Publication: Portabilidad y optimización de una red neuronal para la detección rápida de daños en terremotos usando el toolkit OpenVINO
Loading...
Official URL
Full text at PDC
Publication Date
2020
Authors
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citations
Exportar
Abstract
El contenido de este trabajo se centra en el análisis y mejora de todos los procesos relacionados con la generación e implementación de una red neuronal convolucional. Mediante esta técnica perteneciente al campo del aprendizaje profundo o Deep Learning, ha sido posible diseñar una aplicación que permite detectar si en una imagen se encuentran zonas dañadas por el impacto de un terremoto. Para ello, se emplea una topología de red específica para esta problemática. En concreto, las imágenes utilizadas provienen del terremoto ocurrido en Haití en el año 2010, obtenidas a partir de un satélite de observación terrestre llamado GeoEye-1.
Mediante el registro de diferentes métricas como el tiempo de ejecución y la precisión, se consi- gue mejorar el rendimiento del modelo en las fases de entrenamiento e inferencia usando diferentes dispositivos en cada una de estas fases. Empleando en esta última, el toolkit de OpenVINO desa- rrollado por Intel, como medio para aumentar la eficiencia del modelo. Para ello, siendo necesario el establecimiento de una metodología concreta para su uso.
Finalmente, se establece un aumento de rendimiento en los parámetros analizados en la fase de entrenamiento. Mientras tanto en la parte de inferencia, se constata el amplio nivel de mejora obtenido al utilizar OpenVINO en comparación a la herramienta con el mismo propósito proporcio-nada por Tensorflow. Suponiendo entre un 20 % - 60 % dependiendo de la tecnología y dispositivos empleados. Siendo entre ellos, un stick Movidius NCS2.
The content of this work focus on the analysis and enhancement of all the processes related tothe generation and implementation of a convolutional neural network. Through this technique belonging to the field of Deep Learning, it has been possible to design an application which allows us to detect whether there are damaged zones for an earthquake impact in an image. Therefore, it is used a specific topology for this problem. Specifically, the images used provided by the earthquake occurred in Haití in 2010, obtained from an earth observation satellite called GeoEye-1. Through this registry of different metrics such as execution time and accuracy, it managesto improve the performance of the model in training and inference phases using different devices for each one. Using in the last one, the OpenVINO toolkit developed by Intel for increasing the efficiency of the model. For that, being necessary for the establishment of a methodology for itsusage. Finally, it establishes an increase in the performance of the analyzed parameters in the training phase. Whereas in the inference part, it verifies the wide level of improvement obtained by the usage of OpenVINO in comparison to the other tool with the same purpose provided by Tensorflow. Assuming between a 20 % to 60 % of improvement depending on the technology and devices used. Being one of those, a Movidius stick NCS2 (Neural Compute Stick 2).
The content of this work focus on the analysis and enhancement of all the processes related tothe generation and implementation of a convolutional neural network. Through this technique belonging to the field of Deep Learning, it has been possible to design an application which allows us to detect whether there are damaged zones for an earthquake impact in an image. Therefore, it is used a specific topology for this problem. Specifically, the images used provided by the earthquake occurred in Haití in 2010, obtained from an earth observation satellite called GeoEye-1. Through this registry of different metrics such as execution time and accuracy, it managesto improve the performance of the model in training and inference phases using different devices for each one. Using in the last one, the OpenVINO toolkit developed by Intel for increasing the efficiency of the model. For that, being necessary for the establishment of a methodology for itsusage. Finally, it establishes an increase in the performance of the analyzed parameters in the training phase. Whereas in the inference part, it verifies the wide level of improvement obtained by the usage of OpenVINO in comparison to the other tool with the same purpose provided by Tensorflow. Assuming between a 20 % to 60 % of improvement depending on the technology and devices used. Being one of those, a Movidius stick NCS2 (Neural Compute Stick 2).
Description
Trabajo de Fin de Grado en Ingeniería Informática, Facultad de Informática UCM, Departamento de Arquitectura de Computadores y Automática, curso 2019/2020.