An Adaptive Asymmetric Quantized Compression Method for Deep Neural Network Models

Abstract

The invention discloses a self-adaptive asymmetric quantization deep neural network model compression method, which comprises the following steps of: adaptively quantizing the weight of each layer offloating point of a network into an asymmetric ternary or quaternary value in the training process of each batch before forward propagation starts to calculate during deep neural network training; ina back propagation parameter updating stage, carrying out parameter updating by using the original floating point type network weight; and finally, performing compression storage on the trained quantized deep neural network. According to the method, the parameter redundancy degree of the deep neural network is reduced, the residual parameters are adaptively quantified, the network model is compressed to the greatest extent, and the identification accuracy of the quantization method on the deep network and a big data set is improved.

Description

technical field

[0001] The invention relates to the technical field of deep neural network model compression, in particular to an adaptive asymmetric quantized deep neural network model compression method. Background technique

[0002] In recent years, deep learning has gradually replaced the application of traditional machine learning in daily life. In a series of machine learning tasks such as speech recognition, image classification, and machine translation, deep neural networks have achieved certain results. However, the classic deep neural network model, due to its heavy hierarchical structure, brings millions of floating-point network parameter calculations, making it difficult for most networks to be deployed in mobile devices and embedded devices and maintain good processing performance . How to maximize the compression of neural network parameters and ensure that the recognition performance of the original network is gradually becoming an important research directi...

Images