Optimization Extension And Evaluation Of Deep Learning Framework Based On Embedded Platform

As one of the hotspots in the field of machine learning,deep learning has made a series of breakthroughs in many fields such as speech recognition and image processing.Due to the limited hardware resources of embedded platform and the complex computation and the large memory consumption of deep learning,the embedded platform is faced with a series of challenges in the development of Deep Learning applications.Choosing the appropriate deep learning framework to realize the recognition of 2D images and 3D models on embedded systems is of great research and application value.Based on the NVIDIA Jetson TX1 embedded development platform,this paper chooses the lightweight deep learning framework DarkNet,and optimizes and extends the forward computation part of the convolutional neural network according to the TX1 GPU architecture.The work of this paper is divided into three parts.Firstly,by analyzing the time-consuming of each network layer of the convolutional neural network,we choose the convolutional layer as the optimization goal.Through two ways,direct convolution and cuDNN call,we change the method of re-matrix matrix multiplication of the original framework to achieve memory and speed optimization,and give the performace results and comparison before and after the optimization.At the same time,based on three different convolutional neural networks that are AlexNet,GoogleNet and VGG-16,we compare the three deep learning frameworks of DarkNet,Caffe and MXNet about the application performance of 2D images classification on embedded platform GPU,with summarizing the running characteristics of the three types of frameworks.Lastly,based on DarkNet we realized the 3D convolutional neural network for the fuction of 3D model's classification.And then based on the TX1 GPU we have improved the processing performance and got a very obvious effect in memory saving and speed increasing.