Mixing precision quantification method of neural network

Abstract

A hybrid precision quantization method for a neural network having a first precision and including a plurality of layers and an original final output. The mixed precision quantization method comprises the following steps of: performing quantization of a second precision on one of the layers and the input of the layer; obtaining the output of the layer according to the layer of the second precision and the input of the layer; performing inverse quantization on the output of the layer, and inputting the inverse quantized output of the layer to the next layer; obtaining a final output; obtaining a value of an objective function according to the final output and the original final output; repeating the steps until the value of the objective function corresponding to each of the layers is obtained; determining a quantization precision of each of the layers according to the value of the objective function corresponding to each of the layers; wherein the quantization precision is the first precision, the second precision, the third precision or the fourth precision.

Description

technical field [0001] The present invention relates to a mixed precision quantization method, and in particular to a neural network mixed precision quantization method. Background technique [0002] In the application of neural networks, the prediction process requires a lot of computing resources. Neural network quantization reduces computational cost, but may reduce prediction accuracy. Current quantization methods use the same precision to quantize the entire neural network, but this approach lacks flexibility. Moreover, most of the current quantization methods need to be matched with a large amount of labeled data and integrated into the training process to complete. [0003] In addition, in the current method, if we want to judge the quantization loss of a specific layer in the neural network, only the status of this specific layer is considered, such as the loss of the output of this specific layer, the loss of weight, etc., and this specific layer is not considered...

AI Technical Summary

Problems solved by technology

[0003] In addition, in the current method, if the quantization loss of a specific layer in the neural network is to be judged, only the status of the specific layer is considered, such as the loss of the output of the specific layer, the loss of the weight, etc., and the loss of the specific layer is not considered. The impact on the final result, so the current method cannot achieve the best balance between cost and prediction accuracy

Images