EvoFlow: A Python library for evolving deep neural network architectures in tensorflow

Unai Garciarena, Roberto Santana, Alexander Mendiburu

Abstract: Neuroevolutionary algorithms are one of most effective and extensively applied methods for neural architecture search. While several neuroevolutionary approaches have been proposed, the availability of software that allows a fast development of code to solve problems and test research questions is limited. In this paper we introduce EvoFlow, a Python library for evolving shallow and deep neural network (DNN) architectures. EvoFlow optimizes network structures for DNNs implemented in tensorflow. Single and multi-component DNN architectures are represented by means of descriptors, and the instantiation of the network occurs in the evaluation of the architecture. Genetic operators work by modifying the descriptors. We show how EvoFlow allows efficient architecture optimization of single-component DNNs, such as deep multi-layer perceptrons, but also of multi-component DNNs, such as generative adversarial nets.