The Effect of Pruning Order: A Case Study on CNN Filter Pruning

Progressive structured pruning is a widely adopted strategy for compressing convolutional neural networks, where model capacity is gradually reduced through multiple prune-fine-tune stages. While most existing methods focus on defining effective importance criteria for convolutional filters, the order in which they are removed is typically fixed and the effect of pruning order is largely unexplored. In this work, we investigate different pruning order strategies for structured pruning of convolutional filters. Specifically, we introduce two alternative scheduling policies, tail-based and center-based, that modify the order of filter removal while preserving the same final pruning budget. Unlike classic approaches, which remove the least important filters first, the proposed scheduling strategies modify the pruning order of the components. In this work, we demonstrate the significant impact of pruning scheduling through extensive experiments with ResNet and ResNeXt architectures on CIFAR-100 and ImageNet-1K datasets. In particular, center-based scheduling consistently outperforms the standard progressive strategy, with larger gains at higher pruning rates and on more challenging datasets. Additional analyses show that these improvements are robust with a varying number of pruning stages and importance criteria. These results highlight how changing pruning scheduling is a simple yet effective method to improve structured pruning performance.