Image recognition model rapid training method and system based on many-core processor

Abstract

The invention discloses an image recognition model rapid training method based on a many-core processor, and the method employs the many-core processor, and the image recognition model rapid training method comprises: constructing an image recognition model, and acquiring image data for training, and training the image recognition model parameters by using an improved asynchronous stochastic gradient descent algorithm to obtain a trained image recognition model. The method for training the image recognition model parameters by using the improved asynchronous stochastic gradient descent algorithm comprises: carrying out calculation and learning through calculation nodes, carrying out training on distributed image data, and updating corresponding parameters; receiving, by the parameter server, image recognition model parameters sent by the computing node; and tracking and calculating the updating delay of each calculation node, and if the updating delay is smaller than a threshold value, updating the model parameters; otherwise, discarding the model parameters. A better acceleration effect can be obtained on a many-core processor platform, so that the training speed of the image recognition model is higher, and the training cost is greatly reduced.

Description

technical field [0001] The invention belongs to the technical field of image recognition model training, and in particular relates to a fast training method and system for an image recognition model based on many-core processors. Background technique [0002] In recent years, deep neural network (DNN) has been widely used in many fields due to its excellent algorithm performance, especially in image recognition. DNNs have made great progress because it can perform feature extraction and data fitting by increasing the depth of the model. However, the excellent performance of DNNs comes with a major obstacle—huge computational cost. With the increase of training data scale and model complexity, the training cost of DNNs is getting higher and higher, which will become the bottleneck of image recognition model training. [0003] To shorten the training time of image recognition models, designing a parallel DNN algorithm based on various high-performance computing platforms bec...

AI Technical Summary

Problems solved by technology

SSGD uses the parameter information of all nodes, so the synchronous waiting brought by slow nodes makes the acceleration ratio of data parallelism unsatisfactory

ASGD achieves greater parallelism by reducing the dependencies between nodes, but due to its inherent randomness, it causes a large amount of gradient aging in the network during training, resulting in it taking a long time to reach the same convergence point

Images