Publication: Evaluación temporal del entrenamiento e inferencia de redes neuronales sobre plataformas hardware heterogéneas
Loading...
Official URL
Full text at PDC
Publication Date
2019
Authors
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citations
Exportar
Abstract
Este trabajo evalúa el rendimiento temporal de un conjunto de modelos de aprendizaje automático a partir de datos provenientes de un sistema de captación IoT compuesto por sensores de irradiación solar. Los algoritmos construidos se apoyan en redes neuronales artificiales, que serán implementadas mediante los frameworks para deep learning Tensorflow y Keras utiilzando el lenguaje de programación Python.
Se generará un grupo de modelos capaces de resolver un problema de regresión a partir de datos históricos de radiación solar consiguiendo predecir esta para un tiempo futuro. Para llegar a este objetivo será necesario acometer el entrenamiento e inferencia sobre las distintas redes neuronales. Estos procesos se realizarán sobre distintas arquitecturas hardware heterogéneas formadas por CPUs y GPUs capaces de acelerar los procesos computacionales.
Una vez realizada la computación necesaria, se evaluarán los tiempos requeridos para el procesado de cada uno de los modelos en las diferentes arquitecturas. Se realizará una comparativa para el proceso de entrenamiento y otra para el proceso de inferencia. De este modo se podrá vislumbrar cuándo es conveniente acelerar el cómputo mediante GPUs y cuándo es suficiente con utilizar la CPU en función de los requerimientos temporales de cada contexto.
This work evaluates the temporal performance of a set of Machine learning models based on data from an IoT collection system located in a solar thermal power plant. The constructed algorithms are based on artificial neural networks, which will be implemented through the paradeep learningTensorflow and Kerasutiilil frameworks using the Python programming language. A group of models capable of solving a regression problem will be generated based on historical solar radiation data, predicting this for a future time.To reach this goal it will be necessary to undertake training and inference about the different neural networks. These processes will be carried out on different heterogeneous hard-ware architectures formed by CPUs and GPUs capable of accelerating computational processes. Once the necessary computation has been carried out, the times required for the processing of each of the models in the different architectures will be evaluated. A comparison will be made for the training process and another for the inference process. From this mode you can see when it is convenient to accelerate the computation through GPUs and when it is enough to use the CPU according to the temporal requirements of each context.
This work evaluates the temporal performance of a set of Machine learning models based on data from an IoT collection system located in a solar thermal power plant. The constructed algorithms are based on artificial neural networks, which will be implemented through the paradeep learningTensorflow and Kerasutiilil frameworks using the Python programming language. A group of models capable of solving a regression problem will be generated based on historical solar radiation data, predicting this for a future time.To reach this goal it will be necessary to undertake training and inference about the different neural networks. These processes will be carried out on different heterogeneous hard-ware architectures formed by CPUs and GPUs capable of accelerating computational processes. Once the necessary computation has been carried out, the times required for the processing of each of the models in the different architectures will be evaluated. A comparison will be made for the training process and another for the inference process. From this mode you can see when it is convenient to accelerate the computation through GPUs and when it is enough to use the CPU according to the temporal requirements of each context.
Description
Trabajo de Fin de Máster, Universidad Complutense, Facultad de Informática, Departamento de Arquitectura de Computadores y Automática, Curso 2018/2019