采用可替代滤波器的卷积神经网络模型剪枝方法

将卷积神经网络模型中某一层的所有滤波器抽象到一个欧几里德空间,对其中能被其他滤波器共同表示的滤波器剪枝,降低滤波器冗余,避免精度损失.使用强化学习进行边训练边剪枝,经过微调恢复神经网络模型性能.结果表明:剪枝并微调后的神经网络模型精度损失较小,参数量与浮点计算量显著减少.

Pruning Method of Convolutional Neural Network Using Replaceable Filter

All filters in a certain layer of the convolutional neural network are abstracted into an Euclidean space, Pruning the filter that can be jointly represented by other filters, reducing the redundancy of filter and avoiding the loss of accuracy. Using reinforcement learning to prune while training, the neural network model performance is restored through fine-tuning. The results show that, after pruning and fine-tuning, the loss of neural network model accuracy is smaller, the calculated amount of parameters and floating-point are significantly reduced.