Distributed deep neural network structure conversion method based on splitting-fusion strategy

Abstract

The invention discloses a distributed deep neural network structure conversion method based on a splitting-fusion strategy, and the method comprises the steps of collecting pictures, forming a data set, dividing the data set into a training set and a test set, and preprocessing the pictures in the training set and the test set; performing channel splitting on the network model, expanding the width of the network model, and then adding a fusion layer to obtain an improved network model; and training the improved network model by using the pictures in the training set, and inputting the preprocessed pictures in the test set into the trained network model to obtain a picture classification result. According to the method, the existing network model for the picture classification task is improved aiming at a distributed reasoning scene, and the converted network model can realize higher reasoning speed in a resource-limited distributed system, namely, the picture classification task can be executed in the resource-limited distributed system, a classification result can be more quickly obtained, and the images have higher classification precision.

Description

technical field [0001] The invention belongs to the technical field of artificial intelligence, and relates to a method for converting a distributed deep neural network structure based on a split-merge strategy, which can convert an existing image classification network model into a network model suitable for a distributed system, so as to Perform image classification tasks on resource-constrained distributed systems, and can also be used in other computer vision tasks such as object detection and semantic segmentation. Background technique [0002] Convolutional neural networks have become the mainstream method for image classification tasks, and have achieved the best accuracy in multiple image classification datasets such as CIFAR10 and ImageNet. For the technology to be more widely used, it is necessary to deploy the network model to hardware such as embedded and mobile devices. However, the mainstream convolutional neural network models are usually very large, such as ...

AI Technical Summary

Problems solved by technology

[0003] However, these network models are all designed for a single hardware. With the rapid development of the Internet of Things (IoT) technology, edge computing has become a recent research hotspot and difficulty. Deploying network models in distributed systems has gradually become an extensive Compared with deploying a network on a single hardware, it is more difficult to deploy a network model in a distributed system. There are two main reasons: (1) A distributed system usually contains many edge devices, and the computing power of a single edge node Usually very weak, even the inference of a lightweight network model cannot be done

(2) Communication between different edge nodes will generate additional overhead

But doing so creates new problems

During the entire forward reasoning process, there is no information exchange between different groups in the network model, that is, the phenomenon of "information blocking" appears, which seriously affects the accuracy of the network.

[0006] The parallel strategy of the existing distributed reasoning scheme (Deepthings, Fully Deepthings) is to divide the feature map of each layer in the network model, so that different nodes can reason about different regions of the feature map, but this strategy is not effective in distributed systems. When performing the inference of the image classification task of the convolutional neural network, it will still bring a lot of extra overhead (communication overhead or computing overhead), which will seriously prolong the inference time (the time required to get the classification result ), which has a huge impact on occasions that need to get classification results quickly

[0007] There is a problem of low accuracy in image classification in the prior art

Images