Quantification method, device, storage medium and electronic equipment of neural network

Abstract

The present disclosure relates to a neural network quantization method, device, storage medium, and electronic equipment. The quantization method includes: receiving the first value range interval corresponding to the first quantization bit width output by the previous neural network layer in the neural network input data; map the input data to at least include the second value range interval corresponding to the second quantization bit width to obtain quantized input data; perform calculations according to the quantization bit width corresponding to the quantized input data to obtain quantized output data; quantize The output data is inversely mapped to the first value range interval corresponding to the first quantization bit width to obtain the output data. The disclosure achieves the purpose of improving the reasoning accuracy of the quantized neural network without reducing the data transmission efficiency between hardware.

Description

technical field [0001] The present disclosure relates to the field of computer technology, and in particular, to a quantification method, device, storage medium and electronic device of a neural network. Background technique [0002] Neural networks generally use FP32 floating-point for inference and training. Although floating-point operations have higher precision, floating-point operations consume a lot of computing resources, resulting in low efficiency of neural networks. [0003] Quantization is a method of converting floating-point parameters into fixed-point parameters to reduce operation precision and increase operation speed. Within the range of quantized parameters, converting floating-point operations to integer fixed-point operations for forward inference will not significantly reduce the accuracy of the neural network model, and can speed up the running speed and significantly reduce the inference and training performance of the neural network. Consumption, re...

AI Technical Summary

Problems solved by technology

[0004] For the quantization of the neural network, the larger the quantization bit width of the fixed-point integer after quantization, the higher the accuracy of the neural network model, but too large quantization bit width will also slow down the running speed of the neural network processor chip. , the operation inside the neural network processor chip does not have restrictive requirements on the quantization bit width

[0005] Therefore, there is still room for improvement in the optimization of quantized fixed-point integers for the quantization of neural networks to improve their quantization accuracy.

Images