A nonlinear convolution neural network quantization method

In today's world, convolution neural networks possess tremendous potential and exhibit remarkable recognition rates in computer vision. However, neural network models are often so large that they are not well-suited for mobile devices, making model compression particularly important. Especially in scenarios with limited hardware resources such as the National Grid, model quantification is essential. This paper introduces an INT8 quantization method for convolution neural networks, which transforms the weights and input/output of the convolution neural network into the logarithmic domain through a non-linear formula. This method provides a better representation of the distribution of small values, resulting in less accuracy loss. Meanwhile, model calibration requires only a limited amount of image data, which reduces the time taken for model calibration while ensuring minimal accuracy loss. Experimental results show that our quantization method results in an accuracy loss of less than 1% on ImageNet and CIFAR.