Quantization processing method and device and quantization processing chip

Abstract

The embodiment of the invention provides a quantization processing method and device, and a quantization processing chip, and the method comprises the steps: firstly, determining the initial quantization parameter of each channel based on the distribution characteristics of the network parameters of each channel, and enabling the initial quantization parameter of one channel to be the optimal quantization parameter theoretically meeting the distribution characteristics of the network parameters of the channel; however, the initial quantization parameter may not satisfy a layer quantization hardware deployment condition; therefore, the optimized quantization parameter corresponding to each channel is searched in the search space determined on the basis of the initial quantization parameter, the quantization performance of the searched optimized quantization parameter is close to that of the initial quantization parameter, and a layer quantization hardware deployment condition can be met, so that the method can be applied to a general quantization processing chip, and the quantization performance of the universal quantization processing chip is close to that of a channel quantization mode.

Description

technical field [0001] The present disclosure relates to the technical field of quantization processing, and in particular, to a quantization processing method and device, and a quantization processing chip. Background technique [0002] Quantization processing plays an important role in the accelerated deployment of neural networks. In the related art, there are two quantization processing methods: layer quantization and channel quantization. In the layer quantization method, the network parameters of each channel of the same network layer are quantized with the same quantization parameters; in the channel quantization method, the network parameters of the same network layer are quantized. The network parameters of each channel are quantized using different quantization parameters. The channel quantization method can improve the quantization accuracy, but a dedicated channel quantization processing chip needs to be used for processing, which leads to an increase in the cos...

AI Technical Summary

Problems solved by technology

The channel quantization method can improve the quantization accuracy, but it needs to use a dedicated channel quantization processing chip for processing, which leads to an increase in the cost of quantization processing

Images