MATRIX CAPSULES EM-Tensorflow

A Tensorflow implementation of CapsNet based on paper Matrix Capsules with EM Routing

Status:

1. With configuration A=32, B=8, C=16, D=16, batch_size=128, the code can work on a Tesla P40 GPU at a speed of 8s/iteration. The definitions of A-D can be referred to the paper.
2. With configuration A=B=C=D=32, batch_size=64, the code can work on a Tesla P40 GPU at a speed of 25s/iteration. More optimization on implementation structure is required.
3. Some modification and optimization is implemented to prompt the numerical stability of GMM. Specific explanations can be found in the code.
4. With configuration A=32, B=4, D=4, D=4, batch_size=128, each iteration of training takes around 0.6s on a Tesla P40 GPU.

Current Results on smallNORB:

• Configuration: A=32, B=8, C=16, D=16, batch_size=50, iteration number of EM routing: 2, with Coordinate Addition, spread loss, batch normalization

• Training loss. Variation of loss is suppressed by batch normalization. However there still exists a gap between our best results and the reported results in the original paper.

• Test accuracy(current best result is 87.66%)

Ablation Study on smallNORB:

• Configuration: A=32, B=8, C=16, D=16, batch_size=32, iteration number of EM routing: 2, with Coordinate Addition, spread loss, test accuracy is 79.8%.

Current Results on MNIST:

• Configuration: A=32, B=8, C=16, D=16, batch_size=50, iteration number of EM routing: 2, with Coordinate Addition, spread loss, batch normalization, reconstruction loss.

• Training loss.

• Test accuracy(current best result is 99.3%, only 10% samples are used in test)

Ablation Study on MNIST:

• Configuration: A=32, B=4, C=4, D=4, batch_size=128, iteration number of EM routing: 2, no Coordinate Addition, cross entropy loss, test accuracy is 96.4%.
• Configuration: A=32, B=4, C=4, D=4, batch_size=128, iteration number of EM routing: 2, with Coordinate Addition, cross entropy loss, test accuracy is 96.8%.
• Configuration: A=32, B=8, C=16, D=16, batch_size=32, iteration number of EM routing: 2, with Coordinate Addition, spread loss 99.1%.

To Do List:

1. Experiments on smallNORB as in paper is about to be casted.

Any questions and comments to the code and the original algorithms are welcomed!!! My email: zhangsuofei at njupt.edu.cn

Requirements

• Python >= 3.4
• Numpy
• Tensorflow >= 1.2.0

pip install -r requirement.txt

Usage

Step 1. Clone this repository with git.

$ git clone https://github.com/www0wwwjs1/Matrix-Capsules-EM-Tensorflow.git
$ cd Matrix-Capsules-EM-Tensorflow

Step 2. Download the MNIST dataset, mv and extract it into data/mnist directory.(Be careful the backslash appeared around the curly braces when you copy the wget command to your terminal, remove it)

$ mkdir -p data/mnist
$ wget -c -P data/mnist http://yann.lecun.com/exdb/mnist/{train-images-idx3-ubyte.gz,train-labels-idx1-ubyte.gz,t10k-images-idx3-ubyte.gz,t10k-labels-idx1-ubyte.gz}
$ gunzip data/mnist/*.gz

To install smallNORB, follow instructions in ./data/README.md

Step 3. Start the training(MNIST):

$ python3 train.py "mnist"

Start the training(smallNORB):

$ python3 train.py "smallNORB"

Step 4. View the status of training:

$ tensorboard --logdir=./logdir/train_log/

Open the url tensorboard has shown.

Step 5. Start the test on MNIST:

$ python3 eval.py "mnist"

Start the test on smallNORB:

$ python3 eval.py "smallNORB"

Step 6. View the status of test:

$ tensorboard --logdir=./test_logdir

Open the url tensorboard has shown.

Reference

• naturomics/CapsNet-Tensorflow: the implementation of Hinton's paper Dynamic Routing Between Capsules