Deep convolutional neural network quantization method, system and device, and storage medium

Abstract

The invention provides a deep convolutional neural network quantization method, system and device, and a storage medium, and the method comprises the steps: analyzing a deep learning neural network model, and inserting a pseudo-quantization operator at an operator needing to be quantized; floating point type parameter reasoning is carried out on the constructed test set, and a feature quantization scaling factor is calculated according to the relative entropy divergence; quantizing weight parameters of a convolutional layer or a full connection layer of the deep learning neural network model to obtain a weight quantization scaling factor; and performing quantization reasoning according to the feature quantization scaling factor and the weight quantization scaling factor. According to the method, random rounding is applied to deep neural network model quantization, meanwhile, a channel quantization algorithm is optimized, quantization reasoning is carried out on designed special equipment, and the quantization reasoning precision and efficiency of the deep neural network model are improved.

Description

technical field [0001] The present invention relates to the field of deep learning, and more specifically refers to a method, system, device and storage medium for quantifying deep convolutional neural networks. Background technique [0002] In recent years, with the rapid development of deep learning, in order to improve the inspection accuracy of various AI application scenarios, the size and number of layers of deep neural network structures and the number of various parameters have continued to increase, resulting in greater space requirements for deep learning models. Lower reasoning efficiency. As one of the general deep learning optimization methods, the model is quantized as a deep convolutional neural network and quantized as a fixed-point model, which has smaller storage space and faster reasoning speed, and ensures that the accuracy is within a certain loss range. It is suitable for most Most models and usage scenarios. Model quantization refers to the linear ma...

Images