Deep learning training inference system based on FPGA

Abstract

The invention relates to a deep learning training inference system based on an FPGA. The deep learning training inference system comprises a cloud training device, a sensor device, a result comprehensive display device and an edge training inference device. The cloud training equipment, the sensor equipment and the result comprehensive display equipment are respectively connected with the edge end training reasoning equipment; the cloud training equipment is used for acquiring initial training parameters and transmitting the initial training parameters to the edge end training reasoning equipment; the sensor equipment is used for collecting data and transmitting the collected data to the edge end training reasoning equipment; the edge end training reasoning equipment performs reasoning and online training on the received initial training parameters and data; and the updated training parameters and reasoning results are output to a result comprehensive display device. The method has the advantages that high-speed and high-efficiency reasoning can be realized, the training requirement of edge end online learning can be met, meanwhile, the ability of reasoning training is achieved, and the method has the advantages of being completely autonomous, controllable, capable of being developed in a customized mode, low in power consumption and the like.

Description

technical field [0001] The invention relates to the technical field of deep learning algorithm training, and more particularly, to an FPGA-based deep learning training and reasoning system. Background technique [0002] Deep learning algorithm training generally includes four steps: forward inference, backward error transfer, weight gradient calculation, and weight update. By performing multiple iterative operations on these four steps, the best parameter training results that meet the requirements of the algorithm model are obtained. . The deep learning algorithm inference reads the parameters obtained by training and receives real-time data sources to realize forward inference. The training process values ​​computational accuracy, smaller or more deterministic error entry, the inference process accepts a small loss of accuracy, and cares about speed, cost, and power consumption. [0003] Due to the requirements of device power consumption, independent controllability, cu...

AI Technical Summary

Problems solved by technology

However, since the training process requires operations such as error reverse transfer, weight gradient calculation, and weight update that a

Images