Decentralized training method for heterogeneous edge computing platform
An edge computing and training method technology, applied in the field of artificial intelligence, can solve problems such as increasing model synchronization and model convergence total time, and achieve the effect of reducing accuracy loss and delay loss
Pending Publication Date: 2021-11-16
BEIJING INSTITUTE OF TECHNOLOGYGY
View PDF0 Cites 0 Cited by
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
[0007] In the current decentralized training process, due to the heterogeneity of edge device hardware and software, the training time of different edge computing platforms is highly different, thus increasing the total time of model synchronization and model convergence
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
[0066] Such as Figure 3-5 As shown, the decentralized training method for the heterogeneous edge computing platform includes:
[0067] S1, in the pre-training stage, the aggregated data point generation module converts the local data set X based on the data similarity e Divided into M e (M e >1) subsets, and generate aggregated data points corresponding to each subset
[0068] S11, using data dimensionality reduction technology to reduce the dimension to N e ×d local dataset X e Convert to dimension N e Dimensionality reduction data set X of ×d'(d'e′ , N e Indicates the total number of data points contained in the original data set, d and d' indicate the number of eigenvalues of each data point before and after dimensionality reduction;
[0069] S12, use the data similarity algorithm to divide the dimension to N e × d e′ The dimensionality reduction dataset X e′ Divided into M e subset
[0070] S13, based on dimensionality reduction dataset X e′ The partiti...
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 provides a decentralized training method for a heterogeneous edge computing platform. The method comprises the following steps: 1, in a pre-training stage, dividing a local data set Xe into Me (Me is greater than 1) subsets by an aggregated data point generation module based on data similarity, and generating aggregated data points corresponding to each subset; and S2, in an iteration training stage, using a gossip training balance algorithm to calculate a next epoch training input data proportion of each edge computing platform according to the prediction performance of each edge computing platform, and enabling data points most related to model precision in a local data set to participate in training first by each edge computing platform. Training time differences of different edge computing platforms caused by heterogeneity of hardware and software are balanced, so that some additional delay losses of model synchronization and convergence caused by the training time differences are reduced.
Description
technical field [0001] The invention belongs to the technical field of artificial intelligence, and in particular relates to a decentralized training method for a heterogeneous edge computing platform. Background technique [0002] In recent years, the rapid development of edge computing platforms has extended cloud service capabilities to the edge of the network, such as mobile devices and sensor devices. The popularity of Apache Edgent has promoted the application of deep learning models on edge devices. However, today's edge computing platforms are highly heterogeneous (approximately 2 million device configuration types) for two reasons: First, edge computing platforms have various types of hardware (such as TITAN Xp or RTX 2080 Ti graphics card such as accelerators, or different types of CPUs), and hardware resources of different sizes (such as the number of processor cores, memory size, and I / O bandwidth); second, the hardware resources of the edge computing platform w...
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): G06K9/62G06N3/04G06N3/08
CPCG06N3/084G06N3/045G06F18/23213G06F18/214
Inventor 韩锐刘驰辛高枫李世林
Owner BEIJING INSTITUTE OF TECHNOLOGYGY