Neural network quantization compression method and system

Abstract

The invention provides a neural network quantization compression method, which comprises the following steps: obtaining to-be-compressed neural network data after quantization processing, and carrying out run-length all-zero coding on the neural network data to obtain run-length compressed data, the run-length all-zero coding comprising only carrying out run-length coding on zero characters in the neural network data; and performing normalized Huffman coding on the run length compression data, and reforming a coding result to obtain a normalized Huffman code as a compression result of the neural network data. Aiming at the characteristic that quantized neural network data has sparsity, run-length coding is improved, run-length all-zero coding is provided, and the neural network data can be compressed in a lossless mode more efficiently; the Huffman tree is reformed from top to bottom, storage of a complete Huffman tree structure is omitted, and the complexity of table look-up operation is remarkably reduced.

Description

technical field [0001] The invention relates to the field of neural network operations, and in particular to a method and system for quantizing and compressing neural networks based on hybrid coding. Background technique [0002] In recent years, artificial intelligence has developed rapidly in the context of the double explosion of information volume and hardware computing power, and has increasingly become the main driving force for productivity development and technological innovation. As the main branch of artificial intelligence technology, in order to further improve the accuracy of the model, the neural network algorithm has encountered technical bottlenecks such as complex structure, huge amount of parameters and large amount of calculation, which limits the application of neural network models in the pursuit of throughput rate and energy efficiency ratio. Therefore, computational efficiency becomes the main research goal in the next stage. The most effective neural...

AI Technical Summary

Problems solved by technology

[0016] (1) Compared with the original normal distribution, the frequency of occurrence of each number after retraining is more even, which may lead to an increase in the average code length of Huffman coding. Therefore, the data must be pre-coded before entropy coding. Improve the Huffman encoding of the frequency of input characters

Images