无人驾驶深度学习模型组合剪枝算法

Combined pruning algorithm for unmanned deep learning model

首先对网络模型进行正则化,提高网络泛化能力及模型稀疏度;接着采用组合剪枝算法,先对卷积层卷积核进行删减,再对全连接层权重进行删减。剪枝过程分多次进行,每次剪枝后通过再训练以恢复模型分类识别性能。实验结果表明:提出的组合剪枝算法能够在保证VGG—16网络对车辆数据集分类识别率保持在92.84%的情况下,对网络模型压缩了79%,加速3.38倍。

Firstly,the network model is regularized to improve the generalization ability network and sparsity of model.Then,the combined pruning algorithm is adopted.Convolutional layer convolution kernel is firstly deleted,and then the full connection layer weight is deleted.The pruning process is carried out several times,and each time after pruning,it is retrained to restore the model classification and recognition performance.The experimental results show that the combined pruning algorithm can compress the network model by 79%and accelerate by 3.38 times while ensuring that the VGG-16 network keeps the classification and recognition rate of the vehicle dataset at 92.84%.