This tries to reproduce Capsule Network results in pytorch.

• Pre-reqs are found here
• git clone https://github.com/teju85/dockerfiles
• cd dockerfiles/ubuntu1604
• make pytorch90_70
• cd ../..
• git clone https://github.com/teju85/capsnet

$ ./dockerfiles/scripts/launch -runas user pytorch:latest-90_70 /bin/bash
container$ cd /work/capsnet
container$ python net.py

Use the '-h' option to net.py to know more about its customizations.

$ ./dockerfiles/scripts/launch -runas user pytorch:latest-90_70 /bin/bash
container$ cd /work/capsnet
container$ ./benchmark.sh

• uhat detach - detaching here means to perform backprop only for the last routing iteration
• training epochs of 25 is being used for comparisons below
• accuracy numbers are ratios between 0 and 1
• train and test timings are in seconds and are per epoch
• measurements are all taken using cuda 9.0 with cudnn 7.0

Main paper runs this dataset using an ensemble of 7 models to attain 10.6% test error. In here, we only run one model and that too keeping most of the hyper-params pretty much the same as those with MNIST.

The increase in runtime when compared to Mnist is totally attributable to the difference in input image dimension between these 2 datasets.