Deep neural network compression method and system, storage medium and computer equipment

• Main Authors: LUO QIFENG, ZHOU HENG, LUO HAIXIN, LEI WANG, MENG CHENXU, LIN XIONGFENG, LING XIA, ZENG YIHAO, LI XINHAI, ZENG LINGCHENG, ZENG XINXIONG, QIU TIANYI, LIAO WEIQUAN
• Format: Patent
• Language: Chinese; English
• Published: 19.02.2021
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; PHYSICS

The invention provides a deep neural network compression method and system, a storage medium and computer equipment, and the method comprises the steps: carrying out channel cutting of a convolution layer, carrying out iterative pruning and linear quantification of parameters of a full connection layer, achieving a very good compression effect so that precision loss of a model is smaller. Under the same compression level, model precision loss caused by pruning can be effectively reduced through iterative pruning, linear quantization is further carried out on the basis of iterative pruning, andparameter capacity is reduced while the model precision loss is effectively controlled. 本发明提出了一种深度神经网络压缩方法、系统、储存介质及计算机设备，方法部分通过对卷积层进行通道裁剪，对全连接层参数进行迭代剪枝及线性量化，可以实现很好的压缩效果，而且模型的精度损失较小。在相同的压缩水平下，迭代剪枝能有效降低剪枝带来的模型精度损失，而经过迭代剪枝基础上进一步进行线性量化，还在有效控制模型精度损失的同时降低参数容量。