Neural network quantization method and device, image recognition method and device and computer equipment

Abstract

The invention relates to a neural network quantization method, an image recognition method, a neural network quantization device, an image recognition device, computer equipment and a readable storagemedium. The method comprises the following steps: based on prediction loss and operand loss of a training sample, adjusting network parameters and operation attribute parameters of an initial neuralnetwork to obtain trained target operation attribute parameters, wherein the operation attribute parameter represents the value of the operation attribute of each network layer in the neural network,and the operand loss is positively correlated with the actual operand associated with the operation attribute parameter; and quantizing the neural network by adopting the target operation attribute parameters to obtain the quantized neural network. By adopting the method, the problem that the prediction accuracy of the quantized neural network is seriously reduced can be avoided, and reasonable quantification is realized.

Description

technical field [0001] The present application relates to the technical field of neural networks, in particular to a neural network quantization method, an image recognition method, a neural network quantization device, an image recognition device, computer equipment and a readable storage medium. Background technique [0002] With the development of neural network technology, neural network quantization technology has emerged, mainly for model parameter values ​​and activation values ​​(output values ​​or input values) in each network layer (such as convolutional layer and fully connected layer) in the neural network. , input feature maps, etc. to compress, reduce the bit width of model parameter values, bit width of activation values, and the size (height and width) of input feature maps, etc., so as to realize the compression of the data volume of neural network model files and reduce the neural network model For purposes such as computing resource requirements during the...

AI Technical Summary

Problems solved by technology

[0004] However, although this neural network quantization technique can reduce the computational load of the neural network, it may lead to a serious reduction in the prediction accuracy of the quantized neural network, and there is a problem of unreasonable quantization.

Images