maomran / capsnet

Capsule Network with GMM Dynamic Routing Performance Analysis

EM based Capsnet Performance Analysis

Experiment

Capsule network is a new proposed neural network in GEH18 as an enhancement to Convlutional Neural Network(CNN) and potentially could replace CNNs in different application that require more accurate detection of objects that tends to change in terms of graphical properties such as position, orientation and thickness. In this work, we study the architecture of a capsule network, demonstrate its bottlenecks by profiling its operations.

Read final report in here

Based on the implementation of Matrix-Capsules

Graphs

This figure shows tests accuracy of three benchmarks evaluated which are MNIST, Fashion-MNIST and SmallNORB.

This figure shows activations histogram through the caps layer which shows that activations in capslayer follow the same pattern as CNN layers activation, it gets more sparse when going deeper to the network.

This figure shows the pose matrix (which learns the graphical features of an object) histogram of the learned pose matrix across MNIST and SmallNORB. As observed, pose matrix of smallNORB tends to be more wide compared to MNIST since the dataset represents oriented objects which exploits the problem that CNN struggles to learn.

Profiling Ops

This table shows the most expensive operations in terms of memory usage and execution time

OP Name	Alloc. Bytes	Exec. time
BatchMatMul	297.25MB	263.69ms
Mul	513.12MB	3.32ms
Sum	21.48MB	3.32ms
Sub	163.70MB	2.73ms
Conv2D	14.56MB	1.48ms
Tile	326.25MB	1.04ms

Reproduce

1. Clone this repository with git.

$ git clone https://github.com/maomran/CapsNet.git
$ cd CapsNet

2. Download the MNIST dataset

$ 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

3. Start the training (MNIST):

$ python3 train.py "mnist"

(smallNORB):

$ python3 train.py "smallNORB"

(CNN baseline):

$ python3 train_baseline.py "smallNORB"

4. Start test MNIST:

$ python3 test.py "mnist" "caps"

smallNORB:

$ python3 test.py "smallNORB" "caps"

5. View your profiling on tensorboard.

$ tensorboard --logdir=./test_logdir/{mode}/{dataset}/

capsnet's Issues

Could you tell me your best test accuracy on smallNORB dataset?

I have trid several different configurations based on this implemantation(https://github.com/yl-1993/Matrix-Capsules-EM-PyTorch).
I could not get the similar performance mentioned in the original paper(about 2%).
Thus, could you tell me your best test accuracy on smallNORB? and is there some tips for training Matrix Capsule?