Neural Network Architecture Optimization

This repository allows one to find approximation of the best neural network architecture given 1D, 2D or 3D input.
It also depends on parameters which user can set on their own, but there's default ones. Parameters describe possible choices of number of neurons in a layer, callbacks etc.

Requirements

pip install --user keras numpy sklearn typing h5py warnings datetime tqdm collections deprecated

Basic usage

In order to use:

1. Clone/Download this repository and cd into it.
2. Prepare data x and y (you can use from genetic import prepare_data, or pass in your own data-set).
3. Create a mutator given a population size from genetic import Mutator
4. In order to evolve call Mutator's function evolve with your x y dataset and either validation_split or validation_data argument.

• Please make sure y is in binary class matrix form. keras.utils.np_utils.to_categorical enables to change integer vector to this form.
• If you want to use a different data training set on each generation, you can pass in generator_f and generator_args to Mutator constructor, or call Mutator's set_dataset_generator.

Example 1

from genetic import prepare_data, Mutator

x, y = prepare_data('testing', first_time=False)
m = Mutator(population_size=5)
m.evolve(x, y, validation_split=0.6, generations=5)

Example 2 - with generator

from genetic import prepare_data, Mutator

(x, y), (x_val, y_val) = prepare_data('testing')
m = Mutator(population_size=5, generator_f=prepare_data, generator_args=['testing', False])
m.evolve(x, y, validation_data=(x_val, y_val), generations=5)

More options

• There is an option of starting with preset Networks as starting population. However they need to be initialized by user (from genetic import Network), and passed in to Mutator constructor as starting_population=.
  • Please note that if you are creating network from architecture, you have to look at how architecture describes different layer.
  • Also, when a Network, is created from just architecture, random weights are initialized. To avoid that you can pass in copy_model= with keras.Model.Sequential, or use keras.Model.set_weights function.
• If you want to use different parameters than default, you can either pass them in Mutator constructor, or change them file genetic/program_variables/program_params.py under _mutations field. Please note, that all keys which are there by default have to stay, since program would not work otherwise.