Neural network compression and acceleration method, storage equipment and terminal

Abstract

The invention provides a neural network compression and acceleration method, storage equipment and a terminal. The method comprises the steps that the original neural network is pruned; clustering quantification is carried out on a network weight of the pruned original network, and the original network after clustering quantification is trained to obtain a target neural network; a sparse matrix isused to store the target neural network; an input feature map is converted into an input matrix; and the sparse matrix is multiplied by the input matrix to obtain an output feature map correspondingto the input feature map. Thus, the computing resource and storage space of the neural network are reduced, and thus, the computing cost is reduced.

Description

technical field [0001] The present invention relates to the field of computer technology, in particular, the present invention relates to a neural network compression and acceleration method, storage device and terminal. Background technique [0002] With the development of neural network models, in order to solve increasingly difficult problems such as classification, recognition and detection, deeper and larger neural network models have begun to be applied to such problems. For example, for deep learning algorithms that are currently widely used in artificial intelligence, the deep network structure is deep, and its calculation amount and model are large, so more computing resources and storage space are required. However, in production applications, server computing resources are becoming more and more scarce, the speed requirements are getting higher and higher, and the demand for porting to mobile terminals is becoming more and more urgent. Compression and test accele...

AI Technical Summary

Problems solved by technology

[0005] Aiming at the shortcomings of the existing methods, the present invention proposes a neural network compression and acceleration method, storage device and terminal to solve the problems in the prior art that the calculation resources and storage space of the neural network cannot be effectively reduced, and the calculation cost is relatively high problems to reduce neural network computing resources and storage space, thereby reducing computing costs

Images