| 摘要: |
| 拥有庞大参数量的网络模型很难部署在智能手机、可穿戴智能设备等资源受限的移动设备上。从深度神经网络模型的基本原理出发，在现有压缩算法的基础上，采用优化剪枝策略与参数量化的方法相融合，提出了一种结果导向的数据驱动剪枝算法，利用低精度的量化算法来进一步压缩模型。使用VGGNet作为原始模型，在Kaggle猫狗图像和Oxford102植物样本集上进行微调。实验数据表明，使用本实验改进的方法，模型压缩的存储容量下降到113.1 MB，识别率提高到86.74%。 |
| 关键词: 深度学习 模型压缩 模型剪枝 参数量化 |
| DOI： |
|
| 基金项目:国家自然科学基金资助项目（61772417，61634004，61602377）；陕西省重点研发计划（2017GY-060）；陕西省自然科学基础研究计划项目（2018JM4018） |
|
| DNN Model Compression Algorithms Based on Fusion of Model Pruning and Low-precision Quantization |
| WU Jin,YANG Xue,HU Yiqing |
| (School of Electronic Engineering,Xi′an University of Posts and Telecommunications,Xi′an 710121,China) |
| Abstract: |
| The network model with huge parameters is difficult to be deployed on resource-constrained mobile devices such as smart phones and wearable smart devices.According to the basic principle of the deep neural network model and the existing compression algorithm,a result-oriented data-driven pruning algorithm is proposed,which combines the optimized pruning strategy with the method of parameter quantization.The model is further compressed by using the low-precision quantization algorithm on the basis of the pruning results.VGGNet is used as the original model and fine-tuned on several sample sets including Kaggle cat and dog image set,Oxford102 plant sample set.The experimental data show that,with less loss of model recognition rate,the storage capacity of the model is reduced to 113.1 MB and the operation speed is improved to 86.74%. |
| Key words: deep learning model compression model pruning parameter quantization |
《电讯技术》编辑部 