Normalization quantification method and device, electronic equipment and storage medium

Abstract

The embodiment of the invention provides a normalization quantification method and device, electronic equipment and a storage medium. The method comprises the following steps: obtaining quantized data Qconv serving as input data of a normalization module; calculating a parameter mean value Q [mu] and a standard deviation Q [sigma] required by normalization of the quantized data Qconv; (Qconv-Qmu) * IntA is calculated, SR is calculated again according to a calculation result of the (Qconv-Qmu) * IntA, and the SR is determined according to a re-scaling parameter of the normalization module and a parameter of a subsequent module of the normalization module; and determining a full-fixed-point reasoning result of the neural network according to the result. According to the normalization quantification method provided by the embodiment of the invention, the problem of precision loss caused by normalization in a fixed-point calculation process is avoided by utilizing a thought of first calculation of (Qconv-Qmu) * IntA, so that full fixed-point reasoning of a neural network model with an INGNLN module becomes possible, the reasoning speed of the neural network model is improved, and the reasoning efficiency of the neural network model is improved. And the power consumption of model reasoning is effectively reduced.

Description

technical field [0001] The present invention relates to but not limited to the field of artificial intelligence, and specifically relates to a normalized quantification method, device, electronic equipment and storage medium. Background technique [0002] In neural networks, normalization is a very common and extremely important technology. Currently, several normalization methods commonly used are IN (Instance Normalization, instance normalization), BN (Batch Normalization, batch normalization) , GN (Group Normalization, group normalization), LN (Layer Normalization, layer normalization) each use different usage scenarios. [0003] The formula to normalize the data is: In is the input feature, and Out is the normalized feature. The parameter to be solved here is the mean value μ of the input feature In In and standard deviation σ In . [0004] From the perspective of IN and BN, the biggest difference between the two is that the mean and standard deviation of IN need to...

AI Technical Summary

Problems solved by technology

[0005] In the related deep learning framework (Tensorflow / Pytorch), only the quantitative fusion of BN is supported, but for IN or other normalization methods commonly used in target segmentation and detection tasks, there is no corresponding quantification scheme to support it, making Models with IN, GN, and LN can basically only run in floating-point form, which is less applicable when terminal-side products such as mobile phones have high requirements for power consumption and performance

Images