vanhooser/TF-Genetic

Evolutionary Neural Networks backed by TensorFlow and pure Python

TF-Genetic

###Evolutionary Neural Networks, backed by TensorFlow and pure Python

Based on the work of Maul et al in this paper

Description

This package speeds up the evolution of activation function structure in neural networks. Nets can sometimes become even more accurate to their problem domain when activation functions within each layer are mixed together and not uniformly applied to all neurons.

This package allows for easy simulation of an arbitrary number of layers/neurons/activation functions to find an optimal arrangement.

Dependencies

• TensorFlow (tested 0.11.0rc0)
• Numpy (tested 1.11.2)

Structure

.
├── LICENSE
├── README.md
└── src
    ├── Chromosome.py
    ├── GeneticNetwork.py
    ├── TFGenetic.py
    └── tests.py

Thirty Seconds to TF-Genetic

• Use TensorFlow's included Iris dataset

# Use TensorFlow's built-in Iris dataset
iris_dataset = tf.contrib.learn.datasets.base.load_iris()
input_data = iris_dataset.data

# Auto-encode the data
output_data = iris_dataset.data

• Define valid activation functions for the evolution and colors for each to be plotted as

# Declare valid activation functions for the network,
# and their corresponding colors for plotting
valid_activation_function_list = [
    tf.nn.sigmoid,
    tf.nn.tanh,
    tf.nn.relu,
    tf.nn.softsign,
    tf.nn.elu]

activation_function_colors = [
    'g',
    'r',
    'b',
    'y',
    'c']
    

• Generate a genetic algorithm pool and specify tournament size, population size, etc.

genetic_pool_settings = {
    'populationSize' : 30,
    'tournamentSize' : 4,
    'memberDimensions' : [4, 3, 2, 3, 4],
    'mutationRate' : 0.05,
    'averagesCount' : 1,
    'validActivationFunctions' : valid_activation_function_list,
    'activationFunctionColors' : activation_function_colors,
    'ins' : input_data,
    'outs' : output_data
}

# Declare the genetic pool and initialize properties
genetic_pool = gen.GeneticPool(**genetic_pool_settings)

# Generate population
genetic_pool.generatePopulation()

• For number of generations specified, cycle and generate new individuals

generation_count = 5
for generation_number in range(generation_count):
    genetic_pool.cycle()
    genetic_pool.generation(generation_number)
genetic_pool.printAllSeenChromosomes()
genetic_pool.plotEvolution()

• Profit