基于特征复用的卷积神经网络模型压缩方法

为了在不降低准确率的前提下,减小卷积神经网络模型的体积与计算量,提出一种基于特征复用的卷积神经网络压缩模块--特征复用单元(FR-unit)。首先,针对不同类型的卷积神经网络结构,提出不同的优化方法;然后,在对输入特征图进行卷积操作后,将输入特征与输出特征进行结合;最后,将结合后的特征传递给下一层。通过对低层特征的重复使用,使总的提取的特征数量不发生改变,以保证优化后的网络的准确率不会发生改变。在CIFAR10数据集上进行验证,实验结果表明,优化后的VGG模型体积缩小为优化前的75.4%,预测时间缩短为优化前的43.5%;优化后的Resnet模型体积缩小为优化前的53.1%,预测时间缩短为优化前的60.9%,且在测试集上的准确率均未降低。

Model compression method of convolution neural network based on feature-reuse

In order to reduce the volume and computational complexity of the convolutional neural network model without reducing the accuracy, a compression method of convolutional neural network model based on feature reuse unit called FR-unit (Feature-Reuse unit) was proposed. Firstly, different optimization methods were proposed for different types of convolution neural network structures. Then, after convoluting the input feature map, the input feature was combined with output feature. Finally, the combined feature was transferred to the next layer. Through the reuse of low-level features, the total number of extracted features would not change, so as to ensure that the accuracy of optimized network would not change. The experimental results on CIFAR10 dataset show that, the volume of Visual Geometry Group (VGG) model is reduced to 75.4% and the prediction time is reduced to 43.5% after optimization, the volume of Resnet model is reduced to 53.1% and the prediction time is reduced to 60.9% after optimization, without reducing the accuracy on the test set.