chinmay0301 / capsnet

Attempt at reproducing results from the capsule networks paper

Introduction

This tries to reproduce Capsule Network results in pytorch.

Installation

• Pre-reqs are found here
• git clone https://github.com/teju85/dockerfiles
• cd dockerfiles/ubuntu1604
• make dsb-2018
• cd ../..
• git clone https://github.com/teju85/capsnet

USAGE

$ ./dockerfiles/scripts/launch -runas user dsb:2018 /bin/bash
container$ cd /work/capsnet
container$ env LANG=C.UTF-8 python net.py

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

Accuracy/Perf Numbers

With uhat detach

Detaching uhat means to perform back-prop only for the last routing iteration.

Train epoch:49 time(s):40.344 loss=0.00019366 accuracy:0.9986
Test epoch:49 time(s):2.717 loss=0.00003051 accuracy:0.9938

Achieves 99.38% test accuracy and epoch time of 40s on a P100.

Without uhat detach

Train epoch:49 time(s):47.462 loss=0.00017785 accuracy:0.9996
Test epoch:49 time(s):2.765 loss=0.00002714 accuracy:0.9934

Achieves % test accuracy and epoch time of s on a P100.

Since w/ and w/o detach really doesn't seem to cause huge differences in accuracy, but w/ detach runs ~17% faster than w/o it, detach has been made the default behavior in this repo.