

@article{zhao2021capsule,
  title={Capsule networks with non-iterative cluster routing},
  author={Zhao, Zhihao and Cheng, Samuel},
  journal={Neural Networks},
  volume={143},
  pages={690--697},
  year={2021},
  publisher={Elsevier}
}

Prerequisites

This code is tested with the following packages

• Python 3.7
• Pytorch 1.7.0
• torchvision 0.8.0
• CUDA 10.2

Training CIFAR10

python3 main.py --dataset cifar10 --C 4 --K 8 --D 24 --batch_size 64 --save_dst ./checkpoint

Training CIFAR10 quickly (less parameters)

python3 main.py --dataset cifar10 --C 4 --K 5 --D 6 --batch_size 64 --save_dst ./checkpoint

Performances

model	CIFAR-10	SVHN	smallNORB
Dynamic routing	10.6%	4.3%	2.7%
EM routing	11.9%	-	1.8%
Cluster routing	7.37%	3.37%	1.57%