A flexible and easy-to-follow Pytorch implementation of Hinton's Capsule Network.
There are already many repos containing the code for CapsNet. However, most of them are too tight to customize. And as we all know, Hinton's original paper is only tested on MNIST datasets. We clearly want to do more.
This repo is designed to hold other datasets and configurations. And the most important thing is, we want to make the code flexible. Then, we can tailor the network according to our needs.
Currently, the code supports both MNIST and CIFAR-10 datasets.
Just run Python test_capsnet.py in your terminal. That's all. If you want to change the dataset (MNIST or CIFAR-10), you can easily set the dataset variable.
It is better to run the code on a server with GPUs. Capsule network demands good computing devices. For instance, on my device (Nvidia K80), it will take about 5 minutes for one epoch of the MNIST datasets (batch size = 100).
There are 3 .py files:
capsnet.py: the main class for capsule networkdata_loader.py: the class to hold many classestest_capsnet.py: the training and testing codeThe results on your device may look like the following picture:
