Model optimization and compression method for lightweight neural network

The invention discloses a model optimization and compression method for a lightweight neural network, and the method comprises the steps: firstly adding an improved parameter into an original ReLU function to improve the ReLU function, and updating the improved parameter through employing a chained rule and a momentum method; secondly, replacing a Sigmoid function in the self-gating function by utilizing a ReLU6 function, and fusing the Sigmoid function with the improved ReLU function to obtain a fusion activation function; and finally, according to the obtained elimination threshold, deletinga corresponding number of convolution kernels from bottom to top to compress the convolutional neural network model until the model precision and the parameter quantity are balanced, thereby improving the performance of the lightweight neural network model. 本发明公开了一种面向轻量级神经网络的模型优化和压缩方法，首先向原始ReLU函数中加入改进参数对ReLU函数进行改进，并利用链式法则和动量法对所述改进参数进行更新；其次利用ReLU6函数将自门控函数中的Sigmoid函数进行替换，并与改进后的ReLU函数进行融合，得到融合激活函数；最后，根据获取的剔除阈值，自下而上的删除对应数量的卷