NOTE

This implementation is fork of https://github.com/XifengGuo/CapsNet-Keras , applied to IMDB texts reviews dataset and Rotten Tomotoes dataset

CapsNet-Keras

A Keras implementation of CapsNet in the paper: Sara Sabour, Nicholas Frosst, Geoffrey E Hinton. Dynamic Routing Between Capsules. NIPS 2017

Requirements

• Keras

Usage

Training

Step 1. Install Keras:

$ pip install keras

Step 2. Clone this repository with git.

$ git clone https://github.com/charlieanna/project.git
$ cd CapsNet-Keras

We have analyzed the capsule net on two datasets, imdb and rotten tomotoes dataset. You can check the results of the traning as well as the test results by using the following commands which will run the python files.

Analysis on the rotten tomatoes dataset.

• python rotten_cnn.py
• python rotten_capsulenet.py For convolution layer with capsule net

For convolution layer with the model specified and then capsule net

• python rotten_capsulenet.py --model=LSTM
• python rotten_capsulenet.py --model=GRU
• python rotten_capsulenet.py --model=CuDNNLSTM
• python rotten_capsulenet.py --model=CuDNNGRU

Analysis on the imdb dataset.

• python imdb_cnn.py
• python imdb_capsulenet.py For convolution layer with capsule net

For convolution layer with the model specified and then capsule net

• python imdb_capsulenet.py --model=LSTM
• python imdb_capsulenet.py --model=GRU
• python imdb_capsulenet.py --model=CuDNNLSTM
• python imdb_capsulenet.py --model=CuDNNGRU

Testing

Suppose you have trained a model using the above command, then the trained model will be saved to result/trained_model.h5. Now just launch the following command to get test results.

$ python capsulenet.py --is_training 0 --weights result/trained_model.h5

It will output the testing accuracy.