Low-bit quantization method of depth separable convolution structure

Abstract

The invention discloses a low-bit quantization method for a depth separable convolution structure. The method comprises the following steps: selecting a quantization coefficient to uniformly quantizethe weight of a trained depth separable convolution neural network and feature map data; performing channel-by-channel quantification on the weight in the trained deep separable convolutional neural network; quantifying the depth separable convolutional neural network layer by layer by utilizing the feature map data; performing forward operation in the trained depth separable convolutional neuralnetwork based on the training set, and updating moving average parameters in the batch normalization layer; and fusing the updated moving average parameters and learnable parameters in the batch normalization layer into a network weight quantization coefficient and offset to realize low-bit quantization of the depth separable convolution structure. The low-bit quantization method of the depth separable convolution structure provided by the invention does not need any labeled data participation, is simple in calculation, and improves the accuracy after low-bit quantization compared with a traditional method.

Description

technical field [0001] The invention relates to the fields of deep learning and artificial intelligence, in particular to a low-bit quantization method for a depth-separable convolution structure. Background technique [0002] Since 2012, deep convolutional neural network (CNN) has been widely used in computer vision fields such as image classification, image segmentation, and object detection. When using deep convolutional neural networks to solve problems, people often tend to design more complex networks in order to obtain higher performance. What follows is that the complexity of the model is greatly improved, which brings great challenges to the storage of the embedded model. At the same time, the model reasoning time is getting longer and longer, and the delay is getting bigger and bigger, which brings great challenges to the deep convolutional neural network. The promotion on terminal smart devices has brought great challenges. For example, the weight of the classic...

AI Technical Summary

Problems solved by technology

[0005] The purpose of the present invention is to provide a low-bit quantization method of a depth-separable convolution structure to solve the problem of network precision decline when performing low-bit quantization on the network

Images