Detection model compression method based on semantic segmentation

A semantic segmentation and detection model technology, applied in neural learning methods, biological neural network models, special data processing applications, etc., can solve the problems of not considering the correlation of parameters and low model compression ratio

Active Publication Date: 2019-07-09
北京同方软件有限公司
View PDF6 Cites 1 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

The compression process is still based on a single parameter, without considering t

Method used

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Image

Smart Image Click on the blue labels to locate them in the text.
Viewing Examples
Smart Image
  • Detection model compression method based on semantic segmentation
  • Detection model compression method based on semantic segmentation
  • Detection model compression method based on semantic segmentation

Examples

Experimental program
Comparison scheme
Effect test

Embodiment Construction

[0033] see figure 1 , the present invention a kind of detection model compression method based on semantic segmentation, its steps are:

[0034] (1) Pruning:

[0035] 1) Input the convolution kernel weight.

[0036] 2) Prune the trained network model to remove redundant weights in the model that are less than a certain threshold. The method is to replace these weight values ​​with 0, and obtain the parameter space of sparse weights through the pruning algorithm. And these 0 elements largely become auxiliary information for boundary search during semantic segmentation. can effectively help the semantic segmentation module.

[0037] (2) Semantic Segmentation:

[0038] 1) Perform semantic segmentation on the parameter space, obtain the hyperparameter block and the central vocabulary, and calculate the center position of the hyperparameter block, the method is:

[0039] Initially define the region of the parameter space, such as figure 2 shown.

[0040] Find the lowest...

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to view more

PUM

No PUM Login to view more

Abstract

The invention discloses a detection model compression method based on semantic segmentation, and relates to the fields of artificial intelligence and computer vision. The method comprises the following steps: (1) pruning: 1) inputting a convolution kernel weight; 2) pruning is carried out on the trained network model to obtain a parameter space of sparse weights; (2) semantic segmentation: 1) performing semantic segmentation on the parameter space to obtain a hyper-parameter block and a central vocabulary, and calculating the central position of the hyper-parameter block; and 2) updating the original parameter space by using the central vocabulary. 3) judging whether the change of the current central vocabulary and the previous central vocabulary is smaller than a specified threshold value, if so, continuing to search parameters similar to the central vocabulary, and updating the central vocabulary and returning to the step 2); and ending the updating of the central vocabulary if the current vocabulary is smaller than the threshold. And (3) model storage: storing the hyper-parameter block boundary position, the parameter block center position and the center vocabulary value obtained by training. According to the method, hyper-parameters are used for describing the whole parameter space, overall compression of the parameter space is achieved, and the overall compression ratio ofthe model is increased to the maximum extent.

Description

technical field [0001] The invention relates to the fields of artificial intelligence and computer vision, in particular to a method for compressing a detection model through semantic analysis. Background technique [0002] In the paper "Dynamic Network Surgery for Efficient DNNs" on NIPS in 2016, the deep network model was compressed based on the dynamic pruning algorithm. The optimization algorithm of how to train the network model and compress the network model at the same time is studied. [0003] This dynamic model pruning algorithm mainly includes the following two processes: pruning and connection. Pruning is to cut off unimportant weight parameters, but as the network model training process proceeds, the importance of weights is constantly iterating. Update, so it is impossible to make an intuitive estimate of the importance of these weights, so a connection process is added here. The connection is to estimate the importance of the cut weights, and restore those im...

Claims

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to view more

Application Information

Patent Timeline
no application Login to view more
IPC IPC(8): G06N3/08G06F17/27
CPCG06N3/082G06F40/30
Inventor 刘阳郑全新赵英张磊董小栋孟祥松邓家勇江龙王亚涛
Owner 北京同方软件有限公司
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Try Eureka
PatSnap group products

Browse by: Latest US Patents, China's latest patents, Technical Efficacy Thesaurus, Application Domain, Technology Topic.

© 2024 PatSnap. All rights reserved.Legal|Privacy policy|Modern Slavery Act Transparency Statement|Sitemap