Neural network fixed-point method and device, electronic equipment and readable storage medium

Abstract

The invention provides a neural network fixed-point method and device, electronic equipment and a readable storage medium, and the method comprises the steps: for a to-be-fixed unit of the neural network, based on data distribution of characteristic values output by the to-be-fixed unit under the condition that a fixed-point unit of the neural network is kept in a fixed-point state, determining a fixed-point hyper-parameter of the characteristic value output by the to-be-fixed unit; and based on the fixed-point hyper-parameter, carrying out fixed-point processing on the characteristic value output by the unit to be subjected to fixed-point processing. The method can improve the fixed-point performance of the neural network.

Description

technical field [0001] The present application relates to deep learning technology, and in particular to a neural network fixed-point method, device, electronic equipment and readable storage medium. Background technique [0002] Fixed-point neural network is a commonly used model acceleration algorithm. By quantizing the weight parameters of the model and the output feature values ​​into fixed-bit numbers, such as 8bit, 16bit or extremely low 1, 2bit, it can be well solved in Problems such as complex multiplication and accumulation calculations and huge bandwidth brought by floating-point data on the hardware. The lower the number of bits quantized by the model, the more obvious the acceleration of the model on the hardware. [0003] In the current fixed-point neural network scheme, a certain number of sample pictures are input to the original model at one time, and the forward operation of the model records the data distribution of the eigenvalues ​​output by each layer, ...

AI Technical Summary

Problems solved by technology

[0003] In the current fixed-point neural network scheme, a certain number of sample pictures are input to the original model at one time, and the forward operation of the model records the data distribution of the eigenvalues ​​output by each layer, and then calculates the fixed-points of the eigenvalues ​​output by each layer respectively. The hyperparameters are optimized, and the impact of the fixed-point of the front layer on the back layer is not considered, and the performance of the fixed-point neural network is low.

Images