A three-valued neural network weight processing method and device in embedded equipment

Abstract

The embodiment of the invention discloses a three-valued neural network weight processing method and device in embedded equipment. The method comprises the following steps: loading a neural network model to embedded equipment; grouping the original ternary network weight data in the neural network model, carrying out bitwise AND operation on the grouped weight data to obtain a low 2-bit weight data of each weight data, carrying out leftward shift according to the sequence of the weight data, and carrying out bitwise OR operation on the leftward shifted weight data to obtain compressed weight data; performing memory space development according to the storage space required by the single-layer weight data of the neural network model, performing bitwise and operation on the compressed weightdata corresponding to the single-layer weight data before the single-layer weight data is operated, and performing right shift on the bitwise and operated compressed weight data to obtain original ternary network weight data. And on the premise of ensuring the accuracy and speed of neural network operation, the problem that the three-valued neural network model occupies a relatively large memory space on embedded equipment is solved.

Description

technical field [0001] The embodiments of the present invention relate to the technical field of data processing, and in particular to a method and device for processing ternary neural network weights in embedded devices. Background technique [0002] With the continuous improvement of computer performance, deep learning and artificial neural networks have been widely used in speech recognition, person recognition and detection of various abnormal sounds. Traditionally, speech recognition technology is based on the spectrogram after time-frequency analysis. The spectrogram of a particular speech has a specific structure, but this structure varies with the speaker and the environment. Applying deep learning and artificial neural network to speech technology can use the translation invariance of neural network to overcome the diversity of speech signal itself. [0003] At the same time, the need to run neural network algorithms on embedded devices with low computing and stor...

AI Technical Summary

Problems solved by technology

[0005] To this end, the embodiment of the present invention provides a method and device for processing ternary neural network weights in embedded devices, which solves the problem that the ternary neural network model occupies an embedded device on the premise of ensuring the accuracy and speed of neural network operations. Problems with larger memory spaces

Images