Automatic network compression via evolutionary search-guided channel pruning of deep convolutional neural networks

초록

Deep convolutional neural networks often exhibit parametric and computational redundancy, a common issue in various applications. Channel pruning is widely used to cut down model redundancy while maintaining the network's core structure. However, searching for the optimal compressed network is challenging due to the vast search space and significant computational costs. In this paper, we introduce a novel channel pruning approach that utilizes surrogate-assisted neural architecture search to automatically identify an effective compressed network while adhering to predefined computational constraints. Our method introduces accuracy and efficiency coefficients into the objective function, providing fine-grained control over the trade-off between network accuracy and computational efficiency while searching for the optimal compressed network. The core of our method initiates with the training of a surrogate model employed to approximate the performance outcomes of a candidate network via generating corresponding weight assignments. These candidate networks are subjected to an evolutionary search guided by the help of the weights assigned via the surrogate model, and their interactions are regulated with the objective function. Extensive experimental evaluations showcase the adequate performance of our method in comparison to state-of-the-art techniques when applying channel pruning to ResNet-50, MobileNetV2, and VGG-16 networks. Our method, SAESA, demonstrates strong performance across various benchmarks. On the CIFAR-10 dataset with VGG-16, it achieves 94.06% accuracy using 92 million FLOPs, surpassing the performance of other baseline models. SAESA also achieves leading accuracy of 73.28% on ImageNet with MobileNetV2, while requiring fewer computational resources than recent approaches. Finally, using ResNet-50 on ImageNet, SAESA attains a competitive 76.54% accuracy with improved efficiency, operating at 1980M FLOPs. These findings demonstrate the method's consistent performance and effective trade-off between accuracy and computational cost across different datasets and network architectures.

키워드