Neural network model compression method and system based on mass spectrum data set

Abstract

The embodiment of the invention discloses a neural network model compression method and system based on a mass spectrum data set, and the method comprises the steps: carrying out training of a to-be-compressed neural network model, carrying out pruning of the trained neural network model, carrying out the quantification of the pruned neural network model, and combining a BN layer and a convolutional layer of the quantized neural network model to obtain a neural network model without the BN layer, and quantizing the obtained network model again to obtain a compression model. A large-scale network is used as an input model, unrelated channels are automatically identified and pruned, redundancy on parameter digits of the convolution layer and a full connection layer is removed, the BN layer is discarded, and a model which is equivalent in precision, thin and compact (efficient) is generated.

Description

technical field [0001] The embodiments of the present invention relate to the technical field of machine learning, and in particular to a method and system for compressing neural network models based on mass spectrum data sets. Background technique [0002] Since AlexNet won the ImageNet competition in 2012, convolutional neural networks have become more and more popular in the field of computer vision. A major trend is to make deeper and more complex network models in order to improve accuracy. However, such The model is stretched in terms of scale and speed. In many real scenarios, such as robots, autonomous driving, augmented reality and other tasks, it is difficult to complete the work in real time on a platform with limited computing power. This is the limitation of the large model. Model compression refers to the use of data sets to simplify the trained deep model, and then obtain a lightweight and accurate network. The compressed network has a smaller structure and fe...

AI Technical Summary

Problems solved by technology

[0003] For this reason, embodiments of the present invention provide a method and system for compressing neural network models based on mass spectrometry data sets to solve the problems of existing deep neural network models that are not complex enough, and have high computing and storage costs.

Images