A deep convolutional neural network model adaptive quantization method based on modulus length clustering

Abstract

The invention discloses a deep convolutional neural network model adaptive quantization method, and designs a deep convolutional deep network low-bit quantization algorithm suitable for FPGA calculation, which mainly comprises preprocessing of network model parameters and a grouping adaptive quantization method of a parameter set. the method includes: acquiring dynamic thresholds to perform coarse-grained cutting on the original parameters of the model; constructing an initial clustering center point set suitable for FPGA shift calculation; grouping and clustering the preprocessed model parameters based on a mode length minimization method; finally, overlaying the clustering center point set with the non-null parameter class, achieving self-adaptive low-bit quantization of different networks through optimization; the quantization algorithm is moderate in complexity and quite conforms to the calculation characteristics of the FPGA, hardware resource consumption on the FPGA is reduced, and the model reasoning speed is increased while the model reasoning precision is guaranteed.

Description

technical field [0001] The invention relates to the technical field of deep network model compression, in particular to an adaptive quantization method of a deep convolutional neural network model based on module length clustering. Background technique [0002] With the rapid development of deep learning technology, deep neural networks have achieved leapfrog breakthroughs in computer vision, speech recognition, natural processing and other fields. However, deep learning algorithms have not been widely used in the fields of industry, manufacturing, aerospace and navigation. One of the reasons is that the model of the deep learning network is huge and the amount of calculation is huge. The weight file of a CNN network can easily be hundreds of megabytes, such as AlexNet With 61M parameters and 249MB memory, the memory capacity of complex VGG16 and VGG19 has exceeded 500MB, which means larger storage capacity and more floating-point operations are required. Due to the limited...

AI Technical Summary

Problems solved by technology

Although some of these algorithms can also use shift operations instead of multiplication, their optimization goals (maximum probability criterion) and optimization methods (L1 and L2 regularization) usually cause the neural network parameters to present a centrally symmetrical non-uniform distribution, which will lead to network performance drops more

Although FPGA is not inferior to GPU in parallel computing capability, it is limited by its storage resources and computing resources.

Images