KerasGA 1.0.0

A simple and easy-to-use implementation of Genetic Algorithm for Keras NN models in Python.

KerasGA

A simple and easy-to-use implementation of Genetic Algorithm for Keras NN models in Python.

Features

• create an initial population (of size: population_size) of randomly initialized chromosomes (i.e model weights).
• You can adjust the selection_rate & the mutation_rate.
• Perform the different GA operations (i.e Selection, Crossover, & Mutation).

Examples

Here are a few projects based on this package:

• yahiakr/FlappyAI
• yahiakr/SnakeAI

Usage

• Install KerasGA :

$ pip install KerasGA

• import GeneticAlgorithm from KerasGA and initiate an object :

from KerasGA import GeneticAlgorithm

population_size =  10
GA = GeneticAlgorithm(model, population_size = population_size, selection_rate = 0.1, mutation_rate = 0.2)

PS: model is a Keras model.

• Generate the initial population:

population = GA.initial_population()

• To set the wights of a model you can use .set_weights() built-in function:

for chromosome in population:
	model.set_weights(chromosome)
	# then evaluate the chromosome (i.e assign its final score)

• After calculating the scores for each chromosome, it's time to select the top-performers:

# Selection:
# 'scores' is a list of length = population_size
# 'top_performers' is a list of tuples: (chromosome, it's score)
top_performers = GA.strongest_parents(population,scores)

# Make pairs:
# 'GA.pair' return a tuple of type: (chromosome, it's score)
pairs = []
while len(pairs) != GA.population_size:
	pairs.append( GA.pair(top_performers) )

# Crossover:
base_offsprings =  []
for pair in pairs:
	offsprings = GA.crossover(pair[0][0], pair[1][0])
	# 'offsprings' contains two chromosomes
	base_offsprings.append(offsprings[-1])

# Mutation:
new_population = GA.mutation(base_offsprings)

And that's it :)