Eureka AIR delivers breakthrough ideas for toughest innovation challenges, trusted by R&D personnel around the world.

Pruning method applied to target detection and terminal

A target detection algorithm and target detection technology, applied in the field of computer vision model compression, can solve problems such as loss of precision, achieve the effects of reducing loss of precision, less time-consuming, and simple pruning process

Active Publication Date: 2020-09-01
SANLI VIDEO FREQUENCY SCI & TECH SHENZHEN
View PDF11 Cites 8 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

[0007] Although all the methods proposed above can compress the model, the pruning in the method proposed in the first article is only a fine-grained parameter-level pruning, and it must cooperate with corresponding hardware (such as FPGA) to complete the acceleration. In addition, the quantization scheme and the conversion model format storage scheme must be implemented with fixed hardware, which has considerable limitations; in the method proposed in the second article, the gamma parameter needs to be sparse before pruning, and the network recovery accuracy and pruning should be fine-tuned again. , re-fine-tuning and other steps, and the situation of each data set is different, requires a lot of experience and time, and must rely on the BN layer in the model; although the method proposed in the third article is simple, it has problems when facing complex networks. The problem of loss of precision

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
  • Pruning method applied to target detection and terminal
  • Pruning method applied to target detection and terminal
  • Pruning method applied to target detection and terminal

Examples

Experimental program
Comparison scheme
Effect test

Embodiment 1

[0074] Please refer to figure 1 , a pruning method applied to target detection, including steps:

[0075] S1. Train the preset target detection algorithm until it converges;

[0076] Wherein, the preset target detection algorithm includes but is not limited to currently popular deep learning target detection algorithms, such as algorithms based on deep learning such as Yolov3, SSD, faster rcnn, retinanet, etc.;

[0077] Specifically, you can use the existing data set to train the preset target detection algorithm, train the target detection algorithm to convergence, and use the test standard map of pascalvoc to evaluate the model to obtain objective evaluation data;

[0078] S2. Determine the pruning ratio corresponding to each network layer according to the sum of the norms of the weights of each network layer in the deep neural network model of the target detection algorithm after convergence;

[0079] Determine the weight norm mean value corresponding to each network laye...

Embodiment 2

[0118] Please refer to figure 2 , a pruning terminal 1 applied to target detection, comprising a memory 2, a processor 3 and a computer program stored on the memory 2 and operable on the processor 3, the processor 3 implements the computer program when executing the computer program Each step in the first embodiment.

Embodiment 3

[0120] Test the above pruning method applied to object detection:

[0121] Considering that the compression strengths of different algorithms are inconsistent, this embodiment hopes to compare the loss of precision after compression under the same compression strength. Therefore, this embodiment reproduces the second method and the third method described in the background technology method to prune Yolov3 (only one pruning, no iterative pruning), and set a uniform compression of 0.2percent for each layer of the backbone network darknet53 of the Yolov3 algorithm, that is, each layer retains 0.8 times the original channels, then The original channel and the current pruned channel number are compared according to the order of each layer of darknet53, as shown in Table 1 (only the channels of the convolutional layer are compared, and the shortcut needs to have the same dimension to be added. In order to reduce the complexity, the shortcut is not cut uniformly. layer channel):

[...

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 pruning method applied to target detection and a terminal. The pruning method comprises the following steps: training a preset target detection algorithm until convergence; determining a pruning proportion corresponding to each network layer according to the norm of the weight of each network layer in the converged deep neural network model of the target detection algorithm; clustering the weights of all channels in the network layer corresponding to the pruning proportion according to the pruning proportion, and determining a clustering center corresponding to each network layer; according to the clustering center and the pruning proportion corresponding to each network layer, pruning channels in the network layer where the network layer is located; realizing pruning of channels of each network layer in the deep neural network model based on norm weight clustering; and deleting redundant channels to compress the deep neural network model, thus the pruning process is simple, consumed time is short, dependence is little in the pruning process, dependence on any parameter and a specific layer is not needed, and meanwhile precision loss is reduced while compression is guaranteed.

Description

technical field [0001] The invention relates to the technical field of computer vision model compression, in particular to a pruning method and terminal applied to target detection. Background technique [0002] In computer vision, model compression has always been an important direction for applying deep learning to embedded devices, and model pruning is one of the important sub-directions. Compared with quantization and distillation, pruning is easier to use and more Suitable for deployment and other advantages. Among them, model pruning is to measure the importance of each neuron weight in deep learning by different methods, and subtract unimportant neurons according to the importance of neurons, so as to achieve the purpose of compressing the model. [0003] As a commonly used model compression method, model pruning is widely used to reduce the heavy calculation of deep models. It is a research hotspot in the current academic and industrial circles, such as: [0004] (...

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
Patent Type & Authority Applications(China)
IPC IPC(8): G06N3/08G06N3/04G06K9/62
CPCG06N3/082G06N3/045G06F18/23
Inventor 潘成龙张宇刘东剑
Owner SANLI VIDEO FREQUENCY SCI & TECH SHENZHEN
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
Eureka Blog
Learn More
PatSnap group products