Accelerated execution method of deep learning model in dynamic change network environment

Abstract

The invention discloses an accelerated execution method of a deep learning model in a dynamic change network environment, and the method employs a network flow algorithm to cut a DAG type neural network, allows the DNN to be disposed at an edge end and a cloud end at the same time, limits the data transmission, and accelerates the reasoning. By utilizing the property of undirected graph topping, a two-step method is provided, and the decision time of a minimum cut model is shortened. Compared with other methods, the new time delay measurement method can achieve the effects of reducing the reasoning time delay and improving the throughput. According to the method, the convolutional neural network can be adaptively cut according to the network speed, when the network speed is high, a calculation task is given to a cloud end to be processed as much as possible, and when the network speed is low, the calculation task is calculated at the edge as much as possible, and an intermediate result is transmitted to the cloud end to be processed.

Description

technical field [0001] The invention provides an accelerated execution method of a deep learning model in a dynamically changing network environment, which is mainly applied in the fields of the Internet of Things and edge computing, and involves neural network algorithms, network flows, and Tarjan algorithms. Background technique [0002] Deep learning has made a lot of progress in recent years and has been widely used in many fields. Especially in the field of computer vision, it has brought the speed and accuracy of image recognition and video analysis to a new level. Many IoT devices cooperate with a powerful cloud computing platform to expand many vision applications based on deep learning. For example, in autonomous driving technology, the video stream data generated by the on-board camera is uploaded to the server, and the server performs semantic segmentation and target detection on the images in the video, and then sends the data back to the decision center for aut...

AI Technical Summary

Problems solved by technology

The chain topology cutting method led by Neurosurgeon cannot effectively split these models under these circumstances

2) The work of dividing the neural network of the DAG topology, such as DADS, has a large time complexity

Considering that the segmentation decision is generally made on the resource-constrained terminal device, the usability is affected

3) Existing models are inaccurate in estimating the time of each layer

In fact, many existing machine learning frameworks optimize the activation function, which leads to a large difference between the overall running time of multiple layers and the sum of the running time of each layer alone.

These optimizations lead to the fact that the existing segmentation methods cannot achieve optimal results in real edge environments

Images