Method and apparatus for training classification model and data classification

Patent metadata

Field	Value
Publication number	US-12475687-B2
Application number	US-202217819777-A
Country	US
Kind code	B2
Filing date	Aug 15, 2022
Priority date	Jun 16, 2021
Publication date	Nov 18, 2025
Grant date	Nov 18, 2025

Abstract

A method and an apparatus for training a classification model and data classification includes: obtaining a sample set and a pre-trained classification model, wherein the classification model includes at least two convolutional layers, each convolutional layer is connected to a classification layer through a fully connected layer; inputting the sample set into the classification model, and obtaining a prediction result output by each classification layer, wherein the prediction result includes a prediction probability of a class to which each sample belongs; calculating a probability threshold of each classification layer based on the prediction result output by each classification layer; setting a prediction stopping condition for the classification mode according to the probability threshold of each classification layer.

First claim

What is claimed is:

1 . A method for data classification, applied in intelligent traffic scenarios, comprising:
obtaining a sample set and a pre-trained classification model, wherein the pre-trained classification model comprises at least two convolutional layers, each convolutional layer is connected with a classification layer through a full connected layer; inputting the sample set into the pre-trained classification model to obtain a prediction result output by each classification layer, wherein the prediction result comprises a prediction probability of a category to which each sample belongs; calculating a probability threshold of each classification layer based on the prediction result output by each classification layer; setting a prediction stopping condition of the classification model according to the probability threshold of each classification layer; inputting data to be classified into a classification model, wherein the data to be classified is a image acquired by a sensor; taking a first convolutional layer as a current convolutional layer, and performing following classification steps of: predicting the data through the current convolutional layer, a current fully connected layer and a current classification layer to obtain a maximum prediction probability; if the maximum prediction probability is greater than a probability threshold of the current classification layer, stopping prediction, and using a class corresponding to the maximum prediction probability as a class of the data; and otherwise, inputting an output result of the current convolutional layer to a next convolutional layer, and using the next convolutional layer as the current convolutional layer to continue the above classification steps; wherein calculating a probability threshold of each classification layer based on the prediction result output by each classification layer comprises:
performing determining steps of selecting a target combination from a predetermined data proportion combination set; determining a planning value of each classification layer corresponding to the target combination according to the prediction result output by each classification layer; and calculating a correct rate corresponding to the target combination based on the prediction result output by each classification layer;
repeating the determining steps until traversal of the data proportion combination set is completed, and obtaining the correct rate corresponding to each data proportion combination; and
using the planning value of each classification layer corresponding to the data proportion combination with a maximum correct rate as the probability threshold of each classification layer;
wherein, calculating a correct rate corresponding to the target combination based on the prediction result output by each classification layer comprises:
determining a maximum prediction probability of each sample in each classification layer based on the prediction result output by each classification layer; and
for each classification layer, traversing each sample, and if the maximum prediction probability of the sample in the classification layer is greater than the planning value of the classification layer, accumulating the maximum prediction probability of the sample in the classification layer for the correct rate;
wherein the prediction result comprises a prediction time of each sample; and wherein the method further comprises:
for each classification layer, calculating a total prediction time of the classification layer based on the prediction time of the samples participating in correct rate accumulation;
calculating a total prediction time of the classification model based on the total prediction time of each classification layer; and
if the total prediction time of the classification model is greater than a predetermined time threshold, filtering out the correct rate corresponding to the target combination.

2 . The method according to claim 1 , wherein, calculating the total prediction time of the classification layer based on the prediction time of the samples participating in correct rate accumulation, comprises:
calculating an average prediction time of the classification layer based on the prediction time of the samples participating in correct rate accumulation; and calculating a sum of the average prediction time and a predetermined jitter value as the total prediction time of the classification layer.

3 . The method according to claim 1 , wherein the prediction result comprises a number of operations of each sample; and
the method further comprises:
for each classification layer, calculating a total number of operations of the classification layer based on the number of operations of the samples participating in correct rate accumulation;
calculating a total number of operations of the classification model based on the total number of operations of each classification layer; and
if the total number of operations of the classification model is greater than a predetermined number threshold, filtering out the correct rate corresponding to the target combination.

4 . The method according to claim 1 , wherein the pre-trained classification model is trained by steps of:
obtaining a training data set and a classification model, wherein training data in the training data set has class labels; performing following training steps of selecting training data from the training data set; inputting selected training data into the classification model to obtain a prediction probability output by each classification layer; and calculating a total loss value based on the class labels of the selected training data and the prediction probability output by each classification layer, wherein if the total loss value is less than a target value, the training of the classification model is completed; and if the total loss value is not less than the target value, adjusting relevant parameters of the classification model, and continuing to perform the training steps.

5 . An electronic device, applied in intelligent traffic scenarios, comprising:
at least one processor; and a memory communicatively coupled to the at least one processor; wherein, the memory is stored with instructions executable by the at least one processor, and the at least one processor is configured to:
obtain a sample set and a pre-trained classification model, wherein the pre-trained classification model comprises at least two convolutional layers, each convolutional layer is connected with a classification layer through a full connected layer;
input the sample set into the pre-trained classification model to obtain a prediction result output by each classification layer, wherein the prediction result comprises a prediction probability of a category to which each sample belongs;
calculate a probability threshold of each classification layer based on the prediction result output by each classification layer; and
set a prediction stopping condition of the classification model according to the probability threshold of each classification layer;
input data to be classified into a classification model, wherein the data to be classified is a image acquired by a sensor;
take a first convolutional layer as a current convolutional layer, and performing following classification steps of: predicting the data through the current convolutional layer, a current fully connected layer and a current classification layer to obtain a maximum prediction probability; if the maximum prediction probability is greater than a probability threshold of the current classification layer, stopping prediction, and using a class corresponding to the maximum prediction probability as a class of the data; and
otherwise, input an output result of the current convolutional layer to a

Assignees

• Beijing Baidu Netcom Sci & Tech Co Ltd
• State Key Laboratory Of Internet Of Things For Smart City Univ Of Macau

Inventors

• Wang Kafeng
• Xu Chengzhong
• Xiong Haoyi
• Li Xingjian
• Dou Dejing

Classifications

• G06N3/0985

  Hyperparameter optimisation; Meta-learning; Learning-to-learn · CPC title

• G06V10/764

  using classification, e.g. of video objects · CPC title

• G06V10/82 Primary

  using neural networks · CPC title

• G06V10/778

  Active pattern-learning, e.g. online learning of image or video features · CPC title

• G06F18/2415

  based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate · CPC title