CN-121998122-A - Model training method, device, equipment, medium and computer program product

CN121998122ACN 121998122 ACN121998122 ACN 121998122ACN-121998122-A

Abstract

The invention provides a model training method, a model training device, model training equipment, model training media and model training computer program products, and relates to the technical field of artificial intelligence. The method comprises the steps of obtaining a pulse model which is trained according to a dynamic visual data set collected by a dynamic visual sensor, obtaining an initial quantized pulse model according to a predefined cutoff function, obtaining a reliability estimated value of the initial quantized pulse model according to a first moment of an input variable of the initial quantized pulse model, updating the initial quantized pulse model, and obtaining a quantized pulse model which converts continuous data collected by the dynamic visual sensor into pulse sequence data. The reliability estimation value is obtained through the first moment of the input variable, and the nonlinear estimation is adopted to have closed solution, so that the accurate estimation can be realized, and the prediction consistency of the time sequence data is further improved.

Inventors

WANG ZHAO
ZHANG SHAOQUN
LIU ZIANG
WU YULUN
LV YAN
Tan Ruirui
LIANG SHUANG
ZHANG ZHENG

Assignees

中国移动紫金(江苏)创新研究院有限公司
中国移动通信集团江苏有限公司
中国移动通信集团有限公司
南京大学

Dates

Publication Date: 20260508
Application Date: 20260127

Claims (10)

1. A method of model training, comprising: Acquiring a pulse model which is trained according to a dynamic visual data set and a predefined cut-off function, wherein the dynamic visual data set is environmental change data which is collected by a dynamic visual sensor and recorded by adopting an asynchronous event stream form; Acquiring an initial quantized pulse model according to the pulse model and the truncation function; Acquiring a credibility estimation value of the initial quantized pulse model according to a first moment of an input variable of the initial quantized pulse model; And updating the initial quantized pulse model according to the credibility estimation value to obtain a quantized pulse model, wherein the quantized pulse model is used for converting continuous data collected by a dynamic vision sensor into pulse sequence data.
2. The method of claim 1, wherein obtaining an initial quantized pulse model from the pulse model and the truncation function comprises: Acquiring a first quantization parameter according to the truncation function and a first full-precision parameter of the pulse model; Constructing a quantized pulse model to be trained according to the first quantization parameter; Acquiring a first parameter updating value of the quantized pulse model to be trained according to a first gradient, wherein the first gradient is acquired according to an approximate derivative of the truncated function and a second gradient of the pulse model; And adjusting the quantized pulse model to be trained according to a second quantized parameter to obtain the initial quantized pulse model, wherein the second quantized parameter is obtained according to the first parameter updating value and the cutoff function.
3. The method of claim 1, wherein obtaining the confidence estimate for the initial quantized pulse model based on the first moment of the input variable for the initial quantized pulse model comprises: acquiring a second moment of a pre-activation variable according to the first moment of the input variable; Processing the second moment by adopting an activation function, and converting the pre-activation variable into a post-activation variable; And obtaining a third moment of the activated variable, wherein the third moment comprises the credibility estimated value.
4. The method of claim 1, wherein updating the initial quantized pulse model based on the confidence estimate to obtain a quantized pulse model comprises: Establishing a consistency optimization problem according to the credibility estimation value; and obtaining the quantized pulse model by solving the consistency optimization problem.
5. The method of claim 4, wherein obtaining the quantized pulse model by solving the consistency optimization problem comprises: updating the first full-precision parameters of the pulse model by a gradient descent method to obtain second full-precision parameters; Updating a truncation threshold of the truncation function according to the second full-precision parameter; Intercepting the second full-precision parameter by adopting the interception threshold and the interception function to obtain a third quantized parameter; and acquiring the quantized pulse model according to the third quantization parameter.
6. The method of claim 1, wherein obtaining a pulse model trained from a dynamic visual dataset comprises: constructing an initial pulse model and acquiring the dynamic visual data set; Setting a loss function of the initial pulse model, and establishing a dynamic visual training optimization problem of the initial pulse model; solving the dynamic visual training optimization problem through a substitution gradient to obtain a second parameter updating value of the initial pulse model; and acquiring the pulse model according to the second parameter updating value and the dynamic visual data set.
7. A model training device, comprising: The system comprises a first acquisition module, a second acquisition module and a control module, wherein the first acquisition module is used for acquiring a pulse model which is trained according to a dynamic visual data set and a predefined cut-off function, wherein the dynamic visual data set is environmental change data which is collected by a dynamic visual sensor and recorded by adopting an asynchronous event stream form; the second acquisition module is used for acquiring an initial quantized pulse model according to the pulse model and the truncation function; The third acquisition module is used for acquiring the credibility estimation value of the initial quantized pulse model according to the first moment of the input variable of the initial quantized pulse model; And the fourth acquisition module is used for updating the initial quantized pulse model according to the credibility estimation value to acquire a quantized pulse model, wherein the quantized pulse model is used for converting continuous data collected by the dynamic vision sensor into pulse sequence data.
8. Model training apparatus comprising a processor, a memory and a program or instructions stored on the memory and executable on the processor, characterized in that the processor implements the model training method according to any of claims 1-6 when executing the program or instructions.
9. A readable storage medium having stored thereon a program or instructions which when executed by a processor performs the steps in the model training method according to any of claims 1-6.
10. A computer program product comprising computer instructions which, when executed by a processor, implement the steps of the model training method of any of claims 1-6.

