Method for generating backbone network, apparatus for generating backbone network, device, and storage medium

Patent metadata

Field	Value
Publication number	US-11929871-B2
Application number	US-202217718149-A
Country	US
Kind code	B2
Filing date	Apr 11, 2022
Priority date	Sep 16, 2021
Publication date	Mar 12, 2024
Grant date	Mar 12, 2024

Abstract

The present disclosure provides a method for generating a backbone network, an apparatus for generating a backbone network, a device, and a storage medium. The method includes: acquiring a set of a training image, a set of an inference image, and a set of an initial backbone network; training and inferring, for each initial backbone network in the set of the initial backbone network, the initial backbone network by using the set of the training image and the set of the inference image, to obtain an inference time and an inference accuracy of a trained backbone network in an inference process; determining a basic backbone network based on the inference time and the inference accuracy of the trained backbone network in the inference process; and obtaining a target backbone network based on the basic backbone network and a preset target network.

First claim

What is claimed is:

1. A method for generating a backbone network, comprising:
acquiring a set of training images, a set of inference images, and a set of initial backbone networks;
training and inferring, for each initial backbone network in the set of initial backbone networks, using the set of training images and the set of inference images to obtain an inference time and an inference accuracy of a trained backbone network in an inference process, wherein training and inferring, for each initial backbone network in the set of initial backbone networks, comprises:
training an initial backbone network by using the set of training images to obtain the trained backbone network, and
converting the trained backbone network into an inference network, inferring the set of inference images by using the inference network, and obtaining the inference time and the inference accuracy of the inference network in the inference process, wherein the inference time is a time used for inferring the set of inference images by using the inference network, and the inference accuracy is an accuracy of an inference result of the inference network; and

determining a basic backbone network based on the inference time and the inference accuracy of the trained backbone network in the inference process, wherein determining the basic backbone network comprises:
determining, based on the inference accuracy and the inference time of each inference network converted from the set of initial backbone networks, the basic backbone network from the set of initial backbone networks; and
obtaining a target backbone network based on the basic backbone network and a preset target network.

2. The method according to claim 1 , wherein the determining, based on the inference time and the inference accuracy of each inference network converted from the set of initial backbone networks, the basic backbone network from the set of initial backbone networks comprises:
plotting inference times and inference accuracies of the set of initial backbone networks in the inference process as points in a two-dimensional coordinate system; and
determining a target point from the points in the two-dimensional coordinate system, and determining an initial backbone network of the set of initial backbone networks corresponding to the target point as the basic backbone network.

3. The method according to claim 1 , wherein the obtaining the target backbone network based on the basic backbone network and the preset target network comprises:
acquiring a target network, wherein the target network comprises at least one of an activation function or a fully connected layer; and
adding the target network to the basic backbone network to obtain the target backbone network.

4. The method according to claim 1 , wherein the method further comprises:
updating a convolution kernel size of the target backbone network.

5. The method according to claim 1 , wherein the method further comprises:
adding a Squeeze-and-Excitation module to a predetermined target adding position in the target backbone network to obtain a final backbone network.

6. The method according to claim 5 , wherein the predetermined target adding position is determined through following steps of:
adding the Squeeze-and-Excitation module to different positions in the target backbone network to obtain a corresponding set of first backbone networks;
inferring, for each first backbone network in the corresponding set of first backbone networks, the set of inference images by using the corresponding set of first backbone networks to obtain an inference time and an inference accuracy of each first backbone network in an inference process; and
determining the predetermined target adding position based on the inference time and the inference accuracy of each first backbone network in the inference process.

7. The method according to claim 1 , wherein the set of initial backbone networks comprises at least one initial backbone network; and
the initial backbone network is obtained through following steps of:
acquiring a network block of a lightweight backbone network to obtain a set of the network block; and
randomly combining the network block in the set of the network block to obtain the initial backbone network.

8. The method according to claim 1 , wherein the method further comprises:
acquiring a to-be-classified image;
extracting a feature of the to-be-classified image by using the target backbone network to obtain an image feature; and
classifying the image feature to obtain a classification result.

9. An electronic device, comprising:
at least one processor; and
a memory communicatively connected to the at least one processor; wherein
the memory stores instructions executable by the at least one processor, and the instructions, when executed by the at least one processor, cause the at least one processor to perform operations comprising:
acquiring a set of training images, a set of inference images, and a set of initial backbone networks;
training and inferring, for each initial backbone network in the set of initial backbone networks, using the set of training images and the set of inference images to obtain an inference time and an inference accuracy of a trained backbone network in an inference process, wherein training and inferring for each initial backbone network in the set of initial backbone networks comprises:
training an initial backbone network by using the set of training images to obtain the trained backbone network, and
converting the trained backbone network into an inference network, inferring the set of inference images by using the inference network, and obtaining the inference time and the inference accuracy of the inference network in the inference process, wherein the inference time is a time used for inferring the set of inference images by using the inference network, and the inference accuracy is an accuracy of an inference result of the inference network; and

determining a basic backbone network based on the inference time and the inference accuracy of the trained backbone network in the inference process, wherein determining the basic backbone network comprises:
determining, based on inference accuracy and inference time of each inference network converted from the set of initial backbone networks, the basic backbone network from the set of initial backbone networks; and
obtaining a target backbone network based on the basic backbone network and a preset target network.

10. The electronic device according to claim 9 , wherein the determining, based on the inference time and the inference accuracy of each inference network converted from the set of initial backbone networks, the basic backbone network from the set of initial backbone networks comprises:
plotting inference times and inference accuracies of the set of initial backbone networks in the inference process as points in a two-dimensional coordinate system; and
determining a target point from the points in the two-dimensional coordinate system, and determining an initial backbone network of the set of initial backbone networks corresponding to the target point as the basic backbone network.

11. The electronic device according to claim 9 , wherein the obtaining the target backbone network based on the basic backbone network and the preset target network comprises:
acquiring a target network, wherein the target network comprises at least one of an activation function or a fully connected layer; and
adding the target network to the basic backbone network to obtain the target backbone network.

12. The electronic device according to claim 9 , wherein the ope

Assignees

• Beijing Baidu Netcom Sci & Tech Co Ltd

Inventors

• Cui Cheng
• Gao Tingquan
• Wei Shengyu
• Du Yuning
• Guo Ruoyu
• Lu Bin
• Zhou Ying
• Lyu Xueying
• Liu Qiwen
• Hu Xiaoguang
• Yu Dianhai
• Ma Yanjun

Classifications

• G06N3/082

  modifying the architecture, e.g. adding, deleting or silencing nodes or connections · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/09

  Supervised learning · CPC title

• H04L41/0806 Primary

  for initial configuration or provisioning, e.g. plug-and-play · CPC title

• G06F18/214

  Generating training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title