Method for realizing multi-label model training framework by utilizing missing multi-label data
A technology for model training and implementation methods, applied in neural learning methods, biological neural network models, neural architectures, etc., can solve problems such as underfitting, labor-intensive, overfitting, etc., to reduce the number of calculations and the consumption of computing resources , convenient deployment, the effect of reducing the size
Pending Publication Date: 2020-10-09
CHENGDU REMARK TECH CO LTD +1
View PDF0 Cites 1 Cited by
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
However, obtaining data with multi-label (attribute) annotations requires a lot of manpower; and usually, a larger network structure means better generalization ability
However, if the size of the data set cannot match the size of the model, the network will cause serious overfitting, resulting in a sharp decline in its generalization ability
When training multiple single-attribute models, since the scale of each data set may be different, if a single same network structure is used for training, overfitting and underfitting may occur
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 moreImage
Smart Image Click on the blue labels to locate them in the text.
Smart ImageViewing Examples
Examples
Experimental program
Comparison scheme
Effect test
Embodiment
[0029] Such as figure 1 As shown, this embodiment proposes a method for implementing a multi-label model training framework by using missing multi-label data. The method mainly includes the following steps:
[0030] 1. Data preprocessing: Integrate multiple single-label data sets to form an incomplete multi-label data set.
[0031] 1. Obtain and integrate multiple single-label data sets;
[0032] 2. When integrating multiple single-label datasets, set the other labels of samples from a certain label dataset in the merged dataset to -1, representing missing label values;
[0033] In the supervised learning mode, if a label has i possible values, the specific values of each attribute of the label are numbered in one-to-one correspondence from 0 to i-1. When integrating multiple single-label datasets, the data from one dataset I may not have the labels that other datasets have, so we set the other labels (attributes) of the samples from dataset I in the merged dataset as -1,...
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
Login to View More Abstract
The invention discloses a method for realizing a multi-label model training framework by utilizing missing multi-label data. According to the method, a plurality of single label data sets are integrated to form an incomplete multi-label data set, unified training is carried out by adopting a structure of combining a shared branch network with a non-shared branch network, and simultaneous output training of multiple labels of a target sample is realized. According to the invention, a plurality of labels (attributes) can be simultaneously trained by a single model, and deployment is convenient;the invention further discloses a training frame. The existence of a shared weight branch is adopted; the shared weight branches only need to be subjected to unique forward propagation calculation once, so that the calculation frequency and the calculation resource consumption in the network forward propagation process are remarkably reduced, network parameters contained in the shared feature extractor are shared due to the shared feature extractor, and the size of the model is remarkably reduced.
Description
technical field [0001] The present invention relates to, in particular, an implementation method for realizing a multi-label model training framework by using missing multi-label data. Background technique [0002] Under the framework of supervised learning, using a deep learning model to complete classification and recognition tasks requires each piece of training data to have a label (attribute). At the same time, when a single model is required to be able to predict multiple labels (attributes) at the same time, it is usually necessary to ensure that each label (attribute) of each training sample is complete. However, obtaining data with multi-label (attribute) annotations requires a lot of manpower; and in general, a larger network structure means better generalization ability. But if the size of the dataset does not match the size of the model, the network can suffer from severe overfitting, causing its generalization ability to drop dramatically. When training multip...
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
Login to View More Patent Type & Authority Applications(China)
IPC IPC(8): G06N3/04G06N3/08
CPCG06N3/084G06N3/045
Inventor 肖利喻杨洋王飞
Owner CHENGDU REMARK TECH CO LTD