CapsNet-PyTorch

A PyTorch implementation of CapsNet based on Geoffrey Hinton's paper Dynamic Routing Between Capsules.

Current Status

The current test accuracy = 99.67 % (test error = 0.33), see Results section for details
Trying to find the reason why the test accuracy is lower than the one reported in the paper

Requirements

GPU and NVIDIA driver
PyTorch and other Python modules (see requirements.txt).

This repository also provides Dockerfile for CapsNet training. Check docker directory to know how to setup/use Docker enviroment.

Usage

Step 1. Clone this repository

$ git clone https://github.com/motokimura/capsnet_pytorch.git
$ cd capsnet_pytorch

Step 2. Start the training

$ python main.py

Step 3. Check training status and validation accuracy from TensorBoard

# In another terminal window, 
$ cd capsnet_pytorch
$ tensorboard --logdir ./runs

# Then, open "http://localhost:6006" from your browser and 
# you will see something like the screenshots in the `Results` section.

Some training hyper parameters can be specified from the command line options of main.py.

At default, batch size is 128 both for training and validation, and epoch is set to 100. Learning rate of Adam optimizer is set to 0.001 and is exponentially decayed every epoch with the factor of 0.9.

For more details, type python main.py --help.