Thesis DISEÑO E IMPLEMENTACIÓN DE UNA RED GENERATIVA ADVERSARIA PARA LA CREACIÓN COMPUTACIONAL DE IMÁGENES DE ROSTROS HUMANOS
Loading...
Date
2019-08
Journal Title
Journal ISSN
Volume Title
Program
DEPARTAMENTO DE INFORMÁTICA. INGENIERÍA CIVIL INFORMÁTICA
Campus
Casa Central Valparaíso
Abstract
Desde la proposición de redes generativas adversarias (GAN), diversos trabajos
las han implementado para tareas de generación de imágenes, dentro de las cuales la
generación artificial de imágenes de rostros humanos ha sido el área que ha generado
el mayor interés y donde este trabajo se centra.
A través de un estudio de la literatura fue posible detectar 2 problemas comunes
en el proceso de entrenamiento de la mayoría de las GANs implementadas para
creación artificial de rostros humanos: un alto uso de memoria y el uso de funciones de
pérdida poco efectivas. Ambos problemas tienen severas implicancias en la usabilidad
de este tipo de modelos dado el alto costo computacional de entrenar estos modelos
con sus actuales implementaciones.
Un nuevo enfoque para abordar estos dos problemas es presentado en este trabajo
así como los resultados para validar la calidad de nuestra propuesta.
Los resultados presentados muestran una reducción de aproximadamente 30 % en
uso de memoria sin afectar la calidad de los resultados, disminuyendo las limitantes
del costo computacional en la fase de entrenamiento.
Since the proposition of generative adversarial networks (GAN), several works implemented them for image generation tasks. Among them, the artificial generation of human faces has been the area of most interest and on which this work focuses. Through a study of the literature, it was possible to detect two common problems in the training process of many GANs proposed for human face images generation: high use of resources and less effective loss function. Both problems have several implications in their usability due to the high computational cost of training them with their current implementations. A new approach is proposed in order to address both of these problems as well as the experimental results to validate this proposition. The results presented show a reduction of approximately 30 % of memory usage without affecting the quality of the result, reducing the limitations of the compu- tational cost of the training phase.
Since the proposition of generative adversarial networks (GAN), several works implemented them for image generation tasks. Among them, the artificial generation of human faces has been the area of most interest and on which this work focuses. Through a study of the literature, it was possible to detect two common problems in the training process of many GANs proposed for human face images generation: high use of resources and less effective loss function. Both problems have several implications in their usability due to the high computational cost of training them with their current implementations. A new approach is proposed in order to address both of these problems as well as the experimental results to validate this proposition. The results presented show a reduction of approximately 30 % of memory usage without affecting the quality of the result, reducing the limitations of the compu- tational cost of the training phase.
Description
Keywords
GAN, MACHINE LEARNING