Convolutional Layers Acceleration By Exploring Optimal Filter Structures

CNN models are becoming more and more mature, many of them adopt deeper structures to better accomplish the task objectives, such that the increased computational and storage burdens are unfavorable for the implementation in edge devices. In this paper, we propose an approach to optimize the filter structure by starting from the convolutional filter and finding their minimum structure. The reductions of the filters for the minimum structure in terms of space and channels, the number of model parameters and the computational complexity are effectively reduced. Since the current channel pruning method prunes the same channel for each convolutional layer, which easily leads to a trade-off between the pruning rate and accuracy loss. Instead we propose a new channel pruning approach to find the most suitable required channels for each filter to provide a more detailed pruning method. Experiments conducted on the classification CNN models, such as VGG16 and ResNet56, show that the proposed method can successfully reduce the computations of the models without losing much model accuracy effectively. The proposed method performs well in compressing the model and reducing the number of parameters required by the models for real applications.