GANN

Genetic algorithm with neural networks frameworks developed for my level 4 individual project.

Installation instructions

There are several components needed in order to run this project:

Numpy - to install numpy please visit link to Numpy install guide Theano - to install theano please visit link to Theano install guide

If you have a NVIDIA graphics card you will need to setup CUDA. You can get it from here. If you are a registered NVIDIA developer you might also want to install NVIDIA CUDA Deep Neural Network (cuDNN). It will speed up the training process of neural networks. You can download cuDNN from here

Running instructions

There are three different frameworks in this project. A deep convolutional neural networks framework, a genetic algorithms framework and a combination of both - genetic algorithms with neural networks framework. The first one can be used as specified in '<nn.py>' file.

Neural networks framework

The basic usage is:

Define the network.

net = NNet([
            ConvolutionalLayer(image_shape=(mini_batch_size, 1, 28, 28),
                      filter_shape=(20, 1, 5, 5), activation_fn=ReLU),
            PoolingLayer(image_shape=(mini_batch_size,20, 28, 28),
                        poolsize=(2,2)),
            ConvolutionalLayer(image_shape=(mini_batch_size, 20, 14, 14),
                      filter_shape=(40, 20, 5, 5), activation_fn=ReLU),
            PoolingLayer(image_shape=(mini_batch_size, 40, 14, 14),
                        poolsize=(2,2)),
            FullyConnectedLayer(n_in=40*7*7, n_out=1000, activation_fn=ReLU),
            FullyConnectedLayer(n_in=1000, n_out=1000, activation_fn=ReLU),
            SoftmaxLayer(n_in=1000, n_out=10)],
            mini_batch_size)

Load data needed to train, validate and test the network.

training_data, validation_data, test_data = load_data()

Train the network.

net.train(training_data, epochs, mini_batch_size, learning_rate,
            validation_data, test_data, lmbda=l2)

Genetic algorithms framework

The genetic algorithms framework is very abstract. There isn't an example since we don't use the pure version of it, rather than that we use a mixed version of neural networks and genetic algorithms.

Genetic algorithms with neural networks framework

There is an example file that shows how to run the GANN framework - 'gann.py'. There are several parameters you need to specify as discussed in the dissertation.

parameters["ga"] = {
    "maxGenerations": 1000,
    "threshold": 0.9,
    "population": 20,
    "crossover": 0.2,
    "mutation": 0.1
}

parameters["nn"] = {
    "params" : {
        "layers": [4,10],
        "batchSize": 100,
        "maxEpochs": 5,
        "inputShape": [3,32,32],
        "output": 10,
        "data_split":[0.5, 0.75, 1],
        "learningRate": [0,2],
        "l2": [0,2]
    },
    "definitions" : {
        "convolution": {
            "filters": [1,32],
            "size": [1,8]
        },
        "pooling": {
            "size": [1,4]
        },
        "fullyconnected":{
            "neurons": [1,1000],
            "dropout": [0, 0.5]
        },
        "softmax":{
            "dropout": [0, 0.5]
        }
    }
}

After you define those, you need to initialise the GANN framework by providing those parameters to the constructor, as well as the 'load_data' function you want to use. There are several 'load_data' functions provided for the MNIST, CIFAR and larger images data sets. They can be imported like this

# choose one of the datasets
from NN.CIFARLoader import load_data # CIFAR
from NN.DataLoader import load_data # Larger images
from NN.MNISTLoader import load_data # MNIST

There is one more thing we need to define and this is the callback function which will provide us with the obtained data after the GANN has completed its run. The parameter passed to the function are all generations with all individuals in them.

def processData(generations):
    # do whatever you want to the array
    print(len(generations))

# initialise GANN framework
g = GANN(parameters, load_data, 'test.obj', processData)

After initialising the GANN framework we need to run it by

ga.run()