Method and device for deep learning system of hybrid architecture

Abstract

The invention discloses a method and device for a deep learning system of a hybrid architecture. The method is characterized by comprising the steps in which, when a training data set is updated, a training module re-trains a deep learning network model and stores weights and offset parameters; a server side monitors changes in a parameter file while monitoring the process, encapsulates the changes into a pre-set data structure and notifies a reasoning module; the reasoning module interrupts a reasoning service, reads the weights and offset file content from the server side and updates the network model; and the server side monitors the process, meanwhile processes the input file that needs reasoning and notifies the reasoning module. The system device comprises a server module, a training module, a reasoning module and a bus interface. The training and reasoning hybrid CPU+GPU+CAPI heterogeneous deep learning system of the invention can make full use of resources, gain higher energy efficiency, and realize CAPI direct access to the server memory and real-time online iterations to update the weights and other parameters of the reasoning model.

Description

technical field [0001] The present invention relates to the technical field of circuit design and machine learning, in particular to a method and device for a deep learning system with a hybrid architecture. Background technique [0002] The rapid development of the information technology industry in the 21st century has brought huge benefits and convenience to people. Deep learning applications are divided into two parts: training and reasoning. Taking ImageNet evaluation as an example, the AlexNet model training process requires 8 million pictures with a total of 1,000 categories. After the AlexNet model extracts features and calculates the loss, then through backpropagation such as SGD To update the weight parameters, so as to continuously converge the model, and finally get the ideal network model. The reasoning process is the process of inputting a forward operation through the network model to obtain the accuracy of the final classification (generally choose Top5). T...

AI Technical Summary

Problems solved by technology

[0004] The existing technology has the following deficiencies: 1) The general method adopts the separation of training and reasoning, which needs to maintain two sets of platform environments, and the resources cannot be fully utilized; 2) FPGA / CPLD is completely used for deep learning calculations, and the computing power is not strong enough. Not suitable for large-scale training scenarios; 3) The communication between FPGA / CPLD and the server is generally through DMA, and the time delay between data and CPU server interaction is relatively large

Images