Operation and quantification method of neural network model, electronic equipment and medium

Abstract

The invention relates to the technical field of quantification, and discloses a neural network model operation method, a neural network model quantification method, electronic equipment and a medium. The neural network model comprises nonlinear function operation; the operation method comprises the following steps: in the operation process of the neural network model, obtaining input data of a nonlinear function; quantized data corresponding to the input data are obtained from a quantization mapping table, and in the quantization mapping table, for the input data belonging to the first type of truncation range, the quantized data corresponding to the input data are quantized data corresponding to data with the same numerical value as the input data in the first type of truncation range in the quantization mapping table; for the data belonging to the second type of truncation range, the quantized data corresponding to the input data is the quantized data corresponding to the endpoint data of the first type of truncation range in the quantization mapping table. Based on the scheme, the memory occupied by the quantization mapping table can be effectively reduced, so that the memory occupied by the neural network model is further reduced.

Description

technical field [0001] The present application relates to the technical field of quantification, and in particular, to an operation, quantification method, electronic device and medium of a neural network model. Background technique [0002] At present, with the rapid application of deep learning technology in many fields, a large number of neural network models based on deep learning have appeared. However, the neural network model structure is generally more complex and occupies a large amount of memory, so it is necessary to quantify and compress the neural network model to reduce the memory occupation and speed up the inference speed of the model. [0003] The neural network model includes multiple nodes (or neurons), each node represents a specific output function, called an excitation function or an activation function, and the function type of some nodes is a linear function, that is, the input data and The relationship with the output data is linear, for example, th...

AI Technical Summary

Problems solved by technology

However, using the data within the entire input data range of the nonlinear function and the corresponding quantized data to construct the LUT will take up a large amount of memory.

Images