FMRouting

This repository offers TensorFlow implementations for the paper "An Efficient Agreement Mechanism in CapsNet By Pairwise Product"

Requirements

• Python 3.6
• tensorflow-datasets 1.0.2
• tensorflow-gpu 2.0.0b1

Experiments

Section 4.1

Datasets

cropped-SVHN, Fashion-MNIST, CIFAR10

These benchmark datasets will be downloaded and prepared automatically when running experiments.

Replication

The results in Figure 5 can be replicated by running:

python run_4_1.py --routing=dynamic
python run_4_1.py --routing=FM

The results of dynamic routing and FM agreement will be saved in log/fashion_mnist, log/cifar10, log/svhn_cropped in Tensorboard format, and the accuracy will also be printed at the end of every epoch.

Section 4.2.1

Datasets

CIFAR10

The dataset will be downloaded and prepared automatically when running experiments.

Replication

The results in Table 2 can be replicated by running:

python run_4_2_1.py 

The results will be saved in log/cifar10 in Tensorboard format, and the accuracy will also be printed at the end of every epoch.

Section 4.2.2

Datasets

cropped-SVHN, Fashion-MNIST, CIFAR10

These benchmark datasets will be downloaded and prepared automatically when running experiments.

Replication

The results in Table 3 can be replicated by running:

python run_4_2_2.py 

The results will be saved in log/cifar10 in Tensorboard format, and the accuracy will also be printed at the end of every epoch.

Section 4.3.1

Datasets

ExpandedMNIST, affNIST

This experiment validates the robustness of the network on affine transformations. It uses the same datasets in the section 5.2 of Dynamic Routing Between Capsules , ExpandedMNIST is generated using the code in its official repository, affNIST can be downloaded in its official web, we use its transformed test set as the this experiment's test set.

Replication

The results in Table 4 can be replicated by running:

python run_4_3_1.py 

The results will be saved in log/aff_mnist in Tensorboard format, and the generated digits shown in Table 5 can be found under the Tab "IMAGES" in the Tensorboard page.

Section 4.3.2

Datasets

MultiMNIST

It uses the same dataset in the section 6 of Dynamic Routing Between Capsules , MultiMNIST is generated using the code in its official repository.

Replication

The results in Table 6 can be replicated by running:

python run_4_3_2.py 

The results will be saved in log/multi_mnist in Tensorboard format, and the generated digits shown in Table 6 can be found under the Tab "IMAGES" in the Tensorboard page.

Section 4.4

Datasets

SmallNORB

It uses the same dataset in the section 5 of Matrix Capsules with EM Routing . As the official code of "Matrix Capsules with EM Routing" is not ready, we use the "tfrecord" format of SmallNORB generated by the code in this repository.

Replication

The results in Table 7 can be replicated by running:

python run_4_4.py 

The results will be saved in log/smallNORB in Tensorboard format, and the accuracy will also be printed at the end of every epoch.