Convolutional neural network quantification method for automatically sensing weight distribution in reinforcement learning

Abstract

The invention discloses a convolutional neural network quantification method for automatically sensing weight distribution through reinforcement learning. The method comprises the steps: carrying out the fusion of a parameter of batch processing operation of each layer and the weight of convolution operation, obtaining a fused weight and offset, and obtaining the distribution information of the fused weight of each layer in a floating point convolutional neural network model; automatically searching an optimal weight scaling coefficient of each layer according to the distribution information reinforcement learning of the weight of each layer, and weighting the floating point weight into INT8 type data based on the weight scaling coefficient of each layer; inputting a calibration data set, recording each layer of output feature map when each group of data is input, selecting a mode as a scaling coefficient of each layer of output feature map, and calculating to obtain a scaling coefficient of each layer of bias according to the scaling coefficient of each layer of weight and the scaling coefficient of each layer of output feature map so as to quantize the bias of the floating point into the bias of INT32 type; and constructing a forward reasoning process based on the INT8 type data, the INT32 type bias and the total scaling coefficient to complete quantification.

Description

technical field [0001] The invention belongs to the technical field of artificial intelligence, in particular to a convolutional neural network quantification method for automatic perception of weight distribution through reinforcement learning. Background technique [0002] In recent years, with the development of artificial intelligence technology dominated by convolutional neural networks, more and more computer vision tasks have been well solved, such as image classification, object detection and semantic segmentation. And a current development trend is to deploy high-performance neural network models on end-side platforms and run them in real time (greater than 30 frames) in real scenarios, such as mobile / embedded devices. These platforms are characterized by low memory resources, low processor performance, and limited power consumption. This makes it impossible for the current models with the highest accuracy to be deployed on them and meet real-time requirements due t...

AI Technical Summary

Problems solved by technology

[0006] Aiming at the problem of large precision loss caused by traditional model quantization in the prior art and the dependence on calibration data sets and training data sets in the quantization process, the present invention proposes a convolutional neural network that automatically perceives weight distribution through reinforcement learning Quantization method, the present invention can be used to compress and quantize models when deploying convolutional neural networks on various hardware platforms

Images