Model compression method based on layer number sampling and deep neural network model

Abstract

The invention provides two model compression methods based on layer number sampling from two aspects, each model compression method is based on a plurality of neural networks with completely consistent cascaded sub-network structures, and one model compression method based on layer number sampling comprises the following steps: generating random positive integers uniformly distributed in a preset interval, wherein the extreme value of the interval is not greater than the total number of the sub-networks; selecting the random positive integer number of sub-networks to carry out iteration once, wherein the second model compression method based on layer number sampling comprises the following steps: evaluating the performance of a neural network model, and determining an ideal value of n, and the first n sub-networks are reserved in the neural network. The computing speed can be increased during training and reasoning, computing resources are saved, and the performance of the model is kept. The deep neural network model provided by the invention has corresponding advantages due to the adoption of the model compression method provided by the invention, and is beneficial to implementation in a wider application scene.

Description

technical field [0001] The invention belongs to the technical field of neural networks, and in particular relates to a model compression method of a neural network based on layer number sampling and a corresponding neural network model. Background technique [0002] With the development of artificial intelligence, especially deep learning technology, various intelligent tasks based on neural networks have entered every aspect of ordinary people's daily life. Artificial neural network (ANN), referred to as neural network (NN), is a machine learning model that includes multiple hidden layers. Neural network technology has promoted the development of all walks of life and improved the convenience of people's life and production. [0003] The application of neural network generally has two parts: Training and Inference. The so-called training usually refers to the process of continuously iterating and optimizing the parameters of the neural network according to the given data ...

AI Technical Summary

Problems solved by technology

However, if running online requires high latency, a large model like the BERT model requires huge computing resources and cannot fully meet the needs of the application due to the limitation of its computing speed.

In particular, the above-mentioned major costs are especially unfavorable in the application scenario of smart home

For reasons of economy and space utilization, the terminal devices deployed by users in smart homes are usually small terminal devices with limited computing power, such as smart speakers, smart gateways and small home hosts, etc., which are required for neural network training or reasoning. The cost is unaffordable, which limits the further promotion and application of neural network technology in the lives of ordinary people

Images