CNN-Evolution

This repository contains code for our Natural Computing final project, called Structure-Based Evolution of Convolutional Neural Network Architectures.

We created a system that uses a genetic algorithm to generate convolutional neural networks. These networks are tested on the CIFAR-10 benchmark dataset. Our best network achieved a validation accuracy of 80%.

Modify and run main.py to generate and train networks using the given settings.

Dependencies

• Keras
• TensorFlow
• Matplotlib
• Numpy
• Pandas

Files

• initialization.py: module for creating random network representations (as dictionaries) that are accepted by the genetic algorithm.
• main.py: main file of this project.
• run_benchmark_model.py: module for creating a benchmark CNN based on VGGNET and training it on CIFAR-10.
• run_dict_model.py: module for creating a CNN model in Keras based on a network representation as generated by the genetic algorithm.
• utils.py: module containing utility functions for saving and loading objects using cPickle.
• validate.py: module for validating a given network representation.

Authors

• Lars Kuijpers
• Jeroen Manders
• Timo van Niedek