How to use

In order to use the capabilities of evolutionary_keras in a project is necessary to use the model-classes provided. These classes inherit from the Keras model class and transparently defer to them whenever a Gradient Descent algorithm is used.

As an example, let us consider a project in which we have some neural network constructed with an input layer input_layer and an output layer output_layer. The Keras model would usually be constructed as:

from tensorflow.keras.models import Model
my_model = Model(input_layer, output_layer)

Using evolutionary_keras is as easy as doing:

from evolutionary_keras.models import EvolModel
my_model = EvolModel(input_layer, output_layer)

From that point onward my_model behaves exactly as a normal Keras model implementing the same methods and attributes as well as allowing the usage of Evolutionary Optimizers. For instance, the example below utilizes the Nodal Genetic Algorithm (NGA):

my_model.compile("nga")

Which will use the default parameters of the Nodal Genetic Algorithm (NGA). Subsequent calls to methods such as my_model.fit will use the NGA algorithm to train.

For a more fine-grained usage we can also import the optimizer and instantiate it ourselves:

from evolutionary_keras.optimizers import NGA
my_nga = NGA(population_size = 42, mutation_rate = 0.2)
my_model.compile(my_nga)

Note

Care is due when using a GPU together with evolutionary_keras. At the moment mutant generation is always performed in CPU for all optimizers. If the model is to be evaluated on GPU this will create a big overhead due to constantly copying the weights to the CPU. Only when the evaluation part is big enough that this overhead becomes negligible, evaluating on GPU can be an advantage. To force the tensorflow backend not to use the GPU you can set the environment variable: CUDA_VISIBLE_DEVICES=""