ataalimi / cnn

CNN Implementation

Convolutional Neural Network library

copyright (c) 2014 Vedran Vukotic

contact (gmail): vevukotic

This is a simple library that alows to easily create, train and test CNN architectures. The library is meant for educational purposes and is writtern for clarity - not for speed.

This library was tested with two datasets:

• The MNIST dataset of handwritten digits
• A subsection of the FER-MASTIF TS2010 Croatian traffic signs dataset

Pretrain models are available in ./models. You can test them by calling testMASTIF_TS2010.py or testMNIST.py.

If you're going to learn a network by yourself (learn*.py) have in mind that it's very slow.

Related paper (Convolutional neural networks for Croatian traffic signs recognition): http://un.defined.name/Vukotic_CNNs_for_croatian_traffic_signs.pdf

My Msc thesis (in Croatian): http://un.defined.name/Vukotic_Raspoznavanje_objekata_dubokim_NN.pdf