Neural network design and optimization method based on software and hardware joint learning

Abstract

The invention discloses a neural network design and optimization method based on software and hardware joint learning. The method comprises the following steps: counting a neural network structure rule; carrying out FPGA hardware characteristic prediction; designing an FPGA neural network structure space; and applying a software and hardware joint learning method in a search space, and by combining random search and block supervised search, obtaining a trunk neural network. Based on the design characteristics of the neural network and the hardware characteristics of the FPGA, a search space with prior information is constructed, which is the direction of search establishment; and meanwhile, by combining random search and block supervised search with FPGA model prediction, an efficient neural network model with precision and speed balance is obtained. According to the model, the Top-1 accuracy rate of 77.2% and the speed of 327.67 FPS on an ImageNet data set are achieved on the aspect of ZCU102.

Description

technical field [0001] The invention relates to the technical field of neural network architecture search, in particular to a neural network design and optimization method based on joint learning of software and hardware. Background technique [0002] In the target detection task of autonomous driving, the backbone neural network of the detector undertakes the main feature extraction task, and largely determines the accuracy and speed of the overall detection task. Therefore, it is critical to design a backbone neural network suitable for autonomous driving tasks. [0003] Manually designing neural networks requires solid professional knowledge and a lot of labor, and a new architecture is usually modified from some existing networks through carefully designed manual experiments. This design approach is inefficient and it is difficult to design a network that far exceeds existing advanced networks. Moreover, there are many structural parameters that can be adjusted by the ...

AI Technical Summary

Problems solved by technology

This design approach is inefficient and it is difficult to design a network that far outperforms existing advanced networks

Moreover, there are many structural parameters that can be adjusted by the neural network, and there is no unified design rule. If different task scenarios and operating equipment are considered, the parameters to be considered will further increase

Images