Description

Model training method, device, equipment, medium and computer program product Technical Field The present invention relates to the field of artificial intelligence technology, and in particular, to a model training method, apparatus, device, medium, and computer program product. Background With the rapid development of brain-like intelligence and event-driven perception technologies, dynamic vision systems represented by dynamic vision sensors (Dynamic Vision Sensor, DVS) have shown important application potential in resource-constrained scenes such as intelligent robots, automatic driving, unmanned systems, edge computing and the like due to the advantages of high time resolution, low power consumption, high dynamic range and the like. The dynamic vision sensor outputs environment change information in an asynchronous event stream mode, and the data of the dynamic vision sensor has the remarkable characteristics of time continuity, spatial sparseness, non-framing and the like, and provides new challenges for a traditional frame-based vision modeling method. The impulse neural network (Spiking Neural Network, SNN) takes the pulse sequence as a basic information carrier, has natural consistency with dynamic visual data in terms of time coding, energy efficiency and event-driven processing, and is considered as an important modeling model of dynamic visual perception. However, the SNN still has the defects of low reasoning speed, high-precision numerical value storage (difficult to be deployed on basic hardware), low prediction accuracy, low prediction consistency and the like, and restricts the large-scale application of the SNN in the edge intelligent and embedded systems. Therefore, the light construction and consistency guarantee of the impulse neural network are important problems for realizing high-performance calculation and light deployment of dynamic visual scenes. The quantization technology is a main technology for realizing the lightweight construction of the impulse neural network at present. SNN-Compression-ADMM implements network pruning and model Compression through ADMM optimization. ANN2SNN-Scaling achieves model compression by converting quantized ANN into equivalent SNN. Experimental results show that the method greatly compresses the model size and improves the reasoning speed, but the method still has the defects of high-precision numerical value storage, low prediction accuracy, low prediction consistency and the like. Disclosure of Invention The invention aims to provide a model training method, a device, equipment, a medium and a computer program product, which are used for solving the problems of high-precision numerical value storage, low prediction accuracy and low prediction consistency of a pulse neural network quantization method in a dynamic visual scene in the prior art. To achieve the above object, an embodiment of the present invention provides a model training method, including: Acquiring a pulse model which is trained according to a dynamic visual data set and a predefined cut-off function, wherein the dynamic visual data set is environmental change data which is collected by a dynamic visual sensor and recorded by adopting an asynchronous event stream form; Acquiring an initial quantized pulse model according to the pulse model and the truncation function; Acquiring a credibility estimation value of the initial quantized pulse model according to a first moment of an input variable of the initial quantized pulse model; And updating the initial quantized pulse model according to the credibility estimation value to obtain a quantized pulse model, wherein the quantized pulse model is used for converting continuous data collected by a dynamic vision sensor into pulse sequence data. Optionally, the method, wherein obtaining an initial quantized pulse model according to the pulse model and the truncation function includes: Acquiring a first quantization parameter according to the truncation function and a first full-precision parameter of the pulse model; Constructing a quantized pulse model to be trained according to the first quantization parameter; Acquiring a first parameter updating value of the quantized pulse model to be trained according to a first gradient, wherein the first gradient is acquired according to an approximate derivative of the truncated function and a second gradient of the pulse model; And adjusting the quantized pulse model to be trained according to a second quantized parameter to obtain the initial quantized pulse model, wherein the second quantized parameter is obtained according to the first parameter updating value and the cutoff function. Optionally, the method, wherein obtaining the reliability estimation value of the initial quantized pulse model according to the first moment of the input variable of the initial quantized pulse model includes: acquiring a second moment of a pre-activation variable according to the first moment of the input variab