Downloads provided by UsageCounts
handle: 10017/53370
Este trabajo ofrece una comparativa entre los tiempos de inferencia y la precisión de algoritmos Deep Learning ejecutados en dos arquitecturas bien diferenciadas: un ordenador de sobremesa al uso y el módulo Jetson AGX Xavier de NVIDIA. Para ello, se analizan tres de las principales tareas del mundo Deep Learning como son la clasificación de objetos, la detección de objetos y la segmentación semántica. Para lograr una mayor velocidad de inferencia, se estudian métodos de aceleración de algoritmos empleando principalmente la herramienta de NVIDIA conocida como TensorRT. Esto servirá como base para verificar si el módulo Jetson AGX Xavier, un dispositivo de bajo consumo (alrededor de los 30W) es capaz de ejecutar los algoritmos a una velocidad similar a la de una GPU, un dispositivo con un cosumo mucho mayor (alrededor de los 300W). Además, se comprueba que las pérdidas de precisión derivadas de la aceleración no son significativas. Por otro lado, se realiza un estudio teórico de Deep Learning y del estado del arte de la aceleración de algoritmos, explicados en detalle en el documento. Estos estudios son necesarios para llevar a cabo unos experimentos comparativos que exponen todas las métricas de precisión y de tiempos de inferencia extraídas de los resultados. Finalmente, se incluyen las conclusiones derivadas de todos los aspectos considerados con anterioridad.
This work aims to provide a comparison between the inference time and the accuracy metrics of Deep Learning algorithms executed on two well-differentiated architectures: a standard desktop computer and the NVIDIA Jetson AGX Xavier module. For that matter, three of the main tasks of the Deep Learning world, such as object classification, object detection and semantic segmentation, are analyzed. In order to achieve a higher inference speed, algorithm acceleration methods are studied, mainly using the NVIDIA tool known as TensorRT. This will serve as a basis to verify if the Jetson AGX Xavier module, a low power consumption device (around 30W) is able to run the algorithms at a similar speed as a GPU, a device with a much higher power consumption (around 125W). In addition, the accuracy losses due to the acceleration methods are checked. On the other hand, a theoretical study of Deep Learning and the state-of-the-art of algorithm acceleration is performed and explained in this document. These studies are required in order to carry out an ablation study that exposes all the accuracy metrics and inference times extracted from the results. Finally, the conclusions derived from all of the aspects considered above are included.
Grado en Ingeniería en Electrónica y Automática Industrial
Informática, Deep Learning, TensorRT, Algorithm acceleration, Aprendizaje profundo, Redes neuronales convolucionales, Embedded systems, Convolutional neural networks, Aceleración de algoritmos, Sistemas embebidos, Computer science
Informática, Deep Learning, TensorRT, Algorithm acceleration, Aprendizaje profundo, Redes neuronales convolucionales, Embedded systems, Convolutional neural networks, Aceleración de algoritmos, Sistemas embebidos, Computer science
| selected citations These citations are derived from selected sources. This is an alternative to the "Influence" indicator, which also reflects the overall/total impact of an article in the research community at large, based on the underlying citation network (diachronically). | 0 | |
| popularity This indicator reflects the "current" impact/attention (the "hype") of an article in the research community at large, based on the underlying citation network. | Average | |
| influence This indicator reflects the overall/total impact of an article in the research community at large, based on the underlying citation network (diachronically). | Average | |
| impulse This indicator reflects the initial momentum of an article directly after its publication, based on the underlying citation network. | Average |
| views | 129 | |
| downloads | 261 |

Views provided by UsageCounts
Downloads provided by UsageCounts