Deep neural network calculation acceleration method and device

Abstract

The embodiment of the invention provides a deep neural network calculation acceleration method and device, a terminal and a computer readable storage medium. The method includes the following steps: sampling each input vector that needs to be input into a matrix model to obtain multiple sampling vectors; performing product quantization on each sampling vector according to a preset quantization parameter to obtain a plurality of quantization points; dividing the matrix model into a plurality of matrix blocks according to the quantization parameter; performing calculation on each quantization point and each matrix block to obtain a plurality of pre-calculation tables; and calculating each input vector through each pre-calculation table to obtain a calculation result of the matrix model. According to the embodiment of the invention, the pre-calculation table of the same matrix model only needs to be established once, and all input vectors that need to be calculated by the matrix model canuse the pre-calculation table to perform table lookup calculation, so that the calculation process of the input vectors and the matrix model can be effectively saved, and meanwhile, an original calculation effect of the matrix model can also be maintained.

Description

technical field [0001] The present invention relates to the technical field of data processing, in particular to a method, device, terminal and computer-readable storage medium for accelerating deep neural network calculations. Background technique [0002] Methods for speeding up deep neural networks in the prior art include matrix operations, pruning (pruning) algorithms, and binary networks. Among them, matrix-vector multiplication in matrix operations is memory-constrained, thus limiting the prediction speed of deep neural networks during computation. However, the accuracy loss of the binary network is relatively large. The pruning pruning algorithm requires a high degree of matrix sparsity, and the retraining process is complicated. Therefore, none of the existing calculation methods can well realize the calculation acceleration of the neural network. [0003] The above information disclosed in this Background section is only for enhancement of understanding of the b...

AI Technical Summary

Problems solved by technology

Among them, the matrix-vector multiplication in the matrix operation is memory-constrained, thus limiting the prediction speed of the deep neural network during the calculation process

However, the accuracy loss of the binary network is large

The pruning pruning algorithm requires a high degree of matrix sparsity, and the retraining process is complicated

Therefore, none of the existing calculation methods can well realize the calculation acceleration of the neural network.

Images