Quantification method, device, electronic equipment and medium of a deep neural network

Abstract

The present application discloses a quantization method, device, electronic device and computer-readable storage medium of a deep neural network. The method includes: obtaining the floating-point model of the deep neural network generated by pre-training; calculating each quantization unit in the deep neural network The importance evaluation value of ; the type of quantization unit includes the weight parameter channel and / or hidden layer of the deep neural network, the weight parameter channel is used as the quantization unit for weight parameter quantization, and the hidden layer is used as the quantization unit for activation Output quantization; determine the quantization digits corresponding to each quantization unit; the quantization digits and the importance evaluation value of each quantization unit change positively; quantize the floating-point model according to the quantization digits. This application sets the number of quantization digits related to the importance of information for different weight parameter channels and / or hidden layers, and maximizes quantization compression while ensuring accuracy, reducing data processing and speeding up calculations.

Description

technical field [0001] The present application relates to the technical field of artificial intelligence, and in particular to a quantization method, device, electronic equipment and computer-readable storage medium of a deep neural network. Background technique [0002] With the continuous development of artificial intelligence technology, artificial intelligence technology has gradually been applied in our life. In the field of artificial intelligence technology, deep learning is one of the more typical technologies. Although the capabilities of deep neural networks in image classification and detection are close to or surpass those of humans, in actual deployment, there are still problems such as large models and high computational complexity, which require high hardware costs. In practical applications, neural networks are often deployed on some terminal devices or edge devices, but terminal devices or edge devices generally have low computing power, and memory and powe...

AI Technical Summary

Problems solved by technology

[0003] However, in the prior art, when the deep neural network is quantized after training, it does not take into account the redundant information in the network structure itself generated by training. These redundant information have little influence on the final inference results, but increase Increased computational complexity and speed

Images