Data fixed point method and device

Abstract

A data fixed-point method and device, a data processing method and device, and a data alignment method and device, wherein the data fixed-point method includes: calculating the maximum output value of each input sample in the first target layer of a neural network in a plurality of input samples (S110); Select at least two maximum output values ​​from a plurality of maximum output values ​​as fixed-point reference values ​​(S120); determine the reference integer part bit width (S130) according to each fixed-point reference value; based on the preset output value total bit Width and the bit width of each reference integer part are tested for accuracy, and the reference integer part with the highest accuracy rate is determined as the bit width of the integer part used by the first target layer when the output value is fixed-point (S140). The method selects multiple values ​​from multiple maximum output values ​​of the first target layer as fixed-point reference values, determines the reference integer part bit width corresponding to each fixed-point reference value, and determines the optimal integer part bit width based on the accuracy test Wide, according to the optimal bit width of the integer part, the fixed-point network can transmit more useful information while maintaining high precision, improving the expressiveness and accuracy of the network.

Description

[0001] Copyright statement [0002] The disclosure of this patent document contains material that is protected by copyright. This copyright belongs to the copyright owner. The copyright owner has no objection to the reproduction by anyone of the patent document or the patent disclosure as it exists in the official records and files of the Patent and Trademark Office. technical field [0003] The present application relates to the field of data processing, in particular to a data fixed-point method and device. Background technique [0004] In the current mainstream neural network computing framework, floating point numbers are basically used for training and calculation. Among them, in the backpropagation process of the neural network, the calculation of the gradient needs to be based on the floating-point number representation to ensure sufficient accuracy; each layer of the forward propagation process of the neural network, especially the weight coefficients of the convol...

AI Technical Summary

Problems solved by technology

When the fixed-point position obtained by this fixed-point method is used for fixed-point conversion, although the truncation rate is relatively low, the most useful high-level information will often be eliminated, thereby affecting the expressive ability of the network and causing a decrease in the accuracy of the network

Images