SubNet

Keras implementation of SubNetwork

Results

Compare accuracy and number of parameters

The following results can be reproduced with command:

python main.py --dataset MNIST --subset 1 4
python main.py --dataset MNIST --subset 0 2 6
python main.py --dataset MNIST --subset 0 4 6 7
python main.py --dataset MNIST --subset 0 1 2 3 4 5 6 7 8 9
python main.py --dataset FashionMNIST --subset 5 7 9
python main.py --dataset FashionMNIST --subset 0 1 2 3 4 5 6 7 8 9

subset	SubNetwork		OriginalNetwork		A/B (%)
subset	# Params (A)	Test - acc	# Params (B)	Test - acc	A/B (%)
Using MNISTdataset Network
[1, 4]	21,251	0.999	124,825	0.983	17.02
[0, 2, 6]	29,947	0.992		0.980	23.99
[0, 4, 6, 7]	40,243	0.992		0.981	32.24
ALL	124,825	0.979		0.979	100.
Using FashionMNISTdataset Network
subset for shoes categories [5, 7, 9]	87,147	0.963	330,670	0.962	26.35
ALL	330,670	0.911	330,670	0.911	100.

Average number of nodes according to number of subset elements

The following results can be reproduced with command:

python main.py --dataset MNIST --subset 0 1 2 3 4 5 6 7 8 9 --meanNodes True
python main.py --dataset FashionMNIST --subset 0 1 2 3 4 5 6 7 8 9 --meanNodes True

Average number of Nodes

X-axis: Average number of nodes, Y-aixs: number of subset elements

Usage

Prerequisites

Keras
Python packages: numpy

Command

python main.py --dataset <choose dataset> --subset <subset of total categories>

Example: python main.py --dataset MNIST --subset 0 1 2 3

Arguments

Required:

--dataset: Choose datset. Option: MNIST or FasionMNIST
--subset: Subset elements of total categories. example: --subset 0 1 2

Optional:

--meanNodes: Whether or not to print the average number of nodes. type: bool, Default: False

Acknowledgements

This implementation has been tested with Keras 2.2.4 on Windows 10.

bigdata-inha/SubNet