SR-CapsNet

PyTorch implementation for our paper Self-Routing Capsule Networks in NeurIPS 2019.

[poster]

Prerequisites

• Python >= 3.5.2
• CUDA >= 9.0 supported GPU

Install required packages by:

pip3 install -r requirements.txt

Training

To train a model for CIFAR-10 or SVHN, run:

python3 main.py --dataset=cifar10 --name=resnet_[routing_method] --epochs=350
python3 main.py --dataset=svhn --name=resnet_[routing_method] --epochs=200

routing_method should be one of [avg, max, fc, dynamic_routing, em_routing, self_routing]. This will modify last layers of base model accordingly.

For SmallNORB, run:

python3 main.py --dataset=smallnorb --name=convnet_[routing_method] --epochs=200 --exp=elevation

Here --exp denotes which viewpoint data should be splitted on.

See config.py for more options and their descriptions.

Testing

To test a model, simply run:

python3 main.py --dataset=cifar10 --name=resnet_[routing_method] --is_train=False

You can perform adversarial attacks against a trained model by:

python3 main.py --dataset=cifar10 --name=resnet_[routing_method] --is_train=False --attack=True --attack_type=bim --attack_eps=0.1 --targeted=False

For SmallNORB, you can test against novel viewpoints by:

python3 main.py --dataset=smallnorb --name=convnet_[routing_method] --is_train=False --familiar=False

Citation

@inproceedings{hahn2019,
  title={Self-Routing Capsule Networks},
  author={Hahn, Taeyoung and Pyeon, Myeongjang and Kim, Gunhee},
  booktitle={NeurIPS},
  year={2019}
